1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operations
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/kfence.h>
32 #include <linux/memory.h>
33 #include <linux/math64.h>
34 #include <linux/fault-inject.h>
35 #include <linux/stacktrace.h>
36 #include <linux/prefetch.h>
37 #include <linux/memcontrol.h>
38 #include <linux/random.h>
39 #include <kunit/test.h>
41 #include <linux/debugfs.h>
42 #include <trace/events/kmem.h>
48 * 1. slab_mutex (Global Mutex)
49 * 2. node->list_lock (Spinlock)
50 * 3. kmem_cache->cpu_slab->lock (Local lock)
51 * 4. slab_lock(slab) (Only on some arches or for debugging)
52 * 5. object_map_lock (Only for debugging)
56 * The role of the slab_mutex is to protect the list of all the slabs
57 * and to synchronize major metadata changes to slab cache structures.
58 * Also synchronizes memory hotplug callbacks.
62 * The slab_lock is a wrapper around the page lock, thus it is a bit
65 * The slab_lock is only used for debugging and on arches that do not
66 * have the ability to do a cmpxchg_double. It only protects:
67 * A. slab->freelist -> List of free objects in a slab
68 * B. slab->inuse -> Number of objects in use
69 * C. slab->objects -> Number of objects in slab
70 * D. slab->frozen -> frozen state
74 * If a slab is frozen then it is exempt from list management. It is not
75 * on any list except per cpu partial list. The processor that froze the
76 * slab is the one who can perform list operations on the slab. Other
77 * processors may put objects onto the freelist but the processor that
78 * froze the slab is the only one that can retrieve the objects from the
83 * The list_lock protects the partial and full list on each node and
84 * the partial slab counter. If taken then no new slabs may be added or
85 * removed from the lists nor make the number of partial slabs be modified.
86 * (Note that the total number of slabs is an atomic value that may be
87 * modified without taking the list lock).
89 * The list_lock is a centralized lock and thus we avoid taking it as
90 * much as possible. As long as SLUB does not have to handle partial
91 * slabs, operations can continue without any centralized lock. F.e.
92 * allocating a long series of objects that fill up slabs does not require
95 * cpu_slab->lock local lock
97 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
98 * except the stat counters. This is a percpu structure manipulated only by
99 * the local cpu, so the lock protects against being preempted or interrupted
100 * by an irq. Fast path operations rely on lockless operations instead.
101 * On PREEMPT_RT, the local lock does not actually disable irqs (and thus
102 * prevent the lockless operations), so fastpath operations also need to take
103 * the lock and are no longer lockless.
107 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
108 * are fully lockless when satisfied from the percpu slab (and when
109 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
110 * They also don't disable preemption or migration or irqs. They rely on
111 * the transaction id (tid) field to detect being preempted or moved to
114 * irq, preemption, migration considerations
116 * Interrupts are disabled as part of list_lock or local_lock operations, or
117 * around the slab_lock operation, in order to make the slab allocator safe
118 * to use in the context of an irq.
120 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
121 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
122 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
123 * doesn't have to be revalidated in each section protected by the local lock.
125 * SLUB assigns one slab for allocation to each processor.
126 * Allocations only occur from these slabs called cpu slabs.
128 * Slabs with free elements are kept on a partial list and during regular
129 * operations no list for full slabs is used. If an object in a full slab is
130 * freed then the slab will show up again on the partial lists.
131 * We track full slabs for debugging purposes though because otherwise we
132 * cannot scan all objects.
134 * Slabs are freed when they become empty. Teardown and setup is
135 * minimal so we rely on the page allocators per cpu caches for
136 * fast frees and allocs.
138 * slab->frozen The slab is frozen and exempt from list processing.
139 * This means that the slab is dedicated to a purpose
140 * such as satisfying allocations for a specific
141 * processor. Objects may be freed in the slab while
142 * it is frozen but slab_free will then skip the usual
143 * list operations. It is up to the processor holding
144 * the slab to integrate the slab into the slab lists
145 * when the slab is no longer needed.
147 * One use of this flag is to mark slabs that are
148 * used for allocations. Then such a slab becomes a cpu
149 * slab. The cpu slab may be equipped with an additional
150 * freelist that allows lockless access to
151 * free objects in addition to the regular freelist
152 * that requires the slab lock.
154 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
155 * options set. This moves slab handling out of
156 * the fast path and disables lockless freelists.
160 * We could simply use migrate_disable()/enable() but as long as it's a
161 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
163 #ifndef CONFIG_PREEMPT_RT
164 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
165 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
167 #define slub_get_cpu_ptr(var) \
172 #define slub_put_cpu_ptr(var) \
179 #ifdef CONFIG_SLUB_DEBUG
180 #ifdef CONFIG_SLUB_DEBUG_ON
181 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
183 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
185 #endif /* CONFIG_SLUB_DEBUG */
187 static inline bool kmem_cache_debug(struct kmem_cache *s)
189 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
192 void *fixup_red_left(struct kmem_cache *s, void *p)
194 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
195 p += s->red_left_pad;
200 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
202 #ifdef CONFIG_SLUB_CPU_PARTIAL
203 return !kmem_cache_debug(s);
210 * Issues still to be resolved:
212 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
214 * - Variable sizing of the per node arrays
217 /* Enable to log cmpxchg failures */
218 #undef SLUB_DEBUG_CMPXCHG
221 * Minimum number of partial slabs. These will be left on the partial
222 * lists even if they are empty. kmem_cache_shrink may reclaim them.
224 #define MIN_PARTIAL 5
227 * Maximum number of desirable partial slabs.
228 * The existence of more partial slabs makes kmem_cache_shrink
229 * sort the partial list by the number of objects in use.
231 #define MAX_PARTIAL 10
233 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
234 SLAB_POISON | SLAB_STORE_USER)
237 * These debug flags cannot use CMPXCHG because there might be consistency
238 * issues when checking or reading debug information
240 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
245 * Debugging flags that require metadata to be stored in the slab. These get
246 * disabled when slub_debug=O is used and a cache's min order increases with
249 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
252 #define OO_MASK ((1 << OO_SHIFT) - 1)
253 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
255 /* Internal SLUB flags */
257 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
258 /* Use cmpxchg_double */
259 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
262 * Tracking user of a slab.
264 #define TRACK_ADDRS_COUNT 16
266 unsigned long addr; /* Called from address */
267 #ifdef CONFIG_STACKTRACE
268 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
270 int cpu; /* Was running on cpu */
271 int pid; /* Pid context */
272 unsigned long when; /* When did the operation occur */
275 enum track_item { TRACK_ALLOC, TRACK_FREE };
278 static int sysfs_slab_add(struct kmem_cache *);
279 static int sysfs_slab_alias(struct kmem_cache *, const char *);
281 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
282 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
286 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
287 static void debugfs_slab_add(struct kmem_cache *);
289 static inline void debugfs_slab_add(struct kmem_cache *s) { }
292 static inline void stat(const struct kmem_cache *s, enum stat_item si)
294 #ifdef CONFIG_SLUB_STATS
296 * The rmw is racy on a preemptible kernel but this is acceptable, so
297 * avoid this_cpu_add()'s irq-disable overhead.
299 raw_cpu_inc(s->cpu_slab->stat[si]);
304 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
305 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
306 * differ during memory hotplug/hotremove operations.
307 * Protected by slab_mutex.
309 static nodemask_t slab_nodes;
311 /********************************************************************
312 * Core slab cache functions
313 *******************************************************************/
316 * Returns freelist pointer (ptr). With hardening, this is obfuscated
317 * with an XOR of the address where the pointer is held and a per-cache
320 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
321 unsigned long ptr_addr)
323 #ifdef CONFIG_SLAB_FREELIST_HARDENED
325 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
326 * Normally, this doesn't cause any issues, as both set_freepointer()
327 * and get_freepointer() are called with a pointer with the same tag.
328 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
329 * example, when __free_slub() iterates over objects in a cache, it
330 * passes untagged pointers to check_object(). check_object() in turns
331 * calls get_freepointer() with an untagged pointer, which causes the
332 * freepointer to be restored incorrectly.
334 return (void *)((unsigned long)ptr ^ s->random ^
335 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
341 /* Returns the freelist pointer recorded at location ptr_addr. */
342 static inline void *freelist_dereference(const struct kmem_cache *s,
345 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
346 (unsigned long)ptr_addr);
349 static inline void *get_freepointer(struct kmem_cache *s, void *object)
351 object = kasan_reset_tag(object);
352 return freelist_dereference(s, object + s->offset);
355 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
357 prefetchw(object + s->offset);
360 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
362 unsigned long freepointer_addr;
365 if (!debug_pagealloc_enabled_static())
366 return get_freepointer(s, object);
368 object = kasan_reset_tag(object);
369 freepointer_addr = (unsigned long)object + s->offset;
370 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
371 return freelist_ptr(s, p, freepointer_addr);
374 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
376 unsigned long freeptr_addr = (unsigned long)object + s->offset;
378 #ifdef CONFIG_SLAB_FREELIST_HARDENED
379 BUG_ON(object == fp); /* naive detection of double free or corruption */
382 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
383 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
386 /* Loop over all objects in a slab */
387 #define for_each_object(__p, __s, __addr, __objects) \
388 for (__p = fixup_red_left(__s, __addr); \
389 __p < (__addr) + (__objects) * (__s)->size; \
392 static inline unsigned int order_objects(unsigned int order, unsigned int size)
394 return ((unsigned int)PAGE_SIZE << order) / size;
397 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
400 struct kmem_cache_order_objects x = {
401 (order << OO_SHIFT) + order_objects(order, size)
407 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
409 return x.x >> OO_SHIFT;
412 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
414 return x.x & OO_MASK;
417 #ifdef CONFIG_SLUB_CPU_PARTIAL
418 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
420 unsigned int nr_slabs;
422 s->cpu_partial = nr_objects;
425 * We take the number of objects but actually limit the number of
426 * slabs on the per cpu partial list, in order to limit excessive
427 * growth of the list. For simplicity we assume that the slabs will
430 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
431 s->cpu_partial_slabs = nr_slabs;
435 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
438 #endif /* CONFIG_SLUB_CPU_PARTIAL */
441 * Per slab locking using the pagelock
443 static __always_inline void __slab_lock(struct slab *slab)
445 struct page *page = slab_page(slab);
447 VM_BUG_ON_PAGE(PageTail(page), page);
448 bit_spin_lock(PG_locked, &page->flags);
451 static __always_inline void __slab_unlock(struct slab *slab)
453 struct page *page = slab_page(slab);
455 VM_BUG_ON_PAGE(PageTail(page), page);
456 __bit_spin_unlock(PG_locked, &page->flags);
459 static __always_inline void slab_lock(struct slab *slab, unsigned long *flags)
461 if (IS_ENABLED(CONFIG_PREEMPT_RT))
462 local_irq_save(*flags);
466 static __always_inline void slab_unlock(struct slab *slab, unsigned long *flags)
469 if (IS_ENABLED(CONFIG_PREEMPT_RT))
470 local_irq_restore(*flags);
474 * Interrupts must be disabled (for the fallback code to work right), typically
475 * by an _irqsave() lock variant. Except on PREEMPT_RT where locks are different
476 * so we disable interrupts as part of slab_[un]lock().
478 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
479 void *freelist_old, unsigned long counters_old,
480 void *freelist_new, unsigned long counters_new,
483 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
484 lockdep_assert_irqs_disabled();
485 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
486 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
487 if (s->flags & __CMPXCHG_DOUBLE) {
488 if (cmpxchg_double(&slab->freelist, &slab->counters,
489 freelist_old, counters_old,
490 freelist_new, counters_new))
495 /* init to 0 to prevent spurious warnings */
496 unsigned long flags = 0;
498 slab_lock(slab, &flags);
499 if (slab->freelist == freelist_old &&
500 slab->counters == counters_old) {
501 slab->freelist = freelist_new;
502 slab->counters = counters_new;
503 slab_unlock(slab, &flags);
506 slab_unlock(slab, &flags);
510 stat(s, CMPXCHG_DOUBLE_FAIL);
512 #ifdef SLUB_DEBUG_CMPXCHG
513 pr_info("%s %s: cmpxchg double redo ", n, s->name);
519 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
520 void *freelist_old, unsigned long counters_old,
521 void *freelist_new, unsigned long counters_new,
524 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
525 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
526 if (s->flags & __CMPXCHG_DOUBLE) {
527 if (cmpxchg_double(&slab->freelist, &slab->counters,
528 freelist_old, counters_old,
529 freelist_new, counters_new))
536 local_irq_save(flags);
538 if (slab->freelist == freelist_old &&
539 slab->counters == counters_old) {
540 slab->freelist = freelist_new;
541 slab->counters = counters_new;
543 local_irq_restore(flags);
547 local_irq_restore(flags);
551 stat(s, CMPXCHG_DOUBLE_FAIL);
553 #ifdef SLUB_DEBUG_CMPXCHG
554 pr_info("%s %s: cmpxchg double redo ", n, s->name);
560 #ifdef CONFIG_SLUB_DEBUG
561 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
562 static DEFINE_RAW_SPINLOCK(object_map_lock);
564 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
567 void *addr = slab_address(slab);
570 bitmap_zero(obj_map, slab->objects);
572 for (p = slab->freelist; p; p = get_freepointer(s, p))
573 set_bit(__obj_to_index(s, addr, p), obj_map);
576 #if IS_ENABLED(CONFIG_KUNIT)
577 static bool slab_add_kunit_errors(void)
579 struct kunit_resource *resource;
581 if (likely(!current->kunit_test))
584 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
588 (*(int *)resource->data)++;
589 kunit_put_resource(resource);
593 static inline bool slab_add_kunit_errors(void) { return false; }
597 * Determine a map of objects in use in a slab.
599 * Node listlock must be held to guarantee that the slab does
600 * not vanish from under us.
602 static unsigned long *get_map(struct kmem_cache *s, struct slab *slab)
603 __acquires(&object_map_lock)
605 VM_BUG_ON(!irqs_disabled());
607 raw_spin_lock(&object_map_lock);
609 __fill_map(object_map, s, slab);
614 static void put_map(unsigned long *map) __releases(&object_map_lock)
616 VM_BUG_ON(map != object_map);
617 raw_spin_unlock(&object_map_lock);
620 static inline unsigned int size_from_object(struct kmem_cache *s)
622 if (s->flags & SLAB_RED_ZONE)
623 return s->size - s->red_left_pad;
628 static inline void *restore_red_left(struct kmem_cache *s, void *p)
630 if (s->flags & SLAB_RED_ZONE)
631 p -= s->red_left_pad;
639 #if defined(CONFIG_SLUB_DEBUG_ON)
640 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
642 static slab_flags_t slub_debug;
645 static char *slub_debug_string;
646 static int disable_higher_order_debug;
649 * slub is about to manipulate internal object metadata. This memory lies
650 * outside the range of the allocated object, so accessing it would normally
651 * be reported by kasan as a bounds error. metadata_access_enable() is used
652 * to tell kasan that these accesses are OK.
654 static inline void metadata_access_enable(void)
656 kasan_disable_current();
659 static inline void metadata_access_disable(void)
661 kasan_enable_current();
668 /* Verify that a pointer has an address that is valid within a slab page */
669 static inline int check_valid_pointer(struct kmem_cache *s,
670 struct slab *slab, void *object)
677 base = slab_address(slab);
678 object = kasan_reset_tag(object);
679 object = restore_red_left(s, object);
680 if (object < base || object >= base + slab->objects * s->size ||
681 (object - base) % s->size) {
688 static void print_section(char *level, char *text, u8 *addr,
691 metadata_access_enable();
692 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
693 16, 1, kasan_reset_tag((void *)addr), length, 1);
694 metadata_access_disable();
698 * See comment in calculate_sizes().
700 static inline bool freeptr_outside_object(struct kmem_cache *s)
702 return s->offset >= s->inuse;
706 * Return offset of the end of info block which is inuse + free pointer if
707 * not overlapping with object.
709 static inline unsigned int get_info_end(struct kmem_cache *s)
711 if (freeptr_outside_object(s))
712 return s->inuse + sizeof(void *);
717 static struct track *get_track(struct kmem_cache *s, void *object,
718 enum track_item alloc)
722 p = object + get_info_end(s);
724 return kasan_reset_tag(p + alloc);
727 static void set_track(struct kmem_cache *s, void *object,
728 enum track_item alloc, unsigned long addr)
730 struct track *p = get_track(s, object, alloc);
733 #ifdef CONFIG_STACKTRACE
734 unsigned int nr_entries;
736 metadata_access_enable();
737 nr_entries = stack_trace_save(kasan_reset_tag(p->addrs),
738 TRACK_ADDRS_COUNT, 3);
739 metadata_access_disable();
741 if (nr_entries < TRACK_ADDRS_COUNT)
742 p->addrs[nr_entries] = 0;
745 p->cpu = smp_processor_id();
746 p->pid = current->pid;
749 memset(p, 0, sizeof(struct track));
753 static void init_tracking(struct kmem_cache *s, void *object)
755 if (!(s->flags & SLAB_STORE_USER))
758 set_track(s, object, TRACK_FREE, 0UL);
759 set_track(s, object, TRACK_ALLOC, 0UL);
762 static void print_track(const char *s, struct track *t, unsigned long pr_time)
767 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
768 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
769 #ifdef CONFIG_STACKTRACE
772 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
774 pr_err("\t%pS\n", (void *)t->addrs[i]);
781 void print_tracking(struct kmem_cache *s, void *object)
783 unsigned long pr_time = jiffies;
784 if (!(s->flags & SLAB_STORE_USER))
787 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
788 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
791 static void print_slab_info(const struct slab *slab)
793 struct folio *folio = (struct folio *)slab_folio(slab);
795 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
796 slab, slab->objects, slab->inuse, slab->freelist,
797 folio_flags(folio, 0));
800 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
802 struct va_format vaf;
808 pr_err("=============================================================================\n");
809 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
810 pr_err("-----------------------------------------------------------------------------\n\n");
815 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
817 struct va_format vaf;
820 if (slab_add_kunit_errors())
826 pr_err("FIX %s: %pV\n", s->name, &vaf);
830 static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
832 unsigned int off; /* Offset of last byte */
833 u8 *addr = slab_address(slab);
835 print_tracking(s, p);
837 print_slab_info(slab);
839 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
840 p, p - addr, get_freepointer(s, p));
842 if (s->flags & SLAB_RED_ZONE)
843 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
845 else if (p > addr + 16)
846 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
848 print_section(KERN_ERR, "Object ", p,
849 min_t(unsigned int, s->object_size, PAGE_SIZE));
850 if (s->flags & SLAB_RED_ZONE)
851 print_section(KERN_ERR, "Redzone ", p + s->object_size,
852 s->inuse - s->object_size);
854 off = get_info_end(s);
856 if (s->flags & SLAB_STORE_USER)
857 off += 2 * sizeof(struct track);
859 off += kasan_metadata_size(s);
861 if (off != size_from_object(s))
862 /* Beginning of the filler is the free pointer */
863 print_section(KERN_ERR, "Padding ", p + off,
864 size_from_object(s) - off);
869 static void object_err(struct kmem_cache *s, struct slab *slab,
870 u8 *object, char *reason)
872 if (slab_add_kunit_errors())
875 slab_bug(s, "%s", reason);
876 print_trailer(s, slab, object);
877 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
880 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
881 void **freelist, void *nextfree)
883 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
884 !check_valid_pointer(s, slab, nextfree) && freelist) {
885 object_err(s, slab, *freelist, "Freechain corrupt");
887 slab_fix(s, "Isolate corrupted freechain");
894 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
895 const char *fmt, ...)
900 if (slab_add_kunit_errors())
904 vsnprintf(buf, sizeof(buf), fmt, args);
906 slab_bug(s, "%s", buf);
907 print_slab_info(slab);
909 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
912 static void init_object(struct kmem_cache *s, void *object, u8 val)
914 u8 *p = kasan_reset_tag(object);
916 if (s->flags & SLAB_RED_ZONE)
917 memset(p - s->red_left_pad, val, s->red_left_pad);
919 if (s->flags & __OBJECT_POISON) {
920 memset(p, POISON_FREE, s->object_size - 1);
921 p[s->object_size - 1] = POISON_END;
924 if (s->flags & SLAB_RED_ZONE)
925 memset(p + s->object_size, val, s->inuse - s->object_size);
928 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
929 void *from, void *to)
931 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
932 memset(from, data, to - from);
935 static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
936 u8 *object, char *what,
937 u8 *start, unsigned int value, unsigned int bytes)
941 u8 *addr = slab_address(slab);
943 metadata_access_enable();
944 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
945 metadata_access_disable();
950 while (end > fault && end[-1] == value)
953 if (slab_add_kunit_errors())
956 slab_bug(s, "%s overwritten", what);
957 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
958 fault, end - 1, fault - addr,
960 print_trailer(s, slab, object);
961 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
964 restore_bytes(s, what, value, fault, end);
972 * Bytes of the object to be managed.
973 * If the freepointer may overlay the object then the free
974 * pointer is at the middle of the object.
976 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
979 * object + s->object_size
980 * Padding to reach word boundary. This is also used for Redzoning.
981 * Padding is extended by another word if Redzoning is enabled and
982 * object_size == inuse.
984 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
985 * 0xcc (RED_ACTIVE) for objects in use.
988 * Meta data starts here.
990 * A. Free pointer (if we cannot overwrite object on free)
991 * B. Tracking data for SLAB_STORE_USER
992 * C. Padding to reach required alignment boundary or at minimum
993 * one word if debugging is on to be able to detect writes
994 * before the word boundary.
996 * Padding is done using 0x5a (POISON_INUSE)
999 * Nothing is used beyond s->size.
1001 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1002 * ignored. And therefore no slab options that rely on these boundaries
1003 * may be used with merged slabcaches.
1006 static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1008 unsigned long off = get_info_end(s); /* The end of info */
1010 if (s->flags & SLAB_STORE_USER)
1011 /* We also have user information there */
1012 off += 2 * sizeof(struct track);
1014 off += kasan_metadata_size(s);
1016 if (size_from_object(s) == off)
1019 return check_bytes_and_report(s, slab, p, "Object padding",
1020 p + off, POISON_INUSE, size_from_object(s) - off);
1023 /* Check the pad bytes at the end of a slab page */
1024 static int slab_pad_check(struct kmem_cache *s, struct slab *slab)
1033 if (!(s->flags & SLAB_POISON))
1036 start = slab_address(slab);
1037 length = slab_size(slab);
1038 end = start + length;
1039 remainder = length % s->size;
1043 pad = end - remainder;
1044 metadata_access_enable();
1045 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1046 metadata_access_disable();
1049 while (end > fault && end[-1] == POISON_INUSE)
1052 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1053 fault, end - 1, fault - start);
1054 print_section(KERN_ERR, "Padding ", pad, remainder);
1056 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1060 static int check_object(struct kmem_cache *s, struct slab *slab,
1061 void *object, u8 val)
1064 u8 *endobject = object + s->object_size;
1066 if (s->flags & SLAB_RED_ZONE) {
1067 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1068 object - s->red_left_pad, val, s->red_left_pad))
1071 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1072 endobject, val, s->inuse - s->object_size))
1075 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1076 check_bytes_and_report(s, slab, p, "Alignment padding",
1077 endobject, POISON_INUSE,
1078 s->inuse - s->object_size);
1082 if (s->flags & SLAB_POISON) {
1083 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1084 (!check_bytes_and_report(s, slab, p, "Poison", p,
1085 POISON_FREE, s->object_size - 1) ||
1086 !check_bytes_and_report(s, slab, p, "End Poison",
1087 p + s->object_size - 1, POISON_END, 1)))
1090 * check_pad_bytes cleans up on its own.
1092 check_pad_bytes(s, slab, p);
1095 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1097 * Object and freepointer overlap. Cannot check
1098 * freepointer while object is allocated.
1102 /* Check free pointer validity */
1103 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1104 object_err(s, slab, p, "Freepointer corrupt");
1106 * No choice but to zap it and thus lose the remainder
1107 * of the free objects in this slab. May cause
1108 * another error because the object count is now wrong.
1110 set_freepointer(s, p, NULL);
1116 static int check_slab(struct kmem_cache *s, struct slab *slab)
1120 if (!folio_test_slab(slab_folio(slab))) {
1121 slab_err(s, slab, "Not a valid slab page");
1125 maxobj = order_objects(slab_order(slab), s->size);
1126 if (slab->objects > maxobj) {
1127 slab_err(s, slab, "objects %u > max %u",
1128 slab->objects, maxobj);
1131 if (slab->inuse > slab->objects) {
1132 slab_err(s, slab, "inuse %u > max %u",
1133 slab->inuse, slab->objects);
1136 /* Slab_pad_check fixes things up after itself */
1137 slab_pad_check(s, slab);
1142 * Determine if a certain object in a slab is on the freelist. Must hold the
1143 * slab lock to guarantee that the chains are in a consistent state.
1145 static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1149 void *object = NULL;
1152 fp = slab->freelist;
1153 while (fp && nr <= slab->objects) {
1156 if (!check_valid_pointer(s, slab, fp)) {
1158 object_err(s, slab, object,
1159 "Freechain corrupt");
1160 set_freepointer(s, object, NULL);
1162 slab_err(s, slab, "Freepointer corrupt");
1163 slab->freelist = NULL;
1164 slab->inuse = slab->objects;
1165 slab_fix(s, "Freelist cleared");
1171 fp = get_freepointer(s, object);
1175 max_objects = order_objects(slab_order(slab), s->size);
1176 if (max_objects > MAX_OBJS_PER_PAGE)
1177 max_objects = MAX_OBJS_PER_PAGE;
1179 if (slab->objects != max_objects) {
1180 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1181 slab->objects, max_objects);
1182 slab->objects = max_objects;
1183 slab_fix(s, "Number of objects adjusted");
1185 if (slab->inuse != slab->objects - nr) {
1186 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1187 slab->inuse, slab->objects - nr);
1188 slab->inuse = slab->objects - nr;
1189 slab_fix(s, "Object count adjusted");
1191 return search == NULL;
1194 static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1197 if (s->flags & SLAB_TRACE) {
1198 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1200 alloc ? "alloc" : "free",
1201 object, slab->inuse,
1205 print_section(KERN_INFO, "Object ", (void *)object,
1213 * Tracking of fully allocated slabs for debugging purposes.
1215 static void add_full(struct kmem_cache *s,
1216 struct kmem_cache_node *n, struct slab *slab)
1218 if (!(s->flags & SLAB_STORE_USER))
1221 lockdep_assert_held(&n->list_lock);
1222 list_add(&slab->slab_list, &n->full);
1225 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1227 if (!(s->flags & SLAB_STORE_USER))
1230 lockdep_assert_held(&n->list_lock);
1231 list_del(&slab->slab_list);
1234 /* Tracking of the number of slabs for debugging purposes */
1235 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1237 struct kmem_cache_node *n = get_node(s, node);
1239 return atomic_long_read(&n->nr_slabs);
1242 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1244 return atomic_long_read(&n->nr_slabs);
1247 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1249 struct kmem_cache_node *n = get_node(s, node);
1252 * May be called early in order to allocate a slab for the
1253 * kmem_cache_node structure. Solve the chicken-egg
1254 * dilemma by deferring the increment of the count during
1255 * bootstrap (see early_kmem_cache_node_alloc).
1258 atomic_long_inc(&n->nr_slabs);
1259 atomic_long_add(objects, &n->total_objects);
1262 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1264 struct kmem_cache_node *n = get_node(s, node);
1266 atomic_long_dec(&n->nr_slabs);
1267 atomic_long_sub(objects, &n->total_objects);
1270 /* Object debug checks for alloc/free paths */
1271 static void setup_object_debug(struct kmem_cache *s, struct slab *slab,
1274 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1277 init_object(s, object, SLUB_RED_INACTIVE);
1278 init_tracking(s, object);
1282 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1284 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1287 metadata_access_enable();
1288 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1289 metadata_access_disable();
1292 static inline int alloc_consistency_checks(struct kmem_cache *s,
1293 struct slab *slab, void *object)
1295 if (!check_slab(s, slab))
1298 if (!check_valid_pointer(s, slab, object)) {
1299 object_err(s, slab, object, "Freelist Pointer check fails");
1303 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1309 static noinline int alloc_debug_processing(struct kmem_cache *s,
1311 void *object, unsigned long addr)
1313 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1314 if (!alloc_consistency_checks(s, slab, object))
1318 /* Success perform special debug activities for allocs */
1319 if (s->flags & SLAB_STORE_USER)
1320 set_track(s, object, TRACK_ALLOC, addr);
1321 trace(s, slab, object, 1);
1322 init_object(s, object, SLUB_RED_ACTIVE);
1326 if (folio_test_slab(slab_folio(slab))) {
1328 * If this is a slab page then lets do the best we can
1329 * to avoid issues in the future. Marking all objects
1330 * as used avoids touching the remaining objects.
1332 slab_fix(s, "Marking all objects used");
1333 slab->inuse = slab->objects;
1334 slab->freelist = NULL;
1339 static inline int free_consistency_checks(struct kmem_cache *s,
1340 struct slab *slab, void *object, unsigned long addr)
1342 if (!check_valid_pointer(s, slab, object)) {
1343 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1347 if (on_freelist(s, slab, object)) {
1348 object_err(s, slab, object, "Object already free");
1352 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1355 if (unlikely(s != slab->slab_cache)) {
1356 if (!folio_test_slab(slab_folio(slab))) {
1357 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1359 } else if (!slab->slab_cache) {
1360 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1364 object_err(s, slab, object,
1365 "page slab pointer corrupt.");
1371 /* Supports checking bulk free of a constructed freelist */
1372 static noinline int free_debug_processing(
1373 struct kmem_cache *s, struct slab *slab,
1374 void *head, void *tail, int bulk_cnt,
1377 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
1378 void *object = head;
1380 unsigned long flags, flags2;
1383 spin_lock_irqsave(&n->list_lock, flags);
1384 slab_lock(slab, &flags2);
1386 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1387 if (!check_slab(s, slab))
1394 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1395 if (!free_consistency_checks(s, slab, object, addr))
1399 if (s->flags & SLAB_STORE_USER)
1400 set_track(s, object, TRACK_FREE, addr);
1401 trace(s, slab, object, 0);
1402 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1403 init_object(s, object, SLUB_RED_INACTIVE);
1405 /* Reached end of constructed freelist yet? */
1406 if (object != tail) {
1407 object = get_freepointer(s, object);
1413 if (cnt != bulk_cnt)
1414 slab_err(s, slab, "Bulk freelist count(%d) invalid(%d)\n",
1417 slab_unlock(slab, &flags2);
1418 spin_unlock_irqrestore(&n->list_lock, flags);
1420 slab_fix(s, "Object at 0x%p not freed", object);
1425 * Parse a block of slub_debug options. Blocks are delimited by ';'
1427 * @str: start of block
1428 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1429 * @slabs: return start of list of slabs, or NULL when there's no list
1430 * @init: assume this is initial parsing and not per-kmem-create parsing
1432 * returns the start of next block if there's any, or NULL
1435 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1437 bool higher_order_disable = false;
1439 /* Skip any completely empty blocks */
1440 while (*str && *str == ';')
1445 * No options but restriction on slabs. This means full
1446 * debugging for slabs matching a pattern.
1448 *flags = DEBUG_DEFAULT_FLAGS;
1453 /* Determine which debug features should be switched on */
1454 for (; *str && *str != ',' && *str != ';'; str++) {
1455 switch (tolower(*str)) {
1460 *flags |= SLAB_CONSISTENCY_CHECKS;
1463 *flags |= SLAB_RED_ZONE;
1466 *flags |= SLAB_POISON;
1469 *flags |= SLAB_STORE_USER;
1472 *flags |= SLAB_TRACE;
1475 *flags |= SLAB_FAILSLAB;
1479 * Avoid enabling debugging on caches if its minimum
1480 * order would increase as a result.
1482 higher_order_disable = true;
1486 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1495 /* Skip over the slab list */
1496 while (*str && *str != ';')
1499 /* Skip any completely empty blocks */
1500 while (*str && *str == ';')
1503 if (init && higher_order_disable)
1504 disable_higher_order_debug = 1;
1512 static int __init setup_slub_debug(char *str)
1515 slab_flags_t global_flags;
1518 bool global_slub_debug_changed = false;
1519 bool slab_list_specified = false;
1521 global_flags = DEBUG_DEFAULT_FLAGS;
1522 if (*str++ != '=' || !*str)
1524 * No options specified. Switch on full debugging.
1530 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1533 global_flags = flags;
1534 global_slub_debug_changed = true;
1536 slab_list_specified = true;
1541 * For backwards compatibility, a single list of flags with list of
1542 * slabs means debugging is only changed for those slabs, so the global
1543 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1544 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1545 * long as there is no option specifying flags without a slab list.
1547 if (slab_list_specified) {
1548 if (!global_slub_debug_changed)
1549 global_flags = slub_debug;
1550 slub_debug_string = saved_str;
1553 slub_debug = global_flags;
1554 if (slub_debug != 0 || slub_debug_string)
1555 static_branch_enable(&slub_debug_enabled);
1557 static_branch_disable(&slub_debug_enabled);
1558 if ((static_branch_unlikely(&init_on_alloc) ||
1559 static_branch_unlikely(&init_on_free)) &&
1560 (slub_debug & SLAB_POISON))
1561 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1565 __setup("slub_debug", setup_slub_debug);
1568 * kmem_cache_flags - apply debugging options to the cache
1569 * @object_size: the size of an object without meta data
1570 * @flags: flags to set
1571 * @name: name of the cache
1573 * Debug option(s) are applied to @flags. In addition to the debug
1574 * option(s), if a slab name (or multiple) is specified i.e.
1575 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1576 * then only the select slabs will receive the debug option(s).
1578 slab_flags_t kmem_cache_flags(unsigned int object_size,
1579 slab_flags_t flags, const char *name)
1584 slab_flags_t block_flags;
1585 slab_flags_t slub_debug_local = slub_debug;
1588 * If the slab cache is for debugging (e.g. kmemleak) then
1589 * don't store user (stack trace) information by default,
1590 * but let the user enable it via the command line below.
1592 if (flags & SLAB_NOLEAKTRACE)
1593 slub_debug_local &= ~SLAB_STORE_USER;
1596 next_block = slub_debug_string;
1597 /* Go through all blocks of debug options, see if any matches our slab's name */
1598 while (next_block) {
1599 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1602 /* Found a block that has a slab list, search it */
1607 end = strchrnul(iter, ',');
1608 if (next_block && next_block < end)
1609 end = next_block - 1;
1611 glob = strnchr(iter, end - iter, '*');
1613 cmplen = glob - iter;
1615 cmplen = max_t(size_t, len, (end - iter));
1617 if (!strncmp(name, iter, cmplen)) {
1618 flags |= block_flags;
1622 if (!*end || *end == ';')
1628 return flags | slub_debug_local;
1630 #else /* !CONFIG_SLUB_DEBUG */
1631 static inline void setup_object_debug(struct kmem_cache *s,
1632 struct slab *slab, void *object) {}
1634 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1636 static inline int alloc_debug_processing(struct kmem_cache *s,
1637 struct slab *slab, void *object, unsigned long addr) { return 0; }
1639 static inline int free_debug_processing(
1640 struct kmem_cache *s, struct slab *slab,
1641 void *head, void *tail, int bulk_cnt,
1642 unsigned long addr) { return 0; }
1644 static inline int slab_pad_check(struct kmem_cache *s, struct slab *slab)
1646 static inline int check_object(struct kmem_cache *s, struct slab *slab,
1647 void *object, u8 val) { return 1; }
1648 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1649 struct slab *slab) {}
1650 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1651 struct slab *slab) {}
1652 slab_flags_t kmem_cache_flags(unsigned int object_size,
1653 slab_flags_t flags, const char *name)
1657 #define slub_debug 0
1659 #define disable_higher_order_debug 0
1661 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1663 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1665 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1667 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1670 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1671 void **freelist, void *nextfree)
1675 #endif /* CONFIG_SLUB_DEBUG */
1678 * Hooks for other subsystems that check memory allocations. In a typical
1679 * production configuration these hooks all should produce no code at all.
1681 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1683 ptr = kasan_kmalloc_large(ptr, size, flags);
1684 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1685 kmemleak_alloc(ptr, size, 1, flags);
1689 static __always_inline void kfree_hook(void *x)
1692 kasan_kfree_large(x);
1695 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1698 kmemleak_free_recursive(x, s->flags);
1700 debug_check_no_locks_freed(x, s->object_size);
1702 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1703 debug_check_no_obj_freed(x, s->object_size);
1705 /* Use KCSAN to help debug racy use-after-free. */
1706 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1707 __kcsan_check_access(x, s->object_size,
1708 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1711 * As memory initialization might be integrated into KASAN,
1712 * kasan_slab_free and initialization memset's must be
1713 * kept together to avoid discrepancies in behavior.
1715 * The initialization memset's clear the object and the metadata,
1716 * but don't touch the SLAB redzone.
1721 if (!kasan_has_integrated_init())
1722 memset(kasan_reset_tag(x), 0, s->object_size);
1723 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1724 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1725 s->size - s->inuse - rsize);
1727 /* KASAN might put x into memory quarantine, delaying its reuse. */
1728 return kasan_slab_free(s, x, init);
1731 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1732 void **head, void **tail,
1738 void *old_tail = *tail ? *tail : *head;
1740 if (is_kfence_address(next)) {
1741 slab_free_hook(s, next, false);
1745 /* Head and tail of the reconstructed freelist */
1751 next = get_freepointer(s, object);
1753 /* If object's reuse doesn't have to be delayed */
1754 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1755 /* Move object to the new freelist */
1756 set_freepointer(s, object, *head);
1762 * Adjust the reconstructed freelist depth
1763 * accordingly if object's reuse is delayed.
1767 } while (object != old_tail);
1772 return *head != NULL;
1775 static void *setup_object(struct kmem_cache *s, struct slab *slab,
1778 setup_object_debug(s, slab, object);
1779 object = kasan_init_slab_obj(s, object);
1780 if (unlikely(s->ctor)) {
1781 kasan_unpoison_object_data(s, object);
1783 kasan_poison_object_data(s, object);
1789 * Slab allocation and freeing
1791 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1792 struct kmem_cache_order_objects oo)
1794 struct folio *folio;
1796 unsigned int order = oo_order(oo);
1798 if (node == NUMA_NO_NODE)
1799 folio = (struct folio *)alloc_pages(flags, order);
1801 folio = (struct folio *)__alloc_pages_node(node, flags, order);
1806 slab = folio_slab(folio);
1807 __folio_set_slab(folio);
1808 if (page_is_pfmemalloc(folio_page(folio, 0)))
1809 slab_set_pfmemalloc(slab);
1814 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1815 /* Pre-initialize the random sequence cache */
1816 static int init_cache_random_seq(struct kmem_cache *s)
1818 unsigned int count = oo_objects(s->oo);
1821 /* Bailout if already initialised */
1825 err = cache_random_seq_create(s, count, GFP_KERNEL);
1827 pr_err("SLUB: Unable to initialize free list for %s\n",
1832 /* Transform to an offset on the set of pages */
1833 if (s->random_seq) {
1836 for (i = 0; i < count; i++)
1837 s->random_seq[i] *= s->size;
1842 /* Initialize each random sequence freelist per cache */
1843 static void __init init_freelist_randomization(void)
1845 struct kmem_cache *s;
1847 mutex_lock(&slab_mutex);
1849 list_for_each_entry(s, &slab_caches, list)
1850 init_cache_random_seq(s);
1852 mutex_unlock(&slab_mutex);
1855 /* Get the next entry on the pre-computed freelist randomized */
1856 static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1857 unsigned long *pos, void *start,
1858 unsigned long page_limit,
1859 unsigned long freelist_count)
1864 * If the target page allocation failed, the number of objects on the
1865 * page might be smaller than the usual size defined by the cache.
1868 idx = s->random_seq[*pos];
1870 if (*pos >= freelist_count)
1872 } while (unlikely(idx >= page_limit));
1874 return (char *)start + idx;
1877 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1878 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1883 unsigned long idx, pos, page_limit, freelist_count;
1885 if (slab->objects < 2 || !s->random_seq)
1888 freelist_count = oo_objects(s->oo);
1889 pos = get_random_int() % freelist_count;
1891 page_limit = slab->objects * s->size;
1892 start = fixup_red_left(s, slab_address(slab));
1894 /* First entry is used as the base of the freelist */
1895 cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1897 cur = setup_object(s, slab, cur);
1898 slab->freelist = cur;
1900 for (idx = 1; idx < slab->objects; idx++) {
1901 next = next_freelist_entry(s, slab, &pos, start, page_limit,
1903 next = setup_object(s, slab, next);
1904 set_freepointer(s, cur, next);
1907 set_freepointer(s, cur, NULL);
1912 static inline int init_cache_random_seq(struct kmem_cache *s)
1916 static inline void init_freelist_randomization(void) { }
1917 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1921 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1923 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1926 struct kmem_cache_order_objects oo = s->oo;
1928 void *start, *p, *next;
1932 flags &= gfp_allowed_mask;
1934 flags |= s->allocflags;
1937 * Let the initial higher-order allocation fail under memory pressure
1938 * so we fall-back to the minimum order allocation.
1940 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1941 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1942 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1944 slab = alloc_slab_page(alloc_gfp, node, oo);
1945 if (unlikely(!slab)) {
1949 * Allocation may have failed due to fragmentation.
1950 * Try a lower order alloc if possible
1952 slab = alloc_slab_page(alloc_gfp, node, oo);
1953 if (unlikely(!slab))
1955 stat(s, ORDER_FALLBACK);
1958 slab->objects = oo_objects(oo);
1960 account_slab(slab, oo_order(oo), s, flags);
1962 slab->slab_cache = s;
1964 kasan_poison_slab(slab);
1966 start = slab_address(slab);
1968 setup_slab_debug(s, slab, start);
1970 shuffle = shuffle_freelist(s, slab);
1973 start = fixup_red_left(s, start);
1974 start = setup_object(s, slab, start);
1975 slab->freelist = start;
1976 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
1978 next = setup_object(s, slab, next);
1979 set_freepointer(s, p, next);
1982 set_freepointer(s, p, NULL);
1985 slab->inuse = slab->objects;
1992 inc_slabs_node(s, slab_nid(slab), slab->objects);
1997 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1999 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2000 flags = kmalloc_fix_flags(flags);
2002 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2004 return allocate_slab(s,
2005 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2008 static void __free_slab(struct kmem_cache *s, struct slab *slab)
2010 struct folio *folio = slab_folio(slab);
2011 int order = folio_order(folio);
2012 int pages = 1 << order;
2014 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2017 slab_pad_check(s, slab);
2018 for_each_object(p, s, slab_address(slab), slab->objects)
2019 check_object(s, slab, p, SLUB_RED_INACTIVE);
2022 __slab_clear_pfmemalloc(slab);
2023 __folio_clear_slab(folio);
2024 folio->mapping = NULL;
2025 if (current->reclaim_state)
2026 current->reclaim_state->reclaimed_slab += pages;
2027 unaccount_slab(slab, order, s);
2028 __free_pages(folio_page(folio, 0), order);
2031 static void rcu_free_slab(struct rcu_head *h)
2033 struct slab *slab = container_of(h, struct slab, rcu_head);
2035 __free_slab(slab->slab_cache, slab);
2038 static void free_slab(struct kmem_cache *s, struct slab *slab)
2040 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
2041 call_rcu(&slab->rcu_head, rcu_free_slab);
2043 __free_slab(s, slab);
2046 static void discard_slab(struct kmem_cache *s, struct slab *slab)
2048 dec_slabs_node(s, slab_nid(slab), slab->objects);
2053 * Management of partially allocated slabs.
2056 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2059 if (tail == DEACTIVATE_TO_TAIL)
2060 list_add_tail(&slab->slab_list, &n->partial);
2062 list_add(&slab->slab_list, &n->partial);
2065 static inline void add_partial(struct kmem_cache_node *n,
2066 struct slab *slab, int tail)
2068 lockdep_assert_held(&n->list_lock);
2069 __add_partial(n, slab, tail);
2072 static inline void remove_partial(struct kmem_cache_node *n,
2075 lockdep_assert_held(&n->list_lock);
2076 list_del(&slab->slab_list);
2081 * Remove slab from the partial list, freeze it and
2082 * return the pointer to the freelist.
2084 * Returns a list of objects or NULL if it fails.
2086 static inline void *acquire_slab(struct kmem_cache *s,
2087 struct kmem_cache_node *n, struct slab *slab,
2091 unsigned long counters;
2094 lockdep_assert_held(&n->list_lock);
2097 * Zap the freelist and set the frozen bit.
2098 * The old freelist is the list of objects for the
2099 * per cpu allocation list.
2101 freelist = slab->freelist;
2102 counters = slab->counters;
2103 new.counters = counters;
2105 new.inuse = slab->objects;
2106 new.freelist = NULL;
2108 new.freelist = freelist;
2111 VM_BUG_ON(new.frozen);
2114 if (!__cmpxchg_double_slab(s, slab,
2116 new.freelist, new.counters,
2120 remove_partial(n, slab);
2125 #ifdef CONFIG_SLUB_CPU_PARTIAL
2126 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2128 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2131 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2134 * Try to allocate a partial slab from a specific node.
2136 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2137 struct slab **ret_slab, gfp_t gfpflags)
2139 struct slab *slab, *slab2;
2140 void *object = NULL;
2141 unsigned long flags;
2142 unsigned int partial_slabs = 0;
2145 * Racy check. If we mistakenly see no partial slabs then we
2146 * just allocate an empty slab. If we mistakenly try to get a
2147 * partial slab and there is none available then get_partial()
2150 if (!n || !n->nr_partial)
2153 spin_lock_irqsave(&n->list_lock, flags);
2154 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2157 if (!pfmemalloc_match(slab, gfpflags))
2160 t = acquire_slab(s, n, slab, object == NULL);
2166 stat(s, ALLOC_FROM_PARTIAL);
2169 put_cpu_partial(s, slab, 0);
2170 stat(s, CPU_PARTIAL_NODE);
2173 #ifdef CONFIG_SLUB_CPU_PARTIAL
2174 if (!kmem_cache_has_cpu_partial(s)
2175 || partial_slabs > s->cpu_partial_slabs / 2)
2182 spin_unlock_irqrestore(&n->list_lock, flags);
2187 * Get a slab from somewhere. Search in increasing NUMA distances.
2189 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2190 struct slab **ret_slab)
2193 struct zonelist *zonelist;
2196 enum zone_type highest_zoneidx = gfp_zone(flags);
2198 unsigned int cpuset_mems_cookie;
2201 * The defrag ratio allows a configuration of the tradeoffs between
2202 * inter node defragmentation and node local allocations. A lower
2203 * defrag_ratio increases the tendency to do local allocations
2204 * instead of attempting to obtain partial slabs from other nodes.
2206 * If the defrag_ratio is set to 0 then kmalloc() always
2207 * returns node local objects. If the ratio is higher then kmalloc()
2208 * may return off node objects because partial slabs are obtained
2209 * from other nodes and filled up.
2211 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2212 * (which makes defrag_ratio = 1000) then every (well almost)
2213 * allocation will first attempt to defrag slab caches on other nodes.
2214 * This means scanning over all nodes to look for partial slabs which
2215 * may be expensive if we do it every time we are trying to find a slab
2216 * with available objects.
2218 if (!s->remote_node_defrag_ratio ||
2219 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2223 cpuset_mems_cookie = read_mems_allowed_begin();
2224 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2225 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2226 struct kmem_cache_node *n;
2228 n = get_node(s, zone_to_nid(zone));
2230 if (n && cpuset_zone_allowed(zone, flags) &&
2231 n->nr_partial > s->min_partial) {
2232 object = get_partial_node(s, n, ret_slab, flags);
2235 * Don't check read_mems_allowed_retry()
2236 * here - if mems_allowed was updated in
2237 * parallel, that was a harmless race
2238 * between allocation and the cpuset
2245 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2246 #endif /* CONFIG_NUMA */
2251 * Get a partial slab, lock it and return it.
2253 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2254 struct slab **ret_slab)
2257 int searchnode = node;
2259 if (node == NUMA_NO_NODE)
2260 searchnode = numa_mem_id();
2262 object = get_partial_node(s, get_node(s, searchnode), ret_slab, flags);
2263 if (object || node != NUMA_NO_NODE)
2266 return get_any_partial(s, flags, ret_slab);
2269 #ifdef CONFIG_PREEMPTION
2271 * Calculate the next globally unique transaction for disambiguation
2272 * during cmpxchg. The transactions start with the cpu number and are then
2273 * incremented by CONFIG_NR_CPUS.
2275 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2278 * No preemption supported therefore also no need to check for
2284 static inline unsigned long next_tid(unsigned long tid)
2286 return tid + TID_STEP;
2289 #ifdef SLUB_DEBUG_CMPXCHG
2290 static inline unsigned int tid_to_cpu(unsigned long tid)
2292 return tid % TID_STEP;
2295 static inline unsigned long tid_to_event(unsigned long tid)
2297 return tid / TID_STEP;
2301 static inline unsigned int init_tid(int cpu)
2306 static inline void note_cmpxchg_failure(const char *n,
2307 const struct kmem_cache *s, unsigned long tid)
2309 #ifdef SLUB_DEBUG_CMPXCHG
2310 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2312 pr_info("%s %s: cmpxchg redo ", n, s->name);
2314 #ifdef CONFIG_PREEMPTION
2315 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2316 pr_warn("due to cpu change %d -> %d\n",
2317 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2320 if (tid_to_event(tid) != tid_to_event(actual_tid))
2321 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2322 tid_to_event(tid), tid_to_event(actual_tid));
2324 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2325 actual_tid, tid, next_tid(tid));
2327 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2330 static void init_kmem_cache_cpus(struct kmem_cache *s)
2333 struct kmem_cache_cpu *c;
2335 for_each_possible_cpu(cpu) {
2336 c = per_cpu_ptr(s->cpu_slab, cpu);
2337 local_lock_init(&c->lock);
2338 c->tid = init_tid(cpu);
2343 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2344 * unfreezes the slabs and puts it on the proper list.
2345 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2348 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2351 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE, M_FULL_NOLIST };
2352 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2354 enum slab_modes mode = M_NONE;
2355 void *nextfree, *freelist_iter, *freelist_tail;
2356 int tail = DEACTIVATE_TO_HEAD;
2357 unsigned long flags = 0;
2361 if (slab->freelist) {
2362 stat(s, DEACTIVATE_REMOTE_FREES);
2363 tail = DEACTIVATE_TO_TAIL;
2367 * Stage one: Count the objects on cpu's freelist as free_delta and
2368 * remember the last object in freelist_tail for later splicing.
2370 freelist_tail = NULL;
2371 freelist_iter = freelist;
2372 while (freelist_iter) {
2373 nextfree = get_freepointer(s, freelist_iter);
2376 * If 'nextfree' is invalid, it is possible that the object at
2377 * 'freelist_iter' is already corrupted. So isolate all objects
2378 * starting at 'freelist_iter' by skipping them.
2380 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2383 freelist_tail = freelist_iter;
2386 freelist_iter = nextfree;
2390 * Stage two: Unfreeze the slab while splicing the per-cpu
2391 * freelist to the head of slab's freelist.
2393 * Ensure that the slab is unfrozen while the list presence
2394 * reflects the actual number of objects during unfreeze.
2396 * We first perform cmpxchg holding lock and insert to list
2397 * when it succeed. If there is mismatch then the slab is not
2398 * unfrozen and number of objects in the slab may have changed.
2399 * Then release lock and retry cmpxchg again.
2403 old.freelist = READ_ONCE(slab->freelist);
2404 old.counters = READ_ONCE(slab->counters);
2405 VM_BUG_ON(!old.frozen);
2407 /* Determine target state of the slab */
2408 new.counters = old.counters;
2409 if (freelist_tail) {
2410 new.inuse -= free_delta;
2411 set_freepointer(s, freelist_tail, old.freelist);
2412 new.freelist = freelist;
2414 new.freelist = old.freelist;
2418 if (!new.inuse && n->nr_partial >= s->min_partial) {
2420 } else if (new.freelist) {
2423 * Taking the spinlock removes the possibility that
2424 * acquire_slab() will see a slab that is frozen
2426 spin_lock_irqsave(&n->list_lock, flags);
2427 } else if (kmem_cache_debug_flags(s, SLAB_STORE_USER)) {
2430 * This also ensures that the scanning of full
2431 * slabs from diagnostic functions will not see
2434 spin_lock_irqsave(&n->list_lock, flags);
2436 mode = M_FULL_NOLIST;
2440 if (!cmpxchg_double_slab(s, slab,
2441 old.freelist, old.counters,
2442 new.freelist, new.counters,
2443 "unfreezing slab")) {
2444 if (mode == M_PARTIAL || mode == M_FULL)
2445 spin_unlock_irqrestore(&n->list_lock, flags);
2450 if (mode == M_PARTIAL) {
2451 add_partial(n, slab, tail);
2452 spin_unlock_irqrestore(&n->list_lock, flags);
2454 } else if (mode == M_FREE) {
2455 stat(s, DEACTIVATE_EMPTY);
2456 discard_slab(s, slab);
2458 } else if (mode == M_FULL) {
2459 add_full(s, n, slab);
2460 spin_unlock_irqrestore(&n->list_lock, flags);
2461 stat(s, DEACTIVATE_FULL);
2462 } else if (mode == M_FULL_NOLIST) {
2463 stat(s, DEACTIVATE_FULL);
2467 #ifdef CONFIG_SLUB_CPU_PARTIAL
2468 static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2470 struct kmem_cache_node *n = NULL, *n2 = NULL;
2471 struct slab *slab, *slab_to_discard = NULL;
2472 unsigned long flags = 0;
2474 while (partial_slab) {
2478 slab = partial_slab;
2479 partial_slab = slab->next;
2481 n2 = get_node(s, slab_nid(slab));
2484 spin_unlock_irqrestore(&n->list_lock, flags);
2487 spin_lock_irqsave(&n->list_lock, flags);
2492 old.freelist = slab->freelist;
2493 old.counters = slab->counters;
2494 VM_BUG_ON(!old.frozen);
2496 new.counters = old.counters;
2497 new.freelist = old.freelist;
2501 } while (!__cmpxchg_double_slab(s, slab,
2502 old.freelist, old.counters,
2503 new.freelist, new.counters,
2504 "unfreezing slab"));
2506 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2507 slab->next = slab_to_discard;
2508 slab_to_discard = slab;
2510 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2511 stat(s, FREE_ADD_PARTIAL);
2516 spin_unlock_irqrestore(&n->list_lock, flags);
2518 while (slab_to_discard) {
2519 slab = slab_to_discard;
2520 slab_to_discard = slab_to_discard->next;
2522 stat(s, DEACTIVATE_EMPTY);
2523 discard_slab(s, slab);
2529 * Unfreeze all the cpu partial slabs.
2531 static void unfreeze_partials(struct kmem_cache *s)
2533 struct slab *partial_slab;
2534 unsigned long flags;
2536 local_lock_irqsave(&s->cpu_slab->lock, flags);
2537 partial_slab = this_cpu_read(s->cpu_slab->partial);
2538 this_cpu_write(s->cpu_slab->partial, NULL);
2539 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2542 __unfreeze_partials(s, partial_slab);
2545 static void unfreeze_partials_cpu(struct kmem_cache *s,
2546 struct kmem_cache_cpu *c)
2548 struct slab *partial_slab;
2550 partial_slab = slub_percpu_partial(c);
2554 __unfreeze_partials(s, partial_slab);
2558 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2559 * partial slab slot if available.
2561 * If we did not find a slot then simply move all the partials to the
2562 * per node partial list.
2564 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2566 struct slab *oldslab;
2567 struct slab *slab_to_unfreeze = NULL;
2568 unsigned long flags;
2571 local_lock_irqsave(&s->cpu_slab->lock, flags);
2573 oldslab = this_cpu_read(s->cpu_slab->partial);
2576 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2578 * Partial array is full. Move the existing set to the
2579 * per node partial list. Postpone the actual unfreezing
2580 * outside of the critical section.
2582 slab_to_unfreeze = oldslab;
2585 slabs = oldslab->slabs;
2591 slab->slabs = slabs;
2592 slab->next = oldslab;
2594 this_cpu_write(s->cpu_slab->partial, slab);
2596 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2598 if (slab_to_unfreeze) {
2599 __unfreeze_partials(s, slab_to_unfreeze);
2600 stat(s, CPU_PARTIAL_DRAIN);
2604 #else /* CONFIG_SLUB_CPU_PARTIAL */
2606 static inline void unfreeze_partials(struct kmem_cache *s) { }
2607 static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2608 struct kmem_cache_cpu *c) { }
2610 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2612 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2614 unsigned long flags;
2618 local_lock_irqsave(&s->cpu_slab->lock, flags);
2621 freelist = c->freelist;
2625 c->tid = next_tid(c->tid);
2627 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2630 deactivate_slab(s, slab, freelist);
2631 stat(s, CPUSLAB_FLUSH);
2635 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2637 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2638 void *freelist = c->freelist;
2639 struct slab *slab = c->slab;
2643 c->tid = next_tid(c->tid);
2646 deactivate_slab(s, slab, freelist);
2647 stat(s, CPUSLAB_FLUSH);
2650 unfreeze_partials_cpu(s, c);
2653 struct slub_flush_work {
2654 struct work_struct work;
2655 struct kmem_cache *s;
2662 * Called from CPU work handler with migration disabled.
2664 static void flush_cpu_slab(struct work_struct *w)
2666 struct kmem_cache *s;
2667 struct kmem_cache_cpu *c;
2668 struct slub_flush_work *sfw;
2670 sfw = container_of(w, struct slub_flush_work, work);
2673 c = this_cpu_ptr(s->cpu_slab);
2678 unfreeze_partials(s);
2681 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2683 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2685 return c->slab || slub_percpu_partial(c);
2688 static DEFINE_MUTEX(flush_lock);
2689 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2691 static void flush_all_cpus_locked(struct kmem_cache *s)
2693 struct slub_flush_work *sfw;
2696 lockdep_assert_cpus_held();
2697 mutex_lock(&flush_lock);
2699 for_each_online_cpu(cpu) {
2700 sfw = &per_cpu(slub_flush, cpu);
2701 if (!has_cpu_slab(cpu, s)) {
2705 INIT_WORK(&sfw->work, flush_cpu_slab);
2708 schedule_work_on(cpu, &sfw->work);
2711 for_each_online_cpu(cpu) {
2712 sfw = &per_cpu(slub_flush, cpu);
2715 flush_work(&sfw->work);
2718 mutex_unlock(&flush_lock);
2721 static void flush_all(struct kmem_cache *s)
2724 flush_all_cpus_locked(s);
2729 * Use the cpu notifier to insure that the cpu slabs are flushed when
2732 static int slub_cpu_dead(unsigned int cpu)
2734 struct kmem_cache *s;
2736 mutex_lock(&slab_mutex);
2737 list_for_each_entry(s, &slab_caches, list)
2738 __flush_cpu_slab(s, cpu);
2739 mutex_unlock(&slab_mutex);
2744 * Check if the objects in a per cpu structure fit numa
2745 * locality expectations.
2747 static inline int node_match(struct slab *slab, int node)
2750 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2756 #ifdef CONFIG_SLUB_DEBUG
2757 static int count_free(struct slab *slab)
2759 return slab->objects - slab->inuse;
2762 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2764 return atomic_long_read(&n->total_objects);
2766 #endif /* CONFIG_SLUB_DEBUG */
2768 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2769 static unsigned long count_partial(struct kmem_cache_node *n,
2770 int (*get_count)(struct slab *))
2772 unsigned long flags;
2773 unsigned long x = 0;
2776 spin_lock_irqsave(&n->list_lock, flags);
2777 list_for_each_entry(slab, &n->partial, slab_list)
2778 x += get_count(slab);
2779 spin_unlock_irqrestore(&n->list_lock, flags);
2782 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2784 static noinline void
2785 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2787 #ifdef CONFIG_SLUB_DEBUG
2788 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2789 DEFAULT_RATELIMIT_BURST);
2791 struct kmem_cache_node *n;
2793 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2796 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2797 nid, gfpflags, &gfpflags);
2798 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2799 s->name, s->object_size, s->size, oo_order(s->oo),
2802 if (oo_order(s->min) > get_order(s->object_size))
2803 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2806 for_each_kmem_cache_node(s, node, n) {
2807 unsigned long nr_slabs;
2808 unsigned long nr_objs;
2809 unsigned long nr_free;
2811 nr_free = count_partial(n, count_free);
2812 nr_slabs = node_nr_slabs(n);
2813 nr_objs = node_nr_objs(n);
2815 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2816 node, nr_slabs, nr_objs, nr_free);
2821 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
2823 if (unlikely(slab_test_pfmemalloc(slab)))
2824 return gfp_pfmemalloc_allowed(gfpflags);
2830 * Check the slab->freelist and either transfer the freelist to the
2831 * per cpu freelist or deactivate the slab.
2833 * The slab is still frozen if the return value is not NULL.
2835 * If this function returns NULL then the slab has been unfrozen.
2837 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
2840 unsigned long counters;
2843 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
2846 freelist = slab->freelist;
2847 counters = slab->counters;
2849 new.counters = counters;
2850 VM_BUG_ON(!new.frozen);
2852 new.inuse = slab->objects;
2853 new.frozen = freelist != NULL;
2855 } while (!__cmpxchg_double_slab(s, slab,
2864 * Slow path. The lockless freelist is empty or we need to perform
2867 * Processing is still very fast if new objects have been freed to the
2868 * regular freelist. In that case we simply take over the regular freelist
2869 * as the lockless freelist and zap the regular freelist.
2871 * If that is not working then we fall back to the partial lists. We take the
2872 * first element of the freelist as the object to allocate now and move the
2873 * rest of the freelist to the lockless freelist.
2875 * And if we were unable to get a new slab from the partial slab lists then
2876 * we need to allocate a new slab. This is the slowest path since it involves
2877 * a call to the page allocator and the setup of a new slab.
2879 * Version of __slab_alloc to use when we know that preemption is
2880 * already disabled (which is the case for bulk allocation).
2882 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2883 unsigned long addr, struct kmem_cache_cpu *c)
2887 unsigned long flags;
2889 stat(s, ALLOC_SLOWPATH);
2893 slab = READ_ONCE(c->slab);
2896 * if the node is not online or has no normal memory, just
2897 * ignore the node constraint
2899 if (unlikely(node != NUMA_NO_NODE &&
2900 !node_isset(node, slab_nodes)))
2901 node = NUMA_NO_NODE;
2906 if (unlikely(!node_match(slab, node))) {
2908 * same as above but node_match() being false already
2909 * implies node != NUMA_NO_NODE
2911 if (!node_isset(node, slab_nodes)) {
2912 node = NUMA_NO_NODE;
2915 stat(s, ALLOC_NODE_MISMATCH);
2916 goto deactivate_slab;
2921 * By rights, we should be searching for a slab page that was
2922 * PFMEMALLOC but right now, we are losing the pfmemalloc
2923 * information when the page leaves the per-cpu allocator
2925 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
2926 goto deactivate_slab;
2928 /* must check again c->slab in case we got preempted and it changed */
2929 local_lock_irqsave(&s->cpu_slab->lock, flags);
2930 if (unlikely(slab != c->slab)) {
2931 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2934 freelist = c->freelist;
2938 freelist = get_freelist(s, slab);
2942 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2943 stat(s, DEACTIVATE_BYPASS);
2947 stat(s, ALLOC_REFILL);
2951 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
2954 * freelist is pointing to the list of objects to be used.
2955 * slab is pointing to the slab from which the objects are obtained.
2956 * That slab must be frozen for per cpu allocations to work.
2958 VM_BUG_ON(!c->slab->frozen);
2959 c->freelist = get_freepointer(s, freelist);
2960 c->tid = next_tid(c->tid);
2961 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2966 local_lock_irqsave(&s->cpu_slab->lock, flags);
2967 if (slab != c->slab) {
2968 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2971 freelist = c->freelist;
2974 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2975 deactivate_slab(s, slab, freelist);
2979 if (slub_percpu_partial(c)) {
2980 local_lock_irqsave(&s->cpu_slab->lock, flags);
2981 if (unlikely(c->slab)) {
2982 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2985 if (unlikely(!slub_percpu_partial(c))) {
2986 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2987 /* we were preempted and partial list got empty */
2991 slab = c->slab = slub_percpu_partial(c);
2992 slub_set_percpu_partial(c, slab);
2993 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2994 stat(s, CPU_PARTIAL_ALLOC);
3000 freelist = get_partial(s, gfpflags, node, &slab);
3002 goto check_new_slab;
3004 slub_put_cpu_ptr(s->cpu_slab);
3005 slab = new_slab(s, gfpflags, node);
3006 c = slub_get_cpu_ptr(s->cpu_slab);
3008 if (unlikely(!slab)) {
3009 slab_out_of_memory(s, gfpflags, node);
3014 * No other reference to the slab yet so we can
3015 * muck around with it freely without cmpxchg
3017 freelist = slab->freelist;
3018 slab->freelist = NULL;
3020 stat(s, ALLOC_SLAB);
3024 if (kmem_cache_debug(s)) {
3025 if (!alloc_debug_processing(s, slab, freelist, addr)) {
3026 /* Slab failed checks. Next slab needed */
3030 * For debug case, we don't load freelist so that all
3031 * allocations go through alloc_debug_processing()
3037 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3039 * For !pfmemalloc_match() case we don't load freelist so that
3040 * we don't make further mismatched allocations easier.
3046 local_lock_irqsave(&s->cpu_slab->lock, flags);
3047 if (unlikely(c->slab)) {
3048 void *flush_freelist = c->freelist;
3049 struct slab *flush_slab = c->slab;
3053 c->tid = next_tid(c->tid);
3055 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3057 deactivate_slab(s, flush_slab, flush_freelist);
3059 stat(s, CPUSLAB_FLUSH);
3061 goto retry_load_slab;
3069 deactivate_slab(s, slab, get_freepointer(s, freelist));
3074 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3075 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3078 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3079 unsigned long addr, struct kmem_cache_cpu *c)
3083 #ifdef CONFIG_PREEMPT_COUNT
3085 * We may have been preempted and rescheduled on a different
3086 * cpu before disabling preemption. Need to reload cpu area
3089 c = slub_get_cpu_ptr(s->cpu_slab);
3092 p = ___slab_alloc(s, gfpflags, node, addr, c);
3093 #ifdef CONFIG_PREEMPT_COUNT
3094 slub_put_cpu_ptr(s->cpu_slab);
3100 * If the object has been wiped upon free, make sure it's fully initialized by
3101 * zeroing out freelist pointer.
3103 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3106 if (unlikely(slab_want_init_on_free(s)) && obj)
3107 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3112 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3113 * have the fastpath folded into their functions. So no function call
3114 * overhead for requests that can be satisfied on the fastpath.
3116 * The fastpath works by first checking if the lockless freelist can be used.
3117 * If not then __slab_alloc is called for slow processing.
3119 * Otherwise we can simply pick the next object from the lockless free list.
3121 static __always_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3122 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3125 struct kmem_cache_cpu *c;
3128 struct obj_cgroup *objcg = NULL;
3131 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3135 object = kfence_alloc(s, orig_size, gfpflags);
3136 if (unlikely(object))
3141 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3142 * enabled. We may switch back and forth between cpus while
3143 * reading from one cpu area. That does not matter as long
3144 * as we end up on the original cpu again when doing the cmpxchg.
3146 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3147 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3148 * the tid. If we are preempted and switched to another cpu between the
3149 * two reads, it's OK as the two are still associated with the same cpu
3150 * and cmpxchg later will validate the cpu.
3152 c = raw_cpu_ptr(s->cpu_slab);
3153 tid = READ_ONCE(c->tid);
3156 * Irqless object alloc/free algorithm used here depends on sequence
3157 * of fetching cpu_slab's data. tid should be fetched before anything
3158 * on c to guarantee that object and slab associated with previous tid
3159 * won't be used with current tid. If we fetch tid first, object and
3160 * slab could be one associated with next tid and our alloc/free
3161 * request will be failed. In this case, we will retry. So, no problem.
3166 * The transaction ids are globally unique per cpu and per operation on
3167 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3168 * occurs on the right processor and that there was no operation on the
3169 * linked list in between.
3172 object = c->freelist;
3175 * We cannot use the lockless fastpath on PREEMPT_RT because if a
3176 * slowpath has taken the local_lock_irqsave(), it is not protected
3177 * against a fast path operation in an irq handler. So we need to take
3178 * the slow path which uses local_lock. It is still relatively fast if
3179 * there is a suitable cpu freelist.
3181 if (IS_ENABLED(CONFIG_PREEMPT_RT) ||
3182 unlikely(!object || !slab || !node_match(slab, node))) {
3183 object = __slab_alloc(s, gfpflags, node, addr, c);
3185 void *next_object = get_freepointer_safe(s, object);
3188 * The cmpxchg will only match if there was no additional
3189 * operation and if we are on the right processor.
3191 * The cmpxchg does the following atomically (without lock
3193 * 1. Relocate first pointer to the current per cpu area.
3194 * 2. Verify that tid and freelist have not been changed
3195 * 3. If they were not changed replace tid and freelist
3197 * Since this is without lock semantics the protection is only
3198 * against code executing on this cpu *not* from access by
3201 if (unlikely(!this_cpu_cmpxchg_double(
3202 s->cpu_slab->freelist, s->cpu_slab->tid,
3204 next_object, next_tid(tid)))) {
3206 note_cmpxchg_failure("slab_alloc", s, tid);
3209 prefetch_freepointer(s, next_object);
3210 stat(s, ALLOC_FASTPATH);
3213 maybe_wipe_obj_freeptr(s, object);
3214 init = slab_want_init_on_alloc(gfpflags, s);
3217 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
3222 static __always_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3223 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3225 return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3228 static __always_inline
3229 void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3232 void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3234 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
3240 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3242 return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3244 EXPORT_SYMBOL(kmem_cache_alloc);
3246 void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3249 return __kmem_cache_alloc_lru(s, lru, gfpflags);
3251 EXPORT_SYMBOL(kmem_cache_alloc_lru);
3253 #ifdef CONFIG_TRACING
3254 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3256 void *ret = slab_alloc(s, NULL, gfpflags, _RET_IP_, size);
3257 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
3258 ret = kasan_kmalloc(s, ret, size, gfpflags);
3261 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3265 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3267 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3269 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3270 s->object_size, s->size, gfpflags, node);
3274 EXPORT_SYMBOL(kmem_cache_alloc_node);
3276 #ifdef CONFIG_TRACING
3277 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3279 int node, size_t size)
3281 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
3283 trace_kmalloc_node(_RET_IP_, ret,
3284 size, s->size, gfpflags, node);
3286 ret = kasan_kmalloc(s, ret, size, gfpflags);
3289 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3291 #endif /* CONFIG_NUMA */
3294 * Slow path handling. This may still be called frequently since objects
3295 * have a longer lifetime than the cpu slabs in most processing loads.
3297 * So we still attempt to reduce cache line usage. Just take the slab
3298 * lock and free the item. If there is no additional partial slab
3299 * handling required then we can return immediately.
3301 static void __slab_free(struct kmem_cache *s, struct slab *slab,
3302 void *head, void *tail, int cnt,
3309 unsigned long counters;
3310 struct kmem_cache_node *n = NULL;
3311 unsigned long flags;
3313 stat(s, FREE_SLOWPATH);
3315 if (kfence_free(head))
3318 if (kmem_cache_debug(s) &&
3319 !free_debug_processing(s, slab, head, tail, cnt, addr))
3324 spin_unlock_irqrestore(&n->list_lock, flags);
3327 prior = slab->freelist;
3328 counters = slab->counters;
3329 set_freepointer(s, tail, prior);
3330 new.counters = counters;
3331 was_frozen = new.frozen;
3333 if ((!new.inuse || !prior) && !was_frozen) {
3335 if (kmem_cache_has_cpu_partial(s) && !prior) {
3338 * Slab was on no list before and will be
3340 * We can defer the list move and instead
3345 } else { /* Needs to be taken off a list */
3347 n = get_node(s, slab_nid(slab));
3349 * Speculatively acquire the list_lock.
3350 * If the cmpxchg does not succeed then we may
3351 * drop the list_lock without any processing.
3353 * Otherwise the list_lock will synchronize with
3354 * other processors updating the list of slabs.
3356 spin_lock_irqsave(&n->list_lock, flags);
3361 } while (!cmpxchg_double_slab(s, slab,
3368 if (likely(was_frozen)) {
3370 * The list lock was not taken therefore no list
3371 * activity can be necessary.
3373 stat(s, FREE_FROZEN);
3374 } else if (new.frozen) {
3376 * If we just froze the slab then put it onto the
3377 * per cpu partial list.
3379 put_cpu_partial(s, slab, 1);
3380 stat(s, CPU_PARTIAL_FREE);
3386 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3390 * Objects left in the slab. If it was not on the partial list before
3393 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3394 remove_full(s, n, slab);
3395 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3396 stat(s, FREE_ADD_PARTIAL);
3398 spin_unlock_irqrestore(&n->list_lock, flags);
3404 * Slab on the partial list.
3406 remove_partial(n, slab);
3407 stat(s, FREE_REMOVE_PARTIAL);
3409 /* Slab must be on the full list */
3410 remove_full(s, n, slab);
3413 spin_unlock_irqrestore(&n->list_lock, flags);
3415 discard_slab(s, slab);
3419 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3420 * can perform fastpath freeing without additional function calls.
3422 * The fastpath is only possible if we are freeing to the current cpu slab
3423 * of this processor. This typically the case if we have just allocated
3426 * If fastpath is not possible then fall back to __slab_free where we deal
3427 * with all sorts of special processing.
3429 * Bulk free of a freelist with several objects (all pointing to the
3430 * same slab) possible by specifying head and tail ptr, plus objects
3431 * count (cnt). Bulk free indicated by tail pointer being set.
3433 static __always_inline void do_slab_free(struct kmem_cache *s,
3434 struct slab *slab, void *head, void *tail,
3435 int cnt, unsigned long addr)
3437 void *tail_obj = tail ? : head;
3438 struct kmem_cache_cpu *c;
3441 /* memcg_slab_free_hook() is already called for bulk free. */
3443 memcg_slab_free_hook(s, &head, 1);
3446 * Determine the currently cpus per cpu slab.
3447 * The cpu may change afterward. However that does not matter since
3448 * data is retrieved via this pointer. If we are on the same cpu
3449 * during the cmpxchg then the free will succeed.
3451 c = raw_cpu_ptr(s->cpu_slab);
3452 tid = READ_ONCE(c->tid);
3454 /* Same with comment on barrier() in slab_alloc_node() */
3457 if (likely(slab == c->slab)) {
3458 #ifndef CONFIG_PREEMPT_RT
3459 void **freelist = READ_ONCE(c->freelist);
3461 set_freepointer(s, tail_obj, freelist);
3463 if (unlikely(!this_cpu_cmpxchg_double(
3464 s->cpu_slab->freelist, s->cpu_slab->tid,
3466 head, next_tid(tid)))) {
3468 note_cmpxchg_failure("slab_free", s, tid);
3471 #else /* CONFIG_PREEMPT_RT */
3473 * We cannot use the lockless fastpath on PREEMPT_RT because if
3474 * a slowpath has taken the local_lock_irqsave(), it is not
3475 * protected against a fast path operation in an irq handler. So
3476 * we need to take the local_lock. We shouldn't simply defer to
3477 * __slab_free() as that wouldn't use the cpu freelist at all.
3481 local_lock(&s->cpu_slab->lock);
3482 c = this_cpu_ptr(s->cpu_slab);
3483 if (unlikely(slab != c->slab)) {
3484 local_unlock(&s->cpu_slab->lock);
3488 freelist = c->freelist;
3490 set_freepointer(s, tail_obj, freelist);
3492 c->tid = next_tid(tid);
3494 local_unlock(&s->cpu_slab->lock);
3496 stat(s, FREE_FASTPATH);
3498 __slab_free(s, slab, head, tail_obj, cnt, addr);
3502 static __always_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3503 void *head, void *tail, int cnt,
3507 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3508 * to remove objects, whose reuse must be delayed.
3510 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3511 do_slab_free(s, slab, head, tail, cnt, addr);
3514 #ifdef CONFIG_KASAN_GENERIC
3515 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3517 do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3521 void kmem_cache_free(struct kmem_cache *s, void *x)
3523 s = cache_from_obj(s, x);
3526 trace_kmem_cache_free(_RET_IP_, x, s->name);
3527 slab_free(s, virt_to_slab(x), x, NULL, 1, _RET_IP_);
3529 EXPORT_SYMBOL(kmem_cache_free);
3531 struct detached_freelist {
3536 struct kmem_cache *s;
3539 static inline void free_large_kmalloc(struct folio *folio, void *object)
3541 unsigned int order = folio_order(folio);
3543 if (WARN_ON_ONCE(order == 0))
3544 pr_warn_once("object pointer: 0x%p\n", object);
3547 mod_lruvec_page_state(folio_page(folio, 0), NR_SLAB_UNRECLAIMABLE_B,
3548 -(PAGE_SIZE << order));
3549 __free_pages(folio_page(folio, 0), order);
3553 * This function progressively scans the array with free objects (with
3554 * a limited look ahead) and extract objects belonging to the same
3555 * slab. It builds a detached freelist directly within the given
3556 * slab/objects. This can happen without any need for
3557 * synchronization, because the objects are owned by running process.
3558 * The freelist is build up as a single linked list in the objects.
3559 * The idea is, that this detached freelist can then be bulk
3560 * transferred to the real freelist(s), but only requiring a single
3561 * synchronization primitive. Look ahead in the array is limited due
3562 * to performance reasons.
3565 int build_detached_freelist(struct kmem_cache *s, size_t size,
3566 void **p, struct detached_freelist *df)
3568 size_t first_skipped_index = 0;
3571 struct folio *folio;
3574 /* Always re-init detached_freelist */
3579 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3580 } while (!object && size);
3585 folio = virt_to_folio(object);
3587 /* Handle kalloc'ed objects */
3588 if (unlikely(!folio_test_slab(folio))) {
3589 free_large_kmalloc(folio, object);
3590 p[size] = NULL; /* mark object processed */
3593 /* Derive kmem_cache from object */
3594 slab = folio_slab(folio);
3595 df->s = slab->slab_cache;
3597 slab = folio_slab(folio);
3598 df->s = cache_from_obj(s, object); /* Support for memcg */
3601 if (is_kfence_address(object)) {
3602 slab_free_hook(df->s, object, false);
3603 __kfence_free(object);
3604 p[size] = NULL; /* mark object processed */
3608 /* Start new detached freelist */
3610 set_freepointer(df->s, object, NULL);
3612 df->freelist = object;
3613 p[size] = NULL; /* mark object processed */
3619 continue; /* Skip processed objects */
3621 /* df->slab is always set at this point */
3622 if (df->slab == virt_to_slab(object)) {
3623 /* Opportunity build freelist */
3624 set_freepointer(df->s, object, df->freelist);
3625 df->freelist = object;
3627 p[size] = NULL; /* mark object processed */
3632 /* Limit look ahead search */
3636 if (!first_skipped_index)
3637 first_skipped_index = size + 1;
3640 return first_skipped_index;
3643 /* Note that interrupts must be enabled when calling this function. */
3644 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3649 memcg_slab_free_hook(s, p, size);
3651 struct detached_freelist df;
3653 size = build_detached_freelist(s, size, p, &df);
3657 slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt, _RET_IP_);
3658 } while (likely(size));
3660 EXPORT_SYMBOL(kmem_cache_free_bulk);
3662 /* Note that interrupts must be enabled when calling this function. */
3663 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3666 struct kmem_cache_cpu *c;
3668 struct obj_cgroup *objcg = NULL;
3670 /* memcg and kmem_cache debug support */
3671 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
3675 * Drain objects in the per cpu slab, while disabling local
3676 * IRQs, which protects against PREEMPT and interrupts
3677 * handlers invoking normal fastpath.
3679 c = slub_get_cpu_ptr(s->cpu_slab);
3680 local_lock_irq(&s->cpu_slab->lock);
3682 for (i = 0; i < size; i++) {
3683 void *object = kfence_alloc(s, s->object_size, flags);
3685 if (unlikely(object)) {
3690 object = c->freelist;
3691 if (unlikely(!object)) {
3693 * We may have removed an object from c->freelist using
3694 * the fastpath in the previous iteration; in that case,
3695 * c->tid has not been bumped yet.
3696 * Since ___slab_alloc() may reenable interrupts while
3697 * allocating memory, we should bump c->tid now.
3699 c->tid = next_tid(c->tid);
3701 local_unlock_irq(&s->cpu_slab->lock);
3704 * Invoking slow path likely have side-effect
3705 * of re-populating per CPU c->freelist
3707 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3709 if (unlikely(!p[i]))
3712 c = this_cpu_ptr(s->cpu_slab);
3713 maybe_wipe_obj_freeptr(s, p[i]);
3715 local_lock_irq(&s->cpu_slab->lock);
3717 continue; /* goto for-loop */
3719 c->freelist = get_freepointer(s, object);
3721 maybe_wipe_obj_freeptr(s, p[i]);
3723 c->tid = next_tid(c->tid);
3724 local_unlock_irq(&s->cpu_slab->lock);
3725 slub_put_cpu_ptr(s->cpu_slab);
3728 * memcg and kmem_cache debug support and memory initialization.
3729 * Done outside of the IRQ disabled fastpath loop.
3731 slab_post_alloc_hook(s, objcg, flags, size, p,
3732 slab_want_init_on_alloc(flags, s));
3735 slub_put_cpu_ptr(s->cpu_slab);
3736 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3737 __kmem_cache_free_bulk(s, i, p);
3740 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3744 * Object placement in a slab is made very easy because we always start at
3745 * offset 0. If we tune the size of the object to the alignment then we can
3746 * get the required alignment by putting one properly sized object after
3749 * Notice that the allocation order determines the sizes of the per cpu
3750 * caches. Each processor has always one slab available for allocations.
3751 * Increasing the allocation order reduces the number of times that slabs
3752 * must be moved on and off the partial lists and is therefore a factor in
3757 * Minimum / Maximum order of slab pages. This influences locking overhead
3758 * and slab fragmentation. A higher order reduces the number of partial slabs
3759 * and increases the number of allocations possible without having to
3760 * take the list_lock.
3762 static unsigned int slub_min_order;
3763 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3764 static unsigned int slub_min_objects;
3767 * Calculate the order of allocation given an slab object size.
3769 * The order of allocation has significant impact on performance and other
3770 * system components. Generally order 0 allocations should be preferred since
3771 * order 0 does not cause fragmentation in the page allocator. Larger objects
3772 * be problematic to put into order 0 slabs because there may be too much
3773 * unused space left. We go to a higher order if more than 1/16th of the slab
3776 * In order to reach satisfactory performance we must ensure that a minimum
3777 * number of objects is in one slab. Otherwise we may generate too much
3778 * activity on the partial lists which requires taking the list_lock. This is
3779 * less a concern for large slabs though which are rarely used.
3781 * slub_max_order specifies the order where we begin to stop considering the
3782 * number of objects in a slab as critical. If we reach slub_max_order then
3783 * we try to keep the page order as low as possible. So we accept more waste
3784 * of space in favor of a small page order.
3786 * Higher order allocations also allow the placement of more objects in a
3787 * slab and thereby reduce object handling overhead. If the user has
3788 * requested a higher minimum order then we start with that one instead of
3789 * the smallest order which will fit the object.
3791 static inline unsigned int calc_slab_order(unsigned int size,
3792 unsigned int min_objects, unsigned int max_order,
3793 unsigned int fract_leftover)
3795 unsigned int min_order = slub_min_order;
3798 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3799 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3801 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3802 order <= max_order; order++) {
3804 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3807 rem = slab_size % size;
3809 if (rem <= slab_size / fract_leftover)
3816 static inline int calculate_order(unsigned int size)
3819 unsigned int min_objects;
3820 unsigned int max_objects;
3821 unsigned int nr_cpus;
3824 * Attempt to find best configuration for a slab. This
3825 * works by first attempting to generate a layout with
3826 * the best configuration and backing off gradually.
3828 * First we increase the acceptable waste in a slab. Then
3829 * we reduce the minimum objects required in a slab.
3831 min_objects = slub_min_objects;
3834 * Some architectures will only update present cpus when
3835 * onlining them, so don't trust the number if it's just 1. But
3836 * we also don't want to use nr_cpu_ids always, as on some other
3837 * architectures, there can be many possible cpus, but never
3838 * onlined. Here we compromise between trying to avoid too high
3839 * order on systems that appear larger than they are, and too
3840 * low order on systems that appear smaller than they are.
3842 nr_cpus = num_present_cpus();
3844 nr_cpus = nr_cpu_ids;
3845 min_objects = 4 * (fls(nr_cpus) + 1);
3847 max_objects = order_objects(slub_max_order, size);
3848 min_objects = min(min_objects, max_objects);
3850 while (min_objects > 1) {
3851 unsigned int fraction;
3854 while (fraction >= 4) {
3855 order = calc_slab_order(size, min_objects,
3856 slub_max_order, fraction);
3857 if (order <= slub_max_order)
3865 * We were unable to place multiple objects in a slab. Now
3866 * lets see if we can place a single object there.
3868 order = calc_slab_order(size, 1, slub_max_order, 1);
3869 if (order <= slub_max_order)
3873 * Doh this slab cannot be placed using slub_max_order.
3875 order = calc_slab_order(size, 1, MAX_ORDER, 1);
3876 if (order < MAX_ORDER)
3882 init_kmem_cache_node(struct kmem_cache_node *n)
3885 spin_lock_init(&n->list_lock);
3886 INIT_LIST_HEAD(&n->partial);
3887 #ifdef CONFIG_SLUB_DEBUG
3888 atomic_long_set(&n->nr_slabs, 0);
3889 atomic_long_set(&n->total_objects, 0);
3890 INIT_LIST_HEAD(&n->full);
3894 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3896 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3897 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3900 * Must align to double word boundary for the double cmpxchg
3901 * instructions to work; see __pcpu_double_call_return_bool().
3903 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3904 2 * sizeof(void *));
3909 init_kmem_cache_cpus(s);
3914 static struct kmem_cache *kmem_cache_node;
3917 * No kmalloc_node yet so do it by hand. We know that this is the first
3918 * slab on the node for this slabcache. There are no concurrent accesses
3921 * Note that this function only works on the kmem_cache_node
3922 * when allocating for the kmem_cache_node. This is used for bootstrapping
3923 * memory on a fresh node that has no slab structures yet.
3925 static void early_kmem_cache_node_alloc(int node)
3928 struct kmem_cache_node *n;
3930 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3932 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3935 if (slab_nid(slab) != node) {
3936 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3937 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3942 #ifdef CONFIG_SLUB_DEBUG
3943 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3944 init_tracking(kmem_cache_node, n);
3946 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3947 slab->freelist = get_freepointer(kmem_cache_node, n);
3950 kmem_cache_node->node[node] = n;
3951 init_kmem_cache_node(n);
3952 inc_slabs_node(kmem_cache_node, node, slab->objects);
3955 * No locks need to be taken here as it has just been
3956 * initialized and there is no concurrent access.
3958 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
3961 static void free_kmem_cache_nodes(struct kmem_cache *s)
3964 struct kmem_cache_node *n;
3966 for_each_kmem_cache_node(s, node, n) {
3967 s->node[node] = NULL;
3968 kmem_cache_free(kmem_cache_node, n);
3972 void __kmem_cache_release(struct kmem_cache *s)
3974 cache_random_seq_destroy(s);
3975 free_percpu(s->cpu_slab);
3976 free_kmem_cache_nodes(s);
3979 static int init_kmem_cache_nodes(struct kmem_cache *s)
3983 for_each_node_mask(node, slab_nodes) {
3984 struct kmem_cache_node *n;
3986 if (slab_state == DOWN) {
3987 early_kmem_cache_node_alloc(node);
3990 n = kmem_cache_alloc_node(kmem_cache_node,
3994 free_kmem_cache_nodes(s);
3998 init_kmem_cache_node(n);
4004 static void set_cpu_partial(struct kmem_cache *s)
4006 #ifdef CONFIG_SLUB_CPU_PARTIAL
4007 unsigned int nr_objects;
4010 * cpu_partial determined the maximum number of objects kept in the
4011 * per cpu partial lists of a processor.
4013 * Per cpu partial lists mainly contain slabs that just have one
4014 * object freed. If they are used for allocation then they can be
4015 * filled up again with minimal effort. The slab will never hit the
4016 * per node partial lists and therefore no locking will be required.
4018 * For backwards compatibility reasons, this is determined as number
4019 * of objects, even though we now limit maximum number of pages, see
4020 * slub_set_cpu_partial()
4022 if (!kmem_cache_has_cpu_partial(s))
4024 else if (s->size >= PAGE_SIZE)
4026 else if (s->size >= 1024)
4028 else if (s->size >= 256)
4033 slub_set_cpu_partial(s, nr_objects);
4038 * calculate_sizes() determines the order and the distribution of data within
4041 static int calculate_sizes(struct kmem_cache *s)
4043 slab_flags_t flags = s->flags;
4044 unsigned int size = s->object_size;
4048 * Round up object size to the next word boundary. We can only
4049 * place the free pointer at word boundaries and this determines
4050 * the possible location of the free pointer.
4052 size = ALIGN(size, sizeof(void *));
4054 #ifdef CONFIG_SLUB_DEBUG
4056 * Determine if we can poison the object itself. If the user of
4057 * the slab may touch the object after free or before allocation
4058 * then we should never poison the object itself.
4060 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4062 s->flags |= __OBJECT_POISON;
4064 s->flags &= ~__OBJECT_POISON;
4068 * If we are Redzoning then check if there is some space between the
4069 * end of the object and the free pointer. If not then add an
4070 * additional word to have some bytes to store Redzone information.
4072 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4073 size += sizeof(void *);
4077 * With that we have determined the number of bytes in actual use
4078 * by the object and redzoning.
4082 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4083 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4086 * Relocate free pointer after the object if it is not
4087 * permitted to overwrite the first word of the object on
4090 * This is the case if we do RCU, have a constructor or
4091 * destructor, are poisoning the objects, or are
4092 * redzoning an object smaller than sizeof(void *).
4094 * The assumption that s->offset >= s->inuse means free
4095 * pointer is outside of the object is used in the
4096 * freeptr_outside_object() function. If that is no
4097 * longer true, the function needs to be modified.
4100 size += sizeof(void *);
4103 * Store freelist pointer near middle of object to keep
4104 * it away from the edges of the object to avoid small
4105 * sized over/underflows from neighboring allocations.
4107 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4110 #ifdef CONFIG_SLUB_DEBUG
4111 if (flags & SLAB_STORE_USER)
4113 * Need to store information about allocs and frees after
4116 size += 2 * sizeof(struct track);
4119 kasan_cache_create(s, &size, &s->flags);
4120 #ifdef CONFIG_SLUB_DEBUG
4121 if (flags & SLAB_RED_ZONE) {
4123 * Add some empty padding so that we can catch
4124 * overwrites from earlier objects rather than let
4125 * tracking information or the free pointer be
4126 * corrupted if a user writes before the start
4129 size += sizeof(void *);
4131 s->red_left_pad = sizeof(void *);
4132 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4133 size += s->red_left_pad;
4138 * SLUB stores one object immediately after another beginning from
4139 * offset 0. In order to align the objects we have to simply size
4140 * each object to conform to the alignment.
4142 size = ALIGN(size, s->align);
4144 s->reciprocal_size = reciprocal_value(size);
4145 order = calculate_order(size);
4152 s->allocflags |= __GFP_COMP;
4154 if (s->flags & SLAB_CACHE_DMA)
4155 s->allocflags |= GFP_DMA;
4157 if (s->flags & SLAB_CACHE_DMA32)
4158 s->allocflags |= GFP_DMA32;
4160 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4161 s->allocflags |= __GFP_RECLAIMABLE;
4164 * Determine the number of objects per slab
4166 s->oo = oo_make(order, size);
4167 s->min = oo_make(get_order(size), size);
4168 if (oo_objects(s->oo) > oo_objects(s->max))
4171 return !!oo_objects(s->oo);
4174 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4176 s->flags = kmem_cache_flags(s->size, flags, s->name);
4177 #ifdef CONFIG_SLAB_FREELIST_HARDENED
4178 s->random = get_random_long();
4181 if (!calculate_sizes(s))
4183 if (disable_higher_order_debug) {
4185 * Disable debugging flags that store metadata if the min slab
4188 if (get_order(s->size) > get_order(s->object_size)) {
4189 s->flags &= ~DEBUG_METADATA_FLAGS;
4191 if (!calculate_sizes(s))
4196 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4197 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4198 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4199 /* Enable fast mode */
4200 s->flags |= __CMPXCHG_DOUBLE;
4204 * The larger the object size is, the more slabs we want on the partial
4205 * list to avoid pounding the page allocator excessively.
4207 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4208 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4213 s->remote_node_defrag_ratio = 1000;
4216 /* Initialize the pre-computed randomized freelist if slab is up */
4217 if (slab_state >= UP) {
4218 if (init_cache_random_seq(s))
4222 if (!init_kmem_cache_nodes(s))
4225 if (alloc_kmem_cache_cpus(s))
4229 __kmem_cache_release(s);
4233 static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4236 #ifdef CONFIG_SLUB_DEBUG
4237 void *addr = slab_address(slab);
4238 unsigned long flags;
4242 slab_err(s, slab, text, s->name);
4243 slab_lock(slab, &flags);
4245 map = get_map(s, slab);
4246 for_each_object(p, s, addr, slab->objects) {
4248 if (!test_bit(__obj_to_index(s, addr, p), map)) {
4249 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4250 print_tracking(s, p);
4254 slab_unlock(slab, &flags);
4259 * Attempt to free all partial slabs on a node.
4260 * This is called from __kmem_cache_shutdown(). We must take list_lock
4261 * because sysfs file might still access partial list after the shutdowning.
4263 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4266 struct slab *slab, *h;
4268 BUG_ON(irqs_disabled());
4269 spin_lock_irq(&n->list_lock);
4270 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4272 remove_partial(n, slab);
4273 list_add(&slab->slab_list, &discard);
4275 list_slab_objects(s, slab,
4276 "Objects remaining in %s on __kmem_cache_shutdown()");
4279 spin_unlock_irq(&n->list_lock);
4281 list_for_each_entry_safe(slab, h, &discard, slab_list)
4282 discard_slab(s, slab);
4285 bool __kmem_cache_empty(struct kmem_cache *s)
4288 struct kmem_cache_node *n;
4290 for_each_kmem_cache_node(s, node, n)
4291 if (n->nr_partial || slabs_node(s, node))
4297 * Release all resources used by a slab cache.
4299 int __kmem_cache_shutdown(struct kmem_cache *s)
4302 struct kmem_cache_node *n;
4304 flush_all_cpus_locked(s);
4305 /* Attempt to free all objects */
4306 for_each_kmem_cache_node(s, node, n) {
4308 if (n->nr_partial || slabs_node(s, node))
4314 #ifdef CONFIG_PRINTK
4315 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4318 int __maybe_unused i;
4322 struct kmem_cache *s = slab->slab_cache;
4323 struct track __maybe_unused *trackp;
4325 kpp->kp_ptr = object;
4326 kpp->kp_slab = slab;
4327 kpp->kp_slab_cache = s;
4328 base = slab_address(slab);
4329 objp0 = kasan_reset_tag(object);
4330 #ifdef CONFIG_SLUB_DEBUG
4331 objp = restore_red_left(s, objp0);
4335 objnr = obj_to_index(s, slab, objp);
4336 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4337 objp = base + s->size * objnr;
4338 kpp->kp_objp = objp;
4339 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4340 || (objp - base) % s->size) ||
4341 !(s->flags & SLAB_STORE_USER))
4343 #ifdef CONFIG_SLUB_DEBUG
4344 objp = fixup_red_left(s, objp);
4345 trackp = get_track(s, objp, TRACK_ALLOC);
4346 kpp->kp_ret = (void *)trackp->addr;
4347 #ifdef CONFIG_STACKTRACE
4348 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4349 kpp->kp_stack[i] = (void *)trackp->addrs[i];
4350 if (!kpp->kp_stack[i])
4354 trackp = get_track(s, objp, TRACK_FREE);
4355 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4356 kpp->kp_free_stack[i] = (void *)trackp->addrs[i];
4357 if (!kpp->kp_free_stack[i])
4365 /********************************************************************
4367 *******************************************************************/
4369 static int __init setup_slub_min_order(char *str)
4371 get_option(&str, (int *)&slub_min_order);
4376 __setup("slub_min_order=", setup_slub_min_order);
4378 static int __init setup_slub_max_order(char *str)
4380 get_option(&str, (int *)&slub_max_order);
4381 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4386 __setup("slub_max_order=", setup_slub_max_order);
4388 static int __init setup_slub_min_objects(char *str)
4390 get_option(&str, (int *)&slub_min_objects);
4395 __setup("slub_min_objects=", setup_slub_min_objects);
4397 void *__kmalloc(size_t size, gfp_t flags)
4399 struct kmem_cache *s;
4402 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4403 return kmalloc_large(size, flags);
4405 s = kmalloc_slab(size, flags);
4407 if (unlikely(ZERO_OR_NULL_PTR(s)))
4410 ret = slab_alloc(s, NULL, flags, _RET_IP_, size);
4412 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4414 ret = kasan_kmalloc(s, ret, size, flags);
4418 EXPORT_SYMBOL(__kmalloc);
4421 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4425 unsigned int order = get_order(size);
4427 flags |= __GFP_COMP;
4428 page = alloc_pages_node(node, flags, order);
4430 ptr = page_address(page);
4431 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4432 PAGE_SIZE << order);
4435 return kmalloc_large_node_hook(ptr, size, flags);
4438 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4440 struct kmem_cache *s;
4443 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4444 ret = kmalloc_large_node(size, flags, node);
4446 trace_kmalloc_node(_RET_IP_, ret,
4447 size, PAGE_SIZE << get_order(size),
4453 s = kmalloc_slab(size, flags);
4455 if (unlikely(ZERO_OR_NULL_PTR(s)))
4458 ret = slab_alloc_node(s, NULL, flags, node, _RET_IP_, size);
4460 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4462 ret = kasan_kmalloc(s, ret, size, flags);
4466 EXPORT_SYMBOL(__kmalloc_node);
4467 #endif /* CONFIG_NUMA */
4469 #ifdef CONFIG_HARDENED_USERCOPY
4471 * Rejects incorrectly sized objects and objects that are to be copied
4472 * to/from userspace but do not fall entirely within the containing slab
4473 * cache's usercopy region.
4475 * Returns NULL if check passes, otherwise const char * to name of cache
4476 * to indicate an error.
4478 void __check_heap_object(const void *ptr, unsigned long n,
4479 const struct slab *slab, bool to_user)
4481 struct kmem_cache *s;
4482 unsigned int offset;
4483 bool is_kfence = is_kfence_address(ptr);
4485 ptr = kasan_reset_tag(ptr);
4487 /* Find object and usable object size. */
4488 s = slab->slab_cache;
4490 /* Reject impossible pointers. */
4491 if (ptr < slab_address(slab))
4492 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4495 /* Find offset within object. */
4497 offset = ptr - kfence_object_start(ptr);
4499 offset = (ptr - slab_address(slab)) % s->size;
4501 /* Adjust for redzone and reject if within the redzone. */
4502 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4503 if (offset < s->red_left_pad)
4504 usercopy_abort("SLUB object in left red zone",
4505 s->name, to_user, offset, n);
4506 offset -= s->red_left_pad;
4509 /* Allow address range falling entirely within usercopy region. */
4510 if (offset >= s->useroffset &&
4511 offset - s->useroffset <= s->usersize &&
4512 n <= s->useroffset - offset + s->usersize)
4515 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4517 #endif /* CONFIG_HARDENED_USERCOPY */
4519 size_t __ksize(const void *object)
4521 struct folio *folio;
4523 if (unlikely(object == ZERO_SIZE_PTR))
4526 folio = virt_to_folio(object);
4528 if (unlikely(!folio_test_slab(folio)))
4529 return folio_size(folio);
4531 return slab_ksize(folio_slab(folio)->slab_cache);
4533 EXPORT_SYMBOL(__ksize);
4535 void kfree(const void *x)
4537 struct folio *folio;
4539 void *object = (void *)x;
4541 trace_kfree(_RET_IP_, x);
4543 if (unlikely(ZERO_OR_NULL_PTR(x)))
4546 folio = virt_to_folio(x);
4547 if (unlikely(!folio_test_slab(folio))) {
4548 free_large_kmalloc(folio, object);
4551 slab = folio_slab(folio);
4552 slab_free(slab->slab_cache, slab, object, NULL, 1, _RET_IP_);
4554 EXPORT_SYMBOL(kfree);
4556 #define SHRINK_PROMOTE_MAX 32
4559 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4560 * up most to the head of the partial lists. New allocations will then
4561 * fill those up and thus they can be removed from the partial lists.
4563 * The slabs with the least items are placed last. This results in them
4564 * being allocated from last increasing the chance that the last objects
4565 * are freed in them.
4567 static int __kmem_cache_do_shrink(struct kmem_cache *s)
4571 struct kmem_cache_node *n;
4574 struct list_head discard;
4575 struct list_head promote[SHRINK_PROMOTE_MAX];
4576 unsigned long flags;
4579 for_each_kmem_cache_node(s, node, n) {
4580 INIT_LIST_HEAD(&discard);
4581 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4582 INIT_LIST_HEAD(promote + i);
4584 spin_lock_irqsave(&n->list_lock, flags);
4587 * Build lists of slabs to discard or promote.
4589 * Note that concurrent frees may occur while we hold the
4590 * list_lock. slab->inuse here is the upper limit.
4592 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4593 int free = slab->objects - slab->inuse;
4595 /* Do not reread slab->inuse */
4598 /* We do not keep full slabs on the list */
4601 if (free == slab->objects) {
4602 list_move(&slab->slab_list, &discard);
4604 } else if (free <= SHRINK_PROMOTE_MAX)
4605 list_move(&slab->slab_list, promote + free - 1);
4609 * Promote the slabs filled up most to the head of the
4612 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4613 list_splice(promote + i, &n->partial);
4615 spin_unlock_irqrestore(&n->list_lock, flags);
4617 /* Release empty slabs */
4618 list_for_each_entry_safe(slab, t, &discard, slab_list)
4619 discard_slab(s, slab);
4621 if (slabs_node(s, node))
4628 int __kmem_cache_shrink(struct kmem_cache *s)
4631 return __kmem_cache_do_shrink(s);
4634 static int slab_mem_going_offline_callback(void *arg)
4636 struct kmem_cache *s;
4638 mutex_lock(&slab_mutex);
4639 list_for_each_entry(s, &slab_caches, list) {
4640 flush_all_cpus_locked(s);
4641 __kmem_cache_do_shrink(s);
4643 mutex_unlock(&slab_mutex);
4648 static void slab_mem_offline_callback(void *arg)
4650 struct memory_notify *marg = arg;
4653 offline_node = marg->status_change_nid_normal;
4656 * If the node still has available memory. we need kmem_cache_node
4659 if (offline_node < 0)
4662 mutex_lock(&slab_mutex);
4663 node_clear(offline_node, slab_nodes);
4665 * We no longer free kmem_cache_node structures here, as it would be
4666 * racy with all get_node() users, and infeasible to protect them with
4669 mutex_unlock(&slab_mutex);
4672 static int slab_mem_going_online_callback(void *arg)
4674 struct kmem_cache_node *n;
4675 struct kmem_cache *s;
4676 struct memory_notify *marg = arg;
4677 int nid = marg->status_change_nid_normal;
4681 * If the node's memory is already available, then kmem_cache_node is
4682 * already created. Nothing to do.
4688 * We are bringing a node online. No memory is available yet. We must
4689 * allocate a kmem_cache_node structure in order to bring the node
4692 mutex_lock(&slab_mutex);
4693 list_for_each_entry(s, &slab_caches, list) {
4695 * The structure may already exist if the node was previously
4696 * onlined and offlined.
4698 if (get_node(s, nid))
4701 * XXX: kmem_cache_alloc_node will fallback to other nodes
4702 * since memory is not yet available from the node that
4705 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4710 init_kmem_cache_node(n);
4714 * Any cache created after this point will also have kmem_cache_node
4715 * initialized for the new node.
4717 node_set(nid, slab_nodes);
4719 mutex_unlock(&slab_mutex);
4723 static int slab_memory_callback(struct notifier_block *self,
4724 unsigned long action, void *arg)
4729 case MEM_GOING_ONLINE:
4730 ret = slab_mem_going_online_callback(arg);
4732 case MEM_GOING_OFFLINE:
4733 ret = slab_mem_going_offline_callback(arg);
4736 case MEM_CANCEL_ONLINE:
4737 slab_mem_offline_callback(arg);
4740 case MEM_CANCEL_OFFLINE:
4744 ret = notifier_from_errno(ret);
4750 static struct notifier_block slab_memory_callback_nb = {
4751 .notifier_call = slab_memory_callback,
4752 .priority = SLAB_CALLBACK_PRI,
4755 /********************************************************************
4756 * Basic setup of slabs
4757 *******************************************************************/
4760 * Used for early kmem_cache structures that were allocated using
4761 * the page allocator. Allocate them properly then fix up the pointers
4762 * that may be pointing to the wrong kmem_cache structure.
4765 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4768 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4769 struct kmem_cache_node *n;
4771 memcpy(s, static_cache, kmem_cache->object_size);
4774 * This runs very early, and only the boot processor is supposed to be
4775 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4778 __flush_cpu_slab(s, smp_processor_id());
4779 for_each_kmem_cache_node(s, node, n) {
4782 list_for_each_entry(p, &n->partial, slab_list)
4785 #ifdef CONFIG_SLUB_DEBUG
4786 list_for_each_entry(p, &n->full, slab_list)
4790 list_add(&s->list, &slab_caches);
4794 void __init kmem_cache_init(void)
4796 static __initdata struct kmem_cache boot_kmem_cache,
4797 boot_kmem_cache_node;
4800 if (debug_guardpage_minorder())
4803 /* Print slub debugging pointers without hashing */
4804 if (__slub_debug_enabled())
4805 no_hash_pointers_enable(NULL);
4807 kmem_cache_node = &boot_kmem_cache_node;
4808 kmem_cache = &boot_kmem_cache;
4811 * Initialize the nodemask for which we will allocate per node
4812 * structures. Here we don't need taking slab_mutex yet.
4814 for_each_node_state(node, N_NORMAL_MEMORY)
4815 node_set(node, slab_nodes);
4817 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4818 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4820 register_hotmemory_notifier(&slab_memory_callback_nb);
4822 /* Able to allocate the per node structures */
4823 slab_state = PARTIAL;
4825 create_boot_cache(kmem_cache, "kmem_cache",
4826 offsetof(struct kmem_cache, node) +
4827 nr_node_ids * sizeof(struct kmem_cache_node *),
4828 SLAB_HWCACHE_ALIGN, 0, 0);
4830 kmem_cache = bootstrap(&boot_kmem_cache);
4831 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4833 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4834 setup_kmalloc_cache_index_table();
4835 create_kmalloc_caches(0);
4837 /* Setup random freelists for each cache */
4838 init_freelist_randomization();
4840 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4843 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4845 slub_min_order, slub_max_order, slub_min_objects,
4846 nr_cpu_ids, nr_node_ids);
4849 void __init kmem_cache_init_late(void)
4854 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4855 slab_flags_t flags, void (*ctor)(void *))
4857 struct kmem_cache *s;
4859 s = find_mergeable(size, align, flags, name, ctor);
4864 * Adjust the object sizes so that we clear
4865 * the complete object on kzalloc.
4867 s->object_size = max(s->object_size, size);
4868 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4870 if (sysfs_slab_alias(s, name)) {
4879 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4883 err = kmem_cache_open(s, flags);
4887 /* Mutex is not taken during early boot */
4888 if (slab_state <= UP)
4891 err = sysfs_slab_add(s);
4893 __kmem_cache_release(s);
4897 if (s->flags & SLAB_STORE_USER)
4898 debugfs_slab_add(s);
4903 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4905 struct kmem_cache *s;
4908 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4909 return kmalloc_large(size, gfpflags);
4911 s = kmalloc_slab(size, gfpflags);
4913 if (unlikely(ZERO_OR_NULL_PTR(s)))
4916 ret = slab_alloc(s, NULL, gfpflags, caller, size);
4918 /* Honor the call site pointer we received. */
4919 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4923 EXPORT_SYMBOL(__kmalloc_track_caller);
4926 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4927 int node, unsigned long caller)
4929 struct kmem_cache *s;
4932 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4933 ret = kmalloc_large_node(size, gfpflags, node);
4935 trace_kmalloc_node(caller, ret,
4936 size, PAGE_SIZE << get_order(size),
4942 s = kmalloc_slab(size, gfpflags);
4944 if (unlikely(ZERO_OR_NULL_PTR(s)))
4947 ret = slab_alloc_node(s, NULL, gfpflags, node, caller, size);
4949 /* Honor the call site pointer we received. */
4950 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4954 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4958 static int count_inuse(struct slab *slab)
4963 static int count_total(struct slab *slab)
4965 return slab->objects;
4969 #ifdef CONFIG_SLUB_DEBUG
4970 static void validate_slab(struct kmem_cache *s, struct slab *slab,
4971 unsigned long *obj_map)
4974 void *addr = slab_address(slab);
4975 unsigned long flags;
4977 slab_lock(slab, &flags);
4979 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
4982 /* Now we know that a valid freelist exists */
4983 __fill_map(obj_map, s, slab);
4984 for_each_object(p, s, addr, slab->objects) {
4985 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
4986 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4988 if (!check_object(s, slab, p, val))
4992 slab_unlock(slab, &flags);
4995 static int validate_slab_node(struct kmem_cache *s,
4996 struct kmem_cache_node *n, unsigned long *obj_map)
4998 unsigned long count = 0;
5000 unsigned long flags;
5002 spin_lock_irqsave(&n->list_lock, flags);
5004 list_for_each_entry(slab, &n->partial, slab_list) {
5005 validate_slab(s, slab, obj_map);
5008 if (count != n->nr_partial) {
5009 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5010 s->name, count, n->nr_partial);
5011 slab_add_kunit_errors();
5014 if (!(s->flags & SLAB_STORE_USER))
5017 list_for_each_entry(slab, &n->full, slab_list) {
5018 validate_slab(s, slab, obj_map);
5021 if (count != atomic_long_read(&n->nr_slabs)) {
5022 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5023 s->name, count, atomic_long_read(&n->nr_slabs));
5024 slab_add_kunit_errors();
5028 spin_unlock_irqrestore(&n->list_lock, flags);
5032 long validate_slab_cache(struct kmem_cache *s)
5035 unsigned long count = 0;
5036 struct kmem_cache_node *n;
5037 unsigned long *obj_map;
5039 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5044 for_each_kmem_cache_node(s, node, n)
5045 count += validate_slab_node(s, n, obj_map);
5047 bitmap_free(obj_map);
5051 EXPORT_SYMBOL(validate_slab_cache);
5053 #ifdef CONFIG_DEBUG_FS
5055 * Generate lists of code addresses where slabcache objects are allocated
5060 unsigned long count;
5067 DECLARE_BITMAP(cpus, NR_CPUS);
5073 unsigned long count;
5074 struct location *loc;
5078 static struct dentry *slab_debugfs_root;
5080 static void free_loc_track(struct loc_track *t)
5083 free_pages((unsigned long)t->loc,
5084 get_order(sizeof(struct location) * t->max));
5087 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5092 order = get_order(sizeof(struct location) * max);
5094 l = (void *)__get_free_pages(flags, order);
5099 memcpy(l, t->loc, sizeof(struct location) * t->count);
5107 static int add_location(struct loc_track *t, struct kmem_cache *s,
5108 const struct track *track)
5110 long start, end, pos;
5112 unsigned long caddr;
5113 unsigned long age = jiffies - track->when;
5119 pos = start + (end - start + 1) / 2;
5122 * There is nothing at "end". If we end up there
5123 * we need to add something to before end.
5128 caddr = t->loc[pos].addr;
5129 if (track->addr == caddr) {
5135 if (age < l->min_time)
5137 if (age > l->max_time)
5140 if (track->pid < l->min_pid)
5141 l->min_pid = track->pid;
5142 if (track->pid > l->max_pid)
5143 l->max_pid = track->pid;
5145 cpumask_set_cpu(track->cpu,
5146 to_cpumask(l->cpus));
5148 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5152 if (track->addr < caddr)
5159 * Not found. Insert new tracking element.
5161 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5167 (t->count - pos) * sizeof(struct location));
5170 l->addr = track->addr;
5174 l->min_pid = track->pid;
5175 l->max_pid = track->pid;
5176 cpumask_clear(to_cpumask(l->cpus));
5177 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5178 nodes_clear(l->nodes);
5179 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5183 static void process_slab(struct loc_track *t, struct kmem_cache *s,
5184 struct slab *slab, enum track_item alloc,
5185 unsigned long *obj_map)
5187 void *addr = slab_address(slab);
5190 __fill_map(obj_map, s, slab);
5192 for_each_object(p, s, addr, slab->objects)
5193 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5194 add_location(t, s, get_track(s, p, alloc));
5196 #endif /* CONFIG_DEBUG_FS */
5197 #endif /* CONFIG_SLUB_DEBUG */
5200 enum slab_stat_type {
5201 SL_ALL, /* All slabs */
5202 SL_PARTIAL, /* Only partially allocated slabs */
5203 SL_CPU, /* Only slabs used for cpu caches */
5204 SL_OBJECTS, /* Determine allocated objects not slabs */
5205 SL_TOTAL /* Determine object capacity not slabs */
5208 #define SO_ALL (1 << SL_ALL)
5209 #define SO_PARTIAL (1 << SL_PARTIAL)
5210 #define SO_CPU (1 << SL_CPU)
5211 #define SO_OBJECTS (1 << SL_OBJECTS)
5212 #define SO_TOTAL (1 << SL_TOTAL)
5214 static ssize_t show_slab_objects(struct kmem_cache *s,
5215 char *buf, unsigned long flags)
5217 unsigned long total = 0;
5220 unsigned long *nodes;
5223 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5227 if (flags & SO_CPU) {
5230 for_each_possible_cpu(cpu) {
5231 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5236 slab = READ_ONCE(c->slab);
5240 node = slab_nid(slab);
5241 if (flags & SO_TOTAL)
5243 else if (flags & SO_OBJECTS)
5251 #ifdef CONFIG_SLUB_CPU_PARTIAL
5252 slab = slub_percpu_partial_read_once(c);
5254 node = slab_nid(slab);
5255 if (flags & SO_TOTAL)
5257 else if (flags & SO_OBJECTS)
5269 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5270 * already held which will conflict with an existing lock order:
5272 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5274 * We don't really need mem_hotplug_lock (to hold off
5275 * slab_mem_going_offline_callback) here because slab's memory hot
5276 * unplug code doesn't destroy the kmem_cache->node[] data.
5279 #ifdef CONFIG_SLUB_DEBUG
5280 if (flags & SO_ALL) {
5281 struct kmem_cache_node *n;
5283 for_each_kmem_cache_node(s, node, n) {
5285 if (flags & SO_TOTAL)
5286 x = atomic_long_read(&n->total_objects);
5287 else if (flags & SO_OBJECTS)
5288 x = atomic_long_read(&n->total_objects) -
5289 count_partial(n, count_free);
5291 x = atomic_long_read(&n->nr_slabs);
5298 if (flags & SO_PARTIAL) {
5299 struct kmem_cache_node *n;
5301 for_each_kmem_cache_node(s, node, n) {
5302 if (flags & SO_TOTAL)
5303 x = count_partial(n, count_total);
5304 else if (flags & SO_OBJECTS)
5305 x = count_partial(n, count_inuse);
5313 len += sysfs_emit_at(buf, len, "%lu", total);
5315 for (node = 0; node < nr_node_ids; node++) {
5317 len += sysfs_emit_at(buf, len, " N%d=%lu",
5321 len += sysfs_emit_at(buf, len, "\n");
5327 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5328 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5330 struct slab_attribute {
5331 struct attribute attr;
5332 ssize_t (*show)(struct kmem_cache *s, char *buf);
5333 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5336 #define SLAB_ATTR_RO(_name) \
5337 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5339 #define SLAB_ATTR(_name) \
5340 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5342 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5344 return sysfs_emit(buf, "%u\n", s->size);
5346 SLAB_ATTR_RO(slab_size);
5348 static ssize_t align_show(struct kmem_cache *s, char *buf)
5350 return sysfs_emit(buf, "%u\n", s->align);
5352 SLAB_ATTR_RO(align);
5354 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5356 return sysfs_emit(buf, "%u\n", s->object_size);
5358 SLAB_ATTR_RO(object_size);
5360 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5362 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5364 SLAB_ATTR_RO(objs_per_slab);
5366 static ssize_t order_show(struct kmem_cache *s, char *buf)
5368 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5370 SLAB_ATTR_RO(order);
5372 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5374 return sysfs_emit(buf, "%lu\n", s->min_partial);
5377 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5383 err = kstrtoul(buf, 10, &min);
5387 s->min_partial = min;
5390 SLAB_ATTR(min_partial);
5392 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5394 unsigned int nr_partial = 0;
5395 #ifdef CONFIG_SLUB_CPU_PARTIAL
5396 nr_partial = s->cpu_partial;
5399 return sysfs_emit(buf, "%u\n", nr_partial);
5402 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5405 unsigned int objects;
5408 err = kstrtouint(buf, 10, &objects);
5411 if (objects && !kmem_cache_has_cpu_partial(s))
5414 slub_set_cpu_partial(s, objects);
5418 SLAB_ATTR(cpu_partial);
5420 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5424 return sysfs_emit(buf, "%pS\n", s->ctor);
5428 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5430 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5432 SLAB_ATTR_RO(aliases);
5434 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5436 return show_slab_objects(s, buf, SO_PARTIAL);
5438 SLAB_ATTR_RO(partial);
5440 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5442 return show_slab_objects(s, buf, SO_CPU);
5444 SLAB_ATTR_RO(cpu_slabs);
5446 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5448 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5450 SLAB_ATTR_RO(objects);
5452 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5454 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5456 SLAB_ATTR_RO(objects_partial);
5458 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5462 int cpu __maybe_unused;
5465 #ifdef CONFIG_SLUB_CPU_PARTIAL
5466 for_each_online_cpu(cpu) {
5469 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5472 slabs += slab->slabs;
5476 /* Approximate half-full slabs, see slub_set_cpu_partial() */
5477 objects = (slabs * oo_objects(s->oo)) / 2;
5478 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5480 #if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5481 for_each_online_cpu(cpu) {
5484 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5486 slabs = READ_ONCE(slab->slabs);
5487 objects = (slabs * oo_objects(s->oo)) / 2;
5488 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5489 cpu, objects, slabs);
5493 len += sysfs_emit_at(buf, len, "\n");
5497 SLAB_ATTR_RO(slabs_cpu_partial);
5499 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5501 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5503 SLAB_ATTR_RO(reclaim_account);
5505 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5507 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5509 SLAB_ATTR_RO(hwcache_align);
5511 #ifdef CONFIG_ZONE_DMA
5512 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5514 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5516 SLAB_ATTR_RO(cache_dma);
5519 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5521 return sysfs_emit(buf, "%u\n", s->usersize);
5523 SLAB_ATTR_RO(usersize);
5525 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5527 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5529 SLAB_ATTR_RO(destroy_by_rcu);
5531 #ifdef CONFIG_SLUB_DEBUG
5532 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5534 return show_slab_objects(s, buf, SO_ALL);
5536 SLAB_ATTR_RO(slabs);
5538 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5540 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5542 SLAB_ATTR_RO(total_objects);
5544 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5546 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5548 SLAB_ATTR_RO(sanity_checks);
5550 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5552 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5554 SLAB_ATTR_RO(trace);
5556 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5558 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5561 SLAB_ATTR_RO(red_zone);
5563 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5565 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5568 SLAB_ATTR_RO(poison);
5570 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5572 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5575 SLAB_ATTR_RO(store_user);
5577 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5582 static ssize_t validate_store(struct kmem_cache *s,
5583 const char *buf, size_t length)
5587 if (buf[0] == '1') {
5588 ret = validate_slab_cache(s);
5594 SLAB_ATTR(validate);
5596 #endif /* CONFIG_SLUB_DEBUG */
5598 #ifdef CONFIG_FAILSLAB
5599 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5601 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5603 SLAB_ATTR_RO(failslab);
5606 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5611 static ssize_t shrink_store(struct kmem_cache *s,
5612 const char *buf, size_t length)
5615 kmem_cache_shrink(s);
5623 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5625 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5628 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5629 const char *buf, size_t length)
5634 err = kstrtouint(buf, 10, &ratio);
5640 s->remote_node_defrag_ratio = ratio * 10;
5644 SLAB_ATTR(remote_node_defrag_ratio);
5647 #ifdef CONFIG_SLUB_STATS
5648 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5650 unsigned long sum = 0;
5653 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5658 for_each_online_cpu(cpu) {
5659 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5665 len += sysfs_emit_at(buf, len, "%lu", sum);
5668 for_each_online_cpu(cpu) {
5670 len += sysfs_emit_at(buf, len, " C%d=%u",
5675 len += sysfs_emit_at(buf, len, "\n");
5680 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5684 for_each_online_cpu(cpu)
5685 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5688 #define STAT_ATTR(si, text) \
5689 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5691 return show_stat(s, buf, si); \
5693 static ssize_t text##_store(struct kmem_cache *s, \
5694 const char *buf, size_t length) \
5696 if (buf[0] != '0') \
5698 clear_stat(s, si); \
5703 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5704 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5705 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5706 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5707 STAT_ATTR(FREE_FROZEN, free_frozen);
5708 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5709 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5710 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5711 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5712 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5713 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5714 STAT_ATTR(FREE_SLAB, free_slab);
5715 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5716 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5717 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5718 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5719 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5720 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5721 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5722 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5723 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5724 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5725 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5726 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5727 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5728 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5729 #endif /* CONFIG_SLUB_STATS */
5731 static struct attribute *slab_attrs[] = {
5732 &slab_size_attr.attr,
5733 &object_size_attr.attr,
5734 &objs_per_slab_attr.attr,
5736 &min_partial_attr.attr,
5737 &cpu_partial_attr.attr,
5739 &objects_partial_attr.attr,
5741 &cpu_slabs_attr.attr,
5745 &hwcache_align_attr.attr,
5746 &reclaim_account_attr.attr,
5747 &destroy_by_rcu_attr.attr,
5749 &slabs_cpu_partial_attr.attr,
5750 #ifdef CONFIG_SLUB_DEBUG
5751 &total_objects_attr.attr,
5753 &sanity_checks_attr.attr,
5755 &red_zone_attr.attr,
5757 &store_user_attr.attr,
5758 &validate_attr.attr,
5760 #ifdef CONFIG_ZONE_DMA
5761 &cache_dma_attr.attr,
5764 &remote_node_defrag_ratio_attr.attr,
5766 #ifdef CONFIG_SLUB_STATS
5767 &alloc_fastpath_attr.attr,
5768 &alloc_slowpath_attr.attr,
5769 &free_fastpath_attr.attr,
5770 &free_slowpath_attr.attr,
5771 &free_frozen_attr.attr,
5772 &free_add_partial_attr.attr,
5773 &free_remove_partial_attr.attr,
5774 &alloc_from_partial_attr.attr,
5775 &alloc_slab_attr.attr,
5776 &alloc_refill_attr.attr,
5777 &alloc_node_mismatch_attr.attr,
5778 &free_slab_attr.attr,
5779 &cpuslab_flush_attr.attr,
5780 &deactivate_full_attr.attr,
5781 &deactivate_empty_attr.attr,
5782 &deactivate_to_head_attr.attr,
5783 &deactivate_to_tail_attr.attr,
5784 &deactivate_remote_frees_attr.attr,
5785 &deactivate_bypass_attr.attr,
5786 &order_fallback_attr.attr,
5787 &cmpxchg_double_fail_attr.attr,
5788 &cmpxchg_double_cpu_fail_attr.attr,
5789 &cpu_partial_alloc_attr.attr,
5790 &cpu_partial_free_attr.attr,
5791 &cpu_partial_node_attr.attr,
5792 &cpu_partial_drain_attr.attr,
5794 #ifdef CONFIG_FAILSLAB
5795 &failslab_attr.attr,
5797 &usersize_attr.attr,
5802 static const struct attribute_group slab_attr_group = {
5803 .attrs = slab_attrs,
5806 static ssize_t slab_attr_show(struct kobject *kobj,
5807 struct attribute *attr,
5810 struct slab_attribute *attribute;
5811 struct kmem_cache *s;
5814 attribute = to_slab_attr(attr);
5817 if (!attribute->show)
5820 err = attribute->show(s, buf);
5825 static ssize_t slab_attr_store(struct kobject *kobj,
5826 struct attribute *attr,
5827 const char *buf, size_t len)
5829 struct slab_attribute *attribute;
5830 struct kmem_cache *s;
5833 attribute = to_slab_attr(attr);
5836 if (!attribute->store)
5839 err = attribute->store(s, buf, len);
5843 static void kmem_cache_release(struct kobject *k)
5845 slab_kmem_cache_release(to_slab(k));
5848 static const struct sysfs_ops slab_sysfs_ops = {
5849 .show = slab_attr_show,
5850 .store = slab_attr_store,
5853 static struct kobj_type slab_ktype = {
5854 .sysfs_ops = &slab_sysfs_ops,
5855 .release = kmem_cache_release,
5858 static struct kset *slab_kset;
5860 static inline struct kset *cache_kset(struct kmem_cache *s)
5865 #define ID_STR_LENGTH 64
5867 /* Create a unique string id for a slab cache:
5869 * Format :[flags-]size
5871 static char *create_unique_id(struct kmem_cache *s)
5873 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5880 * First flags affecting slabcache operations. We will only
5881 * get here for aliasable slabs so we do not need to support
5882 * too many flags. The flags here must cover all flags that
5883 * are matched during merging to guarantee that the id is
5886 if (s->flags & SLAB_CACHE_DMA)
5888 if (s->flags & SLAB_CACHE_DMA32)
5890 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5892 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5894 if (s->flags & SLAB_ACCOUNT)
5898 p += sprintf(p, "%07u", s->size);
5900 BUG_ON(p > name + ID_STR_LENGTH - 1);
5904 static int sysfs_slab_add(struct kmem_cache *s)
5908 struct kset *kset = cache_kset(s);
5909 int unmergeable = slab_unmergeable(s);
5912 kobject_init(&s->kobj, &slab_ktype);
5916 if (!unmergeable && disable_higher_order_debug &&
5917 (slub_debug & DEBUG_METADATA_FLAGS))
5922 * Slabcache can never be merged so we can use the name proper.
5923 * This is typically the case for debug situations. In that
5924 * case we can catch duplicate names easily.
5926 sysfs_remove_link(&slab_kset->kobj, s->name);
5930 * Create a unique name for the slab as a target
5933 name = create_unique_id(s);
5936 s->kobj.kset = kset;
5937 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5941 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5946 /* Setup first alias */
5947 sysfs_slab_alias(s, s->name);
5954 kobject_del(&s->kobj);
5958 void sysfs_slab_unlink(struct kmem_cache *s)
5960 if (slab_state >= FULL)
5961 kobject_del(&s->kobj);
5964 void sysfs_slab_release(struct kmem_cache *s)
5966 if (slab_state >= FULL)
5967 kobject_put(&s->kobj);
5971 * Need to buffer aliases during bootup until sysfs becomes
5972 * available lest we lose that information.
5974 struct saved_alias {
5975 struct kmem_cache *s;
5977 struct saved_alias *next;
5980 static struct saved_alias *alias_list;
5982 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5984 struct saved_alias *al;
5986 if (slab_state == FULL) {
5988 * If we have a leftover link then remove it.
5990 sysfs_remove_link(&slab_kset->kobj, name);
5991 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5994 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6000 al->next = alias_list;
6005 static int __init slab_sysfs_init(void)
6007 struct kmem_cache *s;
6010 mutex_lock(&slab_mutex);
6012 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6014 mutex_unlock(&slab_mutex);
6015 pr_err("Cannot register slab subsystem.\n");
6021 list_for_each_entry(s, &slab_caches, list) {
6022 err = sysfs_slab_add(s);
6024 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6028 while (alias_list) {
6029 struct saved_alias *al = alias_list;
6031 alias_list = alias_list->next;
6032 err = sysfs_slab_alias(al->s, al->name);
6034 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6039 mutex_unlock(&slab_mutex);
6043 __initcall(slab_sysfs_init);
6044 #endif /* CONFIG_SYSFS */
6046 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6047 static int slab_debugfs_show(struct seq_file *seq, void *v)
6049 struct loc_track *t = seq->private;
6053 idx = (unsigned long) t->idx;
6054 if (idx < t->count) {
6057 seq_printf(seq, "%7ld ", l->count);
6060 seq_printf(seq, "%pS", (void *)l->addr);
6062 seq_puts(seq, "<not-available>");
6064 if (l->sum_time != l->min_time) {
6065 seq_printf(seq, " age=%ld/%llu/%ld",
6066 l->min_time, div_u64(l->sum_time, l->count),
6069 seq_printf(seq, " age=%ld", l->min_time);
6071 if (l->min_pid != l->max_pid)
6072 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6074 seq_printf(seq, " pid=%ld",
6077 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6078 seq_printf(seq, " cpus=%*pbl",
6079 cpumask_pr_args(to_cpumask(l->cpus)));
6081 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6082 seq_printf(seq, " nodes=%*pbl",
6083 nodemask_pr_args(&l->nodes));
6085 seq_puts(seq, "\n");
6088 if (!idx && !t->count)
6089 seq_puts(seq, "No data\n");
6094 static void slab_debugfs_stop(struct seq_file *seq, void *v)
6098 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6100 struct loc_track *t = seq->private;
6103 if (*ppos <= t->count)
6109 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6111 struct loc_track *t = seq->private;
6117 static const struct seq_operations slab_debugfs_sops = {
6118 .start = slab_debugfs_start,
6119 .next = slab_debugfs_next,
6120 .stop = slab_debugfs_stop,
6121 .show = slab_debugfs_show,
6124 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6127 struct kmem_cache_node *n;
6128 enum track_item alloc;
6130 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6131 sizeof(struct loc_track));
6132 struct kmem_cache *s = file_inode(filep)->i_private;
6133 unsigned long *obj_map;
6138 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6140 seq_release_private(inode, filep);
6144 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6145 alloc = TRACK_ALLOC;
6149 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6150 bitmap_free(obj_map);
6151 seq_release_private(inode, filep);
6155 for_each_kmem_cache_node(s, node, n) {
6156 unsigned long flags;
6159 if (!atomic_long_read(&n->nr_slabs))
6162 spin_lock_irqsave(&n->list_lock, flags);
6163 list_for_each_entry(slab, &n->partial, slab_list)
6164 process_slab(t, s, slab, alloc, obj_map);
6165 list_for_each_entry(slab, &n->full, slab_list)
6166 process_slab(t, s, slab, alloc, obj_map);
6167 spin_unlock_irqrestore(&n->list_lock, flags);
6170 bitmap_free(obj_map);
6174 static int slab_debug_trace_release(struct inode *inode, struct file *file)
6176 struct seq_file *seq = file->private_data;
6177 struct loc_track *t = seq->private;
6180 return seq_release_private(inode, file);
6183 static const struct file_operations slab_debugfs_fops = {
6184 .open = slab_debug_trace_open,
6186 .llseek = seq_lseek,
6187 .release = slab_debug_trace_release,
6190 static void debugfs_slab_add(struct kmem_cache *s)
6192 struct dentry *slab_cache_dir;
6194 if (unlikely(!slab_debugfs_root))
6197 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6199 debugfs_create_file("alloc_traces", 0400,
6200 slab_cache_dir, s, &slab_debugfs_fops);
6202 debugfs_create_file("free_traces", 0400,
6203 slab_cache_dir, s, &slab_debugfs_fops);
6206 void debugfs_slab_release(struct kmem_cache *s)
6208 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6211 static int __init slab_debugfs_init(void)
6213 struct kmem_cache *s;
6215 slab_debugfs_root = debugfs_create_dir("slab", NULL);
6217 list_for_each_entry(s, &slab_caches, list)
6218 if (s->flags & SLAB_STORE_USER)
6219 debugfs_slab_add(s);
6224 __initcall(slab_debugfs_init);
6227 * The /proc/slabinfo ABI
6229 #ifdef CONFIG_SLUB_DEBUG
6230 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6232 unsigned long nr_slabs = 0;
6233 unsigned long nr_objs = 0;
6234 unsigned long nr_free = 0;
6236 struct kmem_cache_node *n;
6238 for_each_kmem_cache_node(s, node, n) {
6239 nr_slabs += node_nr_slabs(n);
6240 nr_objs += node_nr_objs(n);
6241 nr_free += count_partial(n, count_free);
6244 sinfo->active_objs = nr_objs - nr_free;
6245 sinfo->num_objs = nr_objs;
6246 sinfo->active_slabs = nr_slabs;
6247 sinfo->num_slabs = nr_slabs;
6248 sinfo->objects_per_slab = oo_objects(s->oo);
6249 sinfo->cache_order = oo_order(s->oo);
6252 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6256 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6257 size_t count, loff_t *ppos)
6261 #endif /* CONFIG_SLUB_DEBUG */