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 operatios
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/bitops.h>
19 #include <linux/slab.h>
21 #include <linux/proc_fs.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
36 #include <linux/random.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
96 * page->frozen The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 #ifdef CONFIG_SLUB_DEBUG
118 #ifdef CONFIG_SLUB_DEBUG_ON
119 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
121 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
125 static inline bool kmem_cache_debug(struct kmem_cache *s)
127 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
130 void *fixup_red_left(struct kmem_cache *s, void *p)
132 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
133 p += s->red_left_pad;
138 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
140 #ifdef CONFIG_SLUB_CPU_PARTIAL
141 return !kmem_cache_debug(s);
148 * Issues still to be resolved:
150 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
152 * - Variable sizing of the per node arrays
155 /* Enable to test recovery from slab corruption on boot */
156 #undef SLUB_RESILIENCY_TEST
158 /* Enable to log cmpxchg failures */
159 #undef SLUB_DEBUG_CMPXCHG
162 * Mininum number of partial slabs. These will be left on the partial
163 * lists even if they are empty. kmem_cache_shrink may reclaim them.
165 #define MIN_PARTIAL 5
168 * Maximum number of desirable partial slabs.
169 * The existence of more partial slabs makes kmem_cache_shrink
170 * sort the partial list by the number of objects in use.
172 #define MAX_PARTIAL 10
174 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_STORE_USER)
178 * These debug flags cannot use CMPXCHG because there might be consistency
179 * issues when checking or reading debug information
181 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
186 * Debugging flags that require metadata to be stored in the slab. These get
187 * disabled when slub_debug=O is used and a cache's min order increases with
190 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
193 #define OO_MASK ((1 << OO_SHIFT) - 1)
194 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
196 /* Internal SLUB flags */
198 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
199 /* Use cmpxchg_double */
200 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
203 * Tracking user of a slab.
205 #define TRACK_ADDRS_COUNT 16
207 unsigned long addr; /* Called from address */
208 #ifdef CONFIG_STACKTRACE
209 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
211 int cpu; /* Was running on cpu */
212 int pid; /* Pid context */
213 unsigned long when; /* When did the operation occur */
216 enum track_item { TRACK_ALLOC, TRACK_FREE };
219 static int sysfs_slab_add(struct kmem_cache *);
220 static int sysfs_slab_alias(struct kmem_cache *, const char *);
222 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
227 static inline void stat(const struct kmem_cache *s, enum stat_item si)
229 #ifdef CONFIG_SLUB_STATS
231 * The rmw is racy on a preemptible kernel but this is acceptable, so
232 * avoid this_cpu_add()'s irq-disable overhead.
234 raw_cpu_inc(s->cpu_slab->stat[si]);
238 /********************************************************************
239 * Core slab cache functions
240 *******************************************************************/
243 * Returns freelist pointer (ptr). With hardening, this is obfuscated
244 * with an XOR of the address where the pointer is held and a per-cache
247 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
248 unsigned long ptr_addr)
250 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
253 * Normally, this doesn't cause any issues, as both set_freepointer()
254 * and get_freepointer() are called with a pointer with the same tag.
255 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
256 * example, when __free_slub() iterates over objects in a cache, it
257 * passes untagged pointers to check_object(). check_object() in turns
258 * calls get_freepointer() with an untagged pointer, which causes the
259 * freepointer to be restored incorrectly.
261 return (void *)((unsigned long)ptr ^ s->random ^
262 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
268 /* Returns the freelist pointer recorded at location ptr_addr. */
269 static inline void *freelist_dereference(const struct kmem_cache *s,
272 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
273 (unsigned long)ptr_addr);
276 static inline void *get_freepointer(struct kmem_cache *s, void *object)
278 return freelist_dereference(s, object + s->offset);
281 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
283 prefetch(object + s->offset);
286 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
288 unsigned long freepointer_addr;
291 if (!debug_pagealloc_enabled_static())
292 return get_freepointer(s, object);
294 freepointer_addr = (unsigned long)object + s->offset;
295 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
296 return freelist_ptr(s, p, freepointer_addr);
299 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
301 unsigned long freeptr_addr = (unsigned long)object + s->offset;
303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
304 BUG_ON(object == fp); /* naive detection of double free or corruption */
307 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
310 /* Loop over all objects in a slab */
311 #define for_each_object(__p, __s, __addr, __objects) \
312 for (__p = fixup_red_left(__s, __addr); \
313 __p < (__addr) + (__objects) * (__s)->size; \
316 static inline unsigned int order_objects(unsigned int order, unsigned int size)
318 return ((unsigned int)PAGE_SIZE << order) / size;
321 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
324 struct kmem_cache_order_objects x = {
325 (order << OO_SHIFT) + order_objects(order, size)
331 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
333 return x.x >> OO_SHIFT;
336 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
338 return x.x & OO_MASK;
342 * Per slab locking using the pagelock
344 static __always_inline void slab_lock(struct page *page)
346 VM_BUG_ON_PAGE(PageTail(page), page);
347 bit_spin_lock(PG_locked, &page->flags);
350 static __always_inline void slab_unlock(struct page *page)
352 VM_BUG_ON_PAGE(PageTail(page), page);
353 __bit_spin_unlock(PG_locked, &page->flags);
356 /* Interrupts must be disabled (for the fallback code to work right) */
357 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
358 void *freelist_old, unsigned long counters_old,
359 void *freelist_new, unsigned long counters_new,
362 VM_BUG_ON(!irqs_disabled());
363 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
364 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
365 if (s->flags & __CMPXCHG_DOUBLE) {
366 if (cmpxchg_double(&page->freelist, &page->counters,
367 freelist_old, counters_old,
368 freelist_new, counters_new))
374 if (page->freelist == freelist_old &&
375 page->counters == counters_old) {
376 page->freelist = freelist_new;
377 page->counters = counters_new;
385 stat(s, CMPXCHG_DOUBLE_FAIL);
387 #ifdef SLUB_DEBUG_CMPXCHG
388 pr_info("%s %s: cmpxchg double redo ", n, s->name);
394 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
395 void *freelist_old, unsigned long counters_old,
396 void *freelist_new, unsigned long counters_new,
399 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
400 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
401 if (s->flags & __CMPXCHG_DOUBLE) {
402 if (cmpxchg_double(&page->freelist, &page->counters,
403 freelist_old, counters_old,
404 freelist_new, counters_new))
411 local_irq_save(flags);
413 if (page->freelist == freelist_old &&
414 page->counters == counters_old) {
415 page->freelist = freelist_new;
416 page->counters = counters_new;
418 local_irq_restore(flags);
422 local_irq_restore(flags);
426 stat(s, CMPXCHG_DOUBLE_FAIL);
428 #ifdef SLUB_DEBUG_CMPXCHG
429 pr_info("%s %s: cmpxchg double redo ", n, s->name);
435 #ifdef CONFIG_SLUB_DEBUG
436 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
437 static DEFINE_SPINLOCK(object_map_lock);
440 * Determine a map of object in use on a page.
442 * Node listlock must be held to guarantee that the page does
443 * not vanish from under us.
445 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
446 __acquires(&object_map_lock)
449 void *addr = page_address(page);
451 VM_BUG_ON(!irqs_disabled());
453 spin_lock(&object_map_lock);
455 bitmap_zero(object_map, page->objects);
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(__obj_to_index(s, addr, p), object_map);
463 static void put_map(unsigned long *map) __releases(&object_map_lock)
465 VM_BUG_ON(map != object_map);
466 spin_unlock(&object_map_lock);
469 static inline unsigned int size_from_object(struct kmem_cache *s)
471 if (s->flags & SLAB_RED_ZONE)
472 return s->size - s->red_left_pad;
477 static inline void *restore_red_left(struct kmem_cache *s, void *p)
479 if (s->flags & SLAB_RED_ZONE)
480 p -= s->red_left_pad;
488 #if defined(CONFIG_SLUB_DEBUG_ON)
489 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
491 static slab_flags_t slub_debug;
494 static char *slub_debug_string;
495 static int disable_higher_order_debug;
498 * slub is about to manipulate internal object metadata. This memory lies
499 * outside the range of the allocated object, so accessing it would normally
500 * be reported by kasan as a bounds error. metadata_access_enable() is used
501 * to tell kasan that these accesses are OK.
503 static inline void metadata_access_enable(void)
505 kasan_disable_current();
508 static inline void metadata_access_disable(void)
510 kasan_enable_current();
517 /* Verify that a pointer has an address that is valid within a slab page */
518 static inline int check_valid_pointer(struct kmem_cache *s,
519 struct page *page, void *object)
526 base = page_address(page);
527 object = kasan_reset_tag(object);
528 object = restore_red_left(s, object);
529 if (object < base || object >= base + page->objects * s->size ||
530 (object - base) % s->size) {
537 static void print_section(char *level, char *text, u8 *addr,
540 metadata_access_enable();
541 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
543 metadata_access_disable();
547 * See comment in calculate_sizes().
549 static inline bool freeptr_outside_object(struct kmem_cache *s)
551 return s->offset >= s->inuse;
555 * Return offset of the end of info block which is inuse + free pointer if
556 * not overlapping with object.
558 static inline unsigned int get_info_end(struct kmem_cache *s)
560 if (freeptr_outside_object(s))
561 return s->inuse + sizeof(void *);
566 static struct track *get_track(struct kmem_cache *s, void *object,
567 enum track_item alloc)
571 p = object + get_info_end(s);
576 static void set_track(struct kmem_cache *s, void *object,
577 enum track_item alloc, unsigned long addr)
579 struct track *p = get_track(s, object, alloc);
582 #ifdef CONFIG_STACKTRACE
583 unsigned int nr_entries;
585 metadata_access_enable();
586 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
587 metadata_access_disable();
589 if (nr_entries < TRACK_ADDRS_COUNT)
590 p->addrs[nr_entries] = 0;
593 p->cpu = smp_processor_id();
594 p->pid = current->pid;
597 memset(p, 0, sizeof(struct track));
601 static void init_tracking(struct kmem_cache *s, void *object)
603 if (!(s->flags & SLAB_STORE_USER))
606 set_track(s, object, TRACK_FREE, 0UL);
607 set_track(s, object, TRACK_ALLOC, 0UL);
610 static void print_track(const char *s, struct track *t, unsigned long pr_time)
615 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
616 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
617 #ifdef CONFIG_STACKTRACE
620 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
622 pr_err("\t%pS\n", (void *)t->addrs[i]);
629 void print_tracking(struct kmem_cache *s, void *object)
631 unsigned long pr_time = jiffies;
632 if (!(s->flags & SLAB_STORE_USER))
635 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
636 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
639 static void print_page_info(struct page *page)
641 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
642 page, page->objects, page->inuse, page->freelist, page->flags);
646 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
648 struct va_format vaf;
654 pr_err("=============================================================================\n");
655 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
656 pr_err("-----------------------------------------------------------------------------\n\n");
658 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
662 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
664 struct va_format vaf;
670 pr_err("FIX %s: %pV\n", s->name, &vaf);
674 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
675 void **freelist, void *nextfree)
677 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
678 !check_valid_pointer(s, page, nextfree) && freelist) {
679 object_err(s, page, *freelist, "Freechain corrupt");
681 slab_fix(s, "Isolate corrupted freechain");
688 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
690 unsigned int off; /* Offset of last byte */
691 u8 *addr = page_address(page);
693 print_tracking(s, p);
695 print_page_info(page);
697 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
698 p, p - addr, get_freepointer(s, p));
700 if (s->flags & SLAB_RED_ZONE)
701 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
703 else if (p > addr + 16)
704 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
706 print_section(KERN_ERR, "Object ", p,
707 min_t(unsigned int, s->object_size, PAGE_SIZE));
708 if (s->flags & SLAB_RED_ZONE)
709 print_section(KERN_ERR, "Redzone ", p + s->object_size,
710 s->inuse - s->object_size);
712 off = get_info_end(s);
714 if (s->flags & SLAB_STORE_USER)
715 off += 2 * sizeof(struct track);
717 off += kasan_metadata_size(s);
719 if (off != size_from_object(s))
720 /* Beginning of the filler is the free pointer */
721 print_section(KERN_ERR, "Padding ", p + off,
722 size_from_object(s) - off);
727 void object_err(struct kmem_cache *s, struct page *page,
728 u8 *object, char *reason)
730 slab_bug(s, "%s", reason);
731 print_trailer(s, page, object);
734 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
735 const char *fmt, ...)
741 vsnprintf(buf, sizeof(buf), fmt, args);
743 slab_bug(s, "%s", buf);
744 print_page_info(page);
748 static void init_object(struct kmem_cache *s, void *object, u8 val)
752 if (s->flags & SLAB_RED_ZONE)
753 memset(p - s->red_left_pad, val, s->red_left_pad);
755 if (s->flags & __OBJECT_POISON) {
756 memset(p, POISON_FREE, s->object_size - 1);
757 p[s->object_size - 1] = POISON_END;
760 if (s->flags & SLAB_RED_ZONE)
761 memset(p + s->object_size, val, s->inuse - s->object_size);
764 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
765 void *from, void *to)
767 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
768 memset(from, data, to - from);
771 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
772 u8 *object, char *what,
773 u8 *start, unsigned int value, unsigned int bytes)
777 u8 *addr = page_address(page);
779 metadata_access_enable();
780 fault = memchr_inv(start, value, bytes);
781 metadata_access_disable();
786 while (end > fault && end[-1] == value)
789 slab_bug(s, "%s overwritten", what);
790 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
791 fault, end - 1, fault - addr,
793 print_trailer(s, page, object);
795 restore_bytes(s, what, value, fault, end);
803 * Bytes of the object to be managed.
804 * If the freepointer may overlay the object then the free
805 * pointer is at the middle of the object.
807 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
810 * object + s->object_size
811 * Padding to reach word boundary. This is also used for Redzoning.
812 * Padding is extended by another word if Redzoning is enabled and
813 * object_size == inuse.
815 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
816 * 0xcc (RED_ACTIVE) for objects in use.
819 * Meta data starts here.
821 * A. Free pointer (if we cannot overwrite object on free)
822 * B. Tracking data for SLAB_STORE_USER
823 * C. Padding to reach required alignment boundary or at mininum
824 * one word if debugging is on to be able to detect writes
825 * before the word boundary.
827 * Padding is done using 0x5a (POISON_INUSE)
830 * Nothing is used beyond s->size.
832 * If slabcaches are merged then the object_size and inuse boundaries are mostly
833 * ignored. And therefore no slab options that rely on these boundaries
834 * may be used with merged slabcaches.
837 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
839 unsigned long off = get_info_end(s); /* The end of info */
841 if (s->flags & SLAB_STORE_USER)
842 /* We also have user information there */
843 off += 2 * sizeof(struct track);
845 off += kasan_metadata_size(s);
847 if (size_from_object(s) == off)
850 return check_bytes_and_report(s, page, p, "Object padding",
851 p + off, POISON_INUSE, size_from_object(s) - off);
854 /* Check the pad bytes at the end of a slab page */
855 static int slab_pad_check(struct kmem_cache *s, struct page *page)
864 if (!(s->flags & SLAB_POISON))
867 start = page_address(page);
868 length = page_size(page);
869 end = start + length;
870 remainder = length % s->size;
874 pad = end - remainder;
875 metadata_access_enable();
876 fault = memchr_inv(pad, POISON_INUSE, remainder);
877 metadata_access_disable();
880 while (end > fault && end[-1] == POISON_INUSE)
883 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
884 fault, end - 1, fault - start);
885 print_section(KERN_ERR, "Padding ", pad, remainder);
887 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
891 static int check_object(struct kmem_cache *s, struct page *page,
892 void *object, u8 val)
895 u8 *endobject = object + s->object_size;
897 if (s->flags & SLAB_RED_ZONE) {
898 if (!check_bytes_and_report(s, page, object, "Redzone",
899 object - s->red_left_pad, val, s->red_left_pad))
902 if (!check_bytes_and_report(s, page, object, "Redzone",
903 endobject, val, s->inuse - s->object_size))
906 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
907 check_bytes_and_report(s, page, p, "Alignment padding",
908 endobject, POISON_INUSE,
909 s->inuse - s->object_size);
913 if (s->flags & SLAB_POISON) {
914 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
915 (!check_bytes_and_report(s, page, p, "Poison", p,
916 POISON_FREE, s->object_size - 1) ||
917 !check_bytes_and_report(s, page, p, "Poison",
918 p + s->object_size - 1, POISON_END, 1)))
921 * check_pad_bytes cleans up on its own.
923 check_pad_bytes(s, page, p);
926 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
928 * Object and freepointer overlap. Cannot check
929 * freepointer while object is allocated.
933 /* Check free pointer validity */
934 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
935 object_err(s, page, p, "Freepointer corrupt");
937 * No choice but to zap it and thus lose the remainder
938 * of the free objects in this slab. May cause
939 * another error because the object count is now wrong.
941 set_freepointer(s, p, NULL);
947 static int check_slab(struct kmem_cache *s, struct page *page)
951 VM_BUG_ON(!irqs_disabled());
953 if (!PageSlab(page)) {
954 slab_err(s, page, "Not a valid slab page");
958 maxobj = order_objects(compound_order(page), s->size);
959 if (page->objects > maxobj) {
960 slab_err(s, page, "objects %u > max %u",
961 page->objects, maxobj);
964 if (page->inuse > page->objects) {
965 slab_err(s, page, "inuse %u > max %u",
966 page->inuse, page->objects);
969 /* Slab_pad_check fixes things up after itself */
970 slab_pad_check(s, page);
975 * Determine if a certain object on a page is on the freelist. Must hold the
976 * slab lock to guarantee that the chains are in a consistent state.
978 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
986 while (fp && nr <= page->objects) {
989 if (!check_valid_pointer(s, page, fp)) {
991 object_err(s, page, object,
992 "Freechain corrupt");
993 set_freepointer(s, object, NULL);
995 slab_err(s, page, "Freepointer corrupt");
996 page->freelist = NULL;
997 page->inuse = page->objects;
998 slab_fix(s, "Freelist cleared");
1004 fp = get_freepointer(s, object);
1008 max_objects = order_objects(compound_order(page), s->size);
1009 if (max_objects > MAX_OBJS_PER_PAGE)
1010 max_objects = MAX_OBJS_PER_PAGE;
1012 if (page->objects != max_objects) {
1013 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1014 page->objects, max_objects);
1015 page->objects = max_objects;
1016 slab_fix(s, "Number of objects adjusted.");
1018 if (page->inuse != page->objects - nr) {
1019 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1020 page->inuse, page->objects - nr);
1021 page->inuse = page->objects - nr;
1022 slab_fix(s, "Object count adjusted.");
1024 return search == NULL;
1027 static void trace(struct kmem_cache *s, struct page *page, void *object,
1030 if (s->flags & SLAB_TRACE) {
1031 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1033 alloc ? "alloc" : "free",
1034 object, page->inuse,
1038 print_section(KERN_INFO, "Object ", (void *)object,
1046 * Tracking of fully allocated slabs for debugging purposes.
1048 static void add_full(struct kmem_cache *s,
1049 struct kmem_cache_node *n, struct page *page)
1051 if (!(s->flags & SLAB_STORE_USER))
1054 lockdep_assert_held(&n->list_lock);
1055 list_add(&page->slab_list, &n->full);
1058 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1060 if (!(s->flags & SLAB_STORE_USER))
1063 lockdep_assert_held(&n->list_lock);
1064 list_del(&page->slab_list);
1067 /* Tracking of the number of slabs for debugging purposes */
1068 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1070 struct kmem_cache_node *n = get_node(s, node);
1072 return atomic_long_read(&n->nr_slabs);
1075 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1077 return atomic_long_read(&n->nr_slabs);
1080 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1082 struct kmem_cache_node *n = get_node(s, node);
1085 * May be called early in order to allocate a slab for the
1086 * kmem_cache_node structure. Solve the chicken-egg
1087 * dilemma by deferring the increment of the count during
1088 * bootstrap (see early_kmem_cache_node_alloc).
1091 atomic_long_inc(&n->nr_slabs);
1092 atomic_long_add(objects, &n->total_objects);
1095 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1097 struct kmem_cache_node *n = get_node(s, node);
1099 atomic_long_dec(&n->nr_slabs);
1100 atomic_long_sub(objects, &n->total_objects);
1103 /* Object debug checks for alloc/free paths */
1104 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1107 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1110 init_object(s, object, SLUB_RED_INACTIVE);
1111 init_tracking(s, object);
1115 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1117 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1120 metadata_access_enable();
1121 memset(addr, POISON_INUSE, page_size(page));
1122 metadata_access_disable();
1125 static inline int alloc_consistency_checks(struct kmem_cache *s,
1126 struct page *page, void *object)
1128 if (!check_slab(s, page))
1131 if (!check_valid_pointer(s, page, object)) {
1132 object_err(s, page, object, "Freelist Pointer check fails");
1136 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1142 static noinline int alloc_debug_processing(struct kmem_cache *s,
1144 void *object, unsigned long addr)
1146 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1147 if (!alloc_consistency_checks(s, page, object))
1151 /* Success perform special debug activities for allocs */
1152 if (s->flags & SLAB_STORE_USER)
1153 set_track(s, object, TRACK_ALLOC, addr);
1154 trace(s, page, object, 1);
1155 init_object(s, object, SLUB_RED_ACTIVE);
1159 if (PageSlab(page)) {
1161 * If this is a slab page then lets do the best we can
1162 * to avoid issues in the future. Marking all objects
1163 * as used avoids touching the remaining objects.
1165 slab_fix(s, "Marking all objects used");
1166 page->inuse = page->objects;
1167 page->freelist = NULL;
1172 static inline int free_consistency_checks(struct kmem_cache *s,
1173 struct page *page, void *object, unsigned long addr)
1175 if (!check_valid_pointer(s, page, object)) {
1176 slab_err(s, page, "Invalid object pointer 0x%p", object);
1180 if (on_freelist(s, page, object)) {
1181 object_err(s, page, object, "Object already free");
1185 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1188 if (unlikely(s != page->slab_cache)) {
1189 if (!PageSlab(page)) {
1190 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1192 } else if (!page->slab_cache) {
1193 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1197 object_err(s, page, object,
1198 "page slab pointer corrupt.");
1204 /* Supports checking bulk free of a constructed freelist */
1205 static noinline int free_debug_processing(
1206 struct kmem_cache *s, struct page *page,
1207 void *head, void *tail, int bulk_cnt,
1210 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1211 void *object = head;
1213 unsigned long flags;
1216 spin_lock_irqsave(&n->list_lock, flags);
1219 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1220 if (!check_slab(s, page))
1227 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1228 if (!free_consistency_checks(s, page, object, addr))
1232 if (s->flags & SLAB_STORE_USER)
1233 set_track(s, object, TRACK_FREE, addr);
1234 trace(s, page, object, 0);
1235 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1236 init_object(s, object, SLUB_RED_INACTIVE);
1238 /* Reached end of constructed freelist yet? */
1239 if (object != tail) {
1240 object = get_freepointer(s, object);
1246 if (cnt != bulk_cnt)
1247 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1251 spin_unlock_irqrestore(&n->list_lock, flags);
1253 slab_fix(s, "Object at 0x%p not freed", object);
1258 * Parse a block of slub_debug options. Blocks are delimited by ';'
1260 * @str: start of block
1261 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1262 * @slabs: return start of list of slabs, or NULL when there's no list
1263 * @init: assume this is initial parsing and not per-kmem-create parsing
1265 * returns the start of next block if there's any, or NULL
1268 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1270 bool higher_order_disable = false;
1272 /* Skip any completely empty blocks */
1273 while (*str && *str == ';')
1278 * No options but restriction on slabs. This means full
1279 * debugging for slabs matching a pattern.
1281 *flags = DEBUG_DEFAULT_FLAGS;
1286 /* Determine which debug features should be switched on */
1287 for (; *str && *str != ',' && *str != ';'; str++) {
1288 switch (tolower(*str)) {
1293 *flags |= SLAB_CONSISTENCY_CHECKS;
1296 *flags |= SLAB_RED_ZONE;
1299 *flags |= SLAB_POISON;
1302 *flags |= SLAB_STORE_USER;
1305 *flags |= SLAB_TRACE;
1308 *flags |= SLAB_FAILSLAB;
1312 * Avoid enabling debugging on caches if its minimum
1313 * order would increase as a result.
1315 higher_order_disable = true;
1319 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1328 /* Skip over the slab list */
1329 while (*str && *str != ';')
1332 /* Skip any completely empty blocks */
1333 while (*str && *str == ';')
1336 if (init && higher_order_disable)
1337 disable_higher_order_debug = 1;
1345 static int __init setup_slub_debug(char *str)
1350 bool global_slub_debug_changed = false;
1351 bool slab_list_specified = false;
1353 slub_debug = DEBUG_DEFAULT_FLAGS;
1354 if (*str++ != '=' || !*str)
1356 * No options specified. Switch on full debugging.
1362 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1366 global_slub_debug_changed = true;
1368 slab_list_specified = true;
1373 * For backwards compatibility, a single list of flags with list of
1374 * slabs means debugging is only enabled for those slabs, so the global
1375 * slub_debug should be 0. We can extended that to multiple lists as
1376 * long as there is no option specifying flags without a slab list.
1378 if (slab_list_specified) {
1379 if (!global_slub_debug_changed)
1381 slub_debug_string = saved_str;
1384 if (slub_debug != 0 || slub_debug_string)
1385 static_branch_enable(&slub_debug_enabled);
1386 if ((static_branch_unlikely(&init_on_alloc) ||
1387 static_branch_unlikely(&init_on_free)) &&
1388 (slub_debug & SLAB_POISON))
1389 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1393 __setup("slub_debug", setup_slub_debug);
1396 * kmem_cache_flags - apply debugging options to the cache
1397 * @object_size: the size of an object without meta data
1398 * @flags: flags to set
1399 * @name: name of the cache
1400 * @ctor: constructor function
1402 * Debug option(s) are applied to @flags. In addition to the debug
1403 * option(s), if a slab name (or multiple) is specified i.e.
1404 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1405 * then only the select slabs will receive the debug option(s).
1407 slab_flags_t kmem_cache_flags(unsigned int object_size,
1408 slab_flags_t flags, const char *name,
1409 void (*ctor)(void *))
1414 slab_flags_t block_flags;
1416 /* If slub_debug = 0, it folds into the if conditional. */
1417 if (!slub_debug_string)
1418 return flags | slub_debug;
1421 next_block = slub_debug_string;
1422 /* Go through all blocks of debug options, see if any matches our slab's name */
1423 while (next_block) {
1424 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1427 /* Found a block that has a slab list, search it */
1432 end = strchrnul(iter, ',');
1433 if (next_block && next_block < end)
1434 end = next_block - 1;
1436 glob = strnchr(iter, end - iter, '*');
1438 cmplen = glob - iter;
1440 cmplen = max_t(size_t, len, (end - iter));
1442 if (!strncmp(name, iter, cmplen)) {
1443 flags |= block_flags;
1447 if (!*end || *end == ';')
1455 #else /* !CONFIG_SLUB_DEBUG */
1456 static inline void setup_object_debug(struct kmem_cache *s,
1457 struct page *page, void *object) {}
1459 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1461 static inline int alloc_debug_processing(struct kmem_cache *s,
1462 struct page *page, void *object, unsigned long addr) { return 0; }
1464 static inline int free_debug_processing(
1465 struct kmem_cache *s, struct page *page,
1466 void *head, void *tail, int bulk_cnt,
1467 unsigned long addr) { return 0; }
1469 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1471 static inline int check_object(struct kmem_cache *s, struct page *page,
1472 void *object, u8 val) { return 1; }
1473 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1474 struct page *page) {}
1475 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1476 struct page *page) {}
1477 slab_flags_t kmem_cache_flags(unsigned int object_size,
1478 slab_flags_t flags, const char *name,
1479 void (*ctor)(void *))
1483 #define slub_debug 0
1485 #define disable_higher_order_debug 0
1487 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1489 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1491 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1493 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1496 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1497 void **freelist, void *nextfree)
1501 #endif /* CONFIG_SLUB_DEBUG */
1504 * Hooks for other subsystems that check memory allocations. In a typical
1505 * production configuration these hooks all should produce no code at all.
1507 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1509 ptr = kasan_kmalloc_large(ptr, size, flags);
1510 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1511 kmemleak_alloc(ptr, size, 1, flags);
1515 static __always_inline void kfree_hook(void *x)
1518 kasan_kfree_large(x, _RET_IP_);
1521 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1523 kmemleak_free_recursive(x, s->flags);
1526 * Trouble is that we may no longer disable interrupts in the fast path
1527 * So in order to make the debug calls that expect irqs to be
1528 * disabled we need to disable interrupts temporarily.
1530 #ifdef CONFIG_LOCKDEP
1532 unsigned long flags;
1534 local_irq_save(flags);
1535 debug_check_no_locks_freed(x, s->object_size);
1536 local_irq_restore(flags);
1539 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1540 debug_check_no_obj_freed(x, s->object_size);
1542 /* Use KCSAN to help debug racy use-after-free. */
1543 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1544 __kcsan_check_access(x, s->object_size,
1545 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1547 /* KASAN might put x into memory quarantine, delaying its reuse */
1548 return kasan_slab_free(s, x, _RET_IP_);
1551 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1552 void **head, void **tail)
1557 void *old_tail = *tail ? *tail : *head;
1560 /* Head and tail of the reconstructed freelist */
1566 next = get_freepointer(s, object);
1568 if (slab_want_init_on_free(s)) {
1570 * Clear the object and the metadata, but don't touch
1573 memset(object, 0, s->object_size);
1574 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1576 memset((char *)object + s->inuse, 0,
1577 s->size - s->inuse - rsize);
1580 /* If object's reuse doesn't have to be delayed */
1581 if (!slab_free_hook(s, object)) {
1582 /* Move object to the new freelist */
1583 set_freepointer(s, object, *head);
1588 } while (object != old_tail);
1593 return *head != NULL;
1596 static void *setup_object(struct kmem_cache *s, struct page *page,
1599 setup_object_debug(s, page, object);
1600 object = kasan_init_slab_obj(s, object);
1601 if (unlikely(s->ctor)) {
1602 kasan_unpoison_object_data(s, object);
1604 kasan_poison_object_data(s, object);
1610 * Slab allocation and freeing
1612 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1613 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1616 unsigned int order = oo_order(oo);
1618 if (node == NUMA_NO_NODE)
1619 page = alloc_pages(flags, order);
1621 page = __alloc_pages_node(node, flags, order);
1624 account_slab_page(page, order, s);
1629 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1630 /* Pre-initialize the random sequence cache */
1631 static int init_cache_random_seq(struct kmem_cache *s)
1633 unsigned int count = oo_objects(s->oo);
1636 /* Bailout if already initialised */
1640 err = cache_random_seq_create(s, count, GFP_KERNEL);
1642 pr_err("SLUB: Unable to initialize free list for %s\n",
1647 /* Transform to an offset on the set of pages */
1648 if (s->random_seq) {
1651 for (i = 0; i < count; i++)
1652 s->random_seq[i] *= s->size;
1657 /* Initialize each random sequence freelist per cache */
1658 static void __init init_freelist_randomization(void)
1660 struct kmem_cache *s;
1662 mutex_lock(&slab_mutex);
1664 list_for_each_entry(s, &slab_caches, list)
1665 init_cache_random_seq(s);
1667 mutex_unlock(&slab_mutex);
1670 /* Get the next entry on the pre-computed freelist randomized */
1671 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1672 unsigned long *pos, void *start,
1673 unsigned long page_limit,
1674 unsigned long freelist_count)
1679 * If the target page allocation failed, the number of objects on the
1680 * page might be smaller than the usual size defined by the cache.
1683 idx = s->random_seq[*pos];
1685 if (*pos >= freelist_count)
1687 } while (unlikely(idx >= page_limit));
1689 return (char *)start + idx;
1692 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1693 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1698 unsigned long idx, pos, page_limit, freelist_count;
1700 if (page->objects < 2 || !s->random_seq)
1703 freelist_count = oo_objects(s->oo);
1704 pos = get_random_int() % freelist_count;
1706 page_limit = page->objects * s->size;
1707 start = fixup_red_left(s, page_address(page));
1709 /* First entry is used as the base of the freelist */
1710 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1712 cur = setup_object(s, page, cur);
1713 page->freelist = cur;
1715 for (idx = 1; idx < page->objects; idx++) {
1716 next = next_freelist_entry(s, page, &pos, start, page_limit,
1718 next = setup_object(s, page, next);
1719 set_freepointer(s, cur, next);
1722 set_freepointer(s, cur, NULL);
1727 static inline int init_cache_random_seq(struct kmem_cache *s)
1731 static inline void init_freelist_randomization(void) { }
1732 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1736 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1738 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1741 struct kmem_cache_order_objects oo = s->oo;
1743 void *start, *p, *next;
1747 flags &= gfp_allowed_mask;
1749 if (gfpflags_allow_blocking(flags))
1752 flags |= s->allocflags;
1755 * Let the initial higher-order allocation fail under memory pressure
1756 * so we fall-back to the minimum order allocation.
1758 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1759 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1760 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1762 page = alloc_slab_page(s, alloc_gfp, node, oo);
1763 if (unlikely(!page)) {
1767 * Allocation may have failed due to fragmentation.
1768 * Try a lower order alloc if possible
1770 page = alloc_slab_page(s, alloc_gfp, node, oo);
1771 if (unlikely(!page))
1773 stat(s, ORDER_FALLBACK);
1776 page->objects = oo_objects(oo);
1778 page->slab_cache = s;
1779 __SetPageSlab(page);
1780 if (page_is_pfmemalloc(page))
1781 SetPageSlabPfmemalloc(page);
1783 kasan_poison_slab(page);
1785 start = page_address(page);
1787 setup_page_debug(s, page, start);
1789 shuffle = shuffle_freelist(s, page);
1792 start = fixup_red_left(s, start);
1793 start = setup_object(s, page, start);
1794 page->freelist = start;
1795 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1797 next = setup_object(s, page, next);
1798 set_freepointer(s, p, next);
1801 set_freepointer(s, p, NULL);
1804 page->inuse = page->objects;
1808 if (gfpflags_allow_blocking(flags))
1809 local_irq_disable();
1813 inc_slabs_node(s, page_to_nid(page), page->objects);
1818 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1820 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1821 flags = kmalloc_fix_flags(flags);
1823 return allocate_slab(s,
1824 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1827 static void __free_slab(struct kmem_cache *s, struct page *page)
1829 int order = compound_order(page);
1830 int pages = 1 << order;
1832 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1835 slab_pad_check(s, page);
1836 for_each_object(p, s, page_address(page),
1838 check_object(s, page, p, SLUB_RED_INACTIVE);
1841 __ClearPageSlabPfmemalloc(page);
1842 __ClearPageSlab(page);
1844 page->mapping = NULL;
1845 if (current->reclaim_state)
1846 current->reclaim_state->reclaimed_slab += pages;
1847 unaccount_slab_page(page, order, s);
1848 __free_pages(page, order);
1851 static void rcu_free_slab(struct rcu_head *h)
1853 struct page *page = container_of(h, struct page, rcu_head);
1855 __free_slab(page->slab_cache, page);
1858 static void free_slab(struct kmem_cache *s, struct page *page)
1860 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1861 call_rcu(&page->rcu_head, rcu_free_slab);
1863 __free_slab(s, page);
1866 static void discard_slab(struct kmem_cache *s, struct page *page)
1868 dec_slabs_node(s, page_to_nid(page), page->objects);
1873 * Management of partially allocated slabs.
1876 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1879 if (tail == DEACTIVATE_TO_TAIL)
1880 list_add_tail(&page->slab_list, &n->partial);
1882 list_add(&page->slab_list, &n->partial);
1885 static inline void add_partial(struct kmem_cache_node *n,
1886 struct page *page, int tail)
1888 lockdep_assert_held(&n->list_lock);
1889 __add_partial(n, page, tail);
1892 static inline void remove_partial(struct kmem_cache_node *n,
1895 lockdep_assert_held(&n->list_lock);
1896 list_del(&page->slab_list);
1901 * Remove slab from the partial list, freeze it and
1902 * return the pointer to the freelist.
1904 * Returns a list of objects or NULL if it fails.
1906 static inline void *acquire_slab(struct kmem_cache *s,
1907 struct kmem_cache_node *n, struct page *page,
1908 int mode, int *objects)
1911 unsigned long counters;
1914 lockdep_assert_held(&n->list_lock);
1917 * Zap the freelist and set the frozen bit.
1918 * The old freelist is the list of objects for the
1919 * per cpu allocation list.
1921 freelist = page->freelist;
1922 counters = page->counters;
1923 new.counters = counters;
1924 *objects = new.objects - new.inuse;
1926 new.inuse = page->objects;
1927 new.freelist = NULL;
1929 new.freelist = freelist;
1932 VM_BUG_ON(new.frozen);
1935 if (!__cmpxchg_double_slab(s, page,
1937 new.freelist, new.counters,
1941 remove_partial(n, page);
1946 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1947 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1950 * Try to allocate a partial slab from a specific node.
1952 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1953 struct kmem_cache_cpu *c, gfp_t flags)
1955 struct page *page, *page2;
1956 void *object = NULL;
1957 unsigned int available = 0;
1961 * Racy check. If we mistakenly see no partial slabs then we
1962 * just allocate an empty slab. If we mistakenly try to get a
1963 * partial slab and there is none available then get_partials()
1966 if (!n || !n->nr_partial)
1969 spin_lock(&n->list_lock);
1970 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1973 if (!pfmemalloc_match(page, flags))
1976 t = acquire_slab(s, n, page, object == NULL, &objects);
1980 available += objects;
1983 stat(s, ALLOC_FROM_PARTIAL);
1986 put_cpu_partial(s, page, 0);
1987 stat(s, CPU_PARTIAL_NODE);
1989 if (!kmem_cache_has_cpu_partial(s)
1990 || available > slub_cpu_partial(s) / 2)
1994 spin_unlock(&n->list_lock);
1999 * Get a page from somewhere. Search in increasing NUMA distances.
2001 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2002 struct kmem_cache_cpu *c)
2005 struct zonelist *zonelist;
2008 enum zone_type highest_zoneidx = gfp_zone(flags);
2010 unsigned int cpuset_mems_cookie;
2013 * The defrag ratio allows a configuration of the tradeoffs between
2014 * inter node defragmentation and node local allocations. A lower
2015 * defrag_ratio increases the tendency to do local allocations
2016 * instead of attempting to obtain partial slabs from other nodes.
2018 * If the defrag_ratio is set to 0 then kmalloc() always
2019 * returns node local objects. If the ratio is higher then kmalloc()
2020 * may return off node objects because partial slabs are obtained
2021 * from other nodes and filled up.
2023 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2024 * (which makes defrag_ratio = 1000) then every (well almost)
2025 * allocation will first attempt to defrag slab caches on other nodes.
2026 * This means scanning over all nodes to look for partial slabs which
2027 * may be expensive if we do it every time we are trying to find a slab
2028 * with available objects.
2030 if (!s->remote_node_defrag_ratio ||
2031 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2035 cpuset_mems_cookie = read_mems_allowed_begin();
2036 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2037 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2038 struct kmem_cache_node *n;
2040 n = get_node(s, zone_to_nid(zone));
2042 if (n && cpuset_zone_allowed(zone, flags) &&
2043 n->nr_partial > s->min_partial) {
2044 object = get_partial_node(s, n, c, flags);
2047 * Don't check read_mems_allowed_retry()
2048 * here - if mems_allowed was updated in
2049 * parallel, that was a harmless race
2050 * between allocation and the cpuset
2057 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2058 #endif /* CONFIG_NUMA */
2063 * Get a partial page, lock it and return it.
2065 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2066 struct kmem_cache_cpu *c)
2069 int searchnode = node;
2071 if (node == NUMA_NO_NODE)
2072 searchnode = numa_mem_id();
2074 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2075 if (object || node != NUMA_NO_NODE)
2078 return get_any_partial(s, flags, c);
2081 #ifdef CONFIG_PREEMPTION
2083 * Calculate the next globally unique transaction for disambiguation
2084 * during cmpxchg. The transactions start with the cpu number and are then
2085 * incremented by CONFIG_NR_CPUS.
2087 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2090 * No preemption supported therefore also no need to check for
2096 static inline unsigned long next_tid(unsigned long tid)
2098 return tid + TID_STEP;
2101 #ifdef SLUB_DEBUG_CMPXCHG
2102 static inline unsigned int tid_to_cpu(unsigned long tid)
2104 return tid % TID_STEP;
2107 static inline unsigned long tid_to_event(unsigned long tid)
2109 return tid / TID_STEP;
2113 static inline unsigned int init_tid(int cpu)
2118 static inline void note_cmpxchg_failure(const char *n,
2119 const struct kmem_cache *s, unsigned long tid)
2121 #ifdef SLUB_DEBUG_CMPXCHG
2122 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2124 pr_info("%s %s: cmpxchg redo ", n, s->name);
2126 #ifdef CONFIG_PREEMPTION
2127 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2128 pr_warn("due to cpu change %d -> %d\n",
2129 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2132 if (tid_to_event(tid) != tid_to_event(actual_tid))
2133 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2134 tid_to_event(tid), tid_to_event(actual_tid));
2136 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2137 actual_tid, tid, next_tid(tid));
2139 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2142 static void init_kmem_cache_cpus(struct kmem_cache *s)
2146 for_each_possible_cpu(cpu)
2147 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2151 * Remove the cpu slab
2153 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2154 void *freelist, struct kmem_cache_cpu *c)
2156 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2157 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2159 enum slab_modes l = M_NONE, m = M_NONE;
2161 int tail = DEACTIVATE_TO_HEAD;
2165 if (page->freelist) {
2166 stat(s, DEACTIVATE_REMOTE_FREES);
2167 tail = DEACTIVATE_TO_TAIL;
2171 * Stage one: Free all available per cpu objects back
2172 * to the page freelist while it is still frozen. Leave the
2175 * There is no need to take the list->lock because the page
2178 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2180 unsigned long counters;
2183 * If 'nextfree' is invalid, it is possible that the object at
2184 * 'freelist' is already corrupted. So isolate all objects
2185 * starting at 'freelist'.
2187 if (freelist_corrupted(s, page, &freelist, nextfree))
2191 prior = page->freelist;
2192 counters = page->counters;
2193 set_freepointer(s, freelist, prior);
2194 new.counters = counters;
2196 VM_BUG_ON(!new.frozen);
2198 } while (!__cmpxchg_double_slab(s, page,
2200 freelist, new.counters,
2201 "drain percpu freelist"));
2203 freelist = nextfree;
2207 * Stage two: Ensure that the page is unfrozen while the
2208 * list presence reflects the actual number of objects
2211 * We setup the list membership and then perform a cmpxchg
2212 * with the count. If there is a mismatch then the page
2213 * is not unfrozen but the page is on the wrong list.
2215 * Then we restart the process which may have to remove
2216 * the page from the list that we just put it on again
2217 * because the number of objects in the slab may have
2222 old.freelist = page->freelist;
2223 old.counters = page->counters;
2224 VM_BUG_ON(!old.frozen);
2226 /* Determine target state of the slab */
2227 new.counters = old.counters;
2230 set_freepointer(s, freelist, old.freelist);
2231 new.freelist = freelist;
2233 new.freelist = old.freelist;
2237 if (!new.inuse && n->nr_partial >= s->min_partial)
2239 else if (new.freelist) {
2244 * Taking the spinlock removes the possibility
2245 * that acquire_slab() will see a slab page that
2248 spin_lock(&n->list_lock);
2252 if (kmem_cache_debug(s) && !lock) {
2255 * This also ensures that the scanning of full
2256 * slabs from diagnostic functions will not see
2259 spin_lock(&n->list_lock);
2265 remove_partial(n, page);
2266 else if (l == M_FULL)
2267 remove_full(s, n, page);
2270 add_partial(n, page, tail);
2271 else if (m == M_FULL)
2272 add_full(s, n, page);
2276 if (!__cmpxchg_double_slab(s, page,
2277 old.freelist, old.counters,
2278 new.freelist, new.counters,
2283 spin_unlock(&n->list_lock);
2287 else if (m == M_FULL)
2288 stat(s, DEACTIVATE_FULL);
2289 else if (m == M_FREE) {
2290 stat(s, DEACTIVATE_EMPTY);
2291 discard_slab(s, page);
2300 * Unfreeze all the cpu partial slabs.
2302 * This function must be called with interrupts disabled
2303 * for the cpu using c (or some other guarantee must be there
2304 * to guarantee no concurrent accesses).
2306 static void unfreeze_partials(struct kmem_cache *s,
2307 struct kmem_cache_cpu *c)
2309 #ifdef CONFIG_SLUB_CPU_PARTIAL
2310 struct kmem_cache_node *n = NULL, *n2 = NULL;
2311 struct page *page, *discard_page = NULL;
2313 while ((page = slub_percpu_partial(c))) {
2317 slub_set_percpu_partial(c, page);
2319 n2 = get_node(s, page_to_nid(page));
2322 spin_unlock(&n->list_lock);
2325 spin_lock(&n->list_lock);
2330 old.freelist = page->freelist;
2331 old.counters = page->counters;
2332 VM_BUG_ON(!old.frozen);
2334 new.counters = old.counters;
2335 new.freelist = old.freelist;
2339 } while (!__cmpxchg_double_slab(s, page,
2340 old.freelist, old.counters,
2341 new.freelist, new.counters,
2342 "unfreezing slab"));
2344 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2345 page->next = discard_page;
2346 discard_page = page;
2348 add_partial(n, page, DEACTIVATE_TO_TAIL);
2349 stat(s, FREE_ADD_PARTIAL);
2354 spin_unlock(&n->list_lock);
2356 while (discard_page) {
2357 page = discard_page;
2358 discard_page = discard_page->next;
2360 stat(s, DEACTIVATE_EMPTY);
2361 discard_slab(s, page);
2364 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2368 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2369 * partial page slot if available.
2371 * If we did not find a slot then simply move all the partials to the
2372 * per node partial list.
2374 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2376 #ifdef CONFIG_SLUB_CPU_PARTIAL
2377 struct page *oldpage;
2385 oldpage = this_cpu_read(s->cpu_slab->partial);
2388 pobjects = oldpage->pobjects;
2389 pages = oldpage->pages;
2390 if (drain && pobjects > slub_cpu_partial(s)) {
2391 unsigned long flags;
2393 * partial array is full. Move the existing
2394 * set to the per node partial list.
2396 local_irq_save(flags);
2397 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2398 local_irq_restore(flags);
2402 stat(s, CPU_PARTIAL_DRAIN);
2407 pobjects += page->objects - page->inuse;
2409 page->pages = pages;
2410 page->pobjects = pobjects;
2411 page->next = oldpage;
2413 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2415 if (unlikely(!slub_cpu_partial(s))) {
2416 unsigned long flags;
2418 local_irq_save(flags);
2419 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2420 local_irq_restore(flags);
2423 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2426 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2428 stat(s, CPUSLAB_FLUSH);
2429 deactivate_slab(s, c->page, c->freelist, c);
2431 c->tid = next_tid(c->tid);
2437 * Called from IPI handler with interrupts disabled.
2439 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2441 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2446 unfreeze_partials(s, c);
2449 static void flush_cpu_slab(void *d)
2451 struct kmem_cache *s = d;
2453 __flush_cpu_slab(s, smp_processor_id());
2456 static bool has_cpu_slab(int cpu, void *info)
2458 struct kmem_cache *s = info;
2459 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2461 return c->page || slub_percpu_partial(c);
2464 static void flush_all(struct kmem_cache *s)
2466 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2470 * Use the cpu notifier to insure that the cpu slabs are flushed when
2473 static int slub_cpu_dead(unsigned int cpu)
2475 struct kmem_cache *s;
2476 unsigned long flags;
2478 mutex_lock(&slab_mutex);
2479 list_for_each_entry(s, &slab_caches, list) {
2480 local_irq_save(flags);
2481 __flush_cpu_slab(s, cpu);
2482 local_irq_restore(flags);
2484 mutex_unlock(&slab_mutex);
2489 * Check if the objects in a per cpu structure fit numa
2490 * locality expectations.
2492 static inline int node_match(struct page *page, int node)
2495 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2501 #ifdef CONFIG_SLUB_DEBUG
2502 static int count_free(struct page *page)
2504 return page->objects - page->inuse;
2507 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2509 return atomic_long_read(&n->total_objects);
2511 #endif /* CONFIG_SLUB_DEBUG */
2513 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2514 static unsigned long count_partial(struct kmem_cache_node *n,
2515 int (*get_count)(struct page *))
2517 unsigned long flags;
2518 unsigned long x = 0;
2521 spin_lock_irqsave(&n->list_lock, flags);
2522 list_for_each_entry(page, &n->partial, slab_list)
2523 x += get_count(page);
2524 spin_unlock_irqrestore(&n->list_lock, flags);
2527 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2529 static noinline void
2530 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2532 #ifdef CONFIG_SLUB_DEBUG
2533 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2534 DEFAULT_RATELIMIT_BURST);
2536 struct kmem_cache_node *n;
2538 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2541 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2542 nid, gfpflags, &gfpflags);
2543 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2544 s->name, s->object_size, s->size, oo_order(s->oo),
2547 if (oo_order(s->min) > get_order(s->object_size))
2548 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2551 for_each_kmem_cache_node(s, node, n) {
2552 unsigned long nr_slabs;
2553 unsigned long nr_objs;
2554 unsigned long nr_free;
2556 nr_free = count_partial(n, count_free);
2557 nr_slabs = node_nr_slabs(n);
2558 nr_objs = node_nr_objs(n);
2560 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2561 node, nr_slabs, nr_objs, nr_free);
2566 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2567 int node, struct kmem_cache_cpu **pc)
2570 struct kmem_cache_cpu *c = *pc;
2573 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2575 freelist = get_partial(s, flags, node, c);
2580 page = new_slab(s, flags, node);
2582 c = raw_cpu_ptr(s->cpu_slab);
2587 * No other reference to the page yet so we can
2588 * muck around with it freely without cmpxchg
2590 freelist = page->freelist;
2591 page->freelist = NULL;
2593 stat(s, ALLOC_SLAB);
2601 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2603 if (unlikely(PageSlabPfmemalloc(page)))
2604 return gfp_pfmemalloc_allowed(gfpflags);
2610 * Check the page->freelist of a page and either transfer the freelist to the
2611 * per cpu freelist or deactivate the page.
2613 * The page is still frozen if the return value is not NULL.
2615 * If this function returns NULL then the page has been unfrozen.
2617 * This function must be called with interrupt disabled.
2619 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2622 unsigned long counters;
2626 freelist = page->freelist;
2627 counters = page->counters;
2629 new.counters = counters;
2630 VM_BUG_ON(!new.frozen);
2632 new.inuse = page->objects;
2633 new.frozen = freelist != NULL;
2635 } while (!__cmpxchg_double_slab(s, page,
2644 * Slow path. The lockless freelist is empty or we need to perform
2647 * Processing is still very fast if new objects have been freed to the
2648 * regular freelist. In that case we simply take over the regular freelist
2649 * as the lockless freelist and zap the regular freelist.
2651 * If that is not working then we fall back to the partial lists. We take the
2652 * first element of the freelist as the object to allocate now and move the
2653 * rest of the freelist to the lockless freelist.
2655 * And if we were unable to get a new slab from the partial slab lists then
2656 * we need to allocate a new slab. This is the slowest path since it involves
2657 * a call to the page allocator and the setup of a new slab.
2659 * Version of __slab_alloc to use when we know that interrupts are
2660 * already disabled (which is the case for bulk allocation).
2662 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2663 unsigned long addr, struct kmem_cache_cpu *c)
2671 * if the node is not online or has no normal memory, just
2672 * ignore the node constraint
2674 if (unlikely(node != NUMA_NO_NODE &&
2675 !node_state(node, N_NORMAL_MEMORY)))
2676 node = NUMA_NO_NODE;
2681 if (unlikely(!node_match(page, node))) {
2683 * same as above but node_match() being false already
2684 * implies node != NUMA_NO_NODE
2686 if (!node_state(node, N_NORMAL_MEMORY)) {
2687 node = NUMA_NO_NODE;
2690 stat(s, ALLOC_NODE_MISMATCH);
2691 deactivate_slab(s, page, c->freelist, c);
2697 * By rights, we should be searching for a slab page that was
2698 * PFMEMALLOC but right now, we are losing the pfmemalloc
2699 * information when the page leaves the per-cpu allocator
2701 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2702 deactivate_slab(s, page, c->freelist, c);
2706 /* must check again c->freelist in case of cpu migration or IRQ */
2707 freelist = c->freelist;
2711 freelist = get_freelist(s, page);
2715 stat(s, DEACTIVATE_BYPASS);
2719 stat(s, ALLOC_REFILL);
2723 * freelist is pointing to the list of objects to be used.
2724 * page is pointing to the page from which the objects are obtained.
2725 * That page must be frozen for per cpu allocations to work.
2727 VM_BUG_ON(!c->page->frozen);
2728 c->freelist = get_freepointer(s, freelist);
2729 c->tid = next_tid(c->tid);
2734 if (slub_percpu_partial(c)) {
2735 page = c->page = slub_percpu_partial(c);
2736 slub_set_percpu_partial(c, page);
2737 stat(s, CPU_PARTIAL_ALLOC);
2741 freelist = new_slab_objects(s, gfpflags, node, &c);
2743 if (unlikely(!freelist)) {
2744 slab_out_of_memory(s, gfpflags, node);
2749 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2752 /* Only entered in the debug case */
2753 if (kmem_cache_debug(s) &&
2754 !alloc_debug_processing(s, page, freelist, addr))
2755 goto new_slab; /* Slab failed checks. Next slab needed */
2757 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2762 * Another one that disabled interrupt and compensates for possible
2763 * cpu changes by refetching the per cpu area pointer.
2765 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2766 unsigned long addr, struct kmem_cache_cpu *c)
2769 unsigned long flags;
2771 local_irq_save(flags);
2772 #ifdef CONFIG_PREEMPTION
2774 * We may have been preempted and rescheduled on a different
2775 * cpu before disabling interrupts. Need to reload cpu area
2778 c = this_cpu_ptr(s->cpu_slab);
2781 p = ___slab_alloc(s, gfpflags, node, addr, c);
2782 local_irq_restore(flags);
2787 * If the object has been wiped upon free, make sure it's fully initialized by
2788 * zeroing out freelist pointer.
2790 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2793 if (unlikely(slab_want_init_on_free(s)) && obj)
2794 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2798 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2799 * have the fastpath folded into their functions. So no function call
2800 * overhead for requests that can be satisfied on the fastpath.
2802 * The fastpath works by first checking if the lockless freelist can be used.
2803 * If not then __slab_alloc is called for slow processing.
2805 * Otherwise we can simply pick the next object from the lockless free list.
2807 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2808 gfp_t gfpflags, int node, unsigned long addr)
2811 struct kmem_cache_cpu *c;
2814 struct obj_cgroup *objcg = NULL;
2816 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2821 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2822 * enabled. We may switch back and forth between cpus while
2823 * reading from one cpu area. That does not matter as long
2824 * as we end up on the original cpu again when doing the cmpxchg.
2826 * We should guarantee that tid and kmem_cache are retrieved on
2827 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2828 * to check if it is matched or not.
2831 tid = this_cpu_read(s->cpu_slab->tid);
2832 c = raw_cpu_ptr(s->cpu_slab);
2833 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2834 unlikely(tid != READ_ONCE(c->tid)));
2837 * Irqless object alloc/free algorithm used here depends on sequence
2838 * of fetching cpu_slab's data. tid should be fetched before anything
2839 * on c to guarantee that object and page associated with previous tid
2840 * won't be used with current tid. If we fetch tid first, object and
2841 * page could be one associated with next tid and our alloc/free
2842 * request will be failed. In this case, we will retry. So, no problem.
2847 * The transaction ids are globally unique per cpu and per operation on
2848 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2849 * occurs on the right processor and that there was no operation on the
2850 * linked list in between.
2853 object = c->freelist;
2855 if (unlikely(!object || !node_match(page, node))) {
2856 object = __slab_alloc(s, gfpflags, node, addr, c);
2857 stat(s, ALLOC_SLOWPATH);
2859 void *next_object = get_freepointer_safe(s, object);
2862 * The cmpxchg will only match if there was no additional
2863 * operation and if we are on the right processor.
2865 * The cmpxchg does the following atomically (without lock
2867 * 1. Relocate first pointer to the current per cpu area.
2868 * 2. Verify that tid and freelist have not been changed
2869 * 3. If they were not changed replace tid and freelist
2871 * Since this is without lock semantics the protection is only
2872 * against code executing on this cpu *not* from access by
2875 if (unlikely(!this_cpu_cmpxchg_double(
2876 s->cpu_slab->freelist, s->cpu_slab->tid,
2878 next_object, next_tid(tid)))) {
2880 note_cmpxchg_failure("slab_alloc", s, tid);
2883 prefetch_freepointer(s, next_object);
2884 stat(s, ALLOC_FASTPATH);
2887 maybe_wipe_obj_freeptr(s, object);
2889 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2890 memset(object, 0, s->object_size);
2892 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object);
2897 static __always_inline void *slab_alloc(struct kmem_cache *s,
2898 gfp_t gfpflags, unsigned long addr)
2900 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2903 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2905 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2907 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2912 EXPORT_SYMBOL(kmem_cache_alloc);
2914 #ifdef CONFIG_TRACING
2915 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2917 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2918 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2919 ret = kasan_kmalloc(s, ret, size, gfpflags);
2922 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2926 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2928 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2930 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2931 s->object_size, s->size, gfpflags, node);
2935 EXPORT_SYMBOL(kmem_cache_alloc_node);
2937 #ifdef CONFIG_TRACING
2938 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2940 int node, size_t size)
2942 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2944 trace_kmalloc_node(_RET_IP_, ret,
2945 size, s->size, gfpflags, node);
2947 ret = kasan_kmalloc(s, ret, size, gfpflags);
2950 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2952 #endif /* CONFIG_NUMA */
2955 * Slow path handling. This may still be called frequently since objects
2956 * have a longer lifetime than the cpu slabs in most processing loads.
2958 * So we still attempt to reduce cache line usage. Just take the slab
2959 * lock and free the item. If there is no additional partial page
2960 * handling required then we can return immediately.
2962 static void __slab_free(struct kmem_cache *s, struct page *page,
2963 void *head, void *tail, int cnt,
2970 unsigned long counters;
2971 struct kmem_cache_node *n = NULL;
2972 unsigned long flags;
2974 stat(s, FREE_SLOWPATH);
2976 if (kmem_cache_debug(s) &&
2977 !free_debug_processing(s, page, head, tail, cnt, addr))
2982 spin_unlock_irqrestore(&n->list_lock, flags);
2985 prior = page->freelist;
2986 counters = page->counters;
2987 set_freepointer(s, tail, prior);
2988 new.counters = counters;
2989 was_frozen = new.frozen;
2991 if ((!new.inuse || !prior) && !was_frozen) {
2993 if (kmem_cache_has_cpu_partial(s) && !prior) {
2996 * Slab was on no list before and will be
2998 * We can defer the list move and instead
3003 } else { /* Needs to be taken off a list */
3005 n = get_node(s, page_to_nid(page));
3007 * Speculatively acquire the list_lock.
3008 * If the cmpxchg does not succeed then we may
3009 * drop the list_lock without any processing.
3011 * Otherwise the list_lock will synchronize with
3012 * other processors updating the list of slabs.
3014 spin_lock_irqsave(&n->list_lock, flags);
3019 } while (!cmpxchg_double_slab(s, page,
3027 * If we just froze the page then put it onto the
3028 * per cpu partial list.
3030 if (new.frozen && !was_frozen) {
3031 put_cpu_partial(s, page, 1);
3032 stat(s, CPU_PARTIAL_FREE);
3035 * The list lock was not taken therefore no list
3036 * activity can be necessary.
3039 stat(s, FREE_FROZEN);
3043 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3047 * Objects left in the slab. If it was not on the partial list before
3050 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3051 remove_full(s, n, page);
3052 add_partial(n, page, DEACTIVATE_TO_TAIL);
3053 stat(s, FREE_ADD_PARTIAL);
3055 spin_unlock_irqrestore(&n->list_lock, flags);
3061 * Slab on the partial list.
3063 remove_partial(n, page);
3064 stat(s, FREE_REMOVE_PARTIAL);
3066 /* Slab must be on the full list */
3067 remove_full(s, n, page);
3070 spin_unlock_irqrestore(&n->list_lock, flags);
3072 discard_slab(s, page);
3076 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3077 * can perform fastpath freeing without additional function calls.
3079 * The fastpath is only possible if we are freeing to the current cpu slab
3080 * of this processor. This typically the case if we have just allocated
3083 * If fastpath is not possible then fall back to __slab_free where we deal
3084 * with all sorts of special processing.
3086 * Bulk free of a freelist with several objects (all pointing to the
3087 * same page) possible by specifying head and tail ptr, plus objects
3088 * count (cnt). Bulk free indicated by tail pointer being set.
3090 static __always_inline void do_slab_free(struct kmem_cache *s,
3091 struct page *page, void *head, void *tail,
3092 int cnt, unsigned long addr)
3094 void *tail_obj = tail ? : head;
3095 struct kmem_cache_cpu *c;
3098 memcg_slab_free_hook(s, page, head);
3101 * Determine the currently cpus per cpu slab.
3102 * The cpu may change afterward. However that does not matter since
3103 * data is retrieved via this pointer. If we are on the same cpu
3104 * during the cmpxchg then the free will succeed.
3107 tid = this_cpu_read(s->cpu_slab->tid);
3108 c = raw_cpu_ptr(s->cpu_slab);
3109 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3110 unlikely(tid != READ_ONCE(c->tid)));
3112 /* Same with comment on barrier() in slab_alloc_node() */
3115 if (likely(page == c->page)) {
3116 void **freelist = READ_ONCE(c->freelist);
3118 set_freepointer(s, tail_obj, freelist);
3120 if (unlikely(!this_cpu_cmpxchg_double(
3121 s->cpu_slab->freelist, s->cpu_slab->tid,
3123 head, next_tid(tid)))) {
3125 note_cmpxchg_failure("slab_free", s, tid);
3128 stat(s, FREE_FASTPATH);
3130 __slab_free(s, page, head, tail_obj, cnt, addr);
3134 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3135 void *head, void *tail, int cnt,
3139 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3140 * to remove objects, whose reuse must be delayed.
3142 if (slab_free_freelist_hook(s, &head, &tail))
3143 do_slab_free(s, page, head, tail, cnt, addr);
3146 #ifdef CONFIG_KASAN_GENERIC
3147 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3149 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3153 void kmem_cache_free(struct kmem_cache *s, void *x)
3155 s = cache_from_obj(s, x);
3158 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3159 trace_kmem_cache_free(_RET_IP_, x);
3161 EXPORT_SYMBOL(kmem_cache_free);
3163 struct detached_freelist {
3168 struct kmem_cache *s;
3172 * This function progressively scans the array with free objects (with
3173 * a limited look ahead) and extract objects belonging to the same
3174 * page. It builds a detached freelist directly within the given
3175 * page/objects. This can happen without any need for
3176 * synchronization, because the objects are owned by running process.
3177 * The freelist is build up as a single linked list in the objects.
3178 * The idea is, that this detached freelist can then be bulk
3179 * transferred to the real freelist(s), but only requiring a single
3180 * synchronization primitive. Look ahead in the array is limited due
3181 * to performance reasons.
3184 int build_detached_freelist(struct kmem_cache *s, size_t size,
3185 void **p, struct detached_freelist *df)
3187 size_t first_skipped_index = 0;
3192 /* Always re-init detached_freelist */
3197 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3198 } while (!object && size);
3203 page = virt_to_head_page(object);
3205 /* Handle kalloc'ed objects */
3206 if (unlikely(!PageSlab(page))) {
3207 BUG_ON(!PageCompound(page));
3209 __free_pages(page, compound_order(page));
3210 p[size] = NULL; /* mark object processed */
3213 /* Derive kmem_cache from object */
3214 df->s = page->slab_cache;
3216 df->s = cache_from_obj(s, object); /* Support for memcg */
3219 /* Start new detached freelist */
3221 set_freepointer(df->s, object, NULL);
3223 df->freelist = object;
3224 p[size] = NULL; /* mark object processed */
3230 continue; /* Skip processed objects */
3232 /* df->page is always set at this point */
3233 if (df->page == virt_to_head_page(object)) {
3234 /* Opportunity build freelist */
3235 set_freepointer(df->s, object, df->freelist);
3236 df->freelist = object;
3238 p[size] = NULL; /* mark object processed */
3243 /* Limit look ahead search */
3247 if (!first_skipped_index)
3248 first_skipped_index = size + 1;
3251 return first_skipped_index;
3254 /* Note that interrupts must be enabled when calling this function. */
3255 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3261 struct detached_freelist df;
3263 size = build_detached_freelist(s, size, p, &df);
3267 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3268 } while (likely(size));
3270 EXPORT_SYMBOL(kmem_cache_free_bulk);
3272 /* Note that interrupts must be enabled when calling this function. */
3273 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3276 struct kmem_cache_cpu *c;
3278 struct obj_cgroup *objcg = NULL;
3280 /* memcg and kmem_cache debug support */
3281 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3285 * Drain objects in the per cpu slab, while disabling local
3286 * IRQs, which protects against PREEMPT and interrupts
3287 * handlers invoking normal fastpath.
3289 local_irq_disable();
3290 c = this_cpu_ptr(s->cpu_slab);
3292 for (i = 0; i < size; i++) {
3293 void *object = c->freelist;
3295 if (unlikely(!object)) {
3297 * We may have removed an object from c->freelist using
3298 * the fastpath in the previous iteration; in that case,
3299 * c->tid has not been bumped yet.
3300 * Since ___slab_alloc() may reenable interrupts while
3301 * allocating memory, we should bump c->tid now.
3303 c->tid = next_tid(c->tid);
3306 * Invoking slow path likely have side-effect
3307 * of re-populating per CPU c->freelist
3309 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3311 if (unlikely(!p[i]))
3314 c = this_cpu_ptr(s->cpu_slab);
3315 maybe_wipe_obj_freeptr(s, p[i]);
3317 continue; /* goto for-loop */
3319 c->freelist = get_freepointer(s, object);
3321 maybe_wipe_obj_freeptr(s, p[i]);
3323 c->tid = next_tid(c->tid);
3326 /* Clear memory outside IRQ disabled fastpath loop */
3327 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3330 for (j = 0; j < i; j++)
3331 memset(p[j], 0, s->object_size);
3334 /* memcg and kmem_cache debug support */
3335 slab_post_alloc_hook(s, objcg, flags, size, p);
3339 slab_post_alloc_hook(s, objcg, flags, i, p);
3340 __kmem_cache_free_bulk(s, i, p);
3343 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3347 * Object placement in a slab is made very easy because we always start at
3348 * offset 0. If we tune the size of the object to the alignment then we can
3349 * get the required alignment by putting one properly sized object after
3352 * Notice that the allocation order determines the sizes of the per cpu
3353 * caches. Each processor has always one slab available for allocations.
3354 * Increasing the allocation order reduces the number of times that slabs
3355 * must be moved on and off the partial lists and is therefore a factor in
3360 * Mininum / Maximum order of slab pages. This influences locking overhead
3361 * and slab fragmentation. A higher order reduces the number of partial slabs
3362 * and increases the number of allocations possible without having to
3363 * take the list_lock.
3365 static unsigned int slub_min_order;
3366 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3367 static unsigned int slub_min_objects;
3370 * Calculate the order of allocation given an slab object size.
3372 * The order of allocation has significant impact on performance and other
3373 * system components. Generally order 0 allocations should be preferred since
3374 * order 0 does not cause fragmentation in the page allocator. Larger objects
3375 * be problematic to put into order 0 slabs because there may be too much
3376 * unused space left. We go to a higher order if more than 1/16th of the slab
3379 * In order to reach satisfactory performance we must ensure that a minimum
3380 * number of objects is in one slab. Otherwise we may generate too much
3381 * activity on the partial lists which requires taking the list_lock. This is
3382 * less a concern for large slabs though which are rarely used.
3384 * slub_max_order specifies the order where we begin to stop considering the
3385 * number of objects in a slab as critical. If we reach slub_max_order then
3386 * we try to keep the page order as low as possible. So we accept more waste
3387 * of space in favor of a small page order.
3389 * Higher order allocations also allow the placement of more objects in a
3390 * slab and thereby reduce object handling overhead. If the user has
3391 * requested a higher mininum order then we start with that one instead of
3392 * the smallest order which will fit the object.
3394 static inline unsigned int slab_order(unsigned int size,
3395 unsigned int min_objects, unsigned int max_order,
3396 unsigned int fract_leftover)
3398 unsigned int min_order = slub_min_order;
3401 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3402 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3404 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3405 order <= max_order; order++) {
3407 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3410 rem = slab_size % size;
3412 if (rem <= slab_size / fract_leftover)
3419 static inline int calculate_order(unsigned int size)
3422 unsigned int min_objects;
3423 unsigned int max_objects;
3426 * Attempt to find best configuration for a slab. This
3427 * works by first attempting to generate a layout with
3428 * the best configuration and backing off gradually.
3430 * First we increase the acceptable waste in a slab. Then
3431 * we reduce the minimum objects required in a slab.
3433 min_objects = slub_min_objects;
3435 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3436 max_objects = order_objects(slub_max_order, size);
3437 min_objects = min(min_objects, max_objects);
3439 while (min_objects > 1) {
3440 unsigned int fraction;
3443 while (fraction >= 4) {
3444 order = slab_order(size, min_objects,
3445 slub_max_order, fraction);
3446 if (order <= slub_max_order)
3454 * We were unable to place multiple objects in a slab. Now
3455 * lets see if we can place a single object there.
3457 order = slab_order(size, 1, slub_max_order, 1);
3458 if (order <= slub_max_order)
3462 * Doh this slab cannot be placed using slub_max_order.
3464 order = slab_order(size, 1, MAX_ORDER, 1);
3465 if (order < MAX_ORDER)
3471 init_kmem_cache_node(struct kmem_cache_node *n)
3474 spin_lock_init(&n->list_lock);
3475 INIT_LIST_HEAD(&n->partial);
3476 #ifdef CONFIG_SLUB_DEBUG
3477 atomic_long_set(&n->nr_slabs, 0);
3478 atomic_long_set(&n->total_objects, 0);
3479 INIT_LIST_HEAD(&n->full);
3483 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3485 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3486 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3489 * Must align to double word boundary for the double cmpxchg
3490 * instructions to work; see __pcpu_double_call_return_bool().
3492 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3493 2 * sizeof(void *));
3498 init_kmem_cache_cpus(s);
3503 static struct kmem_cache *kmem_cache_node;
3506 * No kmalloc_node yet so do it by hand. We know that this is the first
3507 * slab on the node for this slabcache. There are no concurrent accesses
3510 * Note that this function only works on the kmem_cache_node
3511 * when allocating for the kmem_cache_node. This is used for bootstrapping
3512 * memory on a fresh node that has no slab structures yet.
3514 static void early_kmem_cache_node_alloc(int node)
3517 struct kmem_cache_node *n;
3519 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3521 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3524 if (page_to_nid(page) != node) {
3525 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3526 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3531 #ifdef CONFIG_SLUB_DEBUG
3532 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3533 init_tracking(kmem_cache_node, n);
3535 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3537 page->freelist = get_freepointer(kmem_cache_node, n);
3540 kmem_cache_node->node[node] = n;
3541 init_kmem_cache_node(n);
3542 inc_slabs_node(kmem_cache_node, node, page->objects);
3545 * No locks need to be taken here as it has just been
3546 * initialized and there is no concurrent access.
3548 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3551 static void free_kmem_cache_nodes(struct kmem_cache *s)
3554 struct kmem_cache_node *n;
3556 for_each_kmem_cache_node(s, node, n) {
3557 s->node[node] = NULL;
3558 kmem_cache_free(kmem_cache_node, n);
3562 void __kmem_cache_release(struct kmem_cache *s)
3564 cache_random_seq_destroy(s);
3565 free_percpu(s->cpu_slab);
3566 free_kmem_cache_nodes(s);
3569 static int init_kmem_cache_nodes(struct kmem_cache *s)
3573 for_each_node_state(node, N_NORMAL_MEMORY) {
3574 struct kmem_cache_node *n;
3576 if (slab_state == DOWN) {
3577 early_kmem_cache_node_alloc(node);
3580 n = kmem_cache_alloc_node(kmem_cache_node,
3584 free_kmem_cache_nodes(s);
3588 init_kmem_cache_node(n);
3594 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3596 if (min < MIN_PARTIAL)
3598 else if (min > MAX_PARTIAL)
3600 s->min_partial = min;
3603 static void set_cpu_partial(struct kmem_cache *s)
3605 #ifdef CONFIG_SLUB_CPU_PARTIAL
3607 * cpu_partial determined the maximum number of objects kept in the
3608 * per cpu partial lists of a processor.
3610 * Per cpu partial lists mainly contain slabs that just have one
3611 * object freed. If they are used for allocation then they can be
3612 * filled up again with minimal effort. The slab will never hit the
3613 * per node partial lists and therefore no locking will be required.
3615 * This setting also determines
3617 * A) The number of objects from per cpu partial slabs dumped to the
3618 * per node list when we reach the limit.
3619 * B) The number of objects in cpu partial slabs to extract from the
3620 * per node list when we run out of per cpu objects. We only fetch
3621 * 50% to keep some capacity around for frees.
3623 if (!kmem_cache_has_cpu_partial(s))
3624 slub_set_cpu_partial(s, 0);
3625 else if (s->size >= PAGE_SIZE)
3626 slub_set_cpu_partial(s, 2);
3627 else if (s->size >= 1024)
3628 slub_set_cpu_partial(s, 6);
3629 else if (s->size >= 256)
3630 slub_set_cpu_partial(s, 13);
3632 slub_set_cpu_partial(s, 30);
3637 * calculate_sizes() determines the order and the distribution of data within
3640 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3642 slab_flags_t flags = s->flags;
3643 unsigned int size = s->object_size;
3644 unsigned int freepointer_area;
3648 * Round up object size to the next word boundary. We can only
3649 * place the free pointer at word boundaries and this determines
3650 * the possible location of the free pointer.
3652 size = ALIGN(size, sizeof(void *));
3654 * This is the area of the object where a freepointer can be
3655 * safely written. If redzoning adds more to the inuse size, we
3656 * can't use that portion for writing the freepointer, so
3657 * s->offset must be limited within this for the general case.
3659 freepointer_area = size;
3661 #ifdef CONFIG_SLUB_DEBUG
3663 * Determine if we can poison the object itself. If the user of
3664 * the slab may touch the object after free or before allocation
3665 * then we should never poison the object itself.
3667 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3669 s->flags |= __OBJECT_POISON;
3671 s->flags &= ~__OBJECT_POISON;
3675 * If we are Redzoning then check if there is some space between the
3676 * end of the object and the free pointer. If not then add an
3677 * additional word to have some bytes to store Redzone information.
3679 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3680 size += sizeof(void *);
3684 * With that we have determined the number of bytes in actual use
3685 * by the object. This is the potential offset to the free pointer.
3689 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3692 * Relocate free pointer after the object if it is not
3693 * permitted to overwrite the first word of the object on
3696 * This is the case if we do RCU, have a constructor or
3697 * destructor or are poisoning the objects.
3699 * The assumption that s->offset >= s->inuse means free
3700 * pointer is outside of the object is used in the
3701 * freeptr_outside_object() function. If that is no
3702 * longer true, the function needs to be modified.
3705 size += sizeof(void *);
3706 } else if (freepointer_area > sizeof(void *)) {
3708 * Store freelist pointer near middle of object to keep
3709 * it away from the edges of the object to avoid small
3710 * sized over/underflows from neighboring allocations.
3712 s->offset = ALIGN(freepointer_area / 2, sizeof(void *));
3715 #ifdef CONFIG_SLUB_DEBUG
3716 if (flags & SLAB_STORE_USER)
3718 * Need to store information about allocs and frees after
3721 size += 2 * sizeof(struct track);
3724 kasan_cache_create(s, &size, &s->flags);
3725 #ifdef CONFIG_SLUB_DEBUG
3726 if (flags & SLAB_RED_ZONE) {
3728 * Add some empty padding so that we can catch
3729 * overwrites from earlier objects rather than let
3730 * tracking information or the free pointer be
3731 * corrupted if a user writes before the start
3734 size += sizeof(void *);
3736 s->red_left_pad = sizeof(void *);
3737 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3738 size += s->red_left_pad;
3743 * SLUB stores one object immediately after another beginning from
3744 * offset 0. In order to align the objects we have to simply size
3745 * each object to conform to the alignment.
3747 size = ALIGN(size, s->align);
3749 s->reciprocal_size = reciprocal_value(size);
3750 if (forced_order >= 0)
3751 order = forced_order;
3753 order = calculate_order(size);
3760 s->allocflags |= __GFP_COMP;
3762 if (s->flags & SLAB_CACHE_DMA)
3763 s->allocflags |= GFP_DMA;
3765 if (s->flags & SLAB_CACHE_DMA32)
3766 s->allocflags |= GFP_DMA32;
3768 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3769 s->allocflags |= __GFP_RECLAIMABLE;
3772 * Determine the number of objects per slab
3774 s->oo = oo_make(order, size);
3775 s->min = oo_make(get_order(size), size);
3776 if (oo_objects(s->oo) > oo_objects(s->max))
3779 return !!oo_objects(s->oo);
3782 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3784 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3785 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3786 s->random = get_random_long();
3789 if (!calculate_sizes(s, -1))
3791 if (disable_higher_order_debug) {
3793 * Disable debugging flags that store metadata if the min slab
3796 if (get_order(s->size) > get_order(s->object_size)) {
3797 s->flags &= ~DEBUG_METADATA_FLAGS;
3799 if (!calculate_sizes(s, -1))
3804 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3805 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3806 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3807 /* Enable fast mode */
3808 s->flags |= __CMPXCHG_DOUBLE;
3812 * The larger the object size is, the more pages we want on the partial
3813 * list to avoid pounding the page allocator excessively.
3815 set_min_partial(s, ilog2(s->size) / 2);
3820 s->remote_node_defrag_ratio = 1000;
3823 /* Initialize the pre-computed randomized freelist if slab is up */
3824 if (slab_state >= UP) {
3825 if (init_cache_random_seq(s))
3829 if (!init_kmem_cache_nodes(s))
3832 if (alloc_kmem_cache_cpus(s))
3835 free_kmem_cache_nodes(s);
3840 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3843 #ifdef CONFIG_SLUB_DEBUG
3844 void *addr = page_address(page);
3848 slab_err(s, page, text, s->name);
3851 map = get_map(s, page);
3852 for_each_object(p, s, addr, page->objects) {
3854 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3855 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3856 print_tracking(s, p);
3865 * Attempt to free all partial slabs on a node.
3866 * This is called from __kmem_cache_shutdown(). We must take list_lock
3867 * because sysfs file might still access partial list after the shutdowning.
3869 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3872 struct page *page, *h;
3874 BUG_ON(irqs_disabled());
3875 spin_lock_irq(&n->list_lock);
3876 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3878 remove_partial(n, page);
3879 list_add(&page->slab_list, &discard);
3881 list_slab_objects(s, page,
3882 "Objects remaining in %s on __kmem_cache_shutdown()");
3885 spin_unlock_irq(&n->list_lock);
3887 list_for_each_entry_safe(page, h, &discard, slab_list)
3888 discard_slab(s, page);
3891 bool __kmem_cache_empty(struct kmem_cache *s)
3894 struct kmem_cache_node *n;
3896 for_each_kmem_cache_node(s, node, n)
3897 if (n->nr_partial || slabs_node(s, node))
3903 * Release all resources used by a slab cache.
3905 int __kmem_cache_shutdown(struct kmem_cache *s)
3908 struct kmem_cache_node *n;
3911 /* Attempt to free all objects */
3912 for_each_kmem_cache_node(s, node, n) {
3914 if (n->nr_partial || slabs_node(s, node))
3920 /********************************************************************
3922 *******************************************************************/
3924 static int __init setup_slub_min_order(char *str)
3926 get_option(&str, (int *)&slub_min_order);
3931 __setup("slub_min_order=", setup_slub_min_order);
3933 static int __init setup_slub_max_order(char *str)
3935 get_option(&str, (int *)&slub_max_order);
3936 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3941 __setup("slub_max_order=", setup_slub_max_order);
3943 static int __init setup_slub_min_objects(char *str)
3945 get_option(&str, (int *)&slub_min_objects);
3950 __setup("slub_min_objects=", setup_slub_min_objects);
3952 void *__kmalloc(size_t size, gfp_t flags)
3954 struct kmem_cache *s;
3957 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3958 return kmalloc_large(size, flags);
3960 s = kmalloc_slab(size, flags);
3962 if (unlikely(ZERO_OR_NULL_PTR(s)))
3965 ret = slab_alloc(s, flags, _RET_IP_);
3967 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3969 ret = kasan_kmalloc(s, ret, size, flags);
3973 EXPORT_SYMBOL(__kmalloc);
3976 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3980 unsigned int order = get_order(size);
3982 flags |= __GFP_COMP;
3983 page = alloc_pages_node(node, flags, order);
3985 ptr = page_address(page);
3986 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
3987 PAGE_SIZE << order);
3990 return kmalloc_large_node_hook(ptr, size, flags);
3993 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3995 struct kmem_cache *s;
3998 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3999 ret = kmalloc_large_node(size, flags, node);
4001 trace_kmalloc_node(_RET_IP_, ret,
4002 size, PAGE_SIZE << get_order(size),
4008 s = kmalloc_slab(size, flags);
4010 if (unlikely(ZERO_OR_NULL_PTR(s)))
4013 ret = slab_alloc_node(s, flags, node, _RET_IP_);
4015 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4017 ret = kasan_kmalloc(s, ret, size, flags);
4021 EXPORT_SYMBOL(__kmalloc_node);
4022 #endif /* CONFIG_NUMA */
4024 #ifdef CONFIG_HARDENED_USERCOPY
4026 * Rejects incorrectly sized objects and objects that are to be copied
4027 * to/from userspace but do not fall entirely within the containing slab
4028 * cache's usercopy region.
4030 * Returns NULL if check passes, otherwise const char * to name of cache
4031 * to indicate an error.
4033 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4036 struct kmem_cache *s;
4037 unsigned int offset;
4040 ptr = kasan_reset_tag(ptr);
4042 /* Find object and usable object size. */
4043 s = page->slab_cache;
4045 /* Reject impossible pointers. */
4046 if (ptr < page_address(page))
4047 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4050 /* Find offset within object. */
4051 offset = (ptr - page_address(page)) % s->size;
4053 /* Adjust for redzone and reject if within the redzone. */
4054 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4055 if (offset < s->red_left_pad)
4056 usercopy_abort("SLUB object in left red zone",
4057 s->name, to_user, offset, n);
4058 offset -= s->red_left_pad;
4061 /* Allow address range falling entirely within usercopy region. */
4062 if (offset >= s->useroffset &&
4063 offset - s->useroffset <= s->usersize &&
4064 n <= s->useroffset - offset + s->usersize)
4068 * If the copy is still within the allocated object, produce
4069 * a warning instead of rejecting the copy. This is intended
4070 * to be a temporary method to find any missing usercopy
4073 object_size = slab_ksize(s);
4074 if (usercopy_fallback &&
4075 offset <= object_size && n <= object_size - offset) {
4076 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4080 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4082 #endif /* CONFIG_HARDENED_USERCOPY */
4084 size_t __ksize(const void *object)
4088 if (unlikely(object == ZERO_SIZE_PTR))
4091 page = virt_to_head_page(object);
4093 if (unlikely(!PageSlab(page))) {
4094 WARN_ON(!PageCompound(page));
4095 return page_size(page);
4098 return slab_ksize(page->slab_cache);
4100 EXPORT_SYMBOL(__ksize);
4102 void kfree(const void *x)
4105 void *object = (void *)x;
4107 trace_kfree(_RET_IP_, x);
4109 if (unlikely(ZERO_OR_NULL_PTR(x)))
4112 page = virt_to_head_page(x);
4113 if (unlikely(!PageSlab(page))) {
4114 unsigned int order = compound_order(page);
4116 BUG_ON(!PageCompound(page));
4118 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
4119 -(PAGE_SIZE << order));
4120 __free_pages(page, order);
4123 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4125 EXPORT_SYMBOL(kfree);
4127 #define SHRINK_PROMOTE_MAX 32
4130 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4131 * up most to the head of the partial lists. New allocations will then
4132 * fill those up and thus they can be removed from the partial lists.
4134 * The slabs with the least items are placed last. This results in them
4135 * being allocated from last increasing the chance that the last objects
4136 * are freed in them.
4138 int __kmem_cache_shrink(struct kmem_cache *s)
4142 struct kmem_cache_node *n;
4145 struct list_head discard;
4146 struct list_head promote[SHRINK_PROMOTE_MAX];
4147 unsigned long flags;
4151 for_each_kmem_cache_node(s, node, n) {
4152 INIT_LIST_HEAD(&discard);
4153 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4154 INIT_LIST_HEAD(promote + i);
4156 spin_lock_irqsave(&n->list_lock, flags);
4159 * Build lists of slabs to discard or promote.
4161 * Note that concurrent frees may occur while we hold the
4162 * list_lock. page->inuse here is the upper limit.
4164 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4165 int free = page->objects - page->inuse;
4167 /* Do not reread page->inuse */
4170 /* We do not keep full slabs on the list */
4173 if (free == page->objects) {
4174 list_move(&page->slab_list, &discard);
4176 } else if (free <= SHRINK_PROMOTE_MAX)
4177 list_move(&page->slab_list, promote + free - 1);
4181 * Promote the slabs filled up most to the head of the
4184 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4185 list_splice(promote + i, &n->partial);
4187 spin_unlock_irqrestore(&n->list_lock, flags);
4189 /* Release empty slabs */
4190 list_for_each_entry_safe(page, t, &discard, slab_list)
4191 discard_slab(s, page);
4193 if (slabs_node(s, node))
4200 static int slab_mem_going_offline_callback(void *arg)
4202 struct kmem_cache *s;
4204 mutex_lock(&slab_mutex);
4205 list_for_each_entry(s, &slab_caches, list)
4206 __kmem_cache_shrink(s);
4207 mutex_unlock(&slab_mutex);
4212 static void slab_mem_offline_callback(void *arg)
4214 struct kmem_cache_node *n;
4215 struct kmem_cache *s;
4216 struct memory_notify *marg = arg;
4219 offline_node = marg->status_change_nid_normal;
4222 * If the node still has available memory. we need kmem_cache_node
4225 if (offline_node < 0)
4228 mutex_lock(&slab_mutex);
4229 list_for_each_entry(s, &slab_caches, list) {
4230 n = get_node(s, offline_node);
4233 * if n->nr_slabs > 0, slabs still exist on the node
4234 * that is going down. We were unable to free them,
4235 * and offline_pages() function shouldn't call this
4236 * callback. So, we must fail.
4238 BUG_ON(slabs_node(s, offline_node));
4240 s->node[offline_node] = NULL;
4241 kmem_cache_free(kmem_cache_node, n);
4244 mutex_unlock(&slab_mutex);
4247 static int slab_mem_going_online_callback(void *arg)
4249 struct kmem_cache_node *n;
4250 struct kmem_cache *s;
4251 struct memory_notify *marg = arg;
4252 int nid = marg->status_change_nid_normal;
4256 * If the node's memory is already available, then kmem_cache_node is
4257 * already created. Nothing to do.
4263 * We are bringing a node online. No memory is available yet. We must
4264 * allocate a kmem_cache_node structure in order to bring the node
4267 mutex_lock(&slab_mutex);
4268 list_for_each_entry(s, &slab_caches, list) {
4270 * XXX: kmem_cache_alloc_node will fallback to other nodes
4271 * since memory is not yet available from the node that
4274 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4279 init_kmem_cache_node(n);
4283 mutex_unlock(&slab_mutex);
4287 static int slab_memory_callback(struct notifier_block *self,
4288 unsigned long action, void *arg)
4293 case MEM_GOING_ONLINE:
4294 ret = slab_mem_going_online_callback(arg);
4296 case MEM_GOING_OFFLINE:
4297 ret = slab_mem_going_offline_callback(arg);
4300 case MEM_CANCEL_ONLINE:
4301 slab_mem_offline_callback(arg);
4304 case MEM_CANCEL_OFFLINE:
4308 ret = notifier_from_errno(ret);
4314 static struct notifier_block slab_memory_callback_nb = {
4315 .notifier_call = slab_memory_callback,
4316 .priority = SLAB_CALLBACK_PRI,
4319 /********************************************************************
4320 * Basic setup of slabs
4321 *******************************************************************/
4324 * Used for early kmem_cache structures that were allocated using
4325 * the page allocator. Allocate them properly then fix up the pointers
4326 * that may be pointing to the wrong kmem_cache structure.
4329 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4332 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4333 struct kmem_cache_node *n;
4335 memcpy(s, static_cache, kmem_cache->object_size);
4338 * This runs very early, and only the boot processor is supposed to be
4339 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4342 __flush_cpu_slab(s, smp_processor_id());
4343 for_each_kmem_cache_node(s, node, n) {
4346 list_for_each_entry(p, &n->partial, slab_list)
4349 #ifdef CONFIG_SLUB_DEBUG
4350 list_for_each_entry(p, &n->full, slab_list)
4354 list_add(&s->list, &slab_caches);
4358 void __init kmem_cache_init(void)
4360 static __initdata struct kmem_cache boot_kmem_cache,
4361 boot_kmem_cache_node;
4363 if (debug_guardpage_minorder())
4366 kmem_cache_node = &boot_kmem_cache_node;
4367 kmem_cache = &boot_kmem_cache;
4369 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4370 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4372 register_hotmemory_notifier(&slab_memory_callback_nb);
4374 /* Able to allocate the per node structures */
4375 slab_state = PARTIAL;
4377 create_boot_cache(kmem_cache, "kmem_cache",
4378 offsetof(struct kmem_cache, node) +
4379 nr_node_ids * sizeof(struct kmem_cache_node *),
4380 SLAB_HWCACHE_ALIGN, 0, 0);
4382 kmem_cache = bootstrap(&boot_kmem_cache);
4383 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4385 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4386 setup_kmalloc_cache_index_table();
4387 create_kmalloc_caches(0);
4389 /* Setup random freelists for each cache */
4390 init_freelist_randomization();
4392 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4395 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4397 slub_min_order, slub_max_order, slub_min_objects,
4398 nr_cpu_ids, nr_node_ids);
4401 void __init kmem_cache_init_late(void)
4406 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4407 slab_flags_t flags, void (*ctor)(void *))
4409 struct kmem_cache *s;
4411 s = find_mergeable(size, align, flags, name, ctor);
4416 * Adjust the object sizes so that we clear
4417 * the complete object on kzalloc.
4419 s->object_size = max(s->object_size, size);
4420 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4422 if (sysfs_slab_alias(s, name)) {
4431 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4435 err = kmem_cache_open(s, flags);
4439 /* Mutex is not taken during early boot */
4440 if (slab_state <= UP)
4443 err = sysfs_slab_add(s);
4445 __kmem_cache_release(s);
4450 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4452 struct kmem_cache *s;
4455 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4456 return kmalloc_large(size, gfpflags);
4458 s = kmalloc_slab(size, gfpflags);
4460 if (unlikely(ZERO_OR_NULL_PTR(s)))
4463 ret = slab_alloc(s, gfpflags, caller);
4465 /* Honor the call site pointer we received. */
4466 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4470 EXPORT_SYMBOL(__kmalloc_track_caller);
4473 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4474 int node, unsigned long caller)
4476 struct kmem_cache *s;
4479 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4480 ret = kmalloc_large_node(size, gfpflags, node);
4482 trace_kmalloc_node(caller, ret,
4483 size, PAGE_SIZE << get_order(size),
4489 s = kmalloc_slab(size, gfpflags);
4491 if (unlikely(ZERO_OR_NULL_PTR(s)))
4494 ret = slab_alloc_node(s, gfpflags, node, caller);
4496 /* Honor the call site pointer we received. */
4497 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4501 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4505 static int count_inuse(struct page *page)
4510 static int count_total(struct page *page)
4512 return page->objects;
4516 #ifdef CONFIG_SLUB_DEBUG
4517 static void validate_slab(struct kmem_cache *s, struct page *page)
4520 void *addr = page_address(page);
4525 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4528 /* Now we know that a valid freelist exists */
4529 map = get_map(s, page);
4530 for_each_object(p, s, addr, page->objects) {
4531 u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4532 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4534 if (!check_object(s, page, p, val))
4542 static int validate_slab_node(struct kmem_cache *s,
4543 struct kmem_cache_node *n)
4545 unsigned long count = 0;
4547 unsigned long flags;
4549 spin_lock_irqsave(&n->list_lock, flags);
4551 list_for_each_entry(page, &n->partial, slab_list) {
4552 validate_slab(s, page);
4555 if (count != n->nr_partial)
4556 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4557 s->name, count, n->nr_partial);
4559 if (!(s->flags & SLAB_STORE_USER))
4562 list_for_each_entry(page, &n->full, slab_list) {
4563 validate_slab(s, page);
4566 if (count != atomic_long_read(&n->nr_slabs))
4567 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4568 s->name, count, atomic_long_read(&n->nr_slabs));
4571 spin_unlock_irqrestore(&n->list_lock, flags);
4575 static long validate_slab_cache(struct kmem_cache *s)
4578 unsigned long count = 0;
4579 struct kmem_cache_node *n;
4582 for_each_kmem_cache_node(s, node, n)
4583 count += validate_slab_node(s, n);
4588 * Generate lists of code addresses where slabcache objects are allocated
4593 unsigned long count;
4600 DECLARE_BITMAP(cpus, NR_CPUS);
4606 unsigned long count;
4607 struct location *loc;
4610 static void free_loc_track(struct loc_track *t)
4613 free_pages((unsigned long)t->loc,
4614 get_order(sizeof(struct location) * t->max));
4617 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4622 order = get_order(sizeof(struct location) * max);
4624 l = (void *)__get_free_pages(flags, order);
4629 memcpy(l, t->loc, sizeof(struct location) * t->count);
4637 static int add_location(struct loc_track *t, struct kmem_cache *s,
4638 const struct track *track)
4640 long start, end, pos;
4642 unsigned long caddr;
4643 unsigned long age = jiffies - track->when;
4649 pos = start + (end - start + 1) / 2;
4652 * There is nothing at "end". If we end up there
4653 * we need to add something to before end.
4658 caddr = t->loc[pos].addr;
4659 if (track->addr == caddr) {
4665 if (age < l->min_time)
4667 if (age > l->max_time)
4670 if (track->pid < l->min_pid)
4671 l->min_pid = track->pid;
4672 if (track->pid > l->max_pid)
4673 l->max_pid = track->pid;
4675 cpumask_set_cpu(track->cpu,
4676 to_cpumask(l->cpus));
4678 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4682 if (track->addr < caddr)
4689 * Not found. Insert new tracking element.
4691 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4697 (t->count - pos) * sizeof(struct location));
4700 l->addr = track->addr;
4704 l->min_pid = track->pid;
4705 l->max_pid = track->pid;
4706 cpumask_clear(to_cpumask(l->cpus));
4707 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4708 nodes_clear(l->nodes);
4709 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4713 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4714 struct page *page, enum track_item alloc)
4716 void *addr = page_address(page);
4720 map = get_map(s, page);
4721 for_each_object(p, s, addr, page->objects)
4722 if (!test_bit(__obj_to_index(s, addr, p), map))
4723 add_location(t, s, get_track(s, p, alloc));
4727 static int list_locations(struct kmem_cache *s, char *buf,
4728 enum track_item alloc)
4732 struct loc_track t = { 0, 0, NULL };
4734 struct kmem_cache_node *n;
4736 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4738 return sprintf(buf, "Out of memory\n");
4740 /* Push back cpu slabs */
4743 for_each_kmem_cache_node(s, node, n) {
4744 unsigned long flags;
4747 if (!atomic_long_read(&n->nr_slabs))
4750 spin_lock_irqsave(&n->list_lock, flags);
4751 list_for_each_entry(page, &n->partial, slab_list)
4752 process_slab(&t, s, page, alloc);
4753 list_for_each_entry(page, &n->full, slab_list)
4754 process_slab(&t, s, page, alloc);
4755 spin_unlock_irqrestore(&n->list_lock, flags);
4758 for (i = 0; i < t.count; i++) {
4759 struct location *l = &t.loc[i];
4761 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4763 len += sprintf(buf + len, "%7ld ", l->count);
4766 len += sprintf(buf + len, "%pS", (void *)l->addr);
4768 len += sprintf(buf + len, "<not-available>");
4770 if (l->sum_time != l->min_time) {
4771 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4773 (long)div_u64(l->sum_time, l->count),
4776 len += sprintf(buf + len, " age=%ld",
4779 if (l->min_pid != l->max_pid)
4780 len += sprintf(buf + len, " pid=%ld-%ld",
4781 l->min_pid, l->max_pid);
4783 len += sprintf(buf + len, " pid=%ld",
4786 if (num_online_cpus() > 1 &&
4787 !cpumask_empty(to_cpumask(l->cpus)) &&
4788 len < PAGE_SIZE - 60)
4789 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4791 cpumask_pr_args(to_cpumask(l->cpus)));
4793 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4794 len < PAGE_SIZE - 60)
4795 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4797 nodemask_pr_args(&l->nodes));
4799 len += sprintf(buf + len, "\n");
4804 len += sprintf(buf, "No data\n");
4807 #endif /* CONFIG_SLUB_DEBUG */
4809 #ifdef SLUB_RESILIENCY_TEST
4810 static void __init resiliency_test(void)
4813 int type = KMALLOC_NORMAL;
4815 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4817 pr_err("SLUB resiliency testing\n");
4818 pr_err("-----------------------\n");
4819 pr_err("A. Corruption after allocation\n");
4821 p = kzalloc(16, GFP_KERNEL);
4823 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4826 validate_slab_cache(kmalloc_caches[type][4]);
4828 /* Hmmm... The next two are dangerous */
4829 p = kzalloc(32, GFP_KERNEL);
4830 p[32 + sizeof(void *)] = 0x34;
4831 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4833 pr_err("If allocated object is overwritten then not detectable\n\n");
4835 validate_slab_cache(kmalloc_caches[type][5]);
4836 p = kzalloc(64, GFP_KERNEL);
4837 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4839 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4841 pr_err("If allocated object is overwritten then not detectable\n\n");
4842 validate_slab_cache(kmalloc_caches[type][6]);
4844 pr_err("\nB. Corruption after free\n");
4845 p = kzalloc(128, GFP_KERNEL);
4848 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4849 validate_slab_cache(kmalloc_caches[type][7]);
4851 p = kzalloc(256, GFP_KERNEL);
4854 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4855 validate_slab_cache(kmalloc_caches[type][8]);
4857 p = kzalloc(512, GFP_KERNEL);
4860 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4861 validate_slab_cache(kmalloc_caches[type][9]);
4865 static void resiliency_test(void) {};
4867 #endif /* SLUB_RESILIENCY_TEST */
4870 enum slab_stat_type {
4871 SL_ALL, /* All slabs */
4872 SL_PARTIAL, /* Only partially allocated slabs */
4873 SL_CPU, /* Only slabs used for cpu caches */
4874 SL_OBJECTS, /* Determine allocated objects not slabs */
4875 SL_TOTAL /* Determine object capacity not slabs */
4878 #define SO_ALL (1 << SL_ALL)
4879 #define SO_PARTIAL (1 << SL_PARTIAL)
4880 #define SO_CPU (1 << SL_CPU)
4881 #define SO_OBJECTS (1 << SL_OBJECTS)
4882 #define SO_TOTAL (1 << SL_TOTAL)
4885 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4887 static int __init setup_slub_memcg_sysfs(char *str)
4891 if (get_option(&str, &v) > 0)
4892 memcg_sysfs_enabled = v;
4897 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4900 static ssize_t show_slab_objects(struct kmem_cache *s,
4901 char *buf, unsigned long flags)
4903 unsigned long total = 0;
4906 unsigned long *nodes;
4908 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4912 if (flags & SO_CPU) {
4915 for_each_possible_cpu(cpu) {
4916 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4921 page = READ_ONCE(c->page);
4925 node = page_to_nid(page);
4926 if (flags & SO_TOTAL)
4928 else if (flags & SO_OBJECTS)
4936 page = slub_percpu_partial_read_once(c);
4938 node = page_to_nid(page);
4939 if (flags & SO_TOTAL)
4941 else if (flags & SO_OBJECTS)
4952 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4953 * already held which will conflict with an existing lock order:
4955 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4957 * We don't really need mem_hotplug_lock (to hold off
4958 * slab_mem_going_offline_callback) here because slab's memory hot
4959 * unplug code doesn't destroy the kmem_cache->node[] data.
4962 #ifdef CONFIG_SLUB_DEBUG
4963 if (flags & SO_ALL) {
4964 struct kmem_cache_node *n;
4966 for_each_kmem_cache_node(s, node, n) {
4968 if (flags & SO_TOTAL)
4969 x = atomic_long_read(&n->total_objects);
4970 else if (flags & SO_OBJECTS)
4971 x = atomic_long_read(&n->total_objects) -
4972 count_partial(n, count_free);
4974 x = atomic_long_read(&n->nr_slabs);
4981 if (flags & SO_PARTIAL) {
4982 struct kmem_cache_node *n;
4984 for_each_kmem_cache_node(s, node, n) {
4985 if (flags & SO_TOTAL)
4986 x = count_partial(n, count_total);
4987 else if (flags & SO_OBJECTS)
4988 x = count_partial(n, count_inuse);
4995 x = sprintf(buf, "%lu", total);
4997 for (node = 0; node < nr_node_ids; node++)
4999 x += sprintf(buf + x, " N%d=%lu",
5003 return x + sprintf(buf + x, "\n");
5006 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5007 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5009 struct slab_attribute {
5010 struct attribute attr;
5011 ssize_t (*show)(struct kmem_cache *s, char *buf);
5012 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5015 #define SLAB_ATTR_RO(_name) \
5016 static struct slab_attribute _name##_attr = \
5017 __ATTR(_name, 0400, _name##_show, NULL)
5019 #define SLAB_ATTR(_name) \
5020 static struct slab_attribute _name##_attr = \
5021 __ATTR(_name, 0600, _name##_show, _name##_store)
5023 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5025 return sprintf(buf, "%u\n", s->size);
5027 SLAB_ATTR_RO(slab_size);
5029 static ssize_t align_show(struct kmem_cache *s, char *buf)
5031 return sprintf(buf, "%u\n", s->align);
5033 SLAB_ATTR_RO(align);
5035 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5037 return sprintf(buf, "%u\n", s->object_size);
5039 SLAB_ATTR_RO(object_size);
5041 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5043 return sprintf(buf, "%u\n", oo_objects(s->oo));
5045 SLAB_ATTR_RO(objs_per_slab);
5047 static ssize_t order_show(struct kmem_cache *s, char *buf)
5049 return sprintf(buf, "%u\n", oo_order(s->oo));
5051 SLAB_ATTR_RO(order);
5053 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5055 return sprintf(buf, "%lu\n", s->min_partial);
5058 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5064 err = kstrtoul(buf, 10, &min);
5068 set_min_partial(s, min);
5071 SLAB_ATTR(min_partial);
5073 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5075 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5078 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5081 unsigned int objects;
5084 err = kstrtouint(buf, 10, &objects);
5087 if (objects && !kmem_cache_has_cpu_partial(s))
5090 slub_set_cpu_partial(s, objects);
5094 SLAB_ATTR(cpu_partial);
5096 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5100 return sprintf(buf, "%pS\n", s->ctor);
5104 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5106 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5108 SLAB_ATTR_RO(aliases);
5110 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5112 return show_slab_objects(s, buf, SO_PARTIAL);
5114 SLAB_ATTR_RO(partial);
5116 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5118 return show_slab_objects(s, buf, SO_CPU);
5120 SLAB_ATTR_RO(cpu_slabs);
5122 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5124 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5126 SLAB_ATTR_RO(objects);
5128 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5130 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5132 SLAB_ATTR_RO(objects_partial);
5134 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5141 for_each_online_cpu(cpu) {
5144 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5147 pages += page->pages;
5148 objects += page->pobjects;
5152 len = sprintf(buf, "%d(%d)", objects, pages);
5155 for_each_online_cpu(cpu) {
5158 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5160 if (page && len < PAGE_SIZE - 20)
5161 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5162 page->pobjects, page->pages);
5165 return len + sprintf(buf + len, "\n");
5167 SLAB_ATTR_RO(slabs_cpu_partial);
5169 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5171 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5173 SLAB_ATTR_RO(reclaim_account);
5175 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5177 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5179 SLAB_ATTR_RO(hwcache_align);
5181 #ifdef CONFIG_ZONE_DMA
5182 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5184 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5186 SLAB_ATTR_RO(cache_dma);
5189 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5191 return sprintf(buf, "%u\n", s->usersize);
5193 SLAB_ATTR_RO(usersize);
5195 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5197 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5199 SLAB_ATTR_RO(destroy_by_rcu);
5201 #ifdef CONFIG_SLUB_DEBUG
5202 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5204 return show_slab_objects(s, buf, SO_ALL);
5206 SLAB_ATTR_RO(slabs);
5208 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5210 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5212 SLAB_ATTR_RO(total_objects);
5214 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5216 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5218 SLAB_ATTR_RO(sanity_checks);
5220 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5222 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5224 SLAB_ATTR_RO(trace);
5226 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5228 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5231 SLAB_ATTR_RO(red_zone);
5233 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5235 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5238 SLAB_ATTR_RO(poison);
5240 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5242 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5245 SLAB_ATTR_RO(store_user);
5247 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5252 static ssize_t validate_store(struct kmem_cache *s,
5253 const char *buf, size_t length)
5257 if (buf[0] == '1') {
5258 ret = validate_slab_cache(s);
5264 SLAB_ATTR(validate);
5266 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5268 if (!(s->flags & SLAB_STORE_USER))
5270 return list_locations(s, buf, TRACK_ALLOC);
5272 SLAB_ATTR_RO(alloc_calls);
5274 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5276 if (!(s->flags & SLAB_STORE_USER))
5278 return list_locations(s, buf, TRACK_FREE);
5280 SLAB_ATTR_RO(free_calls);
5281 #endif /* CONFIG_SLUB_DEBUG */
5283 #ifdef CONFIG_FAILSLAB
5284 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5286 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5288 SLAB_ATTR_RO(failslab);
5291 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5296 static ssize_t shrink_store(struct kmem_cache *s,
5297 const char *buf, size_t length)
5300 kmem_cache_shrink(s);
5308 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5310 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5313 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5314 const char *buf, size_t length)
5319 err = kstrtouint(buf, 10, &ratio);
5325 s->remote_node_defrag_ratio = ratio * 10;
5329 SLAB_ATTR(remote_node_defrag_ratio);
5332 #ifdef CONFIG_SLUB_STATS
5333 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5335 unsigned long sum = 0;
5338 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5343 for_each_online_cpu(cpu) {
5344 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5350 len = sprintf(buf, "%lu", sum);
5353 for_each_online_cpu(cpu) {
5354 if (data[cpu] && len < PAGE_SIZE - 20)
5355 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5359 return len + sprintf(buf + len, "\n");
5362 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5366 for_each_online_cpu(cpu)
5367 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5370 #define STAT_ATTR(si, text) \
5371 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5373 return show_stat(s, buf, si); \
5375 static ssize_t text##_store(struct kmem_cache *s, \
5376 const char *buf, size_t length) \
5378 if (buf[0] != '0') \
5380 clear_stat(s, si); \
5385 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5386 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5387 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5388 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5389 STAT_ATTR(FREE_FROZEN, free_frozen);
5390 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5391 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5392 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5393 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5394 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5395 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5396 STAT_ATTR(FREE_SLAB, free_slab);
5397 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5398 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5399 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5400 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5401 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5402 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5403 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5404 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5405 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5406 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5407 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5408 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5409 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5410 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5411 #endif /* CONFIG_SLUB_STATS */
5413 static struct attribute *slab_attrs[] = {
5414 &slab_size_attr.attr,
5415 &object_size_attr.attr,
5416 &objs_per_slab_attr.attr,
5418 &min_partial_attr.attr,
5419 &cpu_partial_attr.attr,
5421 &objects_partial_attr.attr,
5423 &cpu_slabs_attr.attr,
5427 &hwcache_align_attr.attr,
5428 &reclaim_account_attr.attr,
5429 &destroy_by_rcu_attr.attr,
5431 &slabs_cpu_partial_attr.attr,
5432 #ifdef CONFIG_SLUB_DEBUG
5433 &total_objects_attr.attr,
5435 &sanity_checks_attr.attr,
5437 &red_zone_attr.attr,
5439 &store_user_attr.attr,
5440 &validate_attr.attr,
5441 &alloc_calls_attr.attr,
5442 &free_calls_attr.attr,
5444 #ifdef CONFIG_ZONE_DMA
5445 &cache_dma_attr.attr,
5448 &remote_node_defrag_ratio_attr.attr,
5450 #ifdef CONFIG_SLUB_STATS
5451 &alloc_fastpath_attr.attr,
5452 &alloc_slowpath_attr.attr,
5453 &free_fastpath_attr.attr,
5454 &free_slowpath_attr.attr,
5455 &free_frozen_attr.attr,
5456 &free_add_partial_attr.attr,
5457 &free_remove_partial_attr.attr,
5458 &alloc_from_partial_attr.attr,
5459 &alloc_slab_attr.attr,
5460 &alloc_refill_attr.attr,
5461 &alloc_node_mismatch_attr.attr,
5462 &free_slab_attr.attr,
5463 &cpuslab_flush_attr.attr,
5464 &deactivate_full_attr.attr,
5465 &deactivate_empty_attr.attr,
5466 &deactivate_to_head_attr.attr,
5467 &deactivate_to_tail_attr.attr,
5468 &deactivate_remote_frees_attr.attr,
5469 &deactivate_bypass_attr.attr,
5470 &order_fallback_attr.attr,
5471 &cmpxchg_double_fail_attr.attr,
5472 &cmpxchg_double_cpu_fail_attr.attr,
5473 &cpu_partial_alloc_attr.attr,
5474 &cpu_partial_free_attr.attr,
5475 &cpu_partial_node_attr.attr,
5476 &cpu_partial_drain_attr.attr,
5478 #ifdef CONFIG_FAILSLAB
5479 &failslab_attr.attr,
5481 &usersize_attr.attr,
5486 static const struct attribute_group slab_attr_group = {
5487 .attrs = slab_attrs,
5490 static ssize_t slab_attr_show(struct kobject *kobj,
5491 struct attribute *attr,
5494 struct slab_attribute *attribute;
5495 struct kmem_cache *s;
5498 attribute = to_slab_attr(attr);
5501 if (!attribute->show)
5504 err = attribute->show(s, buf);
5509 static ssize_t slab_attr_store(struct kobject *kobj,
5510 struct attribute *attr,
5511 const char *buf, size_t len)
5513 struct slab_attribute *attribute;
5514 struct kmem_cache *s;
5517 attribute = to_slab_attr(attr);
5520 if (!attribute->store)
5523 err = attribute->store(s, buf, len);
5527 static void kmem_cache_release(struct kobject *k)
5529 slab_kmem_cache_release(to_slab(k));
5532 static const struct sysfs_ops slab_sysfs_ops = {
5533 .show = slab_attr_show,
5534 .store = slab_attr_store,
5537 static struct kobj_type slab_ktype = {
5538 .sysfs_ops = &slab_sysfs_ops,
5539 .release = kmem_cache_release,
5542 static struct kset *slab_kset;
5544 static inline struct kset *cache_kset(struct kmem_cache *s)
5549 #define ID_STR_LENGTH 64
5551 /* Create a unique string id for a slab cache:
5553 * Format :[flags-]size
5555 static char *create_unique_id(struct kmem_cache *s)
5557 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5564 * First flags affecting slabcache operations. We will only
5565 * get here for aliasable slabs so we do not need to support
5566 * too many flags. The flags here must cover all flags that
5567 * are matched during merging to guarantee that the id is
5570 if (s->flags & SLAB_CACHE_DMA)
5572 if (s->flags & SLAB_CACHE_DMA32)
5574 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5576 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5578 if (s->flags & SLAB_ACCOUNT)
5582 p += sprintf(p, "%07u", s->size);
5584 BUG_ON(p > name + ID_STR_LENGTH - 1);
5588 static int sysfs_slab_add(struct kmem_cache *s)
5592 struct kset *kset = cache_kset(s);
5593 int unmergeable = slab_unmergeable(s);
5596 kobject_init(&s->kobj, &slab_ktype);
5600 if (!unmergeable && disable_higher_order_debug &&
5601 (slub_debug & DEBUG_METADATA_FLAGS))
5606 * Slabcache can never be merged so we can use the name proper.
5607 * This is typically the case for debug situations. In that
5608 * case we can catch duplicate names easily.
5610 sysfs_remove_link(&slab_kset->kobj, s->name);
5614 * Create a unique name for the slab as a target
5617 name = create_unique_id(s);
5620 s->kobj.kset = kset;
5621 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5623 kobject_put(&s->kobj);
5627 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5632 /* Setup first alias */
5633 sysfs_slab_alias(s, s->name);
5640 kobject_del(&s->kobj);
5644 void sysfs_slab_unlink(struct kmem_cache *s)
5646 if (slab_state >= FULL)
5647 kobject_del(&s->kobj);
5650 void sysfs_slab_release(struct kmem_cache *s)
5652 if (slab_state >= FULL)
5653 kobject_put(&s->kobj);
5657 * Need to buffer aliases during bootup until sysfs becomes
5658 * available lest we lose that information.
5660 struct saved_alias {
5661 struct kmem_cache *s;
5663 struct saved_alias *next;
5666 static struct saved_alias *alias_list;
5668 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5670 struct saved_alias *al;
5672 if (slab_state == FULL) {
5674 * If we have a leftover link then remove it.
5676 sysfs_remove_link(&slab_kset->kobj, name);
5677 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5680 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5686 al->next = alias_list;
5691 static int __init slab_sysfs_init(void)
5693 struct kmem_cache *s;
5696 mutex_lock(&slab_mutex);
5698 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5700 mutex_unlock(&slab_mutex);
5701 pr_err("Cannot register slab subsystem.\n");
5707 list_for_each_entry(s, &slab_caches, list) {
5708 err = sysfs_slab_add(s);
5710 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5714 while (alias_list) {
5715 struct saved_alias *al = alias_list;
5717 alias_list = alias_list->next;
5718 err = sysfs_slab_alias(al->s, al->name);
5720 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5725 mutex_unlock(&slab_mutex);
5730 __initcall(slab_sysfs_init);
5731 #endif /* CONFIG_SYSFS */
5734 * The /proc/slabinfo ABI
5736 #ifdef CONFIG_SLUB_DEBUG
5737 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5739 unsigned long nr_slabs = 0;
5740 unsigned long nr_objs = 0;
5741 unsigned long nr_free = 0;
5743 struct kmem_cache_node *n;
5745 for_each_kmem_cache_node(s, node, n) {
5746 nr_slabs += node_nr_slabs(n);
5747 nr_objs += node_nr_objs(n);
5748 nr_free += count_partial(n, count_free);
5751 sinfo->active_objs = nr_objs - nr_free;
5752 sinfo->num_objs = nr_objs;
5753 sinfo->active_slabs = nr_slabs;
5754 sinfo->num_slabs = nr_slabs;
5755 sinfo->objects_per_slab = oo_objects(s->oo);
5756 sinfo->cache_order = oo_order(s->oo);
5759 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5763 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5764 size_t count, loff_t *ppos)
5768 #endif /* CONFIG_SLUB_DEBUG */