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)
50 * 3. slab_lock(page) (Only on some arches and for debugging)
54 * The role of the slab_mutex is to protect the list of all the slabs
55 * and to synchronize major metadata changes to slab cache structures.
57 * The slab_lock is only used for debugging and on arches that do not
58 * have the ability to do a cmpxchg_double. It only protects:
59 * A. page->freelist -> List of object free in a page
60 * B. page->inuse -> Number of objects in use
61 * C. page->objects -> Number of objects in page
62 * D. page->frozen -> frozen state
64 * If a slab is frozen then it is exempt from list management. It is not
65 * on any list except per cpu partial list. The processor that froze the
66 * slab is the one who can perform list operations on the page. Other
67 * processors may put objects onto the freelist but the processor that
68 * froze the slab is the only one that can retrieve the objects from the
71 * The list_lock protects the partial and full list on each node and
72 * the partial slab counter. If taken then no new slabs may be added or
73 * removed from the lists nor make the number of partial slabs be modified.
74 * (Note that the total number of slabs is an atomic value that may be
75 * modified without taking the list lock).
77 * The list_lock is a centralized lock and thus we avoid taking it as
78 * much as possible. As long as SLUB does not have to handle partial
79 * slabs, operations can continue without any centralized lock. F.e.
80 * allocating a long series of objects that fill up slabs does not require
82 * Interrupts are disabled during allocation and deallocation in order to
83 * make the slab allocator safe to use in the context of an irq. In addition
84 * interrupts are disabled to ensure that the processor does not change
85 * while handling per_cpu slabs, due to kernel preemption.
87 * SLUB assigns one slab for allocation to each processor.
88 * Allocations only occur from these slabs called cpu slabs.
90 * Slabs with free elements are kept on a partial list and during regular
91 * operations no list for full slabs is used. If an object in a full slab is
92 * freed then the slab will show up again on the partial lists.
93 * We track full slabs for debugging purposes though because otherwise we
94 * cannot scan all objects.
96 * Slabs are freed when they become empty. Teardown and setup is
97 * minimal so we rely on the page allocators per cpu caches for
98 * fast frees and allocs.
100 * page->frozen The slab is frozen and exempt from list processing.
101 * This means that the slab is dedicated to a purpose
102 * such as satisfying allocations for a specific
103 * processor. Objects may be freed in the slab while
104 * it is frozen but slab_free will then skip the usual
105 * list operations. It is up to the processor holding
106 * the slab to integrate the slab into the slab lists
107 * when the slab is no longer needed.
109 * One use of this flag is to mark slabs that are
110 * used for allocations. Then such a slab becomes a cpu
111 * slab. The cpu slab may be equipped with an additional
112 * freelist that allows lockless access to
113 * free objects in addition to the regular freelist
114 * that requires the slab lock.
116 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
117 * options set. This moves slab handling out of
118 * the fast path and disables lockless freelists.
121 #ifdef CONFIG_SLUB_DEBUG
122 #ifdef CONFIG_SLUB_DEBUG_ON
123 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
125 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
127 #endif /* CONFIG_SLUB_DEBUG */
129 static inline bool kmem_cache_debug(struct kmem_cache *s)
131 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
134 void *fixup_red_left(struct kmem_cache *s, void *p)
136 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
137 p += s->red_left_pad;
142 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
144 #ifdef CONFIG_SLUB_CPU_PARTIAL
145 return !kmem_cache_debug(s);
152 * Issues still to be resolved:
154 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
156 * - Variable sizing of the per node arrays
159 /* Enable to log cmpxchg failures */
160 #undef SLUB_DEBUG_CMPXCHG
163 * Minimum number of partial slabs. These will be left on the partial
164 * lists even if they are empty. kmem_cache_shrink may reclaim them.
166 #define MIN_PARTIAL 5
169 * Maximum number of desirable partial slabs.
170 * The existence of more partial slabs makes kmem_cache_shrink
171 * sort the partial list by the number of objects in use.
173 #define MAX_PARTIAL 10
175 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_STORE_USER)
179 * These debug flags cannot use CMPXCHG because there might be consistency
180 * issues when checking or reading debug information
182 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
187 * Debugging flags that require metadata to be stored in the slab. These get
188 * disabled when slub_debug=O is used and a cache's min order increases with
191 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
194 #define OO_MASK ((1 << OO_SHIFT) - 1)
195 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
197 /* Internal SLUB flags */
199 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
200 /* Use cmpxchg_double */
201 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
204 * Tracking user of a slab.
206 #define TRACK_ADDRS_COUNT 16
208 unsigned long addr; /* Called from address */
209 #ifdef CONFIG_STACKTRACE
210 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
212 int cpu; /* Was running on cpu */
213 int pid; /* Pid context */
214 unsigned long when; /* When did the operation occur */
217 enum track_item { TRACK_ALLOC, TRACK_FREE };
220 static int sysfs_slab_add(struct kmem_cache *);
221 static int sysfs_slab_alias(struct kmem_cache *, const char *);
223 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
224 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
228 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
229 static void debugfs_slab_add(struct kmem_cache *);
231 static inline void debugfs_slab_add(struct kmem_cache *s) { }
234 static inline void stat(const struct kmem_cache *s, enum stat_item si)
236 #ifdef CONFIG_SLUB_STATS
238 * The rmw is racy on a preemptible kernel but this is acceptable, so
239 * avoid this_cpu_add()'s irq-disable overhead.
241 raw_cpu_inc(s->cpu_slab->stat[si]);
246 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
247 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
248 * differ during memory hotplug/hotremove operations.
249 * Protected by slab_mutex.
251 static nodemask_t slab_nodes;
253 /********************************************************************
254 * Core slab cache functions
255 *******************************************************************/
258 * Returns freelist pointer (ptr). With hardening, this is obfuscated
259 * with an XOR of the address where the pointer is held and a per-cache
262 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
263 unsigned long ptr_addr)
265 #ifdef CONFIG_SLAB_FREELIST_HARDENED
267 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
268 * Normally, this doesn't cause any issues, as both set_freepointer()
269 * and get_freepointer() are called with a pointer with the same tag.
270 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
271 * example, when __free_slub() iterates over objects in a cache, it
272 * passes untagged pointers to check_object(). check_object() in turns
273 * calls get_freepointer() with an untagged pointer, which causes the
274 * freepointer to be restored incorrectly.
276 return (void *)((unsigned long)ptr ^ s->random ^
277 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
283 /* Returns the freelist pointer recorded at location ptr_addr. */
284 static inline void *freelist_dereference(const struct kmem_cache *s,
287 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
288 (unsigned long)ptr_addr);
291 static inline void *get_freepointer(struct kmem_cache *s, void *object)
293 object = kasan_reset_tag(object);
294 return freelist_dereference(s, object + s->offset);
297 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
299 prefetch(object + s->offset);
302 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
304 unsigned long freepointer_addr;
307 if (!debug_pagealloc_enabled_static())
308 return get_freepointer(s, object);
310 object = kasan_reset_tag(object);
311 freepointer_addr = (unsigned long)object + s->offset;
312 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
313 return freelist_ptr(s, p, freepointer_addr);
316 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
318 unsigned long freeptr_addr = (unsigned long)object + s->offset;
320 #ifdef CONFIG_SLAB_FREELIST_HARDENED
321 BUG_ON(object == fp); /* naive detection of double free or corruption */
324 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
325 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
328 /* Loop over all objects in a slab */
329 #define for_each_object(__p, __s, __addr, __objects) \
330 for (__p = fixup_red_left(__s, __addr); \
331 __p < (__addr) + (__objects) * (__s)->size; \
334 static inline unsigned int order_objects(unsigned int order, unsigned int size)
336 return ((unsigned int)PAGE_SIZE << order) / size;
339 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
342 struct kmem_cache_order_objects x = {
343 (order << OO_SHIFT) + order_objects(order, size)
349 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
351 return x.x >> OO_SHIFT;
354 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
356 return x.x & OO_MASK;
360 * Per slab locking using the pagelock
362 static __always_inline void slab_lock(struct page *page)
364 VM_BUG_ON_PAGE(PageTail(page), page);
365 bit_spin_lock(PG_locked, &page->flags);
368 static __always_inline void slab_unlock(struct page *page)
370 VM_BUG_ON_PAGE(PageTail(page), page);
371 __bit_spin_unlock(PG_locked, &page->flags);
374 /* Interrupts must be disabled (for the fallback code to work right) */
375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
380 VM_BUG_ON(!irqs_disabled());
381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 page->counters = counters_new;
403 stat(s, CMPXCHG_DOUBLE_FAIL);
405 #ifdef SLUB_DEBUG_CMPXCHG
406 pr_info("%s %s: cmpxchg double redo ", n, s->name);
412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
429 local_irq_save(flags);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 page->counters = counters_new;
436 local_irq_restore(flags);
440 local_irq_restore(flags);
444 stat(s, CMPXCHG_DOUBLE_FAIL);
446 #ifdef SLUB_DEBUG_CMPXCHG
447 pr_info("%s %s: cmpxchg double redo ", n, s->name);
453 #ifdef CONFIG_SLUB_DEBUG
454 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
455 static DEFINE_SPINLOCK(object_map_lock);
457 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
460 void *addr = page_address(page);
463 bitmap_zero(obj_map, page->objects);
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(__obj_to_index(s, addr, p), obj_map);
469 #if IS_ENABLED(CONFIG_KUNIT)
470 static bool slab_add_kunit_errors(void)
472 struct kunit_resource *resource;
474 if (likely(!current->kunit_test))
477 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
481 (*(int *)resource->data)++;
482 kunit_put_resource(resource);
486 static inline bool slab_add_kunit_errors(void) { return false; }
490 * Determine a map of object in use on a page.
492 * Node listlock must be held to guarantee that the page does
493 * not vanish from under us.
495 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
496 __acquires(&object_map_lock)
498 VM_BUG_ON(!irqs_disabled());
500 spin_lock(&object_map_lock);
502 __fill_map(object_map, s, page);
507 static void put_map(unsigned long *map) __releases(&object_map_lock)
509 VM_BUG_ON(map != object_map);
510 spin_unlock(&object_map_lock);
513 static inline unsigned int size_from_object(struct kmem_cache *s)
515 if (s->flags & SLAB_RED_ZONE)
516 return s->size - s->red_left_pad;
521 static inline void *restore_red_left(struct kmem_cache *s, void *p)
523 if (s->flags & SLAB_RED_ZONE)
524 p -= s->red_left_pad;
532 #if defined(CONFIG_SLUB_DEBUG_ON)
533 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
535 static slab_flags_t slub_debug;
538 static char *slub_debug_string;
539 static int disable_higher_order_debug;
542 * slub is about to manipulate internal object metadata. This memory lies
543 * outside the range of the allocated object, so accessing it would normally
544 * be reported by kasan as a bounds error. metadata_access_enable() is used
545 * to tell kasan that these accesses are OK.
547 static inline void metadata_access_enable(void)
549 kasan_disable_current();
552 static inline void metadata_access_disable(void)
554 kasan_enable_current();
561 /* Verify that a pointer has an address that is valid within a slab page */
562 static inline int check_valid_pointer(struct kmem_cache *s,
563 struct page *page, void *object)
570 base = page_address(page);
571 object = kasan_reset_tag(object);
572 object = restore_red_left(s, object);
573 if (object < base || object >= base + page->objects * s->size ||
574 (object - base) % s->size) {
581 static void print_section(char *level, char *text, u8 *addr,
584 metadata_access_enable();
585 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
586 16, 1, kasan_reset_tag((void *)addr), length, 1);
587 metadata_access_disable();
591 * See comment in calculate_sizes().
593 static inline bool freeptr_outside_object(struct kmem_cache *s)
595 return s->offset >= s->inuse;
599 * Return offset of the end of info block which is inuse + free pointer if
600 * not overlapping with object.
602 static inline unsigned int get_info_end(struct kmem_cache *s)
604 if (freeptr_outside_object(s))
605 return s->inuse + sizeof(void *);
610 static struct track *get_track(struct kmem_cache *s, void *object,
611 enum track_item alloc)
615 p = object + get_info_end(s);
617 return kasan_reset_tag(p + alloc);
620 static void set_track(struct kmem_cache *s, void *object,
621 enum track_item alloc, unsigned long addr)
623 struct track *p = get_track(s, object, alloc);
626 #ifdef CONFIG_STACKTRACE
627 unsigned int nr_entries;
629 metadata_access_enable();
630 nr_entries = stack_trace_save(kasan_reset_tag(p->addrs),
631 TRACK_ADDRS_COUNT, 3);
632 metadata_access_disable();
634 if (nr_entries < TRACK_ADDRS_COUNT)
635 p->addrs[nr_entries] = 0;
638 p->cpu = smp_processor_id();
639 p->pid = current->pid;
642 memset(p, 0, sizeof(struct track));
646 static void init_tracking(struct kmem_cache *s, void *object)
648 if (!(s->flags & SLAB_STORE_USER))
651 set_track(s, object, TRACK_FREE, 0UL);
652 set_track(s, object, TRACK_ALLOC, 0UL);
655 static void print_track(const char *s, struct track *t, unsigned long pr_time)
660 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
661 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
662 #ifdef CONFIG_STACKTRACE
665 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
667 pr_err("\t%pS\n", (void *)t->addrs[i]);
674 void print_tracking(struct kmem_cache *s, void *object)
676 unsigned long pr_time = jiffies;
677 if (!(s->flags & SLAB_STORE_USER))
680 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
681 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
684 static void print_page_info(struct page *page)
686 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%#lx(%pGp)\n",
687 page, page->objects, page->inuse, page->freelist,
688 page->flags, &page->flags);
692 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
694 struct va_format vaf;
700 pr_err("=============================================================================\n");
701 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
702 pr_err("-----------------------------------------------------------------------------\n\n");
707 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
709 struct va_format vaf;
712 if (slab_add_kunit_errors())
718 pr_err("FIX %s: %pV\n", s->name, &vaf);
722 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
723 void **freelist, void *nextfree)
725 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
726 !check_valid_pointer(s, page, nextfree) && freelist) {
727 object_err(s, page, *freelist, "Freechain corrupt");
729 slab_fix(s, "Isolate corrupted freechain");
736 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
738 unsigned int off; /* Offset of last byte */
739 u8 *addr = page_address(page);
741 print_tracking(s, p);
743 print_page_info(page);
745 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
746 p, p - addr, get_freepointer(s, p));
748 if (s->flags & SLAB_RED_ZONE)
749 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
751 else if (p > addr + 16)
752 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
754 print_section(KERN_ERR, "Object ", p,
755 min_t(unsigned int, s->object_size, PAGE_SIZE));
756 if (s->flags & SLAB_RED_ZONE)
757 print_section(KERN_ERR, "Redzone ", p + s->object_size,
758 s->inuse - s->object_size);
760 off = get_info_end(s);
762 if (s->flags & SLAB_STORE_USER)
763 off += 2 * sizeof(struct track);
765 off += kasan_metadata_size(s);
767 if (off != size_from_object(s))
768 /* Beginning of the filler is the free pointer */
769 print_section(KERN_ERR, "Padding ", p + off,
770 size_from_object(s) - off);
775 void object_err(struct kmem_cache *s, struct page *page,
776 u8 *object, char *reason)
778 if (slab_add_kunit_errors())
781 slab_bug(s, "%s", reason);
782 print_trailer(s, page, object);
783 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
786 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
787 const char *fmt, ...)
792 if (slab_add_kunit_errors())
796 vsnprintf(buf, sizeof(buf), fmt, args);
798 slab_bug(s, "%s", buf);
799 print_page_info(page);
801 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
804 static void init_object(struct kmem_cache *s, void *object, u8 val)
806 u8 *p = kasan_reset_tag(object);
808 if (s->flags & SLAB_RED_ZONE)
809 memset(p - s->red_left_pad, val, s->red_left_pad);
811 if (s->flags & __OBJECT_POISON) {
812 memset(p, POISON_FREE, s->object_size - 1);
813 p[s->object_size - 1] = POISON_END;
816 if (s->flags & SLAB_RED_ZONE)
817 memset(p + s->object_size, val, s->inuse - s->object_size);
820 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
821 void *from, void *to)
823 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
824 memset(from, data, to - from);
827 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
828 u8 *object, char *what,
829 u8 *start, unsigned int value, unsigned int bytes)
833 u8 *addr = page_address(page);
835 metadata_access_enable();
836 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
837 metadata_access_disable();
842 while (end > fault && end[-1] == value)
845 if (slab_add_kunit_errors())
848 slab_bug(s, "%s overwritten", what);
849 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
850 fault, end - 1, fault - addr,
852 print_trailer(s, page, object);
853 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
856 restore_bytes(s, what, value, fault, end);
864 * Bytes of the object to be managed.
865 * If the freepointer may overlay the object then the free
866 * pointer is at the middle of the object.
868 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
871 * object + s->object_size
872 * Padding to reach word boundary. This is also used for Redzoning.
873 * Padding is extended by another word if Redzoning is enabled and
874 * object_size == inuse.
876 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
877 * 0xcc (RED_ACTIVE) for objects in use.
880 * Meta data starts here.
882 * A. Free pointer (if we cannot overwrite object on free)
883 * B. Tracking data for SLAB_STORE_USER
884 * C. Padding to reach required alignment boundary or at minimum
885 * one word if debugging is on to be able to detect writes
886 * before the word boundary.
888 * Padding is done using 0x5a (POISON_INUSE)
891 * Nothing is used beyond s->size.
893 * If slabcaches are merged then the object_size and inuse boundaries are mostly
894 * ignored. And therefore no slab options that rely on these boundaries
895 * may be used with merged slabcaches.
898 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
900 unsigned long off = get_info_end(s); /* The end of info */
902 if (s->flags & SLAB_STORE_USER)
903 /* We also have user information there */
904 off += 2 * sizeof(struct track);
906 off += kasan_metadata_size(s);
908 if (size_from_object(s) == off)
911 return check_bytes_and_report(s, page, p, "Object padding",
912 p + off, POISON_INUSE, size_from_object(s) - off);
915 /* Check the pad bytes at the end of a slab page */
916 static int slab_pad_check(struct kmem_cache *s, struct page *page)
925 if (!(s->flags & SLAB_POISON))
928 start = page_address(page);
929 length = page_size(page);
930 end = start + length;
931 remainder = length % s->size;
935 pad = end - remainder;
936 metadata_access_enable();
937 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
938 metadata_access_disable();
941 while (end > fault && end[-1] == POISON_INUSE)
944 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
945 fault, end - 1, fault - start);
946 print_section(KERN_ERR, "Padding ", pad, remainder);
948 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
952 static int check_object(struct kmem_cache *s, struct page *page,
953 void *object, u8 val)
956 u8 *endobject = object + s->object_size;
958 if (s->flags & SLAB_RED_ZONE) {
959 if (!check_bytes_and_report(s, page, object, "Left Redzone",
960 object - s->red_left_pad, val, s->red_left_pad))
963 if (!check_bytes_and_report(s, page, object, "Right Redzone",
964 endobject, val, s->inuse - s->object_size))
967 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
968 check_bytes_and_report(s, page, p, "Alignment padding",
969 endobject, POISON_INUSE,
970 s->inuse - s->object_size);
974 if (s->flags & SLAB_POISON) {
975 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
976 (!check_bytes_and_report(s, page, p, "Poison", p,
977 POISON_FREE, s->object_size - 1) ||
978 !check_bytes_and_report(s, page, p, "End Poison",
979 p + s->object_size - 1, POISON_END, 1)))
982 * check_pad_bytes cleans up on its own.
984 check_pad_bytes(s, page, p);
987 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
989 * Object and freepointer overlap. Cannot check
990 * freepointer while object is allocated.
994 /* Check free pointer validity */
995 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
996 object_err(s, page, p, "Freepointer corrupt");
998 * No choice but to zap it and thus lose the remainder
999 * of the free objects in this slab. May cause
1000 * another error because the object count is now wrong.
1002 set_freepointer(s, p, NULL);
1008 static int check_slab(struct kmem_cache *s, struct page *page)
1012 VM_BUG_ON(!irqs_disabled());
1014 if (!PageSlab(page)) {
1015 slab_err(s, page, "Not a valid slab page");
1019 maxobj = order_objects(compound_order(page), s->size);
1020 if (page->objects > maxobj) {
1021 slab_err(s, page, "objects %u > max %u",
1022 page->objects, maxobj);
1025 if (page->inuse > page->objects) {
1026 slab_err(s, page, "inuse %u > max %u",
1027 page->inuse, page->objects);
1030 /* Slab_pad_check fixes things up after itself */
1031 slab_pad_check(s, page);
1036 * Determine if a certain object on a page is on the freelist. Must hold the
1037 * slab lock to guarantee that the chains are in a consistent state.
1039 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
1043 void *object = NULL;
1046 fp = page->freelist;
1047 while (fp && nr <= page->objects) {
1050 if (!check_valid_pointer(s, page, fp)) {
1052 object_err(s, page, object,
1053 "Freechain corrupt");
1054 set_freepointer(s, object, NULL);
1056 slab_err(s, page, "Freepointer corrupt");
1057 page->freelist = NULL;
1058 page->inuse = page->objects;
1059 slab_fix(s, "Freelist cleared");
1065 fp = get_freepointer(s, object);
1069 max_objects = order_objects(compound_order(page), s->size);
1070 if (max_objects > MAX_OBJS_PER_PAGE)
1071 max_objects = MAX_OBJS_PER_PAGE;
1073 if (page->objects != max_objects) {
1074 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1075 page->objects, max_objects);
1076 page->objects = max_objects;
1077 slab_fix(s, "Number of objects adjusted");
1079 if (page->inuse != page->objects - nr) {
1080 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1081 page->inuse, page->objects - nr);
1082 page->inuse = page->objects - nr;
1083 slab_fix(s, "Object count adjusted");
1085 return search == NULL;
1088 static void trace(struct kmem_cache *s, struct page *page, void *object,
1091 if (s->flags & SLAB_TRACE) {
1092 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1094 alloc ? "alloc" : "free",
1095 object, page->inuse,
1099 print_section(KERN_INFO, "Object ", (void *)object,
1107 * Tracking of fully allocated slabs for debugging purposes.
1109 static void add_full(struct kmem_cache *s,
1110 struct kmem_cache_node *n, struct page *page)
1112 if (!(s->flags & SLAB_STORE_USER))
1115 lockdep_assert_held(&n->list_lock);
1116 list_add(&page->slab_list, &n->full);
1119 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1121 if (!(s->flags & SLAB_STORE_USER))
1124 lockdep_assert_held(&n->list_lock);
1125 list_del(&page->slab_list);
1128 /* Tracking of the number of slabs for debugging purposes */
1129 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1131 struct kmem_cache_node *n = get_node(s, node);
1133 return atomic_long_read(&n->nr_slabs);
1136 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1138 return atomic_long_read(&n->nr_slabs);
1141 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1143 struct kmem_cache_node *n = get_node(s, node);
1146 * May be called early in order to allocate a slab for the
1147 * kmem_cache_node structure. Solve the chicken-egg
1148 * dilemma by deferring the increment of the count during
1149 * bootstrap (see early_kmem_cache_node_alloc).
1152 atomic_long_inc(&n->nr_slabs);
1153 atomic_long_add(objects, &n->total_objects);
1156 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1158 struct kmem_cache_node *n = get_node(s, node);
1160 atomic_long_dec(&n->nr_slabs);
1161 atomic_long_sub(objects, &n->total_objects);
1164 /* Object debug checks for alloc/free paths */
1165 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1168 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1171 init_object(s, object, SLUB_RED_INACTIVE);
1172 init_tracking(s, object);
1176 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1178 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1181 metadata_access_enable();
1182 memset(kasan_reset_tag(addr), POISON_INUSE, page_size(page));
1183 metadata_access_disable();
1186 static inline int alloc_consistency_checks(struct kmem_cache *s,
1187 struct page *page, void *object)
1189 if (!check_slab(s, page))
1192 if (!check_valid_pointer(s, page, object)) {
1193 object_err(s, page, object, "Freelist Pointer check fails");
1197 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1203 static noinline int alloc_debug_processing(struct kmem_cache *s,
1205 void *object, unsigned long addr)
1207 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1208 if (!alloc_consistency_checks(s, page, object))
1212 /* Success perform special debug activities for allocs */
1213 if (s->flags & SLAB_STORE_USER)
1214 set_track(s, object, TRACK_ALLOC, addr);
1215 trace(s, page, object, 1);
1216 init_object(s, object, SLUB_RED_ACTIVE);
1220 if (PageSlab(page)) {
1222 * If this is a slab page then lets do the best we can
1223 * to avoid issues in the future. Marking all objects
1224 * as used avoids touching the remaining objects.
1226 slab_fix(s, "Marking all objects used");
1227 page->inuse = page->objects;
1228 page->freelist = NULL;
1233 static inline int free_consistency_checks(struct kmem_cache *s,
1234 struct page *page, void *object, unsigned long addr)
1236 if (!check_valid_pointer(s, page, object)) {
1237 slab_err(s, page, "Invalid object pointer 0x%p", object);
1241 if (on_freelist(s, page, object)) {
1242 object_err(s, page, object, "Object already free");
1246 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1249 if (unlikely(s != page->slab_cache)) {
1250 if (!PageSlab(page)) {
1251 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1253 } else if (!page->slab_cache) {
1254 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1258 object_err(s, page, object,
1259 "page slab pointer corrupt.");
1265 /* Supports checking bulk free of a constructed freelist */
1266 static noinline int free_debug_processing(
1267 struct kmem_cache *s, struct page *page,
1268 void *head, void *tail, int bulk_cnt,
1271 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1272 void *object = head;
1274 unsigned long flags;
1277 spin_lock_irqsave(&n->list_lock, flags);
1280 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1281 if (!check_slab(s, page))
1288 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1289 if (!free_consistency_checks(s, page, object, addr))
1293 if (s->flags & SLAB_STORE_USER)
1294 set_track(s, object, TRACK_FREE, addr);
1295 trace(s, page, object, 0);
1296 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1297 init_object(s, object, SLUB_RED_INACTIVE);
1299 /* Reached end of constructed freelist yet? */
1300 if (object != tail) {
1301 object = get_freepointer(s, object);
1307 if (cnt != bulk_cnt)
1308 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1312 spin_unlock_irqrestore(&n->list_lock, flags);
1314 slab_fix(s, "Object at 0x%p not freed", object);
1319 * Parse a block of slub_debug options. Blocks are delimited by ';'
1321 * @str: start of block
1322 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1323 * @slabs: return start of list of slabs, or NULL when there's no list
1324 * @init: assume this is initial parsing and not per-kmem-create parsing
1326 * returns the start of next block if there's any, or NULL
1329 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1331 bool higher_order_disable = false;
1333 /* Skip any completely empty blocks */
1334 while (*str && *str == ';')
1339 * No options but restriction on slabs. This means full
1340 * debugging for slabs matching a pattern.
1342 *flags = DEBUG_DEFAULT_FLAGS;
1347 /* Determine which debug features should be switched on */
1348 for (; *str && *str != ',' && *str != ';'; str++) {
1349 switch (tolower(*str)) {
1354 *flags |= SLAB_CONSISTENCY_CHECKS;
1357 *flags |= SLAB_RED_ZONE;
1360 *flags |= SLAB_POISON;
1363 *flags |= SLAB_STORE_USER;
1366 *flags |= SLAB_TRACE;
1369 *flags |= SLAB_FAILSLAB;
1373 * Avoid enabling debugging on caches if its minimum
1374 * order would increase as a result.
1376 higher_order_disable = true;
1380 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1389 /* Skip over the slab list */
1390 while (*str && *str != ';')
1393 /* Skip any completely empty blocks */
1394 while (*str && *str == ';')
1397 if (init && higher_order_disable)
1398 disable_higher_order_debug = 1;
1406 static int __init setup_slub_debug(char *str)
1409 slab_flags_t global_flags;
1412 bool global_slub_debug_changed = false;
1413 bool slab_list_specified = false;
1415 global_flags = DEBUG_DEFAULT_FLAGS;
1416 if (*str++ != '=' || !*str)
1418 * No options specified. Switch on full debugging.
1424 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1427 global_flags = flags;
1428 global_slub_debug_changed = true;
1430 slab_list_specified = true;
1435 * For backwards compatibility, a single list of flags with list of
1436 * slabs means debugging is only changed for those slabs, so the global
1437 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1438 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1439 * long as there is no option specifying flags without a slab list.
1441 if (slab_list_specified) {
1442 if (!global_slub_debug_changed)
1443 global_flags = slub_debug;
1444 slub_debug_string = saved_str;
1447 slub_debug = global_flags;
1448 if (slub_debug != 0 || slub_debug_string)
1449 static_branch_enable(&slub_debug_enabled);
1451 static_branch_disable(&slub_debug_enabled);
1452 if ((static_branch_unlikely(&init_on_alloc) ||
1453 static_branch_unlikely(&init_on_free)) &&
1454 (slub_debug & SLAB_POISON))
1455 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1459 __setup("slub_debug", setup_slub_debug);
1462 * kmem_cache_flags - apply debugging options to the cache
1463 * @object_size: the size of an object without meta data
1464 * @flags: flags to set
1465 * @name: name of the cache
1467 * Debug option(s) are applied to @flags. In addition to the debug
1468 * option(s), if a slab name (or multiple) is specified i.e.
1469 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1470 * then only the select slabs will receive the debug option(s).
1472 slab_flags_t kmem_cache_flags(unsigned int object_size,
1473 slab_flags_t flags, const char *name)
1478 slab_flags_t block_flags;
1479 slab_flags_t slub_debug_local = slub_debug;
1482 * If the slab cache is for debugging (e.g. kmemleak) then
1483 * don't store user (stack trace) information by default,
1484 * but let the user enable it via the command line below.
1486 if (flags & SLAB_NOLEAKTRACE)
1487 slub_debug_local &= ~SLAB_STORE_USER;
1490 next_block = slub_debug_string;
1491 /* Go through all blocks of debug options, see if any matches our slab's name */
1492 while (next_block) {
1493 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1496 /* Found a block that has a slab list, search it */
1501 end = strchrnul(iter, ',');
1502 if (next_block && next_block < end)
1503 end = next_block - 1;
1505 glob = strnchr(iter, end - iter, '*');
1507 cmplen = glob - iter;
1509 cmplen = max_t(size_t, len, (end - iter));
1511 if (!strncmp(name, iter, cmplen)) {
1512 flags |= block_flags;
1516 if (!*end || *end == ';')
1522 return flags | slub_debug_local;
1524 #else /* !CONFIG_SLUB_DEBUG */
1525 static inline void setup_object_debug(struct kmem_cache *s,
1526 struct page *page, void *object) {}
1528 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1530 static inline int alloc_debug_processing(struct kmem_cache *s,
1531 struct page *page, void *object, unsigned long addr) { return 0; }
1533 static inline int free_debug_processing(
1534 struct kmem_cache *s, struct page *page,
1535 void *head, void *tail, int bulk_cnt,
1536 unsigned long addr) { return 0; }
1538 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1540 static inline int check_object(struct kmem_cache *s, struct page *page,
1541 void *object, u8 val) { return 1; }
1542 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1543 struct page *page) {}
1544 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1545 struct page *page) {}
1546 slab_flags_t kmem_cache_flags(unsigned int object_size,
1547 slab_flags_t flags, const char *name)
1551 #define slub_debug 0
1553 #define disable_higher_order_debug 0
1555 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1557 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1559 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1561 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1564 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1565 void **freelist, void *nextfree)
1569 #endif /* CONFIG_SLUB_DEBUG */
1572 * Hooks for other subsystems that check memory allocations. In a typical
1573 * production configuration these hooks all should produce no code at all.
1575 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1577 ptr = kasan_kmalloc_large(ptr, size, flags);
1578 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1579 kmemleak_alloc(ptr, size, 1, flags);
1583 static __always_inline void kfree_hook(void *x)
1586 kasan_kfree_large(x);
1589 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1592 kmemleak_free_recursive(x, s->flags);
1594 debug_check_no_locks_freed(x, s->object_size);
1596 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1597 debug_check_no_obj_freed(x, s->object_size);
1599 /* Use KCSAN to help debug racy use-after-free. */
1600 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1601 __kcsan_check_access(x, s->object_size,
1602 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1605 * As memory initialization might be integrated into KASAN,
1606 * kasan_slab_free and initialization memset's must be
1607 * kept together to avoid discrepancies in behavior.
1609 * The initialization memset's clear the object and the metadata,
1610 * but don't touch the SLAB redzone.
1615 if (!kasan_has_integrated_init())
1616 memset(kasan_reset_tag(x), 0, s->object_size);
1617 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1618 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1619 s->size - s->inuse - rsize);
1621 /* KASAN might put x into memory quarantine, delaying its reuse. */
1622 return kasan_slab_free(s, x, init);
1625 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1626 void **head, void **tail)
1631 void *old_tail = *tail ? *tail : *head;
1633 if (is_kfence_address(next)) {
1634 slab_free_hook(s, next, false);
1638 /* Head and tail of the reconstructed freelist */
1644 next = get_freepointer(s, object);
1646 /* If object's reuse doesn't have to be delayed */
1647 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1648 /* Move object to the new freelist */
1649 set_freepointer(s, object, *head);
1654 } while (object != old_tail);
1659 return *head != NULL;
1662 static void *setup_object(struct kmem_cache *s, struct page *page,
1665 setup_object_debug(s, page, object);
1666 object = kasan_init_slab_obj(s, object);
1667 if (unlikely(s->ctor)) {
1668 kasan_unpoison_object_data(s, object);
1670 kasan_poison_object_data(s, object);
1676 * Slab allocation and freeing
1678 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1679 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1682 unsigned int order = oo_order(oo);
1684 if (node == NUMA_NO_NODE)
1685 page = alloc_pages(flags, order);
1687 page = __alloc_pages_node(node, flags, order);
1692 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1693 /* Pre-initialize the random sequence cache */
1694 static int init_cache_random_seq(struct kmem_cache *s)
1696 unsigned int count = oo_objects(s->oo);
1699 /* Bailout if already initialised */
1703 err = cache_random_seq_create(s, count, GFP_KERNEL);
1705 pr_err("SLUB: Unable to initialize free list for %s\n",
1710 /* Transform to an offset on the set of pages */
1711 if (s->random_seq) {
1714 for (i = 0; i < count; i++)
1715 s->random_seq[i] *= s->size;
1720 /* Initialize each random sequence freelist per cache */
1721 static void __init init_freelist_randomization(void)
1723 struct kmem_cache *s;
1725 mutex_lock(&slab_mutex);
1727 list_for_each_entry(s, &slab_caches, list)
1728 init_cache_random_seq(s);
1730 mutex_unlock(&slab_mutex);
1733 /* Get the next entry on the pre-computed freelist randomized */
1734 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1735 unsigned long *pos, void *start,
1736 unsigned long page_limit,
1737 unsigned long freelist_count)
1742 * If the target page allocation failed, the number of objects on the
1743 * page might be smaller than the usual size defined by the cache.
1746 idx = s->random_seq[*pos];
1748 if (*pos >= freelist_count)
1750 } while (unlikely(idx >= page_limit));
1752 return (char *)start + idx;
1755 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1756 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1761 unsigned long idx, pos, page_limit, freelist_count;
1763 if (page->objects < 2 || !s->random_seq)
1766 freelist_count = oo_objects(s->oo);
1767 pos = get_random_int() % freelist_count;
1769 page_limit = page->objects * s->size;
1770 start = fixup_red_left(s, page_address(page));
1772 /* First entry is used as the base of the freelist */
1773 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1775 cur = setup_object(s, page, cur);
1776 page->freelist = cur;
1778 for (idx = 1; idx < page->objects; idx++) {
1779 next = next_freelist_entry(s, page, &pos, start, page_limit,
1781 next = setup_object(s, page, next);
1782 set_freepointer(s, cur, next);
1785 set_freepointer(s, cur, NULL);
1790 static inline int init_cache_random_seq(struct kmem_cache *s)
1794 static inline void init_freelist_randomization(void) { }
1795 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1799 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1801 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1804 struct kmem_cache_order_objects oo = s->oo;
1806 void *start, *p, *next;
1810 flags &= gfp_allowed_mask;
1812 if (gfpflags_allow_blocking(flags))
1815 flags |= s->allocflags;
1818 * Let the initial higher-order allocation fail under memory pressure
1819 * so we fall-back to the minimum order allocation.
1821 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1822 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1823 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1825 page = alloc_slab_page(s, alloc_gfp, node, oo);
1826 if (unlikely(!page)) {
1830 * Allocation may have failed due to fragmentation.
1831 * Try a lower order alloc if possible
1833 page = alloc_slab_page(s, alloc_gfp, node, oo);
1834 if (unlikely(!page))
1836 stat(s, ORDER_FALLBACK);
1839 page->objects = oo_objects(oo);
1841 account_slab_page(page, oo_order(oo), s, flags);
1843 page->slab_cache = s;
1844 __SetPageSlab(page);
1845 if (page_is_pfmemalloc(page))
1846 SetPageSlabPfmemalloc(page);
1848 kasan_poison_slab(page);
1850 start = page_address(page);
1852 setup_page_debug(s, page, start);
1854 shuffle = shuffle_freelist(s, page);
1857 start = fixup_red_left(s, start);
1858 start = setup_object(s, page, start);
1859 page->freelist = start;
1860 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1862 next = setup_object(s, page, next);
1863 set_freepointer(s, p, next);
1866 set_freepointer(s, p, NULL);
1869 page->inuse = page->objects;
1873 if (gfpflags_allow_blocking(flags))
1874 local_irq_disable();
1878 inc_slabs_node(s, page_to_nid(page), page->objects);
1883 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1885 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1886 flags = kmalloc_fix_flags(flags);
1888 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
1890 return allocate_slab(s,
1891 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1894 static void __free_slab(struct kmem_cache *s, struct page *page)
1896 int order = compound_order(page);
1897 int pages = 1 << order;
1899 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1902 slab_pad_check(s, page);
1903 for_each_object(p, s, page_address(page),
1905 check_object(s, page, p, SLUB_RED_INACTIVE);
1908 __ClearPageSlabPfmemalloc(page);
1909 __ClearPageSlab(page);
1910 /* In union with page->mapping where page allocator expects NULL */
1911 page->slab_cache = NULL;
1912 if (current->reclaim_state)
1913 current->reclaim_state->reclaimed_slab += pages;
1914 unaccount_slab_page(page, order, s);
1915 __free_pages(page, order);
1918 static void rcu_free_slab(struct rcu_head *h)
1920 struct page *page = container_of(h, struct page, rcu_head);
1922 __free_slab(page->slab_cache, page);
1925 static void free_slab(struct kmem_cache *s, struct page *page)
1927 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1928 call_rcu(&page->rcu_head, rcu_free_slab);
1930 __free_slab(s, page);
1933 static void discard_slab(struct kmem_cache *s, struct page *page)
1935 dec_slabs_node(s, page_to_nid(page), page->objects);
1940 * Management of partially allocated slabs.
1943 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1946 if (tail == DEACTIVATE_TO_TAIL)
1947 list_add_tail(&page->slab_list, &n->partial);
1949 list_add(&page->slab_list, &n->partial);
1952 static inline void add_partial(struct kmem_cache_node *n,
1953 struct page *page, int tail)
1955 lockdep_assert_held(&n->list_lock);
1956 __add_partial(n, page, tail);
1959 static inline void remove_partial(struct kmem_cache_node *n,
1962 lockdep_assert_held(&n->list_lock);
1963 list_del(&page->slab_list);
1968 * Remove slab from the partial list, freeze it and
1969 * return the pointer to the freelist.
1971 * Returns a list of objects or NULL if it fails.
1973 static inline void *acquire_slab(struct kmem_cache *s,
1974 struct kmem_cache_node *n, struct page *page,
1975 int mode, int *objects)
1978 unsigned long counters;
1981 lockdep_assert_held(&n->list_lock);
1984 * Zap the freelist and set the frozen bit.
1985 * The old freelist is the list of objects for the
1986 * per cpu allocation list.
1988 freelist = page->freelist;
1989 counters = page->counters;
1990 new.counters = counters;
1991 *objects = new.objects - new.inuse;
1993 new.inuse = page->objects;
1994 new.freelist = NULL;
1996 new.freelist = freelist;
1999 VM_BUG_ON(new.frozen);
2002 if (!__cmpxchg_double_slab(s, page,
2004 new.freelist, new.counters,
2008 remove_partial(n, page);
2013 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
2014 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
2017 * Try to allocate a partial slab from a specific node.
2019 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2020 struct page **ret_page, gfp_t flags)
2022 struct page *page, *page2;
2023 void *object = NULL;
2024 unsigned int available = 0;
2028 * Racy check. If we mistakenly see no partial slabs then we
2029 * just allocate an empty slab. If we mistakenly try to get a
2030 * partial slab and there is none available then get_partial()
2033 if (!n || !n->nr_partial)
2036 spin_lock(&n->list_lock);
2037 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
2040 if (!pfmemalloc_match(page, flags))
2043 t = acquire_slab(s, n, page, object == NULL, &objects);
2047 available += objects;
2050 stat(s, ALLOC_FROM_PARTIAL);
2053 put_cpu_partial(s, page, 0);
2054 stat(s, CPU_PARTIAL_NODE);
2056 if (!kmem_cache_has_cpu_partial(s)
2057 || available > slub_cpu_partial(s) / 2)
2061 spin_unlock(&n->list_lock);
2066 * Get a page from somewhere. Search in increasing NUMA distances.
2068 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2069 struct page **ret_page)
2072 struct zonelist *zonelist;
2075 enum zone_type highest_zoneidx = gfp_zone(flags);
2077 unsigned int cpuset_mems_cookie;
2080 * The defrag ratio allows a configuration of the tradeoffs between
2081 * inter node defragmentation and node local allocations. A lower
2082 * defrag_ratio increases the tendency to do local allocations
2083 * instead of attempting to obtain partial slabs from other nodes.
2085 * If the defrag_ratio is set to 0 then kmalloc() always
2086 * returns node local objects. If the ratio is higher then kmalloc()
2087 * may return off node objects because partial slabs are obtained
2088 * from other nodes and filled up.
2090 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2091 * (which makes defrag_ratio = 1000) then every (well almost)
2092 * allocation will first attempt to defrag slab caches on other nodes.
2093 * This means scanning over all nodes to look for partial slabs which
2094 * may be expensive if we do it every time we are trying to find a slab
2095 * with available objects.
2097 if (!s->remote_node_defrag_ratio ||
2098 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2102 cpuset_mems_cookie = read_mems_allowed_begin();
2103 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2104 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2105 struct kmem_cache_node *n;
2107 n = get_node(s, zone_to_nid(zone));
2109 if (n && cpuset_zone_allowed(zone, flags) &&
2110 n->nr_partial > s->min_partial) {
2111 object = get_partial_node(s, n, ret_page, flags);
2114 * Don't check read_mems_allowed_retry()
2115 * here - if mems_allowed was updated in
2116 * parallel, that was a harmless race
2117 * between allocation and the cpuset
2124 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2125 #endif /* CONFIG_NUMA */
2130 * Get a partial page, lock it and return it.
2132 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2133 struct page **ret_page)
2136 int searchnode = node;
2138 if (node == NUMA_NO_NODE)
2139 searchnode = numa_mem_id();
2141 object = get_partial_node(s, get_node(s, searchnode), ret_page, flags);
2142 if (object || node != NUMA_NO_NODE)
2145 return get_any_partial(s, flags, ret_page);
2148 #ifdef CONFIG_PREEMPTION
2150 * Calculate the next globally unique transaction for disambiguation
2151 * during cmpxchg. The transactions start with the cpu number and are then
2152 * incremented by CONFIG_NR_CPUS.
2154 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2157 * No preemption supported therefore also no need to check for
2163 static inline unsigned long next_tid(unsigned long tid)
2165 return tid + TID_STEP;
2168 #ifdef SLUB_DEBUG_CMPXCHG
2169 static inline unsigned int tid_to_cpu(unsigned long tid)
2171 return tid % TID_STEP;
2174 static inline unsigned long tid_to_event(unsigned long tid)
2176 return tid / TID_STEP;
2180 static inline unsigned int init_tid(int cpu)
2185 static inline void note_cmpxchg_failure(const char *n,
2186 const struct kmem_cache *s, unsigned long tid)
2188 #ifdef SLUB_DEBUG_CMPXCHG
2189 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2191 pr_info("%s %s: cmpxchg redo ", n, s->name);
2193 #ifdef CONFIG_PREEMPTION
2194 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2195 pr_warn("due to cpu change %d -> %d\n",
2196 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2199 if (tid_to_event(tid) != tid_to_event(actual_tid))
2200 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2201 tid_to_event(tid), tid_to_event(actual_tid));
2203 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2204 actual_tid, tid, next_tid(tid));
2206 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2209 static void init_kmem_cache_cpus(struct kmem_cache *s)
2213 for_each_possible_cpu(cpu)
2214 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2218 * Remove the cpu slab
2220 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2221 void *freelist, struct kmem_cache_cpu *c)
2223 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2224 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2225 int lock = 0, free_delta = 0;
2226 enum slab_modes l = M_NONE, m = M_NONE;
2227 void *nextfree, *freelist_iter, *freelist_tail;
2228 int tail = DEACTIVATE_TO_HEAD;
2232 if (page->freelist) {
2233 stat(s, DEACTIVATE_REMOTE_FREES);
2234 tail = DEACTIVATE_TO_TAIL;
2238 * Stage one: Count the objects on cpu's freelist as free_delta and
2239 * remember the last object in freelist_tail for later splicing.
2241 freelist_tail = NULL;
2242 freelist_iter = freelist;
2243 while (freelist_iter) {
2244 nextfree = get_freepointer(s, freelist_iter);
2247 * If 'nextfree' is invalid, it is possible that the object at
2248 * 'freelist_iter' is already corrupted. So isolate all objects
2249 * starting at 'freelist_iter' by skipping them.
2251 if (freelist_corrupted(s, page, &freelist_iter, nextfree))
2254 freelist_tail = freelist_iter;
2257 freelist_iter = nextfree;
2261 * Stage two: Unfreeze the page while splicing the per-cpu
2262 * freelist to the head of page's freelist.
2264 * Ensure that the page is unfrozen while the list presence
2265 * reflects the actual number of objects during unfreeze.
2267 * We setup the list membership and then perform a cmpxchg
2268 * with the count. If there is a mismatch then the page
2269 * is not unfrozen but the page is on the wrong list.
2271 * Then we restart the process which may have to remove
2272 * the page from the list that we just put it on again
2273 * because the number of objects in the slab may have
2278 old.freelist = READ_ONCE(page->freelist);
2279 old.counters = READ_ONCE(page->counters);
2280 VM_BUG_ON(!old.frozen);
2282 /* Determine target state of the slab */
2283 new.counters = old.counters;
2284 if (freelist_tail) {
2285 new.inuse -= free_delta;
2286 set_freepointer(s, freelist_tail, old.freelist);
2287 new.freelist = freelist;
2289 new.freelist = old.freelist;
2293 if (!new.inuse && n->nr_partial >= s->min_partial)
2295 else if (new.freelist) {
2300 * Taking the spinlock removes the possibility
2301 * that acquire_slab() will see a slab page that
2304 spin_lock(&n->list_lock);
2308 if (kmem_cache_debug_flags(s, SLAB_STORE_USER) && !lock) {
2311 * This also ensures that the scanning of full
2312 * slabs from diagnostic functions will not see
2315 spin_lock(&n->list_lock);
2321 remove_partial(n, page);
2322 else if (l == M_FULL)
2323 remove_full(s, n, page);
2326 add_partial(n, page, tail);
2327 else if (m == M_FULL)
2328 add_full(s, n, page);
2332 if (!__cmpxchg_double_slab(s, page,
2333 old.freelist, old.counters,
2334 new.freelist, new.counters,
2339 spin_unlock(&n->list_lock);
2343 else if (m == M_FULL)
2344 stat(s, DEACTIVATE_FULL);
2345 else if (m == M_FREE) {
2346 stat(s, DEACTIVATE_EMPTY);
2347 discard_slab(s, page);
2356 * Unfreeze all the cpu partial slabs.
2358 * This function must be called with interrupts disabled
2359 * for the cpu using c (or some other guarantee must be there
2360 * to guarantee no concurrent accesses).
2362 static void unfreeze_partials(struct kmem_cache *s,
2363 struct kmem_cache_cpu *c)
2365 #ifdef CONFIG_SLUB_CPU_PARTIAL
2366 struct kmem_cache_node *n = NULL, *n2 = NULL;
2367 struct page *page, *discard_page = NULL;
2369 while ((page = slub_percpu_partial(c))) {
2373 slub_set_percpu_partial(c, page);
2375 n2 = get_node(s, page_to_nid(page));
2378 spin_unlock(&n->list_lock);
2381 spin_lock(&n->list_lock);
2386 old.freelist = page->freelist;
2387 old.counters = page->counters;
2388 VM_BUG_ON(!old.frozen);
2390 new.counters = old.counters;
2391 new.freelist = old.freelist;
2395 } while (!__cmpxchg_double_slab(s, page,
2396 old.freelist, old.counters,
2397 new.freelist, new.counters,
2398 "unfreezing slab"));
2400 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2401 page->next = discard_page;
2402 discard_page = page;
2404 add_partial(n, page, DEACTIVATE_TO_TAIL);
2405 stat(s, FREE_ADD_PARTIAL);
2410 spin_unlock(&n->list_lock);
2412 while (discard_page) {
2413 page = discard_page;
2414 discard_page = discard_page->next;
2416 stat(s, DEACTIVATE_EMPTY);
2417 discard_slab(s, page);
2420 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2424 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2425 * partial page slot if available.
2427 * If we did not find a slot then simply move all the partials to the
2428 * per node partial list.
2430 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2432 #ifdef CONFIG_SLUB_CPU_PARTIAL
2433 struct page *oldpage;
2441 oldpage = this_cpu_read(s->cpu_slab->partial);
2444 pobjects = oldpage->pobjects;
2445 pages = oldpage->pages;
2446 if (drain && pobjects > slub_cpu_partial(s)) {
2447 unsigned long flags;
2449 * partial array is full. Move the existing
2450 * set to the per node partial list.
2452 local_irq_save(flags);
2453 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2454 local_irq_restore(flags);
2458 stat(s, CPU_PARTIAL_DRAIN);
2463 pobjects += page->objects - page->inuse;
2465 page->pages = pages;
2466 page->pobjects = pobjects;
2467 page->next = oldpage;
2469 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2472 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2475 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2477 stat(s, CPUSLAB_FLUSH);
2478 deactivate_slab(s, c->page, c->freelist, c);
2480 c->tid = next_tid(c->tid);
2486 * Called from IPI handler with interrupts disabled.
2488 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2490 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2495 unfreeze_partials(s, c);
2498 static void flush_cpu_slab(void *d)
2500 struct kmem_cache *s = d;
2502 __flush_cpu_slab(s, smp_processor_id());
2505 static bool has_cpu_slab(int cpu, void *info)
2507 struct kmem_cache *s = info;
2508 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2510 return c->page || slub_percpu_partial(c);
2513 static void flush_all(struct kmem_cache *s)
2515 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2519 * Use the cpu notifier to insure that the cpu slabs are flushed when
2522 static int slub_cpu_dead(unsigned int cpu)
2524 struct kmem_cache *s;
2525 unsigned long flags;
2527 mutex_lock(&slab_mutex);
2528 list_for_each_entry(s, &slab_caches, list) {
2529 local_irq_save(flags);
2530 __flush_cpu_slab(s, cpu);
2531 local_irq_restore(flags);
2533 mutex_unlock(&slab_mutex);
2538 * Check if the objects in a per cpu structure fit numa
2539 * locality expectations.
2541 static inline int node_match(struct page *page, int node)
2544 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2550 #ifdef CONFIG_SLUB_DEBUG
2551 static int count_free(struct page *page)
2553 return page->objects - page->inuse;
2556 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2558 return atomic_long_read(&n->total_objects);
2560 #endif /* CONFIG_SLUB_DEBUG */
2562 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2563 static unsigned long count_partial(struct kmem_cache_node *n,
2564 int (*get_count)(struct page *))
2566 unsigned long flags;
2567 unsigned long x = 0;
2570 spin_lock_irqsave(&n->list_lock, flags);
2571 list_for_each_entry(page, &n->partial, slab_list)
2572 x += get_count(page);
2573 spin_unlock_irqrestore(&n->list_lock, flags);
2576 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2578 static noinline void
2579 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2581 #ifdef CONFIG_SLUB_DEBUG
2582 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2583 DEFAULT_RATELIMIT_BURST);
2585 struct kmem_cache_node *n;
2587 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2590 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2591 nid, gfpflags, &gfpflags);
2592 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2593 s->name, s->object_size, s->size, oo_order(s->oo),
2596 if (oo_order(s->min) > get_order(s->object_size))
2597 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2600 for_each_kmem_cache_node(s, node, n) {
2601 unsigned long nr_slabs;
2602 unsigned long nr_objs;
2603 unsigned long nr_free;
2605 nr_free = count_partial(n, count_free);
2606 nr_slabs = node_nr_slabs(n);
2607 nr_objs = node_nr_objs(n);
2609 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2610 node, nr_slabs, nr_objs, nr_free);
2615 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2617 if (unlikely(PageSlabPfmemalloc(page)))
2618 return gfp_pfmemalloc_allowed(gfpflags);
2624 * Check the page->freelist of a page and either transfer the freelist to the
2625 * per cpu freelist or deactivate the page.
2627 * The page is still frozen if the return value is not NULL.
2629 * If this function returns NULL then the page has been unfrozen.
2631 * This function must be called with interrupt disabled.
2633 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2636 unsigned long counters;
2640 freelist = page->freelist;
2641 counters = page->counters;
2643 new.counters = counters;
2644 VM_BUG_ON(!new.frozen);
2646 new.inuse = page->objects;
2647 new.frozen = freelist != NULL;
2649 } while (!__cmpxchg_double_slab(s, page,
2658 * Slow path. The lockless freelist is empty or we need to perform
2661 * Processing is still very fast if new objects have been freed to the
2662 * regular freelist. In that case we simply take over the regular freelist
2663 * as the lockless freelist and zap the regular freelist.
2665 * If that is not working then we fall back to the partial lists. We take the
2666 * first element of the freelist as the object to allocate now and move the
2667 * rest of the freelist to the lockless freelist.
2669 * And if we were unable to get a new slab from the partial slab lists then
2670 * we need to allocate a new slab. This is the slowest path since it involves
2671 * a call to the page allocator and the setup of a new slab.
2673 * Version of __slab_alloc to use when we know that preemption is
2674 * already disabled (which is the case for bulk allocation).
2676 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2677 unsigned long addr, struct kmem_cache_cpu *c)
2681 unsigned long flags;
2683 stat(s, ALLOC_SLOWPATH);
2685 local_irq_save(flags);
2689 * if the node is not online or has no normal memory, just
2690 * ignore the node constraint
2692 if (unlikely(node != NUMA_NO_NODE &&
2693 !node_isset(node, slab_nodes)))
2694 node = NUMA_NO_NODE;
2699 if (unlikely(!node_match(page, node))) {
2701 * same as above but node_match() being false already
2702 * implies node != NUMA_NO_NODE
2704 if (!node_isset(node, slab_nodes)) {
2705 node = NUMA_NO_NODE;
2708 stat(s, ALLOC_NODE_MISMATCH);
2709 deactivate_slab(s, page, c->freelist, c);
2715 * By rights, we should be searching for a slab page that was
2716 * PFMEMALLOC but right now, we are losing the pfmemalloc
2717 * information when the page leaves the per-cpu allocator
2719 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2720 deactivate_slab(s, page, c->freelist, c);
2724 /* must check again c->freelist in case of cpu migration or IRQ */
2725 freelist = c->freelist;
2729 freelist = get_freelist(s, page);
2733 stat(s, DEACTIVATE_BYPASS);
2737 stat(s, ALLOC_REFILL);
2741 * freelist is pointing to the list of objects to be used.
2742 * page is pointing to the page from which the objects are obtained.
2743 * That page must be frozen for per cpu allocations to work.
2745 VM_BUG_ON(!c->page->frozen);
2746 c->freelist = get_freepointer(s, freelist);
2747 c->tid = next_tid(c->tid);
2748 local_irq_restore(flags);
2753 if (slub_percpu_partial(c)) {
2754 page = c->page = slub_percpu_partial(c);
2755 slub_set_percpu_partial(c, page);
2756 stat(s, CPU_PARTIAL_ALLOC);
2760 freelist = get_partial(s, gfpflags, node, &page);
2763 goto check_new_page;
2766 put_cpu_ptr(s->cpu_slab);
2767 page = new_slab(s, gfpflags, node);
2768 c = get_cpu_ptr(s->cpu_slab);
2770 if (unlikely(!page)) {
2771 local_irq_restore(flags);
2772 slab_out_of_memory(s, gfpflags, node);
2780 * No other reference to the page yet so we can
2781 * muck around with it freely without cmpxchg
2783 freelist = page->freelist;
2784 page->freelist = NULL;
2786 stat(s, ALLOC_SLAB);
2791 if (kmem_cache_debug(s)) {
2792 if (!alloc_debug_processing(s, page, freelist, addr))
2793 /* Slab failed checks. Next slab needed */
2797 * For debug case, we don't load freelist so that all
2798 * allocations go through alloc_debug_processing()
2803 if (unlikely(!pfmemalloc_match(page, gfpflags)))
2805 * For !pfmemalloc_match() case we don't load freelist so that
2806 * we don't make further mismatched allocations easier.
2814 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2815 local_irq_restore(flags);
2820 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
2821 * disabled. Compensates for possible cpu changes by refetching the per cpu area
2824 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2825 unsigned long addr, struct kmem_cache_cpu *c)
2829 #ifdef CONFIG_PREEMPT_COUNT
2831 * We may have been preempted and rescheduled on a different
2832 * cpu before disabling preemption. Need to reload cpu area
2835 c = get_cpu_ptr(s->cpu_slab);
2838 p = ___slab_alloc(s, gfpflags, node, addr, c);
2839 #ifdef CONFIG_PREEMPT_COUNT
2840 put_cpu_ptr(s->cpu_slab);
2846 * If the object has been wiped upon free, make sure it's fully initialized by
2847 * zeroing out freelist pointer.
2849 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2852 if (unlikely(slab_want_init_on_free(s)) && obj)
2853 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
2858 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2859 * have the fastpath folded into their functions. So no function call
2860 * overhead for requests that can be satisfied on the fastpath.
2862 * The fastpath works by first checking if the lockless freelist can be used.
2863 * If not then __slab_alloc is called for slow processing.
2865 * Otherwise we can simply pick the next object from the lockless free list.
2867 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2868 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
2871 struct kmem_cache_cpu *c;
2874 struct obj_cgroup *objcg = NULL;
2877 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2881 object = kfence_alloc(s, orig_size, gfpflags);
2882 if (unlikely(object))
2887 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2888 * enabled. We may switch back and forth between cpus while
2889 * reading from one cpu area. That does not matter as long
2890 * as we end up on the original cpu again when doing the cmpxchg.
2892 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
2893 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
2894 * the tid. If we are preempted and switched to another cpu between the
2895 * two reads, it's OK as the two are still associated with the same cpu
2896 * and cmpxchg later will validate the cpu.
2898 c = raw_cpu_ptr(s->cpu_slab);
2899 tid = READ_ONCE(c->tid);
2902 * Irqless object alloc/free algorithm used here depends on sequence
2903 * of fetching cpu_slab's data. tid should be fetched before anything
2904 * on c to guarantee that object and page associated with previous tid
2905 * won't be used with current tid. If we fetch tid first, object and
2906 * page could be one associated with next tid and our alloc/free
2907 * request will be failed. In this case, we will retry. So, no problem.
2912 * The transaction ids are globally unique per cpu and per operation on
2913 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2914 * occurs on the right processor and that there was no operation on the
2915 * linked list in between.
2918 object = c->freelist;
2920 if (unlikely(!object || !page || !node_match(page, node))) {
2921 object = __slab_alloc(s, gfpflags, node, addr, c);
2923 void *next_object = get_freepointer_safe(s, object);
2926 * The cmpxchg will only match if there was no additional
2927 * operation and if we are on the right processor.
2929 * The cmpxchg does the following atomically (without lock
2931 * 1. Relocate first pointer to the current per cpu area.
2932 * 2. Verify that tid and freelist have not been changed
2933 * 3. If they were not changed replace tid and freelist
2935 * Since this is without lock semantics the protection is only
2936 * against code executing on this cpu *not* from access by
2939 if (unlikely(!this_cpu_cmpxchg_double(
2940 s->cpu_slab->freelist, s->cpu_slab->tid,
2942 next_object, next_tid(tid)))) {
2944 note_cmpxchg_failure("slab_alloc", s, tid);
2947 prefetch_freepointer(s, next_object);
2948 stat(s, ALLOC_FASTPATH);
2951 maybe_wipe_obj_freeptr(s, object);
2952 init = slab_want_init_on_alloc(gfpflags, s);
2955 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
2960 static __always_inline void *slab_alloc(struct kmem_cache *s,
2961 gfp_t gfpflags, unsigned long addr, size_t orig_size)
2963 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr, orig_size);
2966 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2968 void *ret = slab_alloc(s, gfpflags, _RET_IP_, s->object_size);
2970 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2975 EXPORT_SYMBOL(kmem_cache_alloc);
2977 #ifdef CONFIG_TRACING
2978 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2980 void *ret = slab_alloc(s, gfpflags, _RET_IP_, size);
2981 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2982 ret = kasan_kmalloc(s, ret, size, gfpflags);
2985 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2989 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2991 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, s->object_size);
2993 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2994 s->object_size, s->size, gfpflags, node);
2998 EXPORT_SYMBOL(kmem_cache_alloc_node);
3000 #ifdef CONFIG_TRACING
3001 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3003 int node, size_t size)
3005 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_, size);
3007 trace_kmalloc_node(_RET_IP_, ret,
3008 size, s->size, gfpflags, node);
3010 ret = kasan_kmalloc(s, ret, size, gfpflags);
3013 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3015 #endif /* CONFIG_NUMA */
3018 * Slow path handling. This may still be called frequently since objects
3019 * have a longer lifetime than the cpu slabs in most processing loads.
3021 * So we still attempt to reduce cache line usage. Just take the slab
3022 * lock and free the item. If there is no additional partial page
3023 * handling required then we can return immediately.
3025 static void __slab_free(struct kmem_cache *s, struct page *page,
3026 void *head, void *tail, int cnt,
3033 unsigned long counters;
3034 struct kmem_cache_node *n = NULL;
3035 unsigned long flags;
3037 stat(s, FREE_SLOWPATH);
3039 if (kfence_free(head))
3042 if (kmem_cache_debug(s) &&
3043 !free_debug_processing(s, page, head, tail, cnt, addr))
3048 spin_unlock_irqrestore(&n->list_lock, flags);
3051 prior = page->freelist;
3052 counters = page->counters;
3053 set_freepointer(s, tail, prior);
3054 new.counters = counters;
3055 was_frozen = new.frozen;
3057 if ((!new.inuse || !prior) && !was_frozen) {
3059 if (kmem_cache_has_cpu_partial(s) && !prior) {
3062 * Slab was on no list before and will be
3064 * We can defer the list move and instead
3069 } else { /* Needs to be taken off a list */
3071 n = get_node(s, page_to_nid(page));
3073 * Speculatively acquire the list_lock.
3074 * If the cmpxchg does not succeed then we may
3075 * drop the list_lock without any processing.
3077 * Otherwise the list_lock will synchronize with
3078 * other processors updating the list of slabs.
3080 spin_lock_irqsave(&n->list_lock, flags);
3085 } while (!cmpxchg_double_slab(s, page,
3092 if (likely(was_frozen)) {
3094 * The list lock was not taken therefore no list
3095 * activity can be necessary.
3097 stat(s, FREE_FROZEN);
3098 } else if (new.frozen) {
3100 * If we just froze the page then put it onto the
3101 * per cpu partial list.
3103 put_cpu_partial(s, page, 1);
3104 stat(s, CPU_PARTIAL_FREE);
3110 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3114 * Objects left in the slab. If it was not on the partial list before
3117 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3118 remove_full(s, n, page);
3119 add_partial(n, page, DEACTIVATE_TO_TAIL);
3120 stat(s, FREE_ADD_PARTIAL);
3122 spin_unlock_irqrestore(&n->list_lock, flags);
3128 * Slab on the partial list.
3130 remove_partial(n, page);
3131 stat(s, FREE_REMOVE_PARTIAL);
3133 /* Slab must be on the full list */
3134 remove_full(s, n, page);
3137 spin_unlock_irqrestore(&n->list_lock, flags);
3139 discard_slab(s, page);
3143 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3144 * can perform fastpath freeing without additional function calls.
3146 * The fastpath is only possible if we are freeing to the current cpu slab
3147 * of this processor. This typically the case if we have just allocated
3150 * If fastpath is not possible then fall back to __slab_free where we deal
3151 * with all sorts of special processing.
3153 * Bulk free of a freelist with several objects (all pointing to the
3154 * same page) possible by specifying head and tail ptr, plus objects
3155 * count (cnt). Bulk free indicated by tail pointer being set.
3157 static __always_inline void do_slab_free(struct kmem_cache *s,
3158 struct page *page, void *head, void *tail,
3159 int cnt, unsigned long addr)
3161 void *tail_obj = tail ? : head;
3162 struct kmem_cache_cpu *c;
3165 memcg_slab_free_hook(s, &head, 1);
3168 * Determine the currently cpus per cpu slab.
3169 * The cpu may change afterward. However that does not matter since
3170 * data is retrieved via this pointer. If we are on the same cpu
3171 * during the cmpxchg then the free will succeed.
3173 c = raw_cpu_ptr(s->cpu_slab);
3174 tid = READ_ONCE(c->tid);
3176 /* Same with comment on barrier() in slab_alloc_node() */
3179 if (likely(page == c->page)) {
3180 void **freelist = READ_ONCE(c->freelist);
3182 set_freepointer(s, tail_obj, freelist);
3184 if (unlikely(!this_cpu_cmpxchg_double(
3185 s->cpu_slab->freelist, s->cpu_slab->tid,
3187 head, next_tid(tid)))) {
3189 note_cmpxchg_failure("slab_free", s, tid);
3192 stat(s, FREE_FASTPATH);
3194 __slab_free(s, page, head, tail_obj, cnt, addr);
3198 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3199 void *head, void *tail, int cnt,
3203 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3204 * to remove objects, whose reuse must be delayed.
3206 if (slab_free_freelist_hook(s, &head, &tail))
3207 do_slab_free(s, page, head, tail, cnt, addr);
3210 #ifdef CONFIG_KASAN_GENERIC
3211 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3213 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3217 void kmem_cache_free(struct kmem_cache *s, void *x)
3219 s = cache_from_obj(s, x);
3222 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3223 trace_kmem_cache_free(_RET_IP_, x, s->name);
3225 EXPORT_SYMBOL(kmem_cache_free);
3227 struct detached_freelist {
3232 struct kmem_cache *s;
3235 static inline void free_nonslab_page(struct page *page, void *object)
3237 unsigned int order = compound_order(page);
3239 VM_BUG_ON_PAGE(!PageCompound(page), page);
3241 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B, -(PAGE_SIZE << order));
3242 __free_pages(page, order);
3246 * This function progressively scans the array with free objects (with
3247 * a limited look ahead) and extract objects belonging to the same
3248 * page. It builds a detached freelist directly within the given
3249 * page/objects. This can happen without any need for
3250 * synchronization, because the objects are owned by running process.
3251 * The freelist is build up as a single linked list in the objects.
3252 * The idea is, that this detached freelist can then be bulk
3253 * transferred to the real freelist(s), but only requiring a single
3254 * synchronization primitive. Look ahead in the array is limited due
3255 * to performance reasons.
3258 int build_detached_freelist(struct kmem_cache *s, size_t size,
3259 void **p, struct detached_freelist *df)
3261 size_t first_skipped_index = 0;
3266 /* Always re-init detached_freelist */
3271 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3272 } while (!object && size);
3277 page = virt_to_head_page(object);
3279 /* Handle kalloc'ed objects */
3280 if (unlikely(!PageSlab(page))) {
3281 free_nonslab_page(page, object);
3282 p[size] = NULL; /* mark object processed */
3285 /* Derive kmem_cache from object */
3286 df->s = page->slab_cache;
3288 df->s = cache_from_obj(s, object); /* Support for memcg */
3291 if (is_kfence_address(object)) {
3292 slab_free_hook(df->s, object, false);
3293 __kfence_free(object);
3294 p[size] = NULL; /* mark object processed */
3298 /* Start new detached freelist */
3300 set_freepointer(df->s, object, NULL);
3302 df->freelist = object;
3303 p[size] = NULL; /* mark object processed */
3309 continue; /* Skip processed objects */
3311 /* df->page is always set at this point */
3312 if (df->page == virt_to_head_page(object)) {
3313 /* Opportunity build freelist */
3314 set_freepointer(df->s, object, df->freelist);
3315 df->freelist = object;
3317 p[size] = NULL; /* mark object processed */
3322 /* Limit look ahead search */
3326 if (!first_skipped_index)
3327 first_skipped_index = size + 1;
3330 return first_skipped_index;
3333 /* Note that interrupts must be enabled when calling this function. */
3334 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3339 memcg_slab_free_hook(s, p, size);
3341 struct detached_freelist df;
3343 size = build_detached_freelist(s, size, p, &df);
3347 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
3348 } while (likely(size));
3350 EXPORT_SYMBOL(kmem_cache_free_bulk);
3352 /* Note that interrupts must be enabled when calling this function. */
3353 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3356 struct kmem_cache_cpu *c;
3358 struct obj_cgroup *objcg = NULL;
3360 /* memcg and kmem_cache debug support */
3361 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3365 * Drain objects in the per cpu slab, while disabling local
3366 * IRQs, which protects against PREEMPT and interrupts
3367 * handlers invoking normal fastpath.
3369 c = get_cpu_ptr(s->cpu_slab);
3370 local_irq_disable();
3372 for (i = 0; i < size; i++) {
3373 void *object = kfence_alloc(s, s->object_size, flags);
3375 if (unlikely(object)) {
3380 object = c->freelist;
3381 if (unlikely(!object)) {
3383 * We may have removed an object from c->freelist using
3384 * the fastpath in the previous iteration; in that case,
3385 * c->tid has not been bumped yet.
3386 * Since ___slab_alloc() may reenable interrupts while
3387 * allocating memory, we should bump c->tid now.
3389 c->tid = next_tid(c->tid);
3394 * Invoking slow path likely have side-effect
3395 * of re-populating per CPU c->freelist
3397 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3399 if (unlikely(!p[i]))
3402 c = this_cpu_ptr(s->cpu_slab);
3403 maybe_wipe_obj_freeptr(s, p[i]);
3405 local_irq_disable();
3407 continue; /* goto for-loop */
3409 c->freelist = get_freepointer(s, object);
3411 maybe_wipe_obj_freeptr(s, p[i]);
3413 c->tid = next_tid(c->tid);
3415 put_cpu_ptr(s->cpu_slab);
3418 * memcg and kmem_cache debug support and memory initialization.
3419 * Done outside of the IRQ disabled fastpath loop.
3421 slab_post_alloc_hook(s, objcg, flags, size, p,
3422 slab_want_init_on_alloc(flags, s));
3425 put_cpu_ptr(s->cpu_slab);
3426 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3427 __kmem_cache_free_bulk(s, i, p);
3430 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3434 * Object placement in a slab is made very easy because we always start at
3435 * offset 0. If we tune the size of the object to the alignment then we can
3436 * get the required alignment by putting one properly sized object after
3439 * Notice that the allocation order determines the sizes of the per cpu
3440 * caches. Each processor has always one slab available for allocations.
3441 * Increasing the allocation order reduces the number of times that slabs
3442 * must be moved on and off the partial lists and is therefore a factor in
3447 * Minimum / Maximum order of slab pages. This influences locking overhead
3448 * and slab fragmentation. A higher order reduces the number of partial slabs
3449 * and increases the number of allocations possible without having to
3450 * take the list_lock.
3452 static unsigned int slub_min_order;
3453 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3454 static unsigned int slub_min_objects;
3457 * Calculate the order of allocation given an slab object size.
3459 * The order of allocation has significant impact on performance and other
3460 * system components. Generally order 0 allocations should be preferred since
3461 * order 0 does not cause fragmentation in the page allocator. Larger objects
3462 * be problematic to put into order 0 slabs because there may be too much
3463 * unused space left. We go to a higher order if more than 1/16th of the slab
3466 * In order to reach satisfactory performance we must ensure that a minimum
3467 * number of objects is in one slab. Otherwise we may generate too much
3468 * activity on the partial lists which requires taking the list_lock. This is
3469 * less a concern for large slabs though which are rarely used.
3471 * slub_max_order specifies the order where we begin to stop considering the
3472 * number of objects in a slab as critical. If we reach slub_max_order then
3473 * we try to keep the page order as low as possible. So we accept more waste
3474 * of space in favor of a small page order.
3476 * Higher order allocations also allow the placement of more objects in a
3477 * slab and thereby reduce object handling overhead. If the user has
3478 * requested a higher minimum order then we start with that one instead of
3479 * the smallest order which will fit the object.
3481 static inline unsigned int slab_order(unsigned int size,
3482 unsigned int min_objects, unsigned int max_order,
3483 unsigned int fract_leftover)
3485 unsigned int min_order = slub_min_order;
3488 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3489 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3491 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3492 order <= max_order; order++) {
3494 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3497 rem = slab_size % size;
3499 if (rem <= slab_size / fract_leftover)
3506 static inline int calculate_order(unsigned int size)
3509 unsigned int min_objects;
3510 unsigned int max_objects;
3511 unsigned int nr_cpus;
3514 * Attempt to find best configuration for a slab. This
3515 * works by first attempting to generate a layout with
3516 * the best configuration and backing off gradually.
3518 * First we increase the acceptable waste in a slab. Then
3519 * we reduce the minimum objects required in a slab.
3521 min_objects = slub_min_objects;
3524 * Some architectures will only update present cpus when
3525 * onlining them, so don't trust the number if it's just 1. But
3526 * we also don't want to use nr_cpu_ids always, as on some other
3527 * architectures, there can be many possible cpus, but never
3528 * onlined. Here we compromise between trying to avoid too high
3529 * order on systems that appear larger than they are, and too
3530 * low order on systems that appear smaller than they are.
3532 nr_cpus = num_present_cpus();
3534 nr_cpus = nr_cpu_ids;
3535 min_objects = 4 * (fls(nr_cpus) + 1);
3537 max_objects = order_objects(slub_max_order, size);
3538 min_objects = min(min_objects, max_objects);
3540 while (min_objects > 1) {
3541 unsigned int fraction;
3544 while (fraction >= 4) {
3545 order = slab_order(size, min_objects,
3546 slub_max_order, fraction);
3547 if (order <= slub_max_order)
3555 * We were unable to place multiple objects in a slab. Now
3556 * lets see if we can place a single object there.
3558 order = slab_order(size, 1, slub_max_order, 1);
3559 if (order <= slub_max_order)
3563 * Doh this slab cannot be placed using slub_max_order.
3565 order = slab_order(size, 1, MAX_ORDER, 1);
3566 if (order < MAX_ORDER)
3572 init_kmem_cache_node(struct kmem_cache_node *n)
3575 spin_lock_init(&n->list_lock);
3576 INIT_LIST_HEAD(&n->partial);
3577 #ifdef CONFIG_SLUB_DEBUG
3578 atomic_long_set(&n->nr_slabs, 0);
3579 atomic_long_set(&n->total_objects, 0);
3580 INIT_LIST_HEAD(&n->full);
3584 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3586 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3587 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3590 * Must align to double word boundary for the double cmpxchg
3591 * instructions to work; see __pcpu_double_call_return_bool().
3593 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3594 2 * sizeof(void *));
3599 init_kmem_cache_cpus(s);
3604 static struct kmem_cache *kmem_cache_node;
3607 * No kmalloc_node yet so do it by hand. We know that this is the first
3608 * slab on the node for this slabcache. There are no concurrent accesses
3611 * Note that this function only works on the kmem_cache_node
3612 * when allocating for the kmem_cache_node. This is used for bootstrapping
3613 * memory on a fresh node that has no slab structures yet.
3615 static void early_kmem_cache_node_alloc(int node)
3618 struct kmem_cache_node *n;
3620 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3622 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3625 if (page_to_nid(page) != node) {
3626 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3627 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3632 #ifdef CONFIG_SLUB_DEBUG
3633 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3634 init_tracking(kmem_cache_node, n);
3636 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3637 page->freelist = get_freepointer(kmem_cache_node, n);
3640 kmem_cache_node->node[node] = n;
3641 init_kmem_cache_node(n);
3642 inc_slabs_node(kmem_cache_node, node, page->objects);
3645 * No locks need to be taken here as it has just been
3646 * initialized and there is no concurrent access.
3648 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3651 static void free_kmem_cache_nodes(struct kmem_cache *s)
3654 struct kmem_cache_node *n;
3656 for_each_kmem_cache_node(s, node, n) {
3657 s->node[node] = NULL;
3658 kmem_cache_free(kmem_cache_node, n);
3662 void __kmem_cache_release(struct kmem_cache *s)
3664 cache_random_seq_destroy(s);
3665 free_percpu(s->cpu_slab);
3666 free_kmem_cache_nodes(s);
3669 static int init_kmem_cache_nodes(struct kmem_cache *s)
3673 for_each_node_mask(node, slab_nodes) {
3674 struct kmem_cache_node *n;
3676 if (slab_state == DOWN) {
3677 early_kmem_cache_node_alloc(node);
3680 n = kmem_cache_alloc_node(kmem_cache_node,
3684 free_kmem_cache_nodes(s);
3688 init_kmem_cache_node(n);
3694 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3696 if (min < MIN_PARTIAL)
3698 else if (min > MAX_PARTIAL)
3700 s->min_partial = min;
3703 static void set_cpu_partial(struct kmem_cache *s)
3705 #ifdef CONFIG_SLUB_CPU_PARTIAL
3707 * cpu_partial determined the maximum number of objects kept in the
3708 * per cpu partial lists of a processor.
3710 * Per cpu partial lists mainly contain slabs that just have one
3711 * object freed. If they are used for allocation then they can be
3712 * filled up again with minimal effort. The slab will never hit the
3713 * per node partial lists and therefore no locking will be required.
3715 * This setting also determines
3717 * A) The number of objects from per cpu partial slabs dumped to the
3718 * per node list when we reach the limit.
3719 * B) The number of objects in cpu partial slabs to extract from the
3720 * per node list when we run out of per cpu objects. We only fetch
3721 * 50% to keep some capacity around for frees.
3723 if (!kmem_cache_has_cpu_partial(s))
3724 slub_set_cpu_partial(s, 0);
3725 else if (s->size >= PAGE_SIZE)
3726 slub_set_cpu_partial(s, 2);
3727 else if (s->size >= 1024)
3728 slub_set_cpu_partial(s, 6);
3729 else if (s->size >= 256)
3730 slub_set_cpu_partial(s, 13);
3732 slub_set_cpu_partial(s, 30);
3737 * calculate_sizes() determines the order and the distribution of data within
3740 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3742 slab_flags_t flags = s->flags;
3743 unsigned int size = s->object_size;
3747 * Round up object size to the next word boundary. We can only
3748 * place the free pointer at word boundaries and this determines
3749 * the possible location of the free pointer.
3751 size = ALIGN(size, sizeof(void *));
3753 #ifdef CONFIG_SLUB_DEBUG
3755 * Determine if we can poison the object itself. If the user of
3756 * the slab may touch the object after free or before allocation
3757 * then we should never poison the object itself.
3759 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3761 s->flags |= __OBJECT_POISON;
3763 s->flags &= ~__OBJECT_POISON;
3767 * If we are Redzoning then check if there is some space between the
3768 * end of the object and the free pointer. If not then add an
3769 * additional word to have some bytes to store Redzone information.
3771 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3772 size += sizeof(void *);
3776 * With that we have determined the number of bytes in actual use
3777 * by the object and redzoning.
3781 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3782 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
3785 * Relocate free pointer after the object if it is not
3786 * permitted to overwrite the first word of the object on
3789 * This is the case if we do RCU, have a constructor or
3790 * destructor, are poisoning the objects, or are
3791 * redzoning an object smaller than sizeof(void *).
3793 * The assumption that s->offset >= s->inuse means free
3794 * pointer is outside of the object is used in the
3795 * freeptr_outside_object() function. If that is no
3796 * longer true, the function needs to be modified.
3799 size += sizeof(void *);
3802 * Store freelist pointer near middle of object to keep
3803 * it away from the edges of the object to avoid small
3804 * sized over/underflows from neighboring allocations.
3806 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
3809 #ifdef CONFIG_SLUB_DEBUG
3810 if (flags & SLAB_STORE_USER)
3812 * Need to store information about allocs and frees after
3815 size += 2 * sizeof(struct track);
3818 kasan_cache_create(s, &size, &s->flags);
3819 #ifdef CONFIG_SLUB_DEBUG
3820 if (flags & SLAB_RED_ZONE) {
3822 * Add some empty padding so that we can catch
3823 * overwrites from earlier objects rather than let
3824 * tracking information or the free pointer be
3825 * corrupted if a user writes before the start
3828 size += sizeof(void *);
3830 s->red_left_pad = sizeof(void *);
3831 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3832 size += s->red_left_pad;
3837 * SLUB stores one object immediately after another beginning from
3838 * offset 0. In order to align the objects we have to simply size
3839 * each object to conform to the alignment.
3841 size = ALIGN(size, s->align);
3843 s->reciprocal_size = reciprocal_value(size);
3844 if (forced_order >= 0)
3845 order = forced_order;
3847 order = calculate_order(size);
3854 s->allocflags |= __GFP_COMP;
3856 if (s->flags & SLAB_CACHE_DMA)
3857 s->allocflags |= GFP_DMA;
3859 if (s->flags & SLAB_CACHE_DMA32)
3860 s->allocflags |= GFP_DMA32;
3862 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3863 s->allocflags |= __GFP_RECLAIMABLE;
3866 * Determine the number of objects per slab
3868 s->oo = oo_make(order, size);
3869 s->min = oo_make(get_order(size), size);
3870 if (oo_objects(s->oo) > oo_objects(s->max))
3873 return !!oo_objects(s->oo);
3876 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3878 s->flags = kmem_cache_flags(s->size, flags, s->name);
3879 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3880 s->random = get_random_long();
3883 if (!calculate_sizes(s, -1))
3885 if (disable_higher_order_debug) {
3887 * Disable debugging flags that store metadata if the min slab
3890 if (get_order(s->size) > get_order(s->object_size)) {
3891 s->flags &= ~DEBUG_METADATA_FLAGS;
3893 if (!calculate_sizes(s, -1))
3898 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3899 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3900 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3901 /* Enable fast mode */
3902 s->flags |= __CMPXCHG_DOUBLE;
3906 * The larger the object size is, the more pages we want on the partial
3907 * list to avoid pounding the page allocator excessively.
3909 set_min_partial(s, ilog2(s->size) / 2);
3914 s->remote_node_defrag_ratio = 1000;
3917 /* Initialize the pre-computed randomized freelist if slab is up */
3918 if (slab_state >= UP) {
3919 if (init_cache_random_seq(s))
3923 if (!init_kmem_cache_nodes(s))
3926 if (alloc_kmem_cache_cpus(s))
3929 free_kmem_cache_nodes(s);
3934 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3937 #ifdef CONFIG_SLUB_DEBUG
3938 void *addr = page_address(page);
3942 slab_err(s, page, text, s->name);
3945 map = get_map(s, page);
3946 for_each_object(p, s, addr, page->objects) {
3948 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3949 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
3950 print_tracking(s, p);
3959 * Attempt to free all partial slabs on a node.
3960 * This is called from __kmem_cache_shutdown(). We must take list_lock
3961 * because sysfs file might still access partial list after the shutdowning.
3963 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3966 struct page *page, *h;
3968 BUG_ON(irqs_disabled());
3969 spin_lock_irq(&n->list_lock);
3970 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3972 remove_partial(n, page);
3973 list_add(&page->slab_list, &discard);
3975 list_slab_objects(s, page,
3976 "Objects remaining in %s on __kmem_cache_shutdown()");
3979 spin_unlock_irq(&n->list_lock);
3981 list_for_each_entry_safe(page, h, &discard, slab_list)
3982 discard_slab(s, page);
3985 bool __kmem_cache_empty(struct kmem_cache *s)
3988 struct kmem_cache_node *n;
3990 for_each_kmem_cache_node(s, node, n)
3991 if (n->nr_partial || slabs_node(s, node))
3997 * Release all resources used by a slab cache.
3999 int __kmem_cache_shutdown(struct kmem_cache *s)
4002 struct kmem_cache_node *n;
4005 /* Attempt to free all objects */
4006 for_each_kmem_cache_node(s, node, n) {
4008 if (n->nr_partial || slabs_node(s, node))
4014 #ifdef CONFIG_PRINTK
4015 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
4018 int __maybe_unused i;
4022 struct kmem_cache *s = page->slab_cache;
4023 struct track __maybe_unused *trackp;
4025 kpp->kp_ptr = object;
4026 kpp->kp_page = page;
4027 kpp->kp_slab_cache = s;
4028 base = page_address(page);
4029 objp0 = kasan_reset_tag(object);
4030 #ifdef CONFIG_SLUB_DEBUG
4031 objp = restore_red_left(s, objp0);
4035 objnr = obj_to_index(s, page, objp);
4036 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4037 objp = base + s->size * objnr;
4038 kpp->kp_objp = objp;
4039 if (WARN_ON_ONCE(objp < base || objp >= base + page->objects * s->size || (objp - base) % s->size) ||
4040 !(s->flags & SLAB_STORE_USER))
4042 #ifdef CONFIG_SLUB_DEBUG
4043 objp = fixup_red_left(s, objp);
4044 trackp = get_track(s, objp, TRACK_ALLOC);
4045 kpp->kp_ret = (void *)trackp->addr;
4046 #ifdef CONFIG_STACKTRACE
4047 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4048 kpp->kp_stack[i] = (void *)trackp->addrs[i];
4049 if (!kpp->kp_stack[i])
4053 trackp = get_track(s, objp, TRACK_FREE);
4054 for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4055 kpp->kp_free_stack[i] = (void *)trackp->addrs[i];
4056 if (!kpp->kp_free_stack[i])
4064 /********************************************************************
4066 *******************************************************************/
4068 static int __init setup_slub_min_order(char *str)
4070 get_option(&str, (int *)&slub_min_order);
4075 __setup("slub_min_order=", setup_slub_min_order);
4077 static int __init setup_slub_max_order(char *str)
4079 get_option(&str, (int *)&slub_max_order);
4080 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4085 __setup("slub_max_order=", setup_slub_max_order);
4087 static int __init setup_slub_min_objects(char *str)
4089 get_option(&str, (int *)&slub_min_objects);
4094 __setup("slub_min_objects=", setup_slub_min_objects);
4096 void *__kmalloc(size_t size, gfp_t flags)
4098 struct kmem_cache *s;
4101 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4102 return kmalloc_large(size, flags);
4104 s = kmalloc_slab(size, flags);
4106 if (unlikely(ZERO_OR_NULL_PTR(s)))
4109 ret = slab_alloc(s, flags, _RET_IP_, size);
4111 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4113 ret = kasan_kmalloc(s, ret, size, flags);
4117 EXPORT_SYMBOL(__kmalloc);
4120 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4124 unsigned int order = get_order(size);
4126 flags |= __GFP_COMP;
4127 page = alloc_pages_node(node, flags, order);
4129 ptr = page_address(page);
4130 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4131 PAGE_SIZE << order);
4134 return kmalloc_large_node_hook(ptr, size, flags);
4137 void *__kmalloc_node(size_t size, gfp_t flags, int node)
4139 struct kmem_cache *s;
4142 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4143 ret = kmalloc_large_node(size, flags, node);
4145 trace_kmalloc_node(_RET_IP_, ret,
4146 size, PAGE_SIZE << get_order(size),
4152 s = kmalloc_slab(size, flags);
4154 if (unlikely(ZERO_OR_NULL_PTR(s)))
4157 ret = slab_alloc_node(s, flags, node, _RET_IP_, size);
4159 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4161 ret = kasan_kmalloc(s, ret, size, flags);
4165 EXPORT_SYMBOL(__kmalloc_node);
4166 #endif /* CONFIG_NUMA */
4168 #ifdef CONFIG_HARDENED_USERCOPY
4170 * Rejects incorrectly sized objects and objects that are to be copied
4171 * to/from userspace but do not fall entirely within the containing slab
4172 * cache's usercopy region.
4174 * Returns NULL if check passes, otherwise const char * to name of cache
4175 * to indicate an error.
4177 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4180 struct kmem_cache *s;
4181 unsigned int offset;
4183 bool is_kfence = is_kfence_address(ptr);
4185 ptr = kasan_reset_tag(ptr);
4187 /* Find object and usable object size. */
4188 s = page->slab_cache;
4190 /* Reject impossible pointers. */
4191 if (ptr < page_address(page))
4192 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4195 /* Find offset within object. */
4197 offset = ptr - kfence_object_start(ptr);
4199 offset = (ptr - page_address(page)) % s->size;
4201 /* Adjust for redzone and reject if within the redzone. */
4202 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4203 if (offset < s->red_left_pad)
4204 usercopy_abort("SLUB object in left red zone",
4205 s->name, to_user, offset, n);
4206 offset -= s->red_left_pad;
4209 /* Allow address range falling entirely within usercopy region. */
4210 if (offset >= s->useroffset &&
4211 offset - s->useroffset <= s->usersize &&
4212 n <= s->useroffset - offset + s->usersize)
4216 * If the copy is still within the allocated object, produce
4217 * a warning instead of rejecting the copy. This is intended
4218 * to be a temporary method to find any missing usercopy
4221 object_size = slab_ksize(s);
4222 if (usercopy_fallback &&
4223 offset <= object_size && n <= object_size - offset) {
4224 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4228 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4230 #endif /* CONFIG_HARDENED_USERCOPY */
4232 size_t __ksize(const void *object)
4236 if (unlikely(object == ZERO_SIZE_PTR))
4239 page = virt_to_head_page(object);
4241 if (unlikely(!PageSlab(page))) {
4242 WARN_ON(!PageCompound(page));
4243 return page_size(page);
4246 return slab_ksize(page->slab_cache);
4248 EXPORT_SYMBOL(__ksize);
4250 void kfree(const void *x)
4253 void *object = (void *)x;
4255 trace_kfree(_RET_IP_, x);
4257 if (unlikely(ZERO_OR_NULL_PTR(x)))
4260 page = virt_to_head_page(x);
4261 if (unlikely(!PageSlab(page))) {
4262 free_nonslab_page(page, object);
4265 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4267 EXPORT_SYMBOL(kfree);
4269 #define SHRINK_PROMOTE_MAX 32
4272 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4273 * up most to the head of the partial lists. New allocations will then
4274 * fill those up and thus they can be removed from the partial lists.
4276 * The slabs with the least items are placed last. This results in them
4277 * being allocated from last increasing the chance that the last objects
4278 * are freed in them.
4280 int __kmem_cache_shrink(struct kmem_cache *s)
4284 struct kmem_cache_node *n;
4287 struct list_head discard;
4288 struct list_head promote[SHRINK_PROMOTE_MAX];
4289 unsigned long flags;
4293 for_each_kmem_cache_node(s, node, n) {
4294 INIT_LIST_HEAD(&discard);
4295 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4296 INIT_LIST_HEAD(promote + i);
4298 spin_lock_irqsave(&n->list_lock, flags);
4301 * Build lists of slabs to discard or promote.
4303 * Note that concurrent frees may occur while we hold the
4304 * list_lock. page->inuse here is the upper limit.
4306 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4307 int free = page->objects - page->inuse;
4309 /* Do not reread page->inuse */
4312 /* We do not keep full slabs on the list */
4315 if (free == page->objects) {
4316 list_move(&page->slab_list, &discard);
4318 } else if (free <= SHRINK_PROMOTE_MAX)
4319 list_move(&page->slab_list, promote + free - 1);
4323 * Promote the slabs filled up most to the head of the
4326 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4327 list_splice(promote + i, &n->partial);
4329 spin_unlock_irqrestore(&n->list_lock, flags);
4331 /* Release empty slabs */
4332 list_for_each_entry_safe(page, t, &discard, slab_list)
4333 discard_slab(s, page);
4335 if (slabs_node(s, node))
4342 static int slab_mem_going_offline_callback(void *arg)
4344 struct kmem_cache *s;
4346 mutex_lock(&slab_mutex);
4347 list_for_each_entry(s, &slab_caches, list)
4348 __kmem_cache_shrink(s);
4349 mutex_unlock(&slab_mutex);
4354 static void slab_mem_offline_callback(void *arg)
4356 struct memory_notify *marg = arg;
4359 offline_node = marg->status_change_nid_normal;
4362 * If the node still has available memory. we need kmem_cache_node
4365 if (offline_node < 0)
4368 mutex_lock(&slab_mutex);
4369 node_clear(offline_node, slab_nodes);
4371 * We no longer free kmem_cache_node structures here, as it would be
4372 * racy with all get_node() users, and infeasible to protect them with
4375 mutex_unlock(&slab_mutex);
4378 static int slab_mem_going_online_callback(void *arg)
4380 struct kmem_cache_node *n;
4381 struct kmem_cache *s;
4382 struct memory_notify *marg = arg;
4383 int nid = marg->status_change_nid_normal;
4387 * If the node's memory is already available, then kmem_cache_node is
4388 * already created. Nothing to do.
4394 * We are bringing a node online. No memory is available yet. We must
4395 * allocate a kmem_cache_node structure in order to bring the node
4398 mutex_lock(&slab_mutex);
4399 list_for_each_entry(s, &slab_caches, list) {
4401 * The structure may already exist if the node was previously
4402 * onlined and offlined.
4404 if (get_node(s, nid))
4407 * XXX: kmem_cache_alloc_node will fallback to other nodes
4408 * since memory is not yet available from the node that
4411 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4416 init_kmem_cache_node(n);
4420 * Any cache created after this point will also have kmem_cache_node
4421 * initialized for the new node.
4423 node_set(nid, slab_nodes);
4425 mutex_unlock(&slab_mutex);
4429 static int slab_memory_callback(struct notifier_block *self,
4430 unsigned long action, void *arg)
4435 case MEM_GOING_ONLINE:
4436 ret = slab_mem_going_online_callback(arg);
4438 case MEM_GOING_OFFLINE:
4439 ret = slab_mem_going_offline_callback(arg);
4442 case MEM_CANCEL_ONLINE:
4443 slab_mem_offline_callback(arg);
4446 case MEM_CANCEL_OFFLINE:
4450 ret = notifier_from_errno(ret);
4456 static struct notifier_block slab_memory_callback_nb = {
4457 .notifier_call = slab_memory_callback,
4458 .priority = SLAB_CALLBACK_PRI,
4461 /********************************************************************
4462 * Basic setup of slabs
4463 *******************************************************************/
4466 * Used for early kmem_cache structures that were allocated using
4467 * the page allocator. Allocate them properly then fix up the pointers
4468 * that may be pointing to the wrong kmem_cache structure.
4471 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4474 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4475 struct kmem_cache_node *n;
4477 memcpy(s, static_cache, kmem_cache->object_size);
4480 * This runs very early, and only the boot processor is supposed to be
4481 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4484 __flush_cpu_slab(s, smp_processor_id());
4485 for_each_kmem_cache_node(s, node, n) {
4488 list_for_each_entry(p, &n->partial, slab_list)
4491 #ifdef CONFIG_SLUB_DEBUG
4492 list_for_each_entry(p, &n->full, slab_list)
4496 list_add(&s->list, &slab_caches);
4500 void __init kmem_cache_init(void)
4502 static __initdata struct kmem_cache boot_kmem_cache,
4503 boot_kmem_cache_node;
4506 if (debug_guardpage_minorder())
4509 /* Print slub debugging pointers without hashing */
4510 if (__slub_debug_enabled())
4511 no_hash_pointers_enable(NULL);
4513 kmem_cache_node = &boot_kmem_cache_node;
4514 kmem_cache = &boot_kmem_cache;
4517 * Initialize the nodemask for which we will allocate per node
4518 * structures. Here we don't need taking slab_mutex yet.
4520 for_each_node_state(node, N_NORMAL_MEMORY)
4521 node_set(node, slab_nodes);
4523 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4524 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4526 register_hotmemory_notifier(&slab_memory_callback_nb);
4528 /* Able to allocate the per node structures */
4529 slab_state = PARTIAL;
4531 create_boot_cache(kmem_cache, "kmem_cache",
4532 offsetof(struct kmem_cache, node) +
4533 nr_node_ids * sizeof(struct kmem_cache_node *),
4534 SLAB_HWCACHE_ALIGN, 0, 0);
4536 kmem_cache = bootstrap(&boot_kmem_cache);
4537 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4539 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4540 setup_kmalloc_cache_index_table();
4541 create_kmalloc_caches(0);
4543 /* Setup random freelists for each cache */
4544 init_freelist_randomization();
4546 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4549 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4551 slub_min_order, slub_max_order, slub_min_objects,
4552 nr_cpu_ids, nr_node_ids);
4555 void __init kmem_cache_init_late(void)
4560 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4561 slab_flags_t flags, void (*ctor)(void *))
4563 struct kmem_cache *s;
4565 s = find_mergeable(size, align, flags, name, ctor);
4570 * Adjust the object sizes so that we clear
4571 * the complete object on kzalloc.
4573 s->object_size = max(s->object_size, size);
4574 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4576 if (sysfs_slab_alias(s, name)) {
4585 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4589 err = kmem_cache_open(s, flags);
4593 /* Mutex is not taken during early boot */
4594 if (slab_state <= UP)
4597 err = sysfs_slab_add(s);
4599 __kmem_cache_release(s);
4601 if (s->flags & SLAB_STORE_USER)
4602 debugfs_slab_add(s);
4607 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4609 struct kmem_cache *s;
4612 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4613 return kmalloc_large(size, gfpflags);
4615 s = kmalloc_slab(size, gfpflags);
4617 if (unlikely(ZERO_OR_NULL_PTR(s)))
4620 ret = slab_alloc(s, gfpflags, caller, size);
4622 /* Honor the call site pointer we received. */
4623 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4627 EXPORT_SYMBOL(__kmalloc_track_caller);
4630 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4631 int node, unsigned long caller)
4633 struct kmem_cache *s;
4636 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4637 ret = kmalloc_large_node(size, gfpflags, node);
4639 trace_kmalloc_node(caller, ret,
4640 size, PAGE_SIZE << get_order(size),
4646 s = kmalloc_slab(size, gfpflags);
4648 if (unlikely(ZERO_OR_NULL_PTR(s)))
4651 ret = slab_alloc_node(s, gfpflags, node, caller, size);
4653 /* Honor the call site pointer we received. */
4654 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4658 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4662 static int count_inuse(struct page *page)
4667 static int count_total(struct page *page)
4669 return page->objects;
4673 #ifdef CONFIG_SLUB_DEBUG
4674 static void validate_slab(struct kmem_cache *s, struct page *page,
4675 unsigned long *obj_map)
4678 void *addr = page_address(page);
4682 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4685 /* Now we know that a valid freelist exists */
4686 __fill_map(obj_map, s, page);
4687 for_each_object(p, s, addr, page->objects) {
4688 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
4689 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4691 if (!check_object(s, page, p, val))
4698 static int validate_slab_node(struct kmem_cache *s,
4699 struct kmem_cache_node *n, unsigned long *obj_map)
4701 unsigned long count = 0;
4703 unsigned long flags;
4705 spin_lock_irqsave(&n->list_lock, flags);
4707 list_for_each_entry(page, &n->partial, slab_list) {
4708 validate_slab(s, page, obj_map);
4711 if (count != n->nr_partial) {
4712 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4713 s->name, count, n->nr_partial);
4714 slab_add_kunit_errors();
4717 if (!(s->flags & SLAB_STORE_USER))
4720 list_for_each_entry(page, &n->full, slab_list) {
4721 validate_slab(s, page, obj_map);
4724 if (count != atomic_long_read(&n->nr_slabs)) {
4725 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4726 s->name, count, atomic_long_read(&n->nr_slabs));
4727 slab_add_kunit_errors();
4731 spin_unlock_irqrestore(&n->list_lock, flags);
4735 long validate_slab_cache(struct kmem_cache *s)
4738 unsigned long count = 0;
4739 struct kmem_cache_node *n;
4740 unsigned long *obj_map;
4742 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
4747 for_each_kmem_cache_node(s, node, n)
4748 count += validate_slab_node(s, n, obj_map);
4750 bitmap_free(obj_map);
4754 EXPORT_SYMBOL(validate_slab_cache);
4756 #ifdef CONFIG_DEBUG_FS
4758 * Generate lists of code addresses where slabcache objects are allocated
4763 unsigned long count;
4770 DECLARE_BITMAP(cpus, NR_CPUS);
4776 unsigned long count;
4777 struct location *loc;
4780 static struct dentry *slab_debugfs_root;
4782 static void free_loc_track(struct loc_track *t)
4785 free_pages((unsigned long)t->loc,
4786 get_order(sizeof(struct location) * t->max));
4789 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4794 order = get_order(sizeof(struct location) * max);
4796 l = (void *)__get_free_pages(flags, order);
4801 memcpy(l, t->loc, sizeof(struct location) * t->count);
4809 static int add_location(struct loc_track *t, struct kmem_cache *s,
4810 const struct track *track)
4812 long start, end, pos;
4814 unsigned long caddr;
4815 unsigned long age = jiffies - track->when;
4821 pos = start + (end - start + 1) / 2;
4824 * There is nothing at "end". If we end up there
4825 * we need to add something to before end.
4830 caddr = t->loc[pos].addr;
4831 if (track->addr == caddr) {
4837 if (age < l->min_time)
4839 if (age > l->max_time)
4842 if (track->pid < l->min_pid)
4843 l->min_pid = track->pid;
4844 if (track->pid > l->max_pid)
4845 l->max_pid = track->pid;
4847 cpumask_set_cpu(track->cpu,
4848 to_cpumask(l->cpus));
4850 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4854 if (track->addr < caddr)
4861 * Not found. Insert new tracking element.
4863 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4869 (t->count - pos) * sizeof(struct location));
4872 l->addr = track->addr;
4876 l->min_pid = track->pid;
4877 l->max_pid = track->pid;
4878 cpumask_clear(to_cpumask(l->cpus));
4879 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4880 nodes_clear(l->nodes);
4881 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4885 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4886 struct page *page, enum track_item alloc,
4887 unsigned long *obj_map)
4889 void *addr = page_address(page);
4892 __fill_map(obj_map, s, page);
4894 for_each_object(p, s, addr, page->objects)
4895 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
4896 add_location(t, s, get_track(s, p, alloc));
4898 #endif /* CONFIG_DEBUG_FS */
4899 #endif /* CONFIG_SLUB_DEBUG */
4902 enum slab_stat_type {
4903 SL_ALL, /* All slabs */
4904 SL_PARTIAL, /* Only partially allocated slabs */
4905 SL_CPU, /* Only slabs used for cpu caches */
4906 SL_OBJECTS, /* Determine allocated objects not slabs */
4907 SL_TOTAL /* Determine object capacity not slabs */
4910 #define SO_ALL (1 << SL_ALL)
4911 #define SO_PARTIAL (1 << SL_PARTIAL)
4912 #define SO_CPU (1 << SL_CPU)
4913 #define SO_OBJECTS (1 << SL_OBJECTS)
4914 #define SO_TOTAL (1 << SL_TOTAL)
4916 static ssize_t show_slab_objects(struct kmem_cache *s,
4917 char *buf, unsigned long flags)
4919 unsigned long total = 0;
4922 unsigned long *nodes;
4925 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4929 if (flags & SO_CPU) {
4932 for_each_possible_cpu(cpu) {
4933 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4938 page = READ_ONCE(c->page);
4942 node = page_to_nid(page);
4943 if (flags & SO_TOTAL)
4945 else if (flags & SO_OBJECTS)
4953 page = slub_percpu_partial_read_once(c);
4955 node = page_to_nid(page);
4956 if (flags & SO_TOTAL)
4958 else if (flags & SO_OBJECTS)
4969 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4970 * already held which will conflict with an existing lock order:
4972 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4974 * We don't really need mem_hotplug_lock (to hold off
4975 * slab_mem_going_offline_callback) here because slab's memory hot
4976 * unplug code doesn't destroy the kmem_cache->node[] data.
4979 #ifdef CONFIG_SLUB_DEBUG
4980 if (flags & SO_ALL) {
4981 struct kmem_cache_node *n;
4983 for_each_kmem_cache_node(s, node, n) {
4985 if (flags & SO_TOTAL)
4986 x = atomic_long_read(&n->total_objects);
4987 else if (flags & SO_OBJECTS)
4988 x = atomic_long_read(&n->total_objects) -
4989 count_partial(n, count_free);
4991 x = atomic_long_read(&n->nr_slabs);
4998 if (flags & SO_PARTIAL) {
4999 struct kmem_cache_node *n;
5001 for_each_kmem_cache_node(s, node, n) {
5002 if (flags & SO_TOTAL)
5003 x = count_partial(n, count_total);
5004 else if (flags & SO_OBJECTS)
5005 x = count_partial(n, count_inuse);
5013 len += sysfs_emit_at(buf, len, "%lu", total);
5015 for (node = 0; node < nr_node_ids; node++) {
5017 len += sysfs_emit_at(buf, len, " N%d=%lu",
5021 len += sysfs_emit_at(buf, len, "\n");
5027 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5028 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5030 struct slab_attribute {
5031 struct attribute attr;
5032 ssize_t (*show)(struct kmem_cache *s, char *buf);
5033 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5036 #define SLAB_ATTR_RO(_name) \
5037 static struct slab_attribute _name##_attr = \
5038 __ATTR(_name, 0400, _name##_show, NULL)
5040 #define SLAB_ATTR(_name) \
5041 static struct slab_attribute _name##_attr = \
5042 __ATTR(_name, 0600, _name##_show, _name##_store)
5044 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5046 return sysfs_emit(buf, "%u\n", s->size);
5048 SLAB_ATTR_RO(slab_size);
5050 static ssize_t align_show(struct kmem_cache *s, char *buf)
5052 return sysfs_emit(buf, "%u\n", s->align);
5054 SLAB_ATTR_RO(align);
5056 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5058 return sysfs_emit(buf, "%u\n", s->object_size);
5060 SLAB_ATTR_RO(object_size);
5062 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5064 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5066 SLAB_ATTR_RO(objs_per_slab);
5068 static ssize_t order_show(struct kmem_cache *s, char *buf)
5070 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5072 SLAB_ATTR_RO(order);
5074 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5076 return sysfs_emit(buf, "%lu\n", s->min_partial);
5079 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5085 err = kstrtoul(buf, 10, &min);
5089 set_min_partial(s, min);
5092 SLAB_ATTR(min_partial);
5094 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5096 return sysfs_emit(buf, "%u\n", slub_cpu_partial(s));
5099 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5102 unsigned int objects;
5105 err = kstrtouint(buf, 10, &objects);
5108 if (objects && !kmem_cache_has_cpu_partial(s))
5111 slub_set_cpu_partial(s, objects);
5115 SLAB_ATTR(cpu_partial);
5117 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5121 return sysfs_emit(buf, "%pS\n", s->ctor);
5125 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5127 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5129 SLAB_ATTR_RO(aliases);
5131 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5133 return show_slab_objects(s, buf, SO_PARTIAL);
5135 SLAB_ATTR_RO(partial);
5137 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5139 return show_slab_objects(s, buf, SO_CPU);
5141 SLAB_ATTR_RO(cpu_slabs);
5143 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5145 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5147 SLAB_ATTR_RO(objects);
5149 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5151 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5153 SLAB_ATTR_RO(objects_partial);
5155 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5162 for_each_online_cpu(cpu) {
5165 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5168 pages += page->pages;
5169 objects += page->pobjects;
5173 len += sysfs_emit_at(buf, len, "%d(%d)", objects, pages);
5176 for_each_online_cpu(cpu) {
5179 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5181 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5182 cpu, page->pobjects, page->pages);
5185 len += sysfs_emit_at(buf, len, "\n");
5189 SLAB_ATTR_RO(slabs_cpu_partial);
5191 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5193 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5195 SLAB_ATTR_RO(reclaim_account);
5197 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5199 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5201 SLAB_ATTR_RO(hwcache_align);
5203 #ifdef CONFIG_ZONE_DMA
5204 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5206 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5208 SLAB_ATTR_RO(cache_dma);
5211 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5213 return sysfs_emit(buf, "%u\n", s->usersize);
5215 SLAB_ATTR_RO(usersize);
5217 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5219 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5221 SLAB_ATTR_RO(destroy_by_rcu);
5223 #ifdef CONFIG_SLUB_DEBUG
5224 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5226 return show_slab_objects(s, buf, SO_ALL);
5228 SLAB_ATTR_RO(slabs);
5230 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5232 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5234 SLAB_ATTR_RO(total_objects);
5236 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5238 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5240 SLAB_ATTR_RO(sanity_checks);
5242 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5244 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5246 SLAB_ATTR_RO(trace);
5248 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5250 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5253 SLAB_ATTR_RO(red_zone);
5255 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5257 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5260 SLAB_ATTR_RO(poison);
5262 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5264 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5267 SLAB_ATTR_RO(store_user);
5269 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5274 static ssize_t validate_store(struct kmem_cache *s,
5275 const char *buf, size_t length)
5279 if (buf[0] == '1') {
5280 ret = validate_slab_cache(s);
5286 SLAB_ATTR(validate);
5288 #endif /* CONFIG_SLUB_DEBUG */
5290 #ifdef CONFIG_FAILSLAB
5291 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5293 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5295 SLAB_ATTR_RO(failslab);
5298 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5303 static ssize_t shrink_store(struct kmem_cache *s,
5304 const char *buf, size_t length)
5307 kmem_cache_shrink(s);
5315 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5317 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5320 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5321 const char *buf, size_t length)
5326 err = kstrtouint(buf, 10, &ratio);
5332 s->remote_node_defrag_ratio = ratio * 10;
5336 SLAB_ATTR(remote_node_defrag_ratio);
5339 #ifdef CONFIG_SLUB_STATS
5340 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5342 unsigned long sum = 0;
5345 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5350 for_each_online_cpu(cpu) {
5351 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5357 len += sysfs_emit_at(buf, len, "%lu", sum);
5360 for_each_online_cpu(cpu) {
5362 len += sysfs_emit_at(buf, len, " C%d=%u",
5367 len += sysfs_emit_at(buf, len, "\n");
5372 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5376 for_each_online_cpu(cpu)
5377 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5380 #define STAT_ATTR(si, text) \
5381 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5383 return show_stat(s, buf, si); \
5385 static ssize_t text##_store(struct kmem_cache *s, \
5386 const char *buf, size_t length) \
5388 if (buf[0] != '0') \
5390 clear_stat(s, si); \
5395 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5396 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5397 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5398 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5399 STAT_ATTR(FREE_FROZEN, free_frozen);
5400 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5401 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5402 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5403 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5404 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5405 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5406 STAT_ATTR(FREE_SLAB, free_slab);
5407 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5408 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5409 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5410 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5411 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5412 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5413 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5414 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5415 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5416 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5417 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5418 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5419 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5420 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5421 #endif /* CONFIG_SLUB_STATS */
5423 static struct attribute *slab_attrs[] = {
5424 &slab_size_attr.attr,
5425 &object_size_attr.attr,
5426 &objs_per_slab_attr.attr,
5428 &min_partial_attr.attr,
5429 &cpu_partial_attr.attr,
5431 &objects_partial_attr.attr,
5433 &cpu_slabs_attr.attr,
5437 &hwcache_align_attr.attr,
5438 &reclaim_account_attr.attr,
5439 &destroy_by_rcu_attr.attr,
5441 &slabs_cpu_partial_attr.attr,
5442 #ifdef CONFIG_SLUB_DEBUG
5443 &total_objects_attr.attr,
5445 &sanity_checks_attr.attr,
5447 &red_zone_attr.attr,
5449 &store_user_attr.attr,
5450 &validate_attr.attr,
5452 #ifdef CONFIG_ZONE_DMA
5453 &cache_dma_attr.attr,
5456 &remote_node_defrag_ratio_attr.attr,
5458 #ifdef CONFIG_SLUB_STATS
5459 &alloc_fastpath_attr.attr,
5460 &alloc_slowpath_attr.attr,
5461 &free_fastpath_attr.attr,
5462 &free_slowpath_attr.attr,
5463 &free_frozen_attr.attr,
5464 &free_add_partial_attr.attr,
5465 &free_remove_partial_attr.attr,
5466 &alloc_from_partial_attr.attr,
5467 &alloc_slab_attr.attr,
5468 &alloc_refill_attr.attr,
5469 &alloc_node_mismatch_attr.attr,
5470 &free_slab_attr.attr,
5471 &cpuslab_flush_attr.attr,
5472 &deactivate_full_attr.attr,
5473 &deactivate_empty_attr.attr,
5474 &deactivate_to_head_attr.attr,
5475 &deactivate_to_tail_attr.attr,
5476 &deactivate_remote_frees_attr.attr,
5477 &deactivate_bypass_attr.attr,
5478 &order_fallback_attr.attr,
5479 &cmpxchg_double_fail_attr.attr,
5480 &cmpxchg_double_cpu_fail_attr.attr,
5481 &cpu_partial_alloc_attr.attr,
5482 &cpu_partial_free_attr.attr,
5483 &cpu_partial_node_attr.attr,
5484 &cpu_partial_drain_attr.attr,
5486 #ifdef CONFIG_FAILSLAB
5487 &failslab_attr.attr,
5489 &usersize_attr.attr,
5494 static const struct attribute_group slab_attr_group = {
5495 .attrs = slab_attrs,
5498 static ssize_t slab_attr_show(struct kobject *kobj,
5499 struct attribute *attr,
5502 struct slab_attribute *attribute;
5503 struct kmem_cache *s;
5506 attribute = to_slab_attr(attr);
5509 if (!attribute->show)
5512 err = attribute->show(s, buf);
5517 static ssize_t slab_attr_store(struct kobject *kobj,
5518 struct attribute *attr,
5519 const char *buf, size_t len)
5521 struct slab_attribute *attribute;
5522 struct kmem_cache *s;
5525 attribute = to_slab_attr(attr);
5528 if (!attribute->store)
5531 err = attribute->store(s, buf, len);
5535 static void kmem_cache_release(struct kobject *k)
5537 slab_kmem_cache_release(to_slab(k));
5540 static const struct sysfs_ops slab_sysfs_ops = {
5541 .show = slab_attr_show,
5542 .store = slab_attr_store,
5545 static struct kobj_type slab_ktype = {
5546 .sysfs_ops = &slab_sysfs_ops,
5547 .release = kmem_cache_release,
5550 static struct kset *slab_kset;
5552 static inline struct kset *cache_kset(struct kmem_cache *s)
5557 #define ID_STR_LENGTH 64
5559 /* Create a unique string id for a slab cache:
5561 * Format :[flags-]size
5563 static char *create_unique_id(struct kmem_cache *s)
5565 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5572 * First flags affecting slabcache operations. We will only
5573 * get here for aliasable slabs so we do not need to support
5574 * too many flags. The flags here must cover all flags that
5575 * are matched during merging to guarantee that the id is
5578 if (s->flags & SLAB_CACHE_DMA)
5580 if (s->flags & SLAB_CACHE_DMA32)
5582 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5584 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5586 if (s->flags & SLAB_ACCOUNT)
5590 p += sprintf(p, "%07u", s->size);
5592 BUG_ON(p > name + ID_STR_LENGTH - 1);
5596 static int sysfs_slab_add(struct kmem_cache *s)
5600 struct kset *kset = cache_kset(s);
5601 int unmergeable = slab_unmergeable(s);
5604 kobject_init(&s->kobj, &slab_ktype);
5608 if (!unmergeable && disable_higher_order_debug &&
5609 (slub_debug & DEBUG_METADATA_FLAGS))
5614 * Slabcache can never be merged so we can use the name proper.
5615 * This is typically the case for debug situations. In that
5616 * case we can catch duplicate names easily.
5618 sysfs_remove_link(&slab_kset->kobj, s->name);
5622 * Create a unique name for the slab as a target
5625 name = create_unique_id(s);
5628 s->kobj.kset = kset;
5629 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5633 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5638 /* Setup first alias */
5639 sysfs_slab_alias(s, s->name);
5646 kobject_del(&s->kobj);
5650 void sysfs_slab_unlink(struct kmem_cache *s)
5652 if (slab_state >= FULL)
5653 kobject_del(&s->kobj);
5656 void sysfs_slab_release(struct kmem_cache *s)
5658 if (slab_state >= FULL)
5659 kobject_put(&s->kobj);
5663 * Need to buffer aliases during bootup until sysfs becomes
5664 * available lest we lose that information.
5666 struct saved_alias {
5667 struct kmem_cache *s;
5669 struct saved_alias *next;
5672 static struct saved_alias *alias_list;
5674 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5676 struct saved_alias *al;
5678 if (slab_state == FULL) {
5680 * If we have a leftover link then remove it.
5682 sysfs_remove_link(&slab_kset->kobj, name);
5683 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5686 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5692 al->next = alias_list;
5697 static int __init slab_sysfs_init(void)
5699 struct kmem_cache *s;
5702 mutex_lock(&slab_mutex);
5704 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5706 mutex_unlock(&slab_mutex);
5707 pr_err("Cannot register slab subsystem.\n");
5713 list_for_each_entry(s, &slab_caches, list) {
5714 err = sysfs_slab_add(s);
5716 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5720 while (alias_list) {
5721 struct saved_alias *al = alias_list;
5723 alias_list = alias_list->next;
5724 err = sysfs_slab_alias(al->s, al->name);
5726 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5731 mutex_unlock(&slab_mutex);
5735 __initcall(slab_sysfs_init);
5736 #endif /* CONFIG_SYSFS */
5738 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
5739 static int slab_debugfs_show(struct seq_file *seq, void *v)
5743 unsigned int idx = *(unsigned int *)v;
5744 struct loc_track *t = seq->private;
5746 if (idx < t->count) {
5749 seq_printf(seq, "%7ld ", l->count);
5752 seq_printf(seq, "%pS", (void *)l->addr);
5754 seq_puts(seq, "<not-available>");
5756 if (l->sum_time != l->min_time) {
5757 seq_printf(seq, " age=%ld/%llu/%ld",
5758 l->min_time, div_u64(l->sum_time, l->count),
5761 seq_printf(seq, " age=%ld", l->min_time);
5763 if (l->min_pid != l->max_pid)
5764 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
5766 seq_printf(seq, " pid=%ld",
5769 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
5770 seq_printf(seq, " cpus=%*pbl",
5771 cpumask_pr_args(to_cpumask(l->cpus)));
5773 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
5774 seq_printf(seq, " nodes=%*pbl",
5775 nodemask_pr_args(&l->nodes));
5777 seq_puts(seq, "\n");
5780 if (!idx && !t->count)
5781 seq_puts(seq, "No data\n");
5786 static void slab_debugfs_stop(struct seq_file *seq, void *v)
5790 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
5792 struct loc_track *t = seq->private;
5796 if (*ppos <= t->count)
5802 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
5807 static const struct seq_operations slab_debugfs_sops = {
5808 .start = slab_debugfs_start,
5809 .next = slab_debugfs_next,
5810 .stop = slab_debugfs_stop,
5811 .show = slab_debugfs_show,
5814 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
5817 struct kmem_cache_node *n;
5818 enum track_item alloc;
5820 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
5821 sizeof(struct loc_track));
5822 struct kmem_cache *s = file_inode(filep)->i_private;
5823 unsigned long *obj_map;
5825 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5829 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
5830 alloc = TRACK_ALLOC;
5834 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
5835 bitmap_free(obj_map);
5839 for_each_kmem_cache_node(s, node, n) {
5840 unsigned long flags;
5843 if (!atomic_long_read(&n->nr_slabs))
5846 spin_lock_irqsave(&n->list_lock, flags);
5847 list_for_each_entry(page, &n->partial, slab_list)
5848 process_slab(t, s, page, alloc, obj_map);
5849 list_for_each_entry(page, &n->full, slab_list)
5850 process_slab(t, s, page, alloc, obj_map);
5851 spin_unlock_irqrestore(&n->list_lock, flags);
5854 bitmap_free(obj_map);
5858 static int slab_debug_trace_release(struct inode *inode, struct file *file)
5860 struct seq_file *seq = file->private_data;
5861 struct loc_track *t = seq->private;
5864 return seq_release_private(inode, file);
5867 static const struct file_operations slab_debugfs_fops = {
5868 .open = slab_debug_trace_open,
5870 .llseek = seq_lseek,
5871 .release = slab_debug_trace_release,
5874 static void debugfs_slab_add(struct kmem_cache *s)
5876 struct dentry *slab_cache_dir;
5878 if (unlikely(!slab_debugfs_root))
5881 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
5883 debugfs_create_file("alloc_traces", 0400,
5884 slab_cache_dir, s, &slab_debugfs_fops);
5886 debugfs_create_file("free_traces", 0400,
5887 slab_cache_dir, s, &slab_debugfs_fops);
5890 void debugfs_slab_release(struct kmem_cache *s)
5892 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
5895 static int __init slab_debugfs_init(void)
5897 struct kmem_cache *s;
5899 slab_debugfs_root = debugfs_create_dir("slab", NULL);
5901 list_for_each_entry(s, &slab_caches, list)
5902 if (s->flags & SLAB_STORE_USER)
5903 debugfs_slab_add(s);
5908 __initcall(slab_debugfs_init);
5911 * The /proc/slabinfo ABI
5913 #ifdef CONFIG_SLUB_DEBUG
5914 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5916 unsigned long nr_slabs = 0;
5917 unsigned long nr_objs = 0;
5918 unsigned long nr_free = 0;
5920 struct kmem_cache_node *n;
5922 for_each_kmem_cache_node(s, node, n) {
5923 nr_slabs += node_nr_slabs(n);
5924 nr_objs += node_nr_objs(n);
5925 nr_free += count_partial(n, count_free);
5928 sinfo->active_objs = nr_objs - nr_free;
5929 sinfo->num_objs = nr_objs;
5930 sinfo->active_slabs = nr_slabs;
5931 sinfo->num_slabs = nr_slabs;
5932 sinfo->objects_per_slab = oo_objects(s->oo);
5933 sinfo->cache_order = oo_order(s->oo);
5936 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5940 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5941 size_t count, loff_t *ppos)
5945 #endif /* CONFIG_SLUB_DEBUG */