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/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
37 #include <linux/random.h>
39 #include <trace/events/kmem.h>
45 * 1. slab_mutex (Global Mutex)
47 * 3. slab_lock(page) (Only on some arches and for debugging)
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects:
56 * A. page->freelist -> List of object free in a page
57 * B. page->inuse -> Number of objects in use
58 * C. page->objects -> Number of objects in page
59 * D. page->frozen -> frozen state
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list except per cpu partial list. The processor that froze the
63 * slab is the one who can perform list operations on the page. Other
64 * processors may put objects onto the freelist but the processor that
65 * froze the slab is the only one that can retrieve the objects from the
68 * The list_lock protects the partial and full list on each node and
69 * the partial slab counter. If taken then no new slabs may be added or
70 * removed from the lists nor make the number of partial slabs be modified.
71 * (Note that the total number of slabs is an atomic value that may be
72 * modified without taking the list lock).
74 * The list_lock is a centralized lock and thus we avoid taking it as
75 * much as possible. As long as SLUB does not have to handle partial
76 * slabs, operations can continue without any centralized lock. F.e.
77 * allocating a long series of objects that fill up slabs does not require
79 * Interrupts are disabled during allocation and deallocation in order to
80 * make the slab allocator safe to use in the context of an irq. In addition
81 * interrupts are disabled to ensure that the processor does not change
82 * while handling per_cpu slabs, due to kernel preemption.
84 * SLUB assigns one slab for allocation to each processor.
85 * Allocations only occur from these slabs called cpu slabs.
87 * Slabs with free elements are kept on a partial list and during regular
88 * operations no list for full slabs is used. If an object in a full slab is
89 * freed then the slab will show up again on the partial lists.
90 * We track full slabs for debugging purposes though because otherwise we
91 * cannot scan all objects.
93 * Slabs are freed when they become empty. Teardown and setup is
94 * minimal so we rely on the page allocators per cpu caches for
95 * fast frees and allocs.
97 * page->frozen The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 #ifdef CONFIG_SLUB_DEBUG
119 #ifdef CONFIG_SLUB_DEBUG_ON
120 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
122 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
126 static inline bool kmem_cache_debug(struct kmem_cache *s)
128 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
131 void *fixup_red_left(struct kmem_cache *s, void *p)
133 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
134 p += s->red_left_pad;
139 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
141 #ifdef CONFIG_SLUB_CPU_PARTIAL
142 return !kmem_cache_debug(s);
149 * Issues still to be resolved:
151 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
153 * - Variable sizing of the per node arrays
156 /* Enable to test recovery from slab corruption on boot */
157 #undef SLUB_RESILIENCY_TEST
159 /* Enable to log cmpxchg failures */
160 #undef SLUB_DEBUG_CMPXCHG
163 * Mininum number of partial slabs. These will be left on the partial
164 * lists even if they are empty. kmem_cache_shrink may reclaim them.
166 #define MIN_PARTIAL 5
169 * Maximum number of desirable partial slabs.
170 * The existence of more partial slabs makes kmem_cache_shrink
171 * sort the partial list by the number of objects in use.
173 #define MAX_PARTIAL 10
175 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_STORE_USER)
179 * These debug flags cannot use CMPXCHG because there might be consistency
180 * issues when checking or reading debug information
182 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
187 * Debugging flags that require metadata to be stored in the slab. These get
188 * disabled when slub_debug=O is used and a cache's min order increases with
191 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
194 #define OO_MASK ((1 << OO_SHIFT) - 1)
195 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
197 /* Internal SLUB flags */
199 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
200 /* Use cmpxchg_double */
201 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
204 * Tracking user of a slab.
206 #define TRACK_ADDRS_COUNT 16
208 unsigned long addr; /* Called from address */
209 #ifdef CONFIG_STACKTRACE
210 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
212 int cpu; /* Was running on cpu */
213 int pid; /* Pid context */
214 unsigned long when; /* When did the operation occur */
217 enum track_item { TRACK_ALLOC, TRACK_FREE };
220 static int sysfs_slab_add(struct kmem_cache *);
221 static int sysfs_slab_alias(struct kmem_cache *, const char *);
223 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
224 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s->cpu_slab->stat[si]);
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
244 * Returns freelist pointer (ptr). With hardening, this is obfuscated
245 * with an XOR of the address where the pointer is held and a per-cache
248 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
249 unsigned long ptr_addr)
251 #ifdef CONFIG_SLAB_FREELIST_HARDENED
253 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
254 * Normally, this doesn't cause any issues, as both set_freepointer()
255 * and get_freepointer() are called with a pointer with the same tag.
256 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
257 * example, when __free_slub() iterates over objects in a cache, it
258 * passes untagged pointers to check_object(). check_object() in turns
259 * calls get_freepointer() with an untagged pointer, which causes the
260 * freepointer to be restored incorrectly.
262 return (void *)((unsigned long)ptr ^ s->random ^
263 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
269 /* Returns the freelist pointer recorded at location ptr_addr. */
270 static inline void *freelist_dereference(const struct kmem_cache *s,
273 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
274 (unsigned long)ptr_addr);
277 static inline void *get_freepointer(struct kmem_cache *s, void *object)
279 return freelist_dereference(s, object + s->offset);
282 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
284 prefetch(object + s->offset);
287 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
289 unsigned long freepointer_addr;
292 if (!debug_pagealloc_enabled_static())
293 return get_freepointer(s, object);
295 freepointer_addr = (unsigned long)object + s->offset;
296 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
297 return freelist_ptr(s, p, freepointer_addr);
300 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
302 unsigned long freeptr_addr = (unsigned long)object + s->offset;
304 #ifdef CONFIG_SLAB_FREELIST_HARDENED
305 BUG_ON(object == fp); /* naive detection of double free or corruption */
308 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
311 /* Loop over all objects in a slab */
312 #define for_each_object(__p, __s, __addr, __objects) \
313 for (__p = fixup_red_left(__s, __addr); \
314 __p < (__addr) + (__objects) * (__s)->size; \
317 static inline unsigned int order_objects(unsigned int order, unsigned int size)
319 return ((unsigned int)PAGE_SIZE << order) / size;
322 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
325 struct kmem_cache_order_objects x = {
326 (order << OO_SHIFT) + order_objects(order, size)
332 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
334 return x.x >> OO_SHIFT;
337 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
339 return x.x & OO_MASK;
343 * Per slab locking using the pagelock
345 static __always_inline void slab_lock(struct page *page)
347 VM_BUG_ON_PAGE(PageTail(page), page);
348 bit_spin_lock(PG_locked, &page->flags);
351 static __always_inline void slab_unlock(struct page *page)
353 VM_BUG_ON_PAGE(PageTail(page), page);
354 __bit_spin_unlock(PG_locked, &page->flags);
357 /* Interrupts must be disabled (for the fallback code to work right) */
358 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
359 void *freelist_old, unsigned long counters_old,
360 void *freelist_new, unsigned long counters_new,
363 VM_BUG_ON(!irqs_disabled());
364 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
365 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
366 if (s->flags & __CMPXCHG_DOUBLE) {
367 if (cmpxchg_double(&page->freelist, &page->counters,
368 freelist_old, counters_old,
369 freelist_new, counters_new))
375 if (page->freelist == freelist_old &&
376 page->counters == counters_old) {
377 page->freelist = freelist_new;
378 page->counters = counters_new;
386 stat(s, CMPXCHG_DOUBLE_FAIL);
388 #ifdef SLUB_DEBUG_CMPXCHG
389 pr_info("%s %s: cmpxchg double redo ", n, s->name);
395 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
396 void *freelist_old, unsigned long counters_old,
397 void *freelist_new, unsigned long counters_new,
400 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
401 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
402 if (s->flags & __CMPXCHG_DOUBLE) {
403 if (cmpxchg_double(&page->freelist, &page->counters,
404 freelist_old, counters_old,
405 freelist_new, counters_new))
412 local_irq_save(flags);
414 if (page->freelist == freelist_old &&
415 page->counters == counters_old) {
416 page->freelist = freelist_new;
417 page->counters = counters_new;
419 local_irq_restore(flags);
423 local_irq_restore(flags);
427 stat(s, CMPXCHG_DOUBLE_FAIL);
429 #ifdef SLUB_DEBUG_CMPXCHG
430 pr_info("%s %s: cmpxchg double redo ", n, s->name);
436 #ifdef CONFIG_SLUB_DEBUG
437 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
438 static DEFINE_SPINLOCK(object_map_lock);
441 * Determine a map of object in use on a page.
443 * Node listlock must be held to guarantee that the page does
444 * not vanish from under us.
446 static unsigned long *get_map(struct kmem_cache *s, struct page *page)
447 __acquires(&object_map_lock)
450 void *addr = page_address(page);
452 VM_BUG_ON(!irqs_disabled());
454 spin_lock(&object_map_lock);
456 bitmap_zero(object_map, page->objects);
458 for (p = page->freelist; p; p = get_freepointer(s, p))
459 set_bit(__obj_to_index(s, addr, p), object_map);
464 static void put_map(unsigned long *map) __releases(&object_map_lock)
466 VM_BUG_ON(map != object_map);
467 spin_unlock(&object_map_lock);
470 static inline unsigned int size_from_object(struct kmem_cache *s)
472 if (s->flags & SLAB_RED_ZONE)
473 return s->size - s->red_left_pad;
478 static inline void *restore_red_left(struct kmem_cache *s, void *p)
480 if (s->flags & SLAB_RED_ZONE)
481 p -= s->red_left_pad;
489 #if defined(CONFIG_SLUB_DEBUG_ON)
490 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
492 static slab_flags_t slub_debug;
495 static char *slub_debug_string;
496 static int disable_higher_order_debug;
499 * slub is about to manipulate internal object metadata. This memory lies
500 * outside the range of the allocated object, so accessing it would normally
501 * be reported by kasan as a bounds error. metadata_access_enable() is used
502 * to tell kasan that these accesses are OK.
504 static inline void metadata_access_enable(void)
506 kasan_disable_current();
509 static inline void metadata_access_disable(void)
511 kasan_enable_current();
518 /* Verify that a pointer has an address that is valid within a slab page */
519 static inline int check_valid_pointer(struct kmem_cache *s,
520 struct page *page, void *object)
527 base = page_address(page);
528 object = kasan_reset_tag(object);
529 object = restore_red_left(s, object);
530 if (object < base || object >= base + page->objects * s->size ||
531 (object - base) % s->size) {
538 static void print_section(char *level, char *text, u8 *addr,
541 metadata_access_enable();
542 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
544 metadata_access_disable();
548 * See comment in calculate_sizes().
550 static inline bool freeptr_outside_object(struct kmem_cache *s)
552 return s->offset >= s->inuse;
556 * Return offset of the end of info block which is inuse + free pointer if
557 * not overlapping with object.
559 static inline unsigned int get_info_end(struct kmem_cache *s)
561 if (freeptr_outside_object(s))
562 return s->inuse + sizeof(void *);
567 static struct track *get_track(struct kmem_cache *s, void *object,
568 enum track_item alloc)
572 p = object + get_info_end(s);
577 static void set_track(struct kmem_cache *s, void *object,
578 enum track_item alloc, unsigned long addr)
580 struct track *p = get_track(s, object, alloc);
583 #ifdef CONFIG_STACKTRACE
584 unsigned int nr_entries;
586 metadata_access_enable();
587 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
588 metadata_access_disable();
590 if (nr_entries < TRACK_ADDRS_COUNT)
591 p->addrs[nr_entries] = 0;
594 p->cpu = smp_processor_id();
595 p->pid = current->pid;
598 memset(p, 0, sizeof(struct track));
602 static void init_tracking(struct kmem_cache *s, void *object)
604 if (!(s->flags & SLAB_STORE_USER))
607 set_track(s, object, TRACK_FREE, 0UL);
608 set_track(s, object, TRACK_ALLOC, 0UL);
611 static void print_track(const char *s, struct track *t, unsigned long pr_time)
616 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
617 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
618 #ifdef CONFIG_STACKTRACE
621 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
623 pr_err("\t%pS\n", (void *)t->addrs[i]);
630 void print_tracking(struct kmem_cache *s, void *object)
632 unsigned long pr_time = jiffies;
633 if (!(s->flags & SLAB_STORE_USER))
636 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
637 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
640 static void print_page_info(struct page *page)
642 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
643 page, page->objects, page->inuse, page->freelist, page->flags);
647 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
649 struct va_format vaf;
655 pr_err("=============================================================================\n");
656 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
657 pr_err("-----------------------------------------------------------------------------\n\n");
659 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
663 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
665 struct va_format vaf;
671 pr_err("FIX %s: %pV\n", s->name, &vaf);
675 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
676 void **freelist, void *nextfree)
678 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
679 !check_valid_pointer(s, page, nextfree) && freelist) {
680 object_err(s, page, *freelist, "Freechain corrupt");
682 slab_fix(s, "Isolate corrupted freechain");
689 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
691 unsigned int off; /* Offset of last byte */
692 u8 *addr = page_address(page);
694 print_tracking(s, p);
696 print_page_info(page);
698 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
699 p, p - addr, get_freepointer(s, p));
701 if (s->flags & SLAB_RED_ZONE)
702 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
704 else if (p > addr + 16)
705 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
707 print_section(KERN_ERR, "Object ", p,
708 min_t(unsigned int, s->object_size, PAGE_SIZE));
709 if (s->flags & SLAB_RED_ZONE)
710 print_section(KERN_ERR, "Redzone ", p + s->object_size,
711 s->inuse - s->object_size);
713 off = get_info_end(s);
715 if (s->flags & SLAB_STORE_USER)
716 off += 2 * sizeof(struct track);
718 off += kasan_metadata_size(s);
720 if (off != size_from_object(s))
721 /* Beginning of the filler is the free pointer */
722 print_section(KERN_ERR, "Padding ", p + off,
723 size_from_object(s) - off);
728 void object_err(struct kmem_cache *s, struct page *page,
729 u8 *object, char *reason)
731 slab_bug(s, "%s", reason);
732 print_trailer(s, page, object);
735 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
736 const char *fmt, ...)
742 vsnprintf(buf, sizeof(buf), fmt, args);
744 slab_bug(s, "%s", buf);
745 print_page_info(page);
749 static void init_object(struct kmem_cache *s, void *object, u8 val)
753 if (s->flags & SLAB_RED_ZONE)
754 memset(p - s->red_left_pad, val, s->red_left_pad);
756 if (s->flags & __OBJECT_POISON) {
757 memset(p, POISON_FREE, s->object_size - 1);
758 p[s->object_size - 1] = POISON_END;
761 if (s->flags & SLAB_RED_ZONE)
762 memset(p + s->object_size, val, s->inuse - s->object_size);
765 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
766 void *from, void *to)
768 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
769 memset(from, data, to - from);
772 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
773 u8 *object, char *what,
774 u8 *start, unsigned int value, unsigned int bytes)
778 u8 *addr = page_address(page);
780 metadata_access_enable();
781 fault = memchr_inv(start, value, bytes);
782 metadata_access_disable();
787 while (end > fault && end[-1] == value)
790 slab_bug(s, "%s overwritten", what);
791 pr_err("INFO: 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
792 fault, end - 1, fault - addr,
794 print_trailer(s, page, object);
796 restore_bytes(s, what, value, fault, end);
804 * Bytes of the object to be managed.
805 * If the freepointer may overlay the object then the free
806 * pointer is at the middle of the object.
808 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
811 * object + s->object_size
812 * Padding to reach word boundary. This is also used for Redzoning.
813 * Padding is extended by another word if Redzoning is enabled and
814 * object_size == inuse.
816 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
817 * 0xcc (RED_ACTIVE) for objects in use.
820 * Meta data starts here.
822 * A. Free pointer (if we cannot overwrite object on free)
823 * B. Tracking data for SLAB_STORE_USER
824 * C. Padding to reach required alignment boundary or at mininum
825 * one word if debugging is on to be able to detect writes
826 * before the word boundary.
828 * Padding is done using 0x5a (POISON_INUSE)
831 * Nothing is used beyond s->size.
833 * If slabcaches are merged then the object_size and inuse boundaries are mostly
834 * ignored. And therefore no slab options that rely on these boundaries
835 * may be used with merged slabcaches.
838 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
840 unsigned long off = get_info_end(s); /* The end of info */
842 if (s->flags & SLAB_STORE_USER)
843 /* We also have user information there */
844 off += 2 * sizeof(struct track);
846 off += kasan_metadata_size(s);
848 if (size_from_object(s) == off)
851 return check_bytes_and_report(s, page, p, "Object padding",
852 p + off, POISON_INUSE, size_from_object(s) - off);
855 /* Check the pad bytes at the end of a slab page */
856 static int slab_pad_check(struct kmem_cache *s, struct page *page)
865 if (!(s->flags & SLAB_POISON))
868 start = page_address(page);
869 length = page_size(page);
870 end = start + length;
871 remainder = length % s->size;
875 pad = end - remainder;
876 metadata_access_enable();
877 fault = memchr_inv(pad, POISON_INUSE, remainder);
878 metadata_access_disable();
881 while (end > fault && end[-1] == POISON_INUSE)
884 slab_err(s, page, "Padding overwritten. 0x%p-0x%p @offset=%tu",
885 fault, end - 1, fault - start);
886 print_section(KERN_ERR, "Padding ", pad, remainder);
888 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
892 static int check_object(struct kmem_cache *s, struct page *page,
893 void *object, u8 val)
896 u8 *endobject = object + s->object_size;
898 if (s->flags & SLAB_RED_ZONE) {
899 if (!check_bytes_and_report(s, page, object, "Left Redzone",
900 object - s->red_left_pad, val, s->red_left_pad))
903 if (!check_bytes_and_report(s, page, object, "Right Redzone",
904 endobject, val, s->inuse - s->object_size))
907 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
908 check_bytes_and_report(s, page, p, "Alignment padding",
909 endobject, POISON_INUSE,
910 s->inuse - s->object_size);
914 if (s->flags & SLAB_POISON) {
915 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
916 (!check_bytes_and_report(s, page, p, "Poison", p,
917 POISON_FREE, s->object_size - 1) ||
918 !check_bytes_and_report(s, page, p, "End Poison",
919 p + s->object_size - 1, POISON_END, 1)))
922 * check_pad_bytes cleans up on its own.
924 check_pad_bytes(s, page, p);
927 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
929 * Object and freepointer overlap. Cannot check
930 * freepointer while object is allocated.
934 /* Check free pointer validity */
935 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
936 object_err(s, page, p, "Freepointer corrupt");
938 * No choice but to zap it and thus lose the remainder
939 * of the free objects in this slab. May cause
940 * another error because the object count is now wrong.
942 set_freepointer(s, p, NULL);
948 static int check_slab(struct kmem_cache *s, struct page *page)
952 VM_BUG_ON(!irqs_disabled());
954 if (!PageSlab(page)) {
955 slab_err(s, page, "Not a valid slab page");
959 maxobj = order_objects(compound_order(page), s->size);
960 if (page->objects > maxobj) {
961 slab_err(s, page, "objects %u > max %u",
962 page->objects, maxobj);
965 if (page->inuse > page->objects) {
966 slab_err(s, page, "inuse %u > max %u",
967 page->inuse, page->objects);
970 /* Slab_pad_check fixes things up after itself */
971 slab_pad_check(s, page);
976 * Determine if a certain object on a page is on the freelist. Must hold the
977 * slab lock to guarantee that the chains are in a consistent state.
979 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
987 while (fp && nr <= page->objects) {
990 if (!check_valid_pointer(s, page, fp)) {
992 object_err(s, page, object,
993 "Freechain corrupt");
994 set_freepointer(s, object, NULL);
996 slab_err(s, page, "Freepointer corrupt");
997 page->freelist = NULL;
998 page->inuse = page->objects;
999 slab_fix(s, "Freelist cleared");
1005 fp = get_freepointer(s, object);
1009 max_objects = order_objects(compound_order(page), s->size);
1010 if (max_objects > MAX_OBJS_PER_PAGE)
1011 max_objects = MAX_OBJS_PER_PAGE;
1013 if (page->objects != max_objects) {
1014 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1015 page->objects, max_objects);
1016 page->objects = max_objects;
1017 slab_fix(s, "Number of objects adjusted.");
1019 if (page->inuse != page->objects - nr) {
1020 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1021 page->inuse, page->objects - nr);
1022 page->inuse = page->objects - nr;
1023 slab_fix(s, "Object count adjusted.");
1025 return search == NULL;
1028 static void trace(struct kmem_cache *s, struct page *page, void *object,
1031 if (s->flags & SLAB_TRACE) {
1032 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1034 alloc ? "alloc" : "free",
1035 object, page->inuse,
1039 print_section(KERN_INFO, "Object ", (void *)object,
1047 * Tracking of fully allocated slabs for debugging purposes.
1049 static void add_full(struct kmem_cache *s,
1050 struct kmem_cache_node *n, struct page *page)
1052 if (!(s->flags & SLAB_STORE_USER))
1055 lockdep_assert_held(&n->list_lock);
1056 list_add(&page->slab_list, &n->full);
1059 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1061 if (!(s->flags & SLAB_STORE_USER))
1064 lockdep_assert_held(&n->list_lock);
1065 list_del(&page->slab_list);
1068 /* Tracking of the number of slabs for debugging purposes */
1069 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1071 struct kmem_cache_node *n = get_node(s, node);
1073 return atomic_long_read(&n->nr_slabs);
1076 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1078 return atomic_long_read(&n->nr_slabs);
1081 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1083 struct kmem_cache_node *n = get_node(s, node);
1086 * May be called early in order to allocate a slab for the
1087 * kmem_cache_node structure. Solve the chicken-egg
1088 * dilemma by deferring the increment of the count during
1089 * bootstrap (see early_kmem_cache_node_alloc).
1092 atomic_long_inc(&n->nr_slabs);
1093 atomic_long_add(objects, &n->total_objects);
1096 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1098 struct kmem_cache_node *n = get_node(s, node);
1100 atomic_long_dec(&n->nr_slabs);
1101 atomic_long_sub(objects, &n->total_objects);
1104 /* Object debug checks for alloc/free paths */
1105 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1108 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1111 init_object(s, object, SLUB_RED_INACTIVE);
1112 init_tracking(s, object);
1116 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1118 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1121 metadata_access_enable();
1122 memset(addr, POISON_INUSE, page_size(page));
1123 metadata_access_disable();
1126 static inline int alloc_consistency_checks(struct kmem_cache *s,
1127 struct page *page, void *object)
1129 if (!check_slab(s, page))
1132 if (!check_valid_pointer(s, page, object)) {
1133 object_err(s, page, object, "Freelist Pointer check fails");
1137 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1143 static noinline int alloc_debug_processing(struct kmem_cache *s,
1145 void *object, unsigned long addr)
1147 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1148 if (!alloc_consistency_checks(s, page, object))
1152 /* Success perform special debug activities for allocs */
1153 if (s->flags & SLAB_STORE_USER)
1154 set_track(s, object, TRACK_ALLOC, addr);
1155 trace(s, page, object, 1);
1156 init_object(s, object, SLUB_RED_ACTIVE);
1160 if (PageSlab(page)) {
1162 * If this is a slab page then lets do the best we can
1163 * to avoid issues in the future. Marking all objects
1164 * as used avoids touching the remaining objects.
1166 slab_fix(s, "Marking all objects used");
1167 page->inuse = page->objects;
1168 page->freelist = NULL;
1173 static inline int free_consistency_checks(struct kmem_cache *s,
1174 struct page *page, void *object, unsigned long addr)
1176 if (!check_valid_pointer(s, page, object)) {
1177 slab_err(s, page, "Invalid object pointer 0x%p", object);
1181 if (on_freelist(s, page, object)) {
1182 object_err(s, page, object, "Object already free");
1186 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1189 if (unlikely(s != page->slab_cache)) {
1190 if (!PageSlab(page)) {
1191 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1193 } else if (!page->slab_cache) {
1194 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1198 object_err(s, page, object,
1199 "page slab pointer corrupt.");
1205 /* Supports checking bulk free of a constructed freelist */
1206 static noinline int free_debug_processing(
1207 struct kmem_cache *s, struct page *page,
1208 void *head, void *tail, int bulk_cnt,
1211 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1212 void *object = head;
1214 unsigned long flags;
1217 spin_lock_irqsave(&n->list_lock, flags);
1220 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1221 if (!check_slab(s, page))
1228 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1229 if (!free_consistency_checks(s, page, object, addr))
1233 if (s->flags & SLAB_STORE_USER)
1234 set_track(s, object, TRACK_FREE, addr);
1235 trace(s, page, object, 0);
1236 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1237 init_object(s, object, SLUB_RED_INACTIVE);
1239 /* Reached end of constructed freelist yet? */
1240 if (object != tail) {
1241 object = get_freepointer(s, object);
1247 if (cnt != bulk_cnt)
1248 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1252 spin_unlock_irqrestore(&n->list_lock, flags);
1254 slab_fix(s, "Object at 0x%p not freed", object);
1259 * Parse a block of slub_debug options. Blocks are delimited by ';'
1261 * @str: start of block
1262 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1263 * @slabs: return start of list of slabs, or NULL when there's no list
1264 * @init: assume this is initial parsing and not per-kmem-create parsing
1266 * returns the start of next block if there's any, or NULL
1269 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1271 bool higher_order_disable = false;
1273 /* Skip any completely empty blocks */
1274 while (*str && *str == ';')
1279 * No options but restriction on slabs. This means full
1280 * debugging for slabs matching a pattern.
1282 *flags = DEBUG_DEFAULT_FLAGS;
1287 /* Determine which debug features should be switched on */
1288 for (; *str && *str != ',' && *str != ';'; str++) {
1289 switch (tolower(*str)) {
1294 *flags |= SLAB_CONSISTENCY_CHECKS;
1297 *flags |= SLAB_RED_ZONE;
1300 *flags |= SLAB_POISON;
1303 *flags |= SLAB_STORE_USER;
1306 *flags |= SLAB_TRACE;
1309 *flags |= SLAB_FAILSLAB;
1313 * Avoid enabling debugging on caches if its minimum
1314 * order would increase as a result.
1316 higher_order_disable = true;
1320 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1329 /* Skip over the slab list */
1330 while (*str && *str != ';')
1333 /* Skip any completely empty blocks */
1334 while (*str && *str == ';')
1337 if (init && higher_order_disable)
1338 disable_higher_order_debug = 1;
1346 static int __init setup_slub_debug(char *str)
1351 bool global_slub_debug_changed = false;
1352 bool slab_list_specified = false;
1354 slub_debug = DEBUG_DEFAULT_FLAGS;
1355 if (*str++ != '=' || !*str)
1357 * No options specified. Switch on full debugging.
1363 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1367 global_slub_debug_changed = true;
1369 slab_list_specified = true;
1374 * For backwards compatibility, a single list of flags with list of
1375 * slabs means debugging is only enabled for those slabs, so the global
1376 * slub_debug should be 0. We can extended that to multiple lists as
1377 * long as there is no option specifying flags without a slab list.
1379 if (slab_list_specified) {
1380 if (!global_slub_debug_changed)
1382 slub_debug_string = saved_str;
1385 if (slub_debug != 0 || slub_debug_string)
1386 static_branch_enable(&slub_debug_enabled);
1387 if ((static_branch_unlikely(&init_on_alloc) ||
1388 static_branch_unlikely(&init_on_free)) &&
1389 (slub_debug & SLAB_POISON))
1390 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1394 __setup("slub_debug", setup_slub_debug);
1397 * kmem_cache_flags - apply debugging options to the cache
1398 * @object_size: the size of an object without meta data
1399 * @flags: flags to set
1400 * @name: name of the cache
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)
1413 slab_flags_t block_flags;
1416 next_block = slub_debug_string;
1417 /* Go through all blocks of debug options, see if any matches our slab's name */
1418 while (next_block) {
1419 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1422 /* Found a block that has a slab list, search it */
1427 end = strchrnul(iter, ',');
1428 if (next_block && next_block < end)
1429 end = next_block - 1;
1431 glob = strnchr(iter, end - iter, '*');
1433 cmplen = glob - iter;
1435 cmplen = max_t(size_t, len, (end - iter));
1437 if (!strncmp(name, iter, cmplen)) {
1438 flags |= block_flags;
1442 if (!*end || *end == ';')
1448 return flags | slub_debug;
1450 #else /* !CONFIG_SLUB_DEBUG */
1451 static inline void setup_object_debug(struct kmem_cache *s,
1452 struct page *page, void *object) {}
1454 void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1456 static inline int alloc_debug_processing(struct kmem_cache *s,
1457 struct page *page, void *object, unsigned long addr) { return 0; }
1459 static inline int free_debug_processing(
1460 struct kmem_cache *s, struct page *page,
1461 void *head, void *tail, int bulk_cnt,
1462 unsigned long addr) { return 0; }
1464 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1466 static inline int check_object(struct kmem_cache *s, struct page *page,
1467 void *object, u8 val) { return 1; }
1468 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1469 struct page *page) {}
1470 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1471 struct page *page) {}
1472 slab_flags_t kmem_cache_flags(unsigned int object_size,
1473 slab_flags_t flags, const char *name)
1477 #define slub_debug 0
1479 #define disable_higher_order_debug 0
1481 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1483 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1485 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1487 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1490 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1491 void **freelist, void *nextfree)
1495 #endif /* CONFIG_SLUB_DEBUG */
1498 * Hooks for other subsystems that check memory allocations. In a typical
1499 * production configuration these hooks all should produce no code at all.
1501 static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1503 ptr = kasan_kmalloc_large(ptr, size, flags);
1504 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1505 kmemleak_alloc(ptr, size, 1, flags);
1509 static __always_inline void kfree_hook(void *x)
1512 kasan_kfree_large(x, _RET_IP_);
1515 static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1517 kmemleak_free_recursive(x, s->flags);
1520 * Trouble is that we may no longer disable interrupts in the fast path
1521 * So in order to make the debug calls that expect irqs to be
1522 * disabled we need to disable interrupts temporarily.
1524 #ifdef CONFIG_LOCKDEP
1526 unsigned long flags;
1528 local_irq_save(flags);
1529 debug_check_no_locks_freed(x, s->object_size);
1530 local_irq_restore(flags);
1533 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1534 debug_check_no_obj_freed(x, s->object_size);
1536 /* Use KCSAN to help debug racy use-after-free. */
1537 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1538 __kcsan_check_access(x, s->object_size,
1539 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1541 /* KASAN might put x into memory quarantine, delaying its reuse */
1542 return kasan_slab_free(s, x, _RET_IP_);
1545 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1546 void **head, void **tail,
1552 void *old_tail = *tail ? *tail : *head;
1555 /* Head and tail of the reconstructed freelist */
1561 next = get_freepointer(s, object);
1563 if (slab_want_init_on_free(s)) {
1565 * Clear the object and the metadata, but don't touch
1568 memset(object, 0, s->object_size);
1569 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1571 memset((char *)object + s->inuse, 0,
1572 s->size - s->inuse - rsize);
1575 /* If object's reuse doesn't have to be delayed */
1576 if (!slab_free_hook(s, object)) {
1577 /* Move object to the new freelist */
1578 set_freepointer(s, object, *head);
1584 * Adjust the reconstructed freelist depth
1585 * accordingly if object's reuse is delayed.
1589 } while (object != old_tail);
1594 return *head != NULL;
1597 static void *setup_object(struct kmem_cache *s, struct page *page,
1600 setup_object_debug(s, page, object);
1601 object = kasan_init_slab_obj(s, object);
1602 if (unlikely(s->ctor)) {
1603 kasan_unpoison_object_data(s, object);
1605 kasan_poison_object_data(s, object);
1611 * Slab allocation and freeing
1613 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1614 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1617 unsigned int order = oo_order(oo);
1619 if (node == NUMA_NO_NODE)
1620 page = alloc_pages(flags, order);
1622 page = __alloc_pages_node(node, flags, order);
1625 account_slab_page(page, order, s);
1630 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1631 /* Pre-initialize the random sequence cache */
1632 static int init_cache_random_seq(struct kmem_cache *s)
1634 unsigned int count = oo_objects(s->oo);
1637 /* Bailout if already initialised */
1641 err = cache_random_seq_create(s, count, GFP_KERNEL);
1643 pr_err("SLUB: Unable to initialize free list for %s\n",
1648 /* Transform to an offset on the set of pages */
1649 if (s->random_seq) {
1652 for (i = 0; i < count; i++)
1653 s->random_seq[i] *= s->size;
1658 /* Initialize each random sequence freelist per cache */
1659 static void __init init_freelist_randomization(void)
1661 struct kmem_cache *s;
1663 mutex_lock(&slab_mutex);
1665 list_for_each_entry(s, &slab_caches, list)
1666 init_cache_random_seq(s);
1668 mutex_unlock(&slab_mutex);
1671 /* Get the next entry on the pre-computed freelist randomized */
1672 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1673 unsigned long *pos, void *start,
1674 unsigned long page_limit,
1675 unsigned long freelist_count)
1680 * If the target page allocation failed, the number of objects on the
1681 * page might be smaller than the usual size defined by the cache.
1684 idx = s->random_seq[*pos];
1686 if (*pos >= freelist_count)
1688 } while (unlikely(idx >= page_limit));
1690 return (char *)start + idx;
1693 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1694 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1699 unsigned long idx, pos, page_limit, freelist_count;
1701 if (page->objects < 2 || !s->random_seq)
1704 freelist_count = oo_objects(s->oo);
1705 pos = get_random_int() % freelist_count;
1707 page_limit = page->objects * s->size;
1708 start = fixup_red_left(s, page_address(page));
1710 /* First entry is used as the base of the freelist */
1711 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1713 cur = setup_object(s, page, cur);
1714 page->freelist = cur;
1716 for (idx = 1; idx < page->objects; idx++) {
1717 next = next_freelist_entry(s, page, &pos, start, page_limit,
1719 next = setup_object(s, page, next);
1720 set_freepointer(s, cur, next);
1723 set_freepointer(s, cur, NULL);
1728 static inline int init_cache_random_seq(struct kmem_cache *s)
1732 static inline void init_freelist_randomization(void) { }
1733 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1737 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1739 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1742 struct kmem_cache_order_objects oo = s->oo;
1744 void *start, *p, *next;
1748 flags &= gfp_allowed_mask;
1750 if (gfpflags_allow_blocking(flags))
1753 flags |= s->allocflags;
1756 * Let the initial higher-order allocation fail under memory pressure
1757 * so we fall-back to the minimum order allocation.
1759 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1760 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1761 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1763 page = alloc_slab_page(s, alloc_gfp, node, oo);
1764 if (unlikely(!page)) {
1768 * Allocation may have failed due to fragmentation.
1769 * Try a lower order alloc if possible
1771 page = alloc_slab_page(s, alloc_gfp, node, oo);
1772 if (unlikely(!page))
1774 stat(s, ORDER_FALLBACK);
1777 page->objects = oo_objects(oo);
1779 page->slab_cache = s;
1780 __SetPageSlab(page);
1781 if (page_is_pfmemalloc(page))
1782 SetPageSlabPfmemalloc(page);
1784 kasan_poison_slab(page);
1786 start = page_address(page);
1788 setup_page_debug(s, page, start);
1790 shuffle = shuffle_freelist(s, page);
1793 start = fixup_red_left(s, start);
1794 start = setup_object(s, page, start);
1795 page->freelist = start;
1796 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1798 next = setup_object(s, page, next);
1799 set_freepointer(s, p, next);
1802 set_freepointer(s, p, NULL);
1805 page->inuse = page->objects;
1809 if (gfpflags_allow_blocking(flags))
1810 local_irq_disable();
1814 inc_slabs_node(s, page_to_nid(page), page->objects);
1819 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1821 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1822 flags = kmalloc_fix_flags(flags);
1824 return allocate_slab(s,
1825 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1828 static void __free_slab(struct kmem_cache *s, struct page *page)
1830 int order = compound_order(page);
1831 int pages = 1 << order;
1833 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
1836 slab_pad_check(s, page);
1837 for_each_object(p, s, page_address(page),
1839 check_object(s, page, p, SLUB_RED_INACTIVE);
1842 __ClearPageSlabPfmemalloc(page);
1843 __ClearPageSlab(page);
1845 page->mapping = NULL;
1846 if (current->reclaim_state)
1847 current->reclaim_state->reclaimed_slab += pages;
1848 unaccount_slab_page(page, order, s);
1849 __free_pages(page, order);
1852 static void rcu_free_slab(struct rcu_head *h)
1854 struct page *page = container_of(h, struct page, rcu_head);
1856 __free_slab(page->slab_cache, page);
1859 static void free_slab(struct kmem_cache *s, struct page *page)
1861 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1862 call_rcu(&page->rcu_head, rcu_free_slab);
1864 __free_slab(s, page);
1867 static void discard_slab(struct kmem_cache *s, struct page *page)
1869 dec_slabs_node(s, page_to_nid(page), page->objects);
1874 * Management of partially allocated slabs.
1877 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1880 if (tail == DEACTIVATE_TO_TAIL)
1881 list_add_tail(&page->slab_list, &n->partial);
1883 list_add(&page->slab_list, &n->partial);
1886 static inline void add_partial(struct kmem_cache_node *n,
1887 struct page *page, int tail)
1889 lockdep_assert_held(&n->list_lock);
1890 __add_partial(n, page, tail);
1893 static inline void remove_partial(struct kmem_cache_node *n,
1896 lockdep_assert_held(&n->list_lock);
1897 list_del(&page->slab_list);
1902 * Remove slab from the partial list, freeze it and
1903 * return the pointer to the freelist.
1905 * Returns a list of objects or NULL if it fails.
1907 static inline void *acquire_slab(struct kmem_cache *s,
1908 struct kmem_cache_node *n, struct page *page,
1909 int mode, int *objects)
1912 unsigned long counters;
1915 lockdep_assert_held(&n->list_lock);
1918 * Zap the freelist and set the frozen bit.
1919 * The old freelist is the list of objects for the
1920 * per cpu allocation list.
1922 freelist = page->freelist;
1923 counters = page->counters;
1924 new.counters = counters;
1925 *objects = new.objects - new.inuse;
1927 new.inuse = page->objects;
1928 new.freelist = NULL;
1930 new.freelist = freelist;
1933 VM_BUG_ON(new.frozen);
1936 if (!__cmpxchg_double_slab(s, page,
1938 new.freelist, new.counters,
1942 remove_partial(n, page);
1947 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1948 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1951 * Try to allocate a partial slab from a specific node.
1953 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1954 struct kmem_cache_cpu *c, gfp_t flags)
1956 struct page *page, *page2;
1957 void *object = NULL;
1958 unsigned int available = 0;
1962 * Racy check. If we mistakenly see no partial slabs then we
1963 * just allocate an empty slab. If we mistakenly try to get a
1964 * partial slab and there is none available then get_partial()
1967 if (!n || !n->nr_partial)
1970 spin_lock(&n->list_lock);
1971 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1974 if (!pfmemalloc_match(page, flags))
1977 t = acquire_slab(s, n, page, object == NULL, &objects);
1981 available += objects;
1984 stat(s, ALLOC_FROM_PARTIAL);
1987 put_cpu_partial(s, page, 0);
1988 stat(s, CPU_PARTIAL_NODE);
1990 if (!kmem_cache_has_cpu_partial(s)
1991 || available > slub_cpu_partial(s) / 2)
1995 spin_unlock(&n->list_lock);
2000 * Get a page from somewhere. Search in increasing NUMA distances.
2002 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2003 struct kmem_cache_cpu *c)
2006 struct zonelist *zonelist;
2009 enum zone_type highest_zoneidx = gfp_zone(flags);
2011 unsigned int cpuset_mems_cookie;
2014 * The defrag ratio allows a configuration of the tradeoffs between
2015 * inter node defragmentation and node local allocations. A lower
2016 * defrag_ratio increases the tendency to do local allocations
2017 * instead of attempting to obtain partial slabs from other nodes.
2019 * If the defrag_ratio is set to 0 then kmalloc() always
2020 * returns node local objects. If the ratio is higher then kmalloc()
2021 * may return off node objects because partial slabs are obtained
2022 * from other nodes and filled up.
2024 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2025 * (which makes defrag_ratio = 1000) then every (well almost)
2026 * allocation will first attempt to defrag slab caches on other nodes.
2027 * This means scanning over all nodes to look for partial slabs which
2028 * may be expensive if we do it every time we are trying to find a slab
2029 * with available objects.
2031 if (!s->remote_node_defrag_ratio ||
2032 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2036 cpuset_mems_cookie = read_mems_allowed_begin();
2037 zonelist = node_zonelist(mempolicy_slab_node(), flags);
2038 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2039 struct kmem_cache_node *n;
2041 n = get_node(s, zone_to_nid(zone));
2043 if (n && cpuset_zone_allowed(zone, flags) &&
2044 n->nr_partial > s->min_partial) {
2045 object = get_partial_node(s, n, c, flags);
2048 * Don't check read_mems_allowed_retry()
2049 * here - if mems_allowed was updated in
2050 * parallel, that was a harmless race
2051 * between allocation and the cpuset
2058 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2059 #endif /* CONFIG_NUMA */
2064 * Get a partial page, lock it and return it.
2066 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2067 struct kmem_cache_cpu *c)
2070 int searchnode = node;
2072 if (node == NUMA_NO_NODE)
2073 searchnode = numa_mem_id();
2075 object = get_partial_node(s, get_node(s, searchnode), c, flags);
2076 if (object || node != NUMA_NO_NODE)
2079 return get_any_partial(s, flags, c);
2082 #ifdef CONFIG_PREEMPTION
2084 * Calculate the next globally unique transaction for disambiguation
2085 * during cmpxchg. The transactions start with the cpu number and are then
2086 * incremented by CONFIG_NR_CPUS.
2088 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2091 * No preemption supported therefore also no need to check for
2097 static inline unsigned long next_tid(unsigned long tid)
2099 return tid + TID_STEP;
2102 #ifdef SLUB_DEBUG_CMPXCHG
2103 static inline unsigned int tid_to_cpu(unsigned long tid)
2105 return tid % TID_STEP;
2108 static inline unsigned long tid_to_event(unsigned long tid)
2110 return tid / TID_STEP;
2114 static inline unsigned int init_tid(int cpu)
2119 static inline void note_cmpxchg_failure(const char *n,
2120 const struct kmem_cache *s, unsigned long tid)
2122 #ifdef SLUB_DEBUG_CMPXCHG
2123 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2125 pr_info("%s %s: cmpxchg redo ", n, s->name);
2127 #ifdef CONFIG_PREEMPTION
2128 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2129 pr_warn("due to cpu change %d -> %d\n",
2130 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2133 if (tid_to_event(tid) != tid_to_event(actual_tid))
2134 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2135 tid_to_event(tid), tid_to_event(actual_tid));
2137 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2138 actual_tid, tid, next_tid(tid));
2140 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2143 static void init_kmem_cache_cpus(struct kmem_cache *s)
2147 for_each_possible_cpu(cpu)
2148 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2152 * Remove the cpu slab
2154 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2155 void *freelist, struct kmem_cache_cpu *c)
2157 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2158 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2160 enum slab_modes l = M_NONE, m = M_NONE;
2162 int tail = DEACTIVATE_TO_HEAD;
2166 if (page->freelist) {
2167 stat(s, DEACTIVATE_REMOTE_FREES);
2168 tail = DEACTIVATE_TO_TAIL;
2172 * Stage one: Free all available per cpu objects back
2173 * to the page freelist while it is still frozen. Leave the
2176 * There is no need to take the list->lock because the page
2179 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2181 unsigned long counters;
2184 * If 'nextfree' is invalid, it is possible that the object at
2185 * 'freelist' is already corrupted. So isolate all objects
2186 * starting at 'freelist'.
2188 if (freelist_corrupted(s, page, &freelist, nextfree))
2192 prior = page->freelist;
2193 counters = page->counters;
2194 set_freepointer(s, freelist, prior);
2195 new.counters = counters;
2197 VM_BUG_ON(!new.frozen);
2199 } while (!__cmpxchg_double_slab(s, page,
2201 freelist, new.counters,
2202 "drain percpu freelist"));
2204 freelist = nextfree;
2208 * Stage two: Ensure that the page is unfrozen while the
2209 * list presence reflects the actual number of objects
2212 * We setup the list membership and then perform a cmpxchg
2213 * with the count. If there is a mismatch then the page
2214 * is not unfrozen but the page is on the wrong list.
2216 * Then we restart the process which may have to remove
2217 * the page from the list that we just put it on again
2218 * because the number of objects in the slab may have
2223 old.freelist = page->freelist;
2224 old.counters = page->counters;
2225 VM_BUG_ON(!old.frozen);
2227 /* Determine target state of the slab */
2228 new.counters = old.counters;
2231 set_freepointer(s, freelist, old.freelist);
2232 new.freelist = freelist;
2234 new.freelist = old.freelist;
2238 if (!new.inuse && n->nr_partial >= s->min_partial)
2240 else if (new.freelist) {
2245 * Taking the spinlock removes the possibility
2246 * that acquire_slab() will see a slab page that
2249 spin_lock(&n->list_lock);
2253 #ifdef CONFIG_SLUB_DEBUG
2254 if ((s->flags & SLAB_STORE_USER) && !lock) {
2257 * This also ensures that the scanning of full
2258 * slabs from diagnostic functions will not see
2261 spin_lock(&n->list_lock);
2268 remove_partial(n, page);
2269 else if (l == M_FULL)
2270 remove_full(s, n, page);
2273 add_partial(n, page, tail);
2274 else if (m == M_FULL)
2275 add_full(s, n, page);
2279 if (!__cmpxchg_double_slab(s, page,
2280 old.freelist, old.counters,
2281 new.freelist, new.counters,
2286 spin_unlock(&n->list_lock);
2290 else if (m == M_FULL)
2291 stat(s, DEACTIVATE_FULL);
2292 else if (m == M_FREE) {
2293 stat(s, DEACTIVATE_EMPTY);
2294 discard_slab(s, page);
2300 c->tid = next_tid(c->tid);
2304 * Unfreeze all the cpu partial slabs.
2306 * This function must be called with interrupts disabled
2307 * for the cpu using c (or some other guarantee must be there
2308 * to guarantee no concurrent accesses).
2310 static void unfreeze_partials(struct kmem_cache *s,
2311 struct kmem_cache_cpu *c)
2313 #ifdef CONFIG_SLUB_CPU_PARTIAL
2314 struct kmem_cache_node *n = NULL, *n2 = NULL;
2315 struct page *page, *discard_page = NULL;
2317 while ((page = slub_percpu_partial(c))) {
2321 slub_set_percpu_partial(c, page);
2323 n2 = get_node(s, page_to_nid(page));
2326 spin_unlock(&n->list_lock);
2329 spin_lock(&n->list_lock);
2334 old.freelist = page->freelist;
2335 old.counters = page->counters;
2336 VM_BUG_ON(!old.frozen);
2338 new.counters = old.counters;
2339 new.freelist = old.freelist;
2343 } while (!__cmpxchg_double_slab(s, page,
2344 old.freelist, old.counters,
2345 new.freelist, new.counters,
2346 "unfreezing slab"));
2348 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2349 page->next = discard_page;
2350 discard_page = page;
2352 add_partial(n, page, DEACTIVATE_TO_TAIL);
2353 stat(s, FREE_ADD_PARTIAL);
2358 spin_unlock(&n->list_lock);
2360 while (discard_page) {
2361 page = discard_page;
2362 discard_page = discard_page->next;
2364 stat(s, DEACTIVATE_EMPTY);
2365 discard_slab(s, page);
2368 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2372 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2373 * partial page slot if available.
2375 * If we did not find a slot then simply move all the partials to the
2376 * per node partial list.
2378 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2380 #ifdef CONFIG_SLUB_CPU_PARTIAL
2381 struct page *oldpage;
2389 oldpage = this_cpu_read(s->cpu_slab->partial);
2392 pobjects = oldpage->pobjects;
2393 pages = oldpage->pages;
2394 if (drain && pobjects > slub_cpu_partial(s)) {
2395 unsigned long flags;
2397 * partial array is full. Move the existing
2398 * set to the per node partial list.
2400 local_irq_save(flags);
2401 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2402 local_irq_restore(flags);
2406 stat(s, CPU_PARTIAL_DRAIN);
2411 pobjects += page->objects - page->inuse;
2413 page->pages = pages;
2414 page->pobjects = pobjects;
2415 page->next = oldpage;
2417 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2419 if (unlikely(!slub_cpu_partial(s))) {
2420 unsigned long flags;
2422 local_irq_save(flags);
2423 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2424 local_irq_restore(flags);
2427 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2430 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2432 stat(s, CPUSLAB_FLUSH);
2433 deactivate_slab(s, c->page, c->freelist, c);
2439 * Called from IPI handler with interrupts disabled.
2441 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2443 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2448 unfreeze_partials(s, c);
2451 static void flush_cpu_slab(void *d)
2453 struct kmem_cache *s = d;
2455 __flush_cpu_slab(s, smp_processor_id());
2458 static bool has_cpu_slab(int cpu, void *info)
2460 struct kmem_cache *s = info;
2461 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2463 return c->page || slub_percpu_partial(c);
2466 static void flush_all(struct kmem_cache *s)
2468 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1);
2472 * Use the cpu notifier to insure that the cpu slabs are flushed when
2475 static int slub_cpu_dead(unsigned int cpu)
2477 struct kmem_cache *s;
2478 unsigned long flags;
2480 mutex_lock(&slab_mutex);
2481 list_for_each_entry(s, &slab_caches, list) {
2482 local_irq_save(flags);
2483 __flush_cpu_slab(s, cpu);
2484 local_irq_restore(flags);
2486 mutex_unlock(&slab_mutex);
2491 * Check if the objects in a per cpu structure fit numa
2492 * locality expectations.
2494 static inline int node_match(struct page *page, int node)
2497 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2503 #ifdef CONFIG_SLUB_DEBUG
2504 static int count_free(struct page *page)
2506 return page->objects - page->inuse;
2509 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2511 return atomic_long_read(&n->total_objects);
2513 #endif /* CONFIG_SLUB_DEBUG */
2515 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2516 static unsigned long count_partial(struct kmem_cache_node *n,
2517 int (*get_count)(struct page *))
2519 unsigned long flags;
2520 unsigned long x = 0;
2523 spin_lock_irqsave(&n->list_lock, flags);
2524 list_for_each_entry(page, &n->partial, slab_list)
2525 x += get_count(page);
2526 spin_unlock_irqrestore(&n->list_lock, flags);
2529 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2531 static noinline void
2532 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2534 #ifdef CONFIG_SLUB_DEBUG
2535 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2536 DEFAULT_RATELIMIT_BURST);
2538 struct kmem_cache_node *n;
2540 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2543 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2544 nid, gfpflags, &gfpflags);
2545 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2546 s->name, s->object_size, s->size, oo_order(s->oo),
2549 if (oo_order(s->min) > get_order(s->object_size))
2550 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2553 for_each_kmem_cache_node(s, node, n) {
2554 unsigned long nr_slabs;
2555 unsigned long nr_objs;
2556 unsigned long nr_free;
2558 nr_free = count_partial(n, count_free);
2559 nr_slabs = node_nr_slabs(n);
2560 nr_objs = node_nr_objs(n);
2562 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2563 node, nr_slabs, nr_objs, nr_free);
2568 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2569 int node, struct kmem_cache_cpu **pc)
2572 struct kmem_cache_cpu *c = *pc;
2575 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2577 freelist = get_partial(s, flags, node, c);
2582 page = new_slab(s, flags, node);
2584 c = raw_cpu_ptr(s->cpu_slab);
2589 * No other reference to the page yet so we can
2590 * muck around with it freely without cmpxchg
2592 freelist = page->freelist;
2593 page->freelist = NULL;
2595 stat(s, ALLOC_SLAB);
2603 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2605 if (unlikely(PageSlabPfmemalloc(page)))
2606 return gfp_pfmemalloc_allowed(gfpflags);
2612 * Check the page->freelist of a page and either transfer the freelist to the
2613 * per cpu freelist or deactivate the page.
2615 * The page is still frozen if the return value is not NULL.
2617 * If this function returns NULL then the page has been unfrozen.
2619 * This function must be called with interrupt disabled.
2621 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2624 unsigned long counters;
2628 freelist = page->freelist;
2629 counters = page->counters;
2631 new.counters = counters;
2632 VM_BUG_ON(!new.frozen);
2634 new.inuse = page->objects;
2635 new.frozen = freelist != NULL;
2637 } while (!__cmpxchg_double_slab(s, page,
2646 * Slow path. The lockless freelist is empty or we need to perform
2649 * Processing is still very fast if new objects have been freed to the
2650 * regular freelist. In that case we simply take over the regular freelist
2651 * as the lockless freelist and zap the regular freelist.
2653 * If that is not working then we fall back to the partial lists. We take the
2654 * first element of the freelist as the object to allocate now and move the
2655 * rest of the freelist to the lockless freelist.
2657 * And if we were unable to get a new slab from the partial slab lists then
2658 * we need to allocate a new slab. This is the slowest path since it involves
2659 * a call to the page allocator and the setup of a new slab.
2661 * Version of __slab_alloc to use when we know that interrupts are
2662 * already disabled (which is the case for bulk allocation).
2664 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2665 unsigned long addr, struct kmem_cache_cpu *c)
2670 stat(s, ALLOC_SLOWPATH);
2675 * if the node is not online or has no normal memory, just
2676 * ignore the node constraint
2678 if (unlikely(node != NUMA_NO_NODE &&
2679 !node_state(node, N_NORMAL_MEMORY)))
2680 node = NUMA_NO_NODE;
2685 if (unlikely(!node_match(page, node))) {
2687 * same as above but node_match() being false already
2688 * implies node != NUMA_NO_NODE
2690 if (!node_state(node, N_NORMAL_MEMORY)) {
2691 node = NUMA_NO_NODE;
2694 stat(s, ALLOC_NODE_MISMATCH);
2695 deactivate_slab(s, page, c->freelist, c);
2701 * By rights, we should be searching for a slab page that was
2702 * PFMEMALLOC but right now, we are losing the pfmemalloc
2703 * information when the page leaves the per-cpu allocator
2705 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2706 deactivate_slab(s, page, c->freelist, c);
2710 /* must check again c->freelist in case of cpu migration or IRQ */
2711 freelist = c->freelist;
2715 freelist = get_freelist(s, page);
2719 c->tid = next_tid(c->tid);
2720 stat(s, DEACTIVATE_BYPASS);
2724 stat(s, ALLOC_REFILL);
2728 * freelist is pointing to the list of objects to be used.
2729 * page is pointing to the page from which the objects are obtained.
2730 * That page must be frozen for per cpu allocations to work.
2732 VM_BUG_ON(!c->page->frozen);
2733 c->freelist = get_freepointer(s, freelist);
2734 c->tid = next_tid(c->tid);
2739 if (slub_percpu_partial(c)) {
2740 page = c->page = slub_percpu_partial(c);
2741 slub_set_percpu_partial(c, page);
2742 stat(s, CPU_PARTIAL_ALLOC);
2746 freelist = new_slab_objects(s, gfpflags, node, &c);
2748 if (unlikely(!freelist)) {
2749 slab_out_of_memory(s, gfpflags, node);
2754 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2757 /* Only entered in the debug case */
2758 if (kmem_cache_debug(s) &&
2759 !alloc_debug_processing(s, page, freelist, addr))
2760 goto new_slab; /* Slab failed checks. Next slab needed */
2762 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2767 * Another one that disabled interrupt and compensates for possible
2768 * cpu changes by refetching the per cpu area pointer.
2770 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2771 unsigned long addr, struct kmem_cache_cpu *c)
2774 unsigned long flags;
2776 local_irq_save(flags);
2777 #ifdef CONFIG_PREEMPTION
2779 * We may have been preempted and rescheduled on a different
2780 * cpu before disabling interrupts. Need to reload cpu area
2783 c = this_cpu_ptr(s->cpu_slab);
2786 p = ___slab_alloc(s, gfpflags, node, addr, c);
2787 local_irq_restore(flags);
2792 * If the object has been wiped upon free, make sure it's fully initialized by
2793 * zeroing out freelist pointer.
2795 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2798 if (unlikely(slab_want_init_on_free(s)) && obj)
2799 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2803 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2804 * have the fastpath folded into their functions. So no function call
2805 * overhead for requests that can be satisfied on the fastpath.
2807 * The fastpath works by first checking if the lockless freelist can be used.
2808 * If not then __slab_alloc is called for slow processing.
2810 * Otherwise we can simply pick the next object from the lockless free list.
2812 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2813 gfp_t gfpflags, int node, unsigned long addr)
2816 struct kmem_cache_cpu *c;
2819 struct obj_cgroup *objcg = NULL;
2821 s = slab_pre_alloc_hook(s, &objcg, 1, gfpflags);
2826 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2827 * enabled. We may switch back and forth between cpus while
2828 * reading from one cpu area. That does not matter as long
2829 * as we end up on the original cpu again when doing the cmpxchg.
2831 * We should guarantee that tid and kmem_cache are retrieved on
2832 * the same cpu. It could be different if CONFIG_PREEMPTION so we need
2833 * to check if it is matched or not.
2836 tid = this_cpu_read(s->cpu_slab->tid);
2837 c = raw_cpu_ptr(s->cpu_slab);
2838 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
2839 unlikely(tid != READ_ONCE(c->tid)));
2842 * Irqless object alloc/free algorithm used here depends on sequence
2843 * of fetching cpu_slab's data. tid should be fetched before anything
2844 * on c to guarantee that object and page associated with previous tid
2845 * won't be used with current tid. If we fetch tid first, object and
2846 * page could be one associated with next tid and our alloc/free
2847 * request will be failed. In this case, we will retry. So, no problem.
2852 * The transaction ids are globally unique per cpu and per operation on
2853 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2854 * occurs on the right processor and that there was no operation on the
2855 * linked list in between.
2858 object = c->freelist;
2860 if (unlikely(!object || !page || !node_match(page, node))) {
2861 object = __slab_alloc(s, gfpflags, node, addr, c);
2863 void *next_object = get_freepointer_safe(s, object);
2866 * The cmpxchg will only match if there was no additional
2867 * operation and if we are on the right processor.
2869 * The cmpxchg does the following atomically (without lock
2871 * 1. Relocate first pointer to the current per cpu area.
2872 * 2. Verify that tid and freelist have not been changed
2873 * 3. If they were not changed replace tid and freelist
2875 * Since this is without lock semantics the protection is only
2876 * against code executing on this cpu *not* from access by
2879 if (unlikely(!this_cpu_cmpxchg_double(
2880 s->cpu_slab->freelist, s->cpu_slab->tid,
2882 next_object, next_tid(tid)))) {
2884 note_cmpxchg_failure("slab_alloc", s, tid);
2887 prefetch_freepointer(s, next_object);
2888 stat(s, ALLOC_FASTPATH);
2891 maybe_wipe_obj_freeptr(s, object);
2893 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2894 memset(object, 0, s->object_size);
2896 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object);
2901 static __always_inline void *slab_alloc(struct kmem_cache *s,
2902 gfp_t gfpflags, unsigned long addr)
2904 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2907 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2909 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2911 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2916 EXPORT_SYMBOL(kmem_cache_alloc);
2918 #ifdef CONFIG_TRACING
2919 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2921 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2922 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2923 ret = kasan_kmalloc(s, ret, size, gfpflags);
2926 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2930 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2932 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2934 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2935 s->object_size, s->size, gfpflags, node);
2939 EXPORT_SYMBOL(kmem_cache_alloc_node);
2941 #ifdef CONFIG_TRACING
2942 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2944 int node, size_t size)
2946 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2948 trace_kmalloc_node(_RET_IP_, ret,
2949 size, s->size, gfpflags, node);
2951 ret = kasan_kmalloc(s, ret, size, gfpflags);
2954 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2956 #endif /* CONFIG_NUMA */
2959 * Slow path handling. This may still be called frequently since objects
2960 * have a longer lifetime than the cpu slabs in most processing loads.
2962 * So we still attempt to reduce cache line usage. Just take the slab
2963 * lock and free the item. If there is no additional partial page
2964 * handling required then we can return immediately.
2966 static void __slab_free(struct kmem_cache *s, struct page *page,
2967 void *head, void *tail, int cnt,
2974 unsigned long counters;
2975 struct kmem_cache_node *n = NULL;
2976 unsigned long flags;
2978 stat(s, FREE_SLOWPATH);
2980 if (kmem_cache_debug(s) &&
2981 !free_debug_processing(s, page, head, tail, cnt, addr))
2986 spin_unlock_irqrestore(&n->list_lock, flags);
2989 prior = page->freelist;
2990 counters = page->counters;
2991 set_freepointer(s, tail, prior);
2992 new.counters = counters;
2993 was_frozen = new.frozen;
2995 if ((!new.inuse || !prior) && !was_frozen) {
2997 if (kmem_cache_has_cpu_partial(s) && !prior) {
3000 * Slab was on no list before and will be
3002 * We can defer the list move and instead
3007 } else { /* Needs to be taken off a list */
3009 n = get_node(s, page_to_nid(page));
3011 * Speculatively acquire the list_lock.
3012 * If the cmpxchg does not succeed then we may
3013 * drop the list_lock without any processing.
3015 * Otherwise the list_lock will synchronize with
3016 * other processors updating the list of slabs.
3018 spin_lock_irqsave(&n->list_lock, flags);
3023 } while (!cmpxchg_double_slab(s, page,
3030 if (likely(was_frozen)) {
3032 * The list lock was not taken therefore no list
3033 * activity can be necessary.
3035 stat(s, FREE_FROZEN);
3036 } else if (new.frozen) {
3038 * If we just froze the page then put it onto the
3039 * per cpu partial list.
3041 put_cpu_partial(s, page, 1);
3042 stat(s, CPU_PARTIAL_FREE);
3048 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3052 * Objects left in the slab. If it was not on the partial list before
3055 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3056 remove_full(s, n, page);
3057 add_partial(n, page, DEACTIVATE_TO_TAIL);
3058 stat(s, FREE_ADD_PARTIAL);
3060 spin_unlock_irqrestore(&n->list_lock, flags);
3066 * Slab on the partial list.
3068 remove_partial(n, page);
3069 stat(s, FREE_REMOVE_PARTIAL);
3071 /* Slab must be on the full list */
3072 remove_full(s, n, page);
3075 spin_unlock_irqrestore(&n->list_lock, flags);
3077 discard_slab(s, page);
3081 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3082 * can perform fastpath freeing without additional function calls.
3084 * The fastpath is only possible if we are freeing to the current cpu slab
3085 * of this processor. This typically the case if we have just allocated
3088 * If fastpath is not possible then fall back to __slab_free where we deal
3089 * with all sorts of special processing.
3091 * Bulk free of a freelist with several objects (all pointing to the
3092 * same page) possible by specifying head and tail ptr, plus objects
3093 * count (cnt). Bulk free indicated by tail pointer being set.
3095 static __always_inline void do_slab_free(struct kmem_cache *s,
3096 struct page *page, void *head, void *tail,
3097 int cnt, unsigned long addr)
3099 void *tail_obj = tail ? : head;
3100 struct kmem_cache_cpu *c;
3103 /* memcg_slab_free_hook() is already called for bulk free. */
3105 memcg_slab_free_hook(s, &head, 1);
3108 * Determine the currently cpus per cpu slab.
3109 * The cpu may change afterward. However that does not matter since
3110 * data is retrieved via this pointer. If we are on the same cpu
3111 * during the cmpxchg then the free will succeed.
3114 tid = this_cpu_read(s->cpu_slab->tid);
3115 c = raw_cpu_ptr(s->cpu_slab);
3116 } while (IS_ENABLED(CONFIG_PREEMPTION) &&
3117 unlikely(tid != READ_ONCE(c->tid)));
3119 /* Same with comment on barrier() in slab_alloc_node() */
3122 if (likely(page == c->page)) {
3123 void **freelist = READ_ONCE(c->freelist);
3125 set_freepointer(s, tail_obj, freelist);
3127 if (unlikely(!this_cpu_cmpxchg_double(
3128 s->cpu_slab->freelist, s->cpu_slab->tid,
3130 head, next_tid(tid)))) {
3132 note_cmpxchg_failure("slab_free", s, tid);
3135 stat(s, FREE_FASTPATH);
3137 __slab_free(s, page, head, tail_obj, cnt, addr);
3141 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3142 void *head, void *tail, int cnt,
3146 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3147 * to remove objects, whose reuse must be delayed.
3149 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3150 do_slab_free(s, page, head, tail, cnt, addr);
3153 #ifdef CONFIG_KASAN_GENERIC
3154 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3156 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3160 void kmem_cache_free(struct kmem_cache *s, void *x)
3162 s = cache_from_obj(s, x);
3165 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3166 trace_kmem_cache_free(_RET_IP_, x);
3168 EXPORT_SYMBOL(kmem_cache_free);
3170 struct detached_freelist {
3175 struct kmem_cache *s;
3179 * This function progressively scans the array with free objects (with
3180 * a limited look ahead) and extract objects belonging to the same
3181 * page. It builds a detached freelist directly within the given
3182 * page/objects. This can happen without any need for
3183 * synchronization, because the objects are owned by running process.
3184 * The freelist is build up as a single linked list in the objects.
3185 * The idea is, that this detached freelist can then be bulk
3186 * transferred to the real freelist(s), but only requiring a single
3187 * synchronization primitive. Look ahead in the array is limited due
3188 * to performance reasons.
3191 int build_detached_freelist(struct kmem_cache *s, size_t size,
3192 void **p, struct detached_freelist *df)
3194 size_t first_skipped_index = 0;
3199 /* Always re-init detached_freelist */
3204 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3205 } while (!object && size);
3210 page = virt_to_head_page(object);
3212 /* Handle kalloc'ed objects */
3213 if (unlikely(!PageSlab(page))) {
3214 BUG_ON(!PageCompound(page));
3216 __free_pages(page, compound_order(page));
3217 p[size] = NULL; /* mark object processed */
3220 /* Derive kmem_cache from object */
3221 df->s = page->slab_cache;
3223 df->s = cache_from_obj(s, object); /* Support for memcg */
3226 /* Start new detached freelist */
3228 set_freepointer(df->s, object, NULL);
3230 df->freelist = object;
3231 p[size] = NULL; /* mark object processed */
3237 continue; /* Skip processed objects */
3239 /* df->page is always set at this point */
3240 if (df->page == virt_to_head_page(object)) {
3241 /* Opportunity build freelist */
3242 set_freepointer(df->s, object, df->freelist);
3243 df->freelist = object;
3245 p[size] = NULL; /* mark object processed */
3250 /* Limit look ahead search */
3254 if (!first_skipped_index)
3255 first_skipped_index = size + 1;
3258 return first_skipped_index;
3261 /* Note that interrupts must be enabled when calling this function. */
3262 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3267 memcg_slab_free_hook(s, p, size);
3269 struct detached_freelist df;
3271 size = build_detached_freelist(s, size, p, &df);
3275 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3276 } while (likely(size));
3278 EXPORT_SYMBOL(kmem_cache_free_bulk);
3280 /* Note that interrupts must be enabled when calling this function. */
3281 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3284 struct kmem_cache_cpu *c;
3286 struct obj_cgroup *objcg = NULL;
3288 /* memcg and kmem_cache debug support */
3289 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3293 * Drain objects in the per cpu slab, while disabling local
3294 * IRQs, which protects against PREEMPT and interrupts
3295 * handlers invoking normal fastpath.
3297 local_irq_disable();
3298 c = this_cpu_ptr(s->cpu_slab);
3300 for (i = 0; i < size; i++) {
3301 void *object = c->freelist;
3303 if (unlikely(!object)) {
3305 * We may have removed an object from c->freelist using
3306 * the fastpath in the previous iteration; in that case,
3307 * c->tid has not been bumped yet.
3308 * Since ___slab_alloc() may reenable interrupts while
3309 * allocating memory, we should bump c->tid now.
3311 c->tid = next_tid(c->tid);
3314 * Invoking slow path likely have side-effect
3315 * of re-populating per CPU c->freelist
3317 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3319 if (unlikely(!p[i]))
3322 c = this_cpu_ptr(s->cpu_slab);
3323 maybe_wipe_obj_freeptr(s, p[i]);
3325 continue; /* goto for-loop */
3327 c->freelist = get_freepointer(s, object);
3329 maybe_wipe_obj_freeptr(s, p[i]);
3331 c->tid = next_tid(c->tid);
3334 /* Clear memory outside IRQ disabled fastpath loop */
3335 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3338 for (j = 0; j < i; j++)
3339 memset(p[j], 0, s->object_size);
3342 /* memcg and kmem_cache debug support */
3343 slab_post_alloc_hook(s, objcg, flags, size, p);
3347 slab_post_alloc_hook(s, objcg, flags, i, p);
3348 __kmem_cache_free_bulk(s, i, p);
3351 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3355 * Object placement in a slab is made very easy because we always start at
3356 * offset 0. If we tune the size of the object to the alignment then we can
3357 * get the required alignment by putting one properly sized object after
3360 * Notice that the allocation order determines the sizes of the per cpu
3361 * caches. Each processor has always one slab available for allocations.
3362 * Increasing the allocation order reduces the number of times that slabs
3363 * must be moved on and off the partial lists and is therefore a factor in
3368 * Mininum / Maximum order of slab pages. This influences locking overhead
3369 * and slab fragmentation. A higher order reduces the number of partial slabs
3370 * and increases the number of allocations possible without having to
3371 * take the list_lock.
3373 static unsigned int slub_min_order;
3374 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3375 static unsigned int slub_min_objects;
3378 * Calculate the order of allocation given an slab object size.
3380 * The order of allocation has significant impact on performance and other
3381 * system components. Generally order 0 allocations should be preferred since
3382 * order 0 does not cause fragmentation in the page allocator. Larger objects
3383 * be problematic to put into order 0 slabs because there may be too much
3384 * unused space left. We go to a higher order if more than 1/16th of the slab
3387 * In order to reach satisfactory performance we must ensure that a minimum
3388 * number of objects is in one slab. Otherwise we may generate too much
3389 * activity on the partial lists which requires taking the list_lock. This is
3390 * less a concern for large slabs though which are rarely used.
3392 * slub_max_order specifies the order where we begin to stop considering the
3393 * number of objects in a slab as critical. If we reach slub_max_order then
3394 * we try to keep the page order as low as possible. So we accept more waste
3395 * of space in favor of a small page order.
3397 * Higher order allocations also allow the placement of more objects in a
3398 * slab and thereby reduce object handling overhead. If the user has
3399 * requested a higher mininum order then we start with that one instead of
3400 * the smallest order which will fit the object.
3402 static inline unsigned int slab_order(unsigned int size,
3403 unsigned int min_objects, unsigned int max_order,
3404 unsigned int fract_leftover)
3406 unsigned int min_order = slub_min_order;
3409 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3410 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3412 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3413 order <= max_order; order++) {
3415 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3418 rem = slab_size % size;
3420 if (rem <= slab_size / fract_leftover)
3427 static inline int calculate_order(unsigned int size)
3430 unsigned int min_objects;
3431 unsigned int max_objects;
3434 * Attempt to find best configuration for a slab. This
3435 * works by first attempting to generate a layout with
3436 * the best configuration and backing off gradually.
3438 * First we increase the acceptable waste in a slab. Then
3439 * we reduce the minimum objects required in a slab.
3441 min_objects = slub_min_objects;
3443 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3444 max_objects = order_objects(slub_max_order, size);
3445 min_objects = min(min_objects, max_objects);
3447 while (min_objects > 1) {
3448 unsigned int fraction;
3451 while (fraction >= 4) {
3452 order = slab_order(size, min_objects,
3453 slub_max_order, fraction);
3454 if (order <= slub_max_order)
3462 * We were unable to place multiple objects in a slab. Now
3463 * lets see if we can place a single object there.
3465 order = slab_order(size, 1, slub_max_order, 1);
3466 if (order <= slub_max_order)
3470 * Doh this slab cannot be placed using slub_max_order.
3472 order = slab_order(size, 1, MAX_ORDER, 1);
3473 if (order < MAX_ORDER)
3479 init_kmem_cache_node(struct kmem_cache_node *n)
3482 spin_lock_init(&n->list_lock);
3483 INIT_LIST_HEAD(&n->partial);
3484 #ifdef CONFIG_SLUB_DEBUG
3485 atomic_long_set(&n->nr_slabs, 0);
3486 atomic_long_set(&n->total_objects, 0);
3487 INIT_LIST_HEAD(&n->full);
3491 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3493 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3494 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3497 * Must align to double word boundary for the double cmpxchg
3498 * instructions to work; see __pcpu_double_call_return_bool().
3500 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3501 2 * sizeof(void *));
3506 init_kmem_cache_cpus(s);
3511 static struct kmem_cache *kmem_cache_node;
3514 * No kmalloc_node yet so do it by hand. We know that this is the first
3515 * slab on the node for this slabcache. There are no concurrent accesses
3518 * Note that this function only works on the kmem_cache_node
3519 * when allocating for the kmem_cache_node. This is used for bootstrapping
3520 * memory on a fresh node that has no slab structures yet.
3522 static void early_kmem_cache_node_alloc(int node)
3525 struct kmem_cache_node *n;
3527 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3529 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3532 if (page_to_nid(page) != node) {
3533 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3534 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3539 #ifdef CONFIG_SLUB_DEBUG
3540 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3541 init_tracking(kmem_cache_node, n);
3543 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3545 page->freelist = get_freepointer(kmem_cache_node, n);
3548 kmem_cache_node->node[node] = n;
3549 init_kmem_cache_node(n);
3550 inc_slabs_node(kmem_cache_node, node, page->objects);
3553 * No locks need to be taken here as it has just been
3554 * initialized and there is no concurrent access.
3556 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3559 static void free_kmem_cache_nodes(struct kmem_cache *s)
3562 struct kmem_cache_node *n;
3564 for_each_kmem_cache_node(s, node, n) {
3565 s->node[node] = NULL;
3566 kmem_cache_free(kmem_cache_node, n);
3570 void __kmem_cache_release(struct kmem_cache *s)
3572 cache_random_seq_destroy(s);
3573 free_percpu(s->cpu_slab);
3574 free_kmem_cache_nodes(s);
3577 static int init_kmem_cache_nodes(struct kmem_cache *s)
3581 for_each_node_state(node, N_NORMAL_MEMORY) {
3582 struct kmem_cache_node *n;
3584 if (slab_state == DOWN) {
3585 early_kmem_cache_node_alloc(node);
3588 n = kmem_cache_alloc_node(kmem_cache_node,
3592 free_kmem_cache_nodes(s);
3596 init_kmem_cache_node(n);
3602 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3604 if (min < MIN_PARTIAL)
3606 else if (min > MAX_PARTIAL)
3608 s->min_partial = min;
3611 static void set_cpu_partial(struct kmem_cache *s)
3613 #ifdef CONFIG_SLUB_CPU_PARTIAL
3615 * cpu_partial determined the maximum number of objects kept in the
3616 * per cpu partial lists of a processor.
3618 * Per cpu partial lists mainly contain slabs that just have one
3619 * object freed. If they are used for allocation then they can be
3620 * filled up again with minimal effort. The slab will never hit the
3621 * per node partial lists and therefore no locking will be required.
3623 * This setting also determines
3625 * A) The number of objects from per cpu partial slabs dumped to the
3626 * per node list when we reach the limit.
3627 * B) The number of objects in cpu partial slabs to extract from the
3628 * per node list when we run out of per cpu objects. We only fetch
3629 * 50% to keep some capacity around for frees.
3631 if (!kmem_cache_has_cpu_partial(s))
3632 slub_set_cpu_partial(s, 0);
3633 else if (s->size >= PAGE_SIZE)
3634 slub_set_cpu_partial(s, 2);
3635 else if (s->size >= 1024)
3636 slub_set_cpu_partial(s, 6);
3637 else if (s->size >= 256)
3638 slub_set_cpu_partial(s, 13);
3640 slub_set_cpu_partial(s, 30);
3645 * calculate_sizes() determines the order and the distribution of data within
3648 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3650 slab_flags_t flags = s->flags;
3651 unsigned int size = s->object_size;
3655 * Round up object size to the next word boundary. We can only
3656 * place the free pointer at word boundaries and this determines
3657 * the possible location of the free pointer.
3659 size = ALIGN(size, sizeof(void *));
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 and redzoning.
3689 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3690 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
3693 * Relocate free pointer after the object if it is not
3694 * permitted to overwrite the first word of the object on
3697 * This is the case if we do RCU, have a constructor or
3698 * destructor, are poisoning the objects, or are
3699 * redzoning an object smaller than sizeof(void *).
3701 * The assumption that s->offset >= s->inuse means free
3702 * pointer is outside of the object is used in the
3703 * freeptr_outside_object() function. If that is no
3704 * longer true, the function needs to be modified.
3707 size += sizeof(void *);
3710 * Store freelist pointer near middle of object to keep
3711 * it away from the edges of the object to avoid small
3712 * sized over/underflows from neighboring allocations.
3714 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
3717 #ifdef CONFIG_SLUB_DEBUG
3718 if (flags & SLAB_STORE_USER)
3720 * Need to store information about allocs and frees after
3723 size += 2 * sizeof(struct track);
3726 kasan_cache_create(s, &size, &s->flags);
3727 #ifdef CONFIG_SLUB_DEBUG
3728 if (flags & SLAB_RED_ZONE) {
3730 * Add some empty padding so that we can catch
3731 * overwrites from earlier objects rather than let
3732 * tracking information or the free pointer be
3733 * corrupted if a user writes before the start
3736 size += sizeof(void *);
3738 s->red_left_pad = sizeof(void *);
3739 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3740 size += s->red_left_pad;
3745 * SLUB stores one object immediately after another beginning from
3746 * offset 0. In order to align the objects we have to simply size
3747 * each object to conform to the alignment.
3749 size = ALIGN(size, s->align);
3751 s->reciprocal_size = reciprocal_value(size);
3752 if (forced_order >= 0)
3753 order = forced_order;
3755 order = calculate_order(size);
3762 s->allocflags |= __GFP_COMP;
3764 if (s->flags & SLAB_CACHE_DMA)
3765 s->allocflags |= GFP_DMA;
3767 if (s->flags & SLAB_CACHE_DMA32)
3768 s->allocflags |= GFP_DMA32;
3770 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3771 s->allocflags |= __GFP_RECLAIMABLE;
3774 * Determine the number of objects per slab
3776 s->oo = oo_make(order, size);
3777 s->min = oo_make(get_order(size), size);
3778 if (oo_objects(s->oo) > oo_objects(s->max))
3781 return !!oo_objects(s->oo);
3784 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3786 s->flags = kmem_cache_flags(s->size, flags, s->name);
3787 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3788 s->random = get_random_long();
3791 if (!calculate_sizes(s, -1))
3793 if (disable_higher_order_debug) {
3795 * Disable debugging flags that store metadata if the min slab
3798 if (get_order(s->size) > get_order(s->object_size)) {
3799 s->flags &= ~DEBUG_METADATA_FLAGS;
3801 if (!calculate_sizes(s, -1))
3806 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3807 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3808 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3809 /* Enable fast mode */
3810 s->flags |= __CMPXCHG_DOUBLE;
3814 * The larger the object size is, the more pages we want on the partial
3815 * list to avoid pounding the page allocator excessively.
3817 set_min_partial(s, ilog2(s->size) / 2);
3822 s->remote_node_defrag_ratio = 1000;
3825 /* Initialize the pre-computed randomized freelist if slab is up */
3826 if (slab_state >= UP) {
3827 if (init_cache_random_seq(s))
3831 if (!init_kmem_cache_nodes(s))
3834 if (alloc_kmem_cache_cpus(s))
3838 __kmem_cache_release(s);
3842 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3845 #ifdef CONFIG_SLUB_DEBUG
3846 void *addr = page_address(page);
3850 slab_err(s, page, text, s->name);
3853 map = get_map(s, page);
3854 for_each_object(p, s, addr, page->objects) {
3856 if (!test_bit(__obj_to_index(s, addr, p), map)) {
3857 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3858 print_tracking(s, p);
3867 * Attempt to free all partial slabs on a node.
3868 * This is called from __kmem_cache_shutdown(). We must take list_lock
3869 * because sysfs file might still access partial list after the shutdowning.
3871 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3874 struct page *page, *h;
3876 BUG_ON(irqs_disabled());
3877 spin_lock_irq(&n->list_lock);
3878 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3880 remove_partial(n, page);
3881 list_add(&page->slab_list, &discard);
3883 list_slab_objects(s, page,
3884 "Objects remaining in %s on __kmem_cache_shutdown()");
3887 spin_unlock_irq(&n->list_lock);
3889 list_for_each_entry_safe(page, h, &discard, slab_list)
3890 discard_slab(s, page);
3893 bool __kmem_cache_empty(struct kmem_cache *s)
3896 struct kmem_cache_node *n;
3898 for_each_kmem_cache_node(s, node, n)
3899 if (n->nr_partial || slabs_node(s, node))
3905 * Release all resources used by a slab cache.
3907 int __kmem_cache_shutdown(struct kmem_cache *s)
3910 struct kmem_cache_node *n;
3913 /* Attempt to free all objects */
3914 for_each_kmem_cache_node(s, node, n) {
3916 if (n->nr_partial || slabs_node(s, node))
3922 /********************************************************************
3924 *******************************************************************/
3926 static int __init setup_slub_min_order(char *str)
3928 get_option(&str, (int *)&slub_min_order);
3933 __setup("slub_min_order=", setup_slub_min_order);
3935 static int __init setup_slub_max_order(char *str)
3937 get_option(&str, (int *)&slub_max_order);
3938 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3943 __setup("slub_max_order=", setup_slub_max_order);
3945 static int __init setup_slub_min_objects(char *str)
3947 get_option(&str, (int *)&slub_min_objects);
3952 __setup("slub_min_objects=", setup_slub_min_objects);
3954 void *__kmalloc(size_t size, gfp_t flags)
3956 struct kmem_cache *s;
3959 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3960 return kmalloc_large(size, flags);
3962 s = kmalloc_slab(size, flags);
3964 if (unlikely(ZERO_OR_NULL_PTR(s)))
3967 ret = slab_alloc(s, flags, _RET_IP_);
3969 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3971 ret = kasan_kmalloc(s, ret, size, flags);
3975 EXPORT_SYMBOL(__kmalloc);
3978 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3982 unsigned int order = get_order(size);
3984 flags |= __GFP_COMP;
3985 page = alloc_pages_node(node, flags, order);
3987 ptr = page_address(page);
3988 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
3989 PAGE_SIZE << order);
3992 return kmalloc_large_node_hook(ptr, size, flags);
3995 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3997 struct kmem_cache *s;
4000 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4001 ret = kmalloc_large_node(size, flags, node);
4003 trace_kmalloc_node(_RET_IP_, ret,
4004 size, PAGE_SIZE << get_order(size),
4010 s = kmalloc_slab(size, flags);
4012 if (unlikely(ZERO_OR_NULL_PTR(s)))
4015 ret = slab_alloc_node(s, flags, node, _RET_IP_);
4017 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4019 ret = kasan_kmalloc(s, ret, size, flags);
4023 EXPORT_SYMBOL(__kmalloc_node);
4024 #endif /* CONFIG_NUMA */
4026 #ifdef CONFIG_HARDENED_USERCOPY
4028 * Rejects incorrectly sized objects and objects that are to be copied
4029 * to/from userspace but do not fall entirely within the containing slab
4030 * cache's usercopy region.
4032 * Returns NULL if check passes, otherwise const char * to name of cache
4033 * to indicate an error.
4035 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4038 struct kmem_cache *s;
4039 unsigned int offset;
4042 ptr = kasan_reset_tag(ptr);
4044 /* Find object and usable object size. */
4045 s = page->slab_cache;
4047 /* Reject impossible pointers. */
4048 if (ptr < page_address(page))
4049 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4052 /* Find offset within object. */
4053 offset = (ptr - page_address(page)) % s->size;
4055 /* Adjust for redzone and reject if within the redzone. */
4056 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4057 if (offset < s->red_left_pad)
4058 usercopy_abort("SLUB object in left red zone",
4059 s->name, to_user, offset, n);
4060 offset -= s->red_left_pad;
4063 /* Allow address range falling entirely within usercopy region. */
4064 if (offset >= s->useroffset &&
4065 offset - s->useroffset <= s->usersize &&
4066 n <= s->useroffset - offset + s->usersize)
4070 * If the copy is still within the allocated object, produce
4071 * a warning instead of rejecting the copy. This is intended
4072 * to be a temporary method to find any missing usercopy
4075 object_size = slab_ksize(s);
4076 if (usercopy_fallback &&
4077 offset <= object_size && n <= object_size - offset) {
4078 usercopy_warn("SLUB object", s->name, to_user, offset, n);
4082 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4084 #endif /* CONFIG_HARDENED_USERCOPY */
4086 size_t __ksize(const void *object)
4090 if (unlikely(object == ZERO_SIZE_PTR))
4093 page = virt_to_head_page(object);
4095 if (unlikely(!PageSlab(page))) {
4096 WARN_ON(!PageCompound(page));
4097 return page_size(page);
4100 return slab_ksize(page->slab_cache);
4102 EXPORT_SYMBOL(__ksize);
4104 void kfree(const void *x)
4107 void *object = (void *)x;
4109 trace_kfree(_RET_IP_, x);
4111 if (unlikely(ZERO_OR_NULL_PTR(x)))
4114 page = virt_to_head_page(x);
4115 if (unlikely(!PageSlab(page))) {
4116 unsigned int order = compound_order(page);
4118 BUG_ON(!PageCompound(page));
4120 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4121 -(PAGE_SIZE << order));
4122 __free_pages(page, order);
4125 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
4127 EXPORT_SYMBOL(kfree);
4129 #define SHRINK_PROMOTE_MAX 32
4132 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4133 * up most to the head of the partial lists. New allocations will then
4134 * fill those up and thus they can be removed from the partial lists.
4136 * The slabs with the least items are placed last. This results in them
4137 * being allocated from last increasing the chance that the last objects
4138 * are freed in them.
4140 int __kmem_cache_shrink(struct kmem_cache *s)
4144 struct kmem_cache_node *n;
4147 struct list_head discard;
4148 struct list_head promote[SHRINK_PROMOTE_MAX];
4149 unsigned long flags;
4153 for_each_kmem_cache_node(s, node, n) {
4154 INIT_LIST_HEAD(&discard);
4155 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4156 INIT_LIST_HEAD(promote + i);
4158 spin_lock_irqsave(&n->list_lock, flags);
4161 * Build lists of slabs to discard or promote.
4163 * Note that concurrent frees may occur while we hold the
4164 * list_lock. page->inuse here is the upper limit.
4166 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
4167 int free = page->objects - page->inuse;
4169 /* Do not reread page->inuse */
4172 /* We do not keep full slabs on the list */
4175 if (free == page->objects) {
4176 list_move(&page->slab_list, &discard);
4178 } else if (free <= SHRINK_PROMOTE_MAX)
4179 list_move(&page->slab_list, promote + free - 1);
4183 * Promote the slabs filled up most to the head of the
4186 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4187 list_splice(promote + i, &n->partial);
4189 spin_unlock_irqrestore(&n->list_lock, flags);
4191 /* Release empty slabs */
4192 list_for_each_entry_safe(page, t, &discard, slab_list)
4193 discard_slab(s, page);
4195 if (slabs_node(s, node))
4202 static int slab_mem_going_offline_callback(void *arg)
4204 struct kmem_cache *s;
4206 mutex_lock(&slab_mutex);
4207 list_for_each_entry(s, &slab_caches, list)
4208 __kmem_cache_shrink(s);
4209 mutex_unlock(&slab_mutex);
4214 static void slab_mem_offline_callback(void *arg)
4216 struct kmem_cache_node *n;
4217 struct kmem_cache *s;
4218 struct memory_notify *marg = arg;
4221 offline_node = marg->status_change_nid_normal;
4224 * If the node still has available memory. we need kmem_cache_node
4227 if (offline_node < 0)
4230 mutex_lock(&slab_mutex);
4231 list_for_each_entry(s, &slab_caches, list) {
4232 n = get_node(s, offline_node);
4235 * if n->nr_slabs > 0, slabs still exist on the node
4236 * that is going down. We were unable to free them,
4237 * and offline_pages() function shouldn't call this
4238 * callback. So, we must fail.
4240 BUG_ON(slabs_node(s, offline_node));
4242 s->node[offline_node] = NULL;
4243 kmem_cache_free(kmem_cache_node, n);
4246 mutex_unlock(&slab_mutex);
4249 static int slab_mem_going_online_callback(void *arg)
4251 struct kmem_cache_node *n;
4252 struct kmem_cache *s;
4253 struct memory_notify *marg = arg;
4254 int nid = marg->status_change_nid_normal;
4258 * If the node's memory is already available, then kmem_cache_node is
4259 * already created. Nothing to do.
4265 * We are bringing a node online. No memory is available yet. We must
4266 * allocate a kmem_cache_node structure in order to bring the node
4269 mutex_lock(&slab_mutex);
4270 list_for_each_entry(s, &slab_caches, list) {
4272 * XXX: kmem_cache_alloc_node will fallback to other nodes
4273 * since memory is not yet available from the node that
4276 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4281 init_kmem_cache_node(n);
4285 mutex_unlock(&slab_mutex);
4289 static int slab_memory_callback(struct notifier_block *self,
4290 unsigned long action, void *arg)
4295 case MEM_GOING_ONLINE:
4296 ret = slab_mem_going_online_callback(arg);
4298 case MEM_GOING_OFFLINE:
4299 ret = slab_mem_going_offline_callback(arg);
4302 case MEM_CANCEL_ONLINE:
4303 slab_mem_offline_callback(arg);
4306 case MEM_CANCEL_OFFLINE:
4310 ret = notifier_from_errno(ret);
4316 static struct notifier_block slab_memory_callback_nb = {
4317 .notifier_call = slab_memory_callback,
4318 .priority = SLAB_CALLBACK_PRI,
4321 /********************************************************************
4322 * Basic setup of slabs
4323 *******************************************************************/
4326 * Used for early kmem_cache structures that were allocated using
4327 * the page allocator. Allocate them properly then fix up the pointers
4328 * that may be pointing to the wrong kmem_cache structure.
4331 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4334 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4335 struct kmem_cache_node *n;
4337 memcpy(s, static_cache, kmem_cache->object_size);
4340 * This runs very early, and only the boot processor is supposed to be
4341 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4344 __flush_cpu_slab(s, smp_processor_id());
4345 for_each_kmem_cache_node(s, node, n) {
4348 list_for_each_entry(p, &n->partial, slab_list)
4351 #ifdef CONFIG_SLUB_DEBUG
4352 list_for_each_entry(p, &n->full, slab_list)
4356 list_add(&s->list, &slab_caches);
4360 void __init kmem_cache_init(void)
4362 static __initdata struct kmem_cache boot_kmem_cache,
4363 boot_kmem_cache_node;
4365 if (debug_guardpage_minorder())
4368 kmem_cache_node = &boot_kmem_cache_node;
4369 kmem_cache = &boot_kmem_cache;
4371 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4372 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4374 register_hotmemory_notifier(&slab_memory_callback_nb);
4376 /* Able to allocate the per node structures */
4377 slab_state = PARTIAL;
4379 create_boot_cache(kmem_cache, "kmem_cache",
4380 offsetof(struct kmem_cache, node) +
4381 nr_node_ids * sizeof(struct kmem_cache_node *),
4382 SLAB_HWCACHE_ALIGN, 0, 0);
4384 kmem_cache = bootstrap(&boot_kmem_cache);
4385 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4387 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4388 setup_kmalloc_cache_index_table();
4389 create_kmalloc_caches(0);
4391 /* Setup random freelists for each cache */
4392 init_freelist_randomization();
4394 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4397 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4399 slub_min_order, slub_max_order, slub_min_objects,
4400 nr_cpu_ids, nr_node_ids);
4403 void __init kmem_cache_init_late(void)
4408 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4409 slab_flags_t flags, void (*ctor)(void *))
4411 struct kmem_cache *s;
4413 s = find_mergeable(size, align, flags, name, ctor);
4418 * Adjust the object sizes so that we clear
4419 * the complete object on kzalloc.
4421 s->object_size = max(s->object_size, size);
4422 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4424 if (sysfs_slab_alias(s, name)) {
4433 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4437 err = kmem_cache_open(s, flags);
4441 /* Mutex is not taken during early boot */
4442 if (slab_state <= UP)
4445 err = sysfs_slab_add(s);
4447 __kmem_cache_release(s);
4452 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4454 struct kmem_cache *s;
4457 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4458 return kmalloc_large(size, gfpflags);
4460 s = kmalloc_slab(size, gfpflags);
4462 if (unlikely(ZERO_OR_NULL_PTR(s)))
4465 ret = slab_alloc(s, gfpflags, caller);
4467 /* Honor the call site pointer we received. */
4468 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4472 EXPORT_SYMBOL(__kmalloc_track_caller);
4475 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4476 int node, unsigned long caller)
4478 struct kmem_cache *s;
4481 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4482 ret = kmalloc_large_node(size, gfpflags, node);
4484 trace_kmalloc_node(caller, ret,
4485 size, PAGE_SIZE << get_order(size),
4491 s = kmalloc_slab(size, gfpflags);
4493 if (unlikely(ZERO_OR_NULL_PTR(s)))
4496 ret = slab_alloc_node(s, gfpflags, node, caller);
4498 /* Honor the call site pointer we received. */
4499 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4503 EXPORT_SYMBOL(__kmalloc_node_track_caller);
4507 static int count_inuse(struct page *page)
4512 static int count_total(struct page *page)
4514 return page->objects;
4518 #ifdef CONFIG_SLUB_DEBUG
4519 static void validate_slab(struct kmem_cache *s, struct page *page)
4522 void *addr = page_address(page);
4527 if (!check_slab(s, page) || !on_freelist(s, page, NULL))
4530 /* Now we know that a valid freelist exists */
4531 map = get_map(s, page);
4532 for_each_object(p, s, addr, page->objects) {
4533 u8 val = test_bit(__obj_to_index(s, addr, p), map) ?
4534 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4536 if (!check_object(s, page, p, val))
4544 static int validate_slab_node(struct kmem_cache *s,
4545 struct kmem_cache_node *n)
4547 unsigned long count = 0;
4549 unsigned long flags;
4551 spin_lock_irqsave(&n->list_lock, flags);
4553 list_for_each_entry(page, &n->partial, slab_list) {
4554 validate_slab(s, page);
4557 if (count != n->nr_partial)
4558 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4559 s->name, count, n->nr_partial);
4561 if (!(s->flags & SLAB_STORE_USER))
4564 list_for_each_entry(page, &n->full, slab_list) {
4565 validate_slab(s, page);
4568 if (count != atomic_long_read(&n->nr_slabs))
4569 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4570 s->name, count, atomic_long_read(&n->nr_slabs));
4573 spin_unlock_irqrestore(&n->list_lock, flags);
4577 static long validate_slab_cache(struct kmem_cache *s)
4580 unsigned long count = 0;
4581 struct kmem_cache_node *n;
4584 for_each_kmem_cache_node(s, node, n)
4585 count += validate_slab_node(s, n);
4590 * Generate lists of code addresses where slabcache objects are allocated
4595 unsigned long count;
4602 DECLARE_BITMAP(cpus, NR_CPUS);
4608 unsigned long count;
4609 struct location *loc;
4612 static void free_loc_track(struct loc_track *t)
4615 free_pages((unsigned long)t->loc,
4616 get_order(sizeof(struct location) * t->max));
4619 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4624 order = get_order(sizeof(struct location) * max);
4626 l = (void *)__get_free_pages(flags, order);
4631 memcpy(l, t->loc, sizeof(struct location) * t->count);
4639 static int add_location(struct loc_track *t, struct kmem_cache *s,
4640 const struct track *track)
4642 long start, end, pos;
4644 unsigned long caddr;
4645 unsigned long age = jiffies - track->when;
4651 pos = start + (end - start + 1) / 2;
4654 * There is nothing at "end". If we end up there
4655 * we need to add something to before end.
4660 caddr = t->loc[pos].addr;
4661 if (track->addr == caddr) {
4667 if (age < l->min_time)
4669 if (age > l->max_time)
4672 if (track->pid < l->min_pid)
4673 l->min_pid = track->pid;
4674 if (track->pid > l->max_pid)
4675 l->max_pid = track->pid;
4677 cpumask_set_cpu(track->cpu,
4678 to_cpumask(l->cpus));
4680 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4684 if (track->addr < caddr)
4691 * Not found. Insert new tracking element.
4693 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4699 (t->count - pos) * sizeof(struct location));
4702 l->addr = track->addr;
4706 l->min_pid = track->pid;
4707 l->max_pid = track->pid;
4708 cpumask_clear(to_cpumask(l->cpus));
4709 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4710 nodes_clear(l->nodes);
4711 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4715 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4716 struct page *page, enum track_item alloc)
4718 void *addr = page_address(page);
4722 map = get_map(s, page);
4723 for_each_object(p, s, addr, page->objects)
4724 if (!test_bit(__obj_to_index(s, addr, p), map))
4725 add_location(t, s, get_track(s, p, alloc));
4729 static int list_locations(struct kmem_cache *s, char *buf,
4730 enum track_item alloc)
4734 struct loc_track t = { 0, 0, NULL };
4736 struct kmem_cache_node *n;
4738 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4740 return sprintf(buf, "Out of memory\n");
4742 /* Push back cpu slabs */
4745 for_each_kmem_cache_node(s, node, n) {
4746 unsigned long flags;
4749 if (!atomic_long_read(&n->nr_slabs))
4752 spin_lock_irqsave(&n->list_lock, flags);
4753 list_for_each_entry(page, &n->partial, slab_list)
4754 process_slab(&t, s, page, alloc);
4755 list_for_each_entry(page, &n->full, slab_list)
4756 process_slab(&t, s, page, alloc);
4757 spin_unlock_irqrestore(&n->list_lock, flags);
4760 for (i = 0; i < t.count; i++) {
4761 struct location *l = &t.loc[i];
4763 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4765 len += sprintf(buf + len, "%7ld ", l->count);
4768 len += sprintf(buf + len, "%pS", (void *)l->addr);
4770 len += sprintf(buf + len, "<not-available>");
4772 if (l->sum_time != l->min_time) {
4773 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4775 (long)div_u64(l->sum_time, l->count),
4778 len += sprintf(buf + len, " age=%ld",
4781 if (l->min_pid != l->max_pid)
4782 len += sprintf(buf + len, " pid=%ld-%ld",
4783 l->min_pid, l->max_pid);
4785 len += sprintf(buf + len, " pid=%ld",
4788 if (num_online_cpus() > 1 &&
4789 !cpumask_empty(to_cpumask(l->cpus)) &&
4790 len < PAGE_SIZE - 60)
4791 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4793 cpumask_pr_args(to_cpumask(l->cpus)));
4795 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4796 len < PAGE_SIZE - 60)
4797 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4799 nodemask_pr_args(&l->nodes));
4801 len += sprintf(buf + len, "\n");
4806 len += sprintf(buf, "No data\n");
4809 #endif /* CONFIG_SLUB_DEBUG */
4811 #ifdef SLUB_RESILIENCY_TEST
4812 static void __init resiliency_test(void)
4815 int type = KMALLOC_NORMAL;
4817 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4819 pr_err("SLUB resiliency testing\n");
4820 pr_err("-----------------------\n");
4821 pr_err("A. Corruption after allocation\n");
4823 p = kzalloc(16, GFP_KERNEL);
4825 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4828 validate_slab_cache(kmalloc_caches[type][4]);
4830 /* Hmmm... The next two are dangerous */
4831 p = kzalloc(32, GFP_KERNEL);
4832 p[32 + sizeof(void *)] = 0x34;
4833 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4835 pr_err("If allocated object is overwritten then not detectable\n\n");
4837 validate_slab_cache(kmalloc_caches[type][5]);
4838 p = kzalloc(64, GFP_KERNEL);
4839 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4841 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4843 pr_err("If allocated object is overwritten then not detectable\n\n");
4844 validate_slab_cache(kmalloc_caches[type][6]);
4846 pr_err("\nB. Corruption after free\n");
4847 p = kzalloc(128, GFP_KERNEL);
4850 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4851 validate_slab_cache(kmalloc_caches[type][7]);
4853 p = kzalloc(256, GFP_KERNEL);
4856 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4857 validate_slab_cache(kmalloc_caches[type][8]);
4859 p = kzalloc(512, GFP_KERNEL);
4862 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4863 validate_slab_cache(kmalloc_caches[type][9]);
4867 static void resiliency_test(void) {};
4869 #endif /* SLUB_RESILIENCY_TEST */
4872 enum slab_stat_type {
4873 SL_ALL, /* All slabs */
4874 SL_PARTIAL, /* Only partially allocated slabs */
4875 SL_CPU, /* Only slabs used for cpu caches */
4876 SL_OBJECTS, /* Determine allocated objects not slabs */
4877 SL_TOTAL /* Determine object capacity not slabs */
4880 #define SO_ALL (1 << SL_ALL)
4881 #define SO_PARTIAL (1 << SL_PARTIAL)
4882 #define SO_CPU (1 << SL_CPU)
4883 #define SO_OBJECTS (1 << SL_OBJECTS)
4884 #define SO_TOTAL (1 << SL_TOTAL)
4887 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4889 static int __init setup_slub_memcg_sysfs(char *str)
4893 if (get_option(&str, &v) > 0)
4894 memcg_sysfs_enabled = v;
4899 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4902 static ssize_t show_slab_objects(struct kmem_cache *s,
4903 char *buf, unsigned long flags)
4905 unsigned long total = 0;
4908 unsigned long *nodes;
4910 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4914 if (flags & SO_CPU) {
4917 for_each_possible_cpu(cpu) {
4918 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4923 page = READ_ONCE(c->page);
4927 node = page_to_nid(page);
4928 if (flags & SO_TOTAL)
4930 else if (flags & SO_OBJECTS)
4938 page = slub_percpu_partial_read_once(c);
4940 node = page_to_nid(page);
4941 if (flags & SO_TOTAL)
4943 else if (flags & SO_OBJECTS)
4954 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4955 * already held which will conflict with an existing lock order:
4957 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4959 * We don't really need mem_hotplug_lock (to hold off
4960 * slab_mem_going_offline_callback) here because slab's memory hot
4961 * unplug code doesn't destroy the kmem_cache->node[] data.
4964 #ifdef CONFIG_SLUB_DEBUG
4965 if (flags & SO_ALL) {
4966 struct kmem_cache_node *n;
4968 for_each_kmem_cache_node(s, node, n) {
4970 if (flags & SO_TOTAL)
4971 x = atomic_long_read(&n->total_objects);
4972 else if (flags & SO_OBJECTS)
4973 x = atomic_long_read(&n->total_objects) -
4974 count_partial(n, count_free);
4976 x = atomic_long_read(&n->nr_slabs);
4983 if (flags & SO_PARTIAL) {
4984 struct kmem_cache_node *n;
4986 for_each_kmem_cache_node(s, node, n) {
4987 if (flags & SO_TOTAL)
4988 x = count_partial(n, count_total);
4989 else if (flags & SO_OBJECTS)
4990 x = count_partial(n, count_inuse);
4997 x = sprintf(buf, "%lu", total);
4999 for (node = 0; node < nr_node_ids; node++)
5001 x += sprintf(buf + x, " N%d=%lu",
5005 return x + sprintf(buf + x, "\n");
5008 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5009 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5011 struct slab_attribute {
5012 struct attribute attr;
5013 ssize_t (*show)(struct kmem_cache *s, char *buf);
5014 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5017 #define SLAB_ATTR_RO(_name) \
5018 static struct slab_attribute _name##_attr = \
5019 __ATTR(_name, 0400, _name##_show, NULL)
5021 #define SLAB_ATTR(_name) \
5022 static struct slab_attribute _name##_attr = \
5023 __ATTR(_name, 0600, _name##_show, _name##_store)
5025 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5027 return sprintf(buf, "%u\n", s->size);
5029 SLAB_ATTR_RO(slab_size);
5031 static ssize_t align_show(struct kmem_cache *s, char *buf)
5033 return sprintf(buf, "%u\n", s->align);
5035 SLAB_ATTR_RO(align);
5037 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5039 return sprintf(buf, "%u\n", s->object_size);
5041 SLAB_ATTR_RO(object_size);
5043 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5045 return sprintf(buf, "%u\n", oo_objects(s->oo));
5047 SLAB_ATTR_RO(objs_per_slab);
5049 static ssize_t order_show(struct kmem_cache *s, char *buf)
5051 return sprintf(buf, "%u\n", oo_order(s->oo));
5053 SLAB_ATTR_RO(order);
5055 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5057 return sprintf(buf, "%lu\n", s->min_partial);
5060 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5066 err = kstrtoul(buf, 10, &min);
5070 set_min_partial(s, min);
5073 SLAB_ATTR(min_partial);
5075 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5077 return sprintf(buf, "%u\n", slub_cpu_partial(s));
5080 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5083 unsigned int objects;
5086 err = kstrtouint(buf, 10, &objects);
5089 if (objects && !kmem_cache_has_cpu_partial(s))
5092 slub_set_cpu_partial(s, objects);
5096 SLAB_ATTR(cpu_partial);
5098 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5102 return sprintf(buf, "%pS\n", s->ctor);
5106 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5108 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5110 SLAB_ATTR_RO(aliases);
5112 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5114 return show_slab_objects(s, buf, SO_PARTIAL);
5116 SLAB_ATTR_RO(partial);
5118 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5120 return show_slab_objects(s, buf, SO_CPU);
5122 SLAB_ATTR_RO(cpu_slabs);
5124 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5126 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5128 SLAB_ATTR_RO(objects);
5130 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5132 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5134 SLAB_ATTR_RO(objects_partial);
5136 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5143 for_each_online_cpu(cpu) {
5146 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5149 pages += page->pages;
5150 objects += page->pobjects;
5154 len = sprintf(buf, "%d(%d)", objects, pages);
5157 for_each_online_cpu(cpu) {
5160 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5162 if (page && len < PAGE_SIZE - 20)
5163 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5164 page->pobjects, page->pages);
5167 return len + sprintf(buf + len, "\n");
5169 SLAB_ATTR_RO(slabs_cpu_partial);
5171 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5173 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5175 SLAB_ATTR_RO(reclaim_account);
5177 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5179 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5181 SLAB_ATTR_RO(hwcache_align);
5183 #ifdef CONFIG_ZONE_DMA
5184 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5186 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5188 SLAB_ATTR_RO(cache_dma);
5191 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5193 return sprintf(buf, "%u\n", s->usersize);
5195 SLAB_ATTR_RO(usersize);
5197 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5199 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5201 SLAB_ATTR_RO(destroy_by_rcu);
5203 #ifdef CONFIG_SLUB_DEBUG
5204 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5206 return show_slab_objects(s, buf, SO_ALL);
5208 SLAB_ATTR_RO(slabs);
5210 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5212 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5214 SLAB_ATTR_RO(total_objects);
5216 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5218 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5220 SLAB_ATTR_RO(sanity_checks);
5222 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5224 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5226 SLAB_ATTR_RO(trace);
5228 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5230 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5233 SLAB_ATTR_RO(red_zone);
5235 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5237 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5240 SLAB_ATTR_RO(poison);
5242 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5244 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5247 SLAB_ATTR_RO(store_user);
5249 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5254 static ssize_t validate_store(struct kmem_cache *s,
5255 const char *buf, size_t length)
5259 if (buf[0] == '1') {
5260 ret = validate_slab_cache(s);
5266 SLAB_ATTR(validate);
5268 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5270 if (!(s->flags & SLAB_STORE_USER))
5272 return list_locations(s, buf, TRACK_ALLOC);
5274 SLAB_ATTR_RO(alloc_calls);
5276 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5278 if (!(s->flags & SLAB_STORE_USER))
5280 return list_locations(s, buf, TRACK_FREE);
5282 SLAB_ATTR_RO(free_calls);
5283 #endif /* CONFIG_SLUB_DEBUG */
5285 #ifdef CONFIG_FAILSLAB
5286 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5288 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5290 SLAB_ATTR_RO(failslab);
5293 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5298 static ssize_t shrink_store(struct kmem_cache *s,
5299 const char *buf, size_t length)
5302 kmem_cache_shrink(s);
5310 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5312 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5315 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5316 const char *buf, size_t length)
5321 err = kstrtouint(buf, 10, &ratio);
5327 s->remote_node_defrag_ratio = ratio * 10;
5331 SLAB_ATTR(remote_node_defrag_ratio);
5334 #ifdef CONFIG_SLUB_STATS
5335 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5337 unsigned long sum = 0;
5340 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5345 for_each_online_cpu(cpu) {
5346 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5352 len = sprintf(buf, "%lu", sum);
5355 for_each_online_cpu(cpu) {
5356 if (data[cpu] && len < PAGE_SIZE - 20)
5357 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5361 return len + sprintf(buf + len, "\n");
5364 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5368 for_each_online_cpu(cpu)
5369 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5372 #define STAT_ATTR(si, text) \
5373 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5375 return show_stat(s, buf, si); \
5377 static ssize_t text##_store(struct kmem_cache *s, \
5378 const char *buf, size_t length) \
5380 if (buf[0] != '0') \
5382 clear_stat(s, si); \
5387 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5388 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5389 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5390 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5391 STAT_ATTR(FREE_FROZEN, free_frozen);
5392 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5393 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5394 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5395 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5396 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5397 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5398 STAT_ATTR(FREE_SLAB, free_slab);
5399 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5400 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5401 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5402 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5403 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5404 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5405 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5406 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5407 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5408 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5409 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5410 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5411 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5412 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5413 #endif /* CONFIG_SLUB_STATS */
5415 static struct attribute *slab_attrs[] = {
5416 &slab_size_attr.attr,
5417 &object_size_attr.attr,
5418 &objs_per_slab_attr.attr,
5420 &min_partial_attr.attr,
5421 &cpu_partial_attr.attr,
5423 &objects_partial_attr.attr,
5425 &cpu_slabs_attr.attr,
5429 &hwcache_align_attr.attr,
5430 &reclaim_account_attr.attr,
5431 &destroy_by_rcu_attr.attr,
5433 &slabs_cpu_partial_attr.attr,
5434 #ifdef CONFIG_SLUB_DEBUG
5435 &total_objects_attr.attr,
5437 &sanity_checks_attr.attr,
5439 &red_zone_attr.attr,
5441 &store_user_attr.attr,
5442 &validate_attr.attr,
5443 &alloc_calls_attr.attr,
5444 &free_calls_attr.attr,
5446 #ifdef CONFIG_ZONE_DMA
5447 &cache_dma_attr.attr,
5450 &remote_node_defrag_ratio_attr.attr,
5452 #ifdef CONFIG_SLUB_STATS
5453 &alloc_fastpath_attr.attr,
5454 &alloc_slowpath_attr.attr,
5455 &free_fastpath_attr.attr,
5456 &free_slowpath_attr.attr,
5457 &free_frozen_attr.attr,
5458 &free_add_partial_attr.attr,
5459 &free_remove_partial_attr.attr,
5460 &alloc_from_partial_attr.attr,
5461 &alloc_slab_attr.attr,
5462 &alloc_refill_attr.attr,
5463 &alloc_node_mismatch_attr.attr,
5464 &free_slab_attr.attr,
5465 &cpuslab_flush_attr.attr,
5466 &deactivate_full_attr.attr,
5467 &deactivate_empty_attr.attr,
5468 &deactivate_to_head_attr.attr,
5469 &deactivate_to_tail_attr.attr,
5470 &deactivate_remote_frees_attr.attr,
5471 &deactivate_bypass_attr.attr,
5472 &order_fallback_attr.attr,
5473 &cmpxchg_double_fail_attr.attr,
5474 &cmpxchg_double_cpu_fail_attr.attr,
5475 &cpu_partial_alloc_attr.attr,
5476 &cpu_partial_free_attr.attr,
5477 &cpu_partial_node_attr.attr,
5478 &cpu_partial_drain_attr.attr,
5480 #ifdef CONFIG_FAILSLAB
5481 &failslab_attr.attr,
5483 &usersize_attr.attr,
5488 static const struct attribute_group slab_attr_group = {
5489 .attrs = slab_attrs,
5492 static ssize_t slab_attr_show(struct kobject *kobj,
5493 struct attribute *attr,
5496 struct slab_attribute *attribute;
5497 struct kmem_cache *s;
5500 attribute = to_slab_attr(attr);
5503 if (!attribute->show)
5506 err = attribute->show(s, buf);
5511 static ssize_t slab_attr_store(struct kobject *kobj,
5512 struct attribute *attr,
5513 const char *buf, size_t len)
5515 struct slab_attribute *attribute;
5516 struct kmem_cache *s;
5519 attribute = to_slab_attr(attr);
5522 if (!attribute->store)
5525 err = attribute->store(s, buf, len);
5529 static void kmem_cache_release(struct kobject *k)
5531 slab_kmem_cache_release(to_slab(k));
5534 static const struct sysfs_ops slab_sysfs_ops = {
5535 .show = slab_attr_show,
5536 .store = slab_attr_store,
5539 static struct kobj_type slab_ktype = {
5540 .sysfs_ops = &slab_sysfs_ops,
5541 .release = kmem_cache_release,
5544 static struct kset *slab_kset;
5546 static inline struct kset *cache_kset(struct kmem_cache *s)
5551 #define ID_STR_LENGTH 64
5553 /* Create a unique string id for a slab cache:
5555 * Format :[flags-]size
5557 static char *create_unique_id(struct kmem_cache *s)
5559 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5563 return ERR_PTR(-ENOMEM);
5567 * First flags affecting slabcache operations. We will only
5568 * get here for aliasable slabs so we do not need to support
5569 * too many flags. The flags here must cover all flags that
5570 * are matched during merging to guarantee that the id is
5573 if (s->flags & SLAB_CACHE_DMA)
5575 if (s->flags & SLAB_CACHE_DMA32)
5577 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5579 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5581 if (s->flags & SLAB_ACCOUNT)
5585 p += sprintf(p, "%07u", s->size);
5587 BUG_ON(p > name + ID_STR_LENGTH - 1);
5591 static int sysfs_slab_add(struct kmem_cache *s)
5595 struct kset *kset = cache_kset(s);
5596 int unmergeable = slab_unmergeable(s);
5599 kobject_init(&s->kobj, &slab_ktype);
5603 if (!unmergeable && disable_higher_order_debug &&
5604 (slub_debug & DEBUG_METADATA_FLAGS))
5609 * Slabcache can never be merged so we can use the name proper.
5610 * This is typically the case for debug situations. In that
5611 * case we can catch duplicate names easily.
5613 sysfs_remove_link(&slab_kset->kobj, s->name);
5617 * Create a unique name for the slab as a target
5620 name = create_unique_id(s);
5622 return PTR_ERR(name);
5625 s->kobj.kset = kset;
5626 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5630 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5635 /* Setup first alias */
5636 sysfs_slab_alias(s, s->name);
5643 kobject_del(&s->kobj);
5647 void sysfs_slab_unlink(struct kmem_cache *s)
5649 if (slab_state >= FULL)
5650 kobject_del(&s->kobj);
5653 void sysfs_slab_release(struct kmem_cache *s)
5655 if (slab_state >= FULL)
5656 kobject_put(&s->kobj);
5660 * Need to buffer aliases during bootup until sysfs becomes
5661 * available lest we lose that information.
5663 struct saved_alias {
5664 struct kmem_cache *s;
5666 struct saved_alias *next;
5669 static struct saved_alias *alias_list;
5671 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5673 struct saved_alias *al;
5675 if (slab_state == FULL) {
5677 * If we have a leftover link then remove it.
5679 sysfs_remove_link(&slab_kset->kobj, name);
5680 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5683 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5689 al->next = alias_list;
5694 static int __init slab_sysfs_init(void)
5696 struct kmem_cache *s;
5699 mutex_lock(&slab_mutex);
5701 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
5703 mutex_unlock(&slab_mutex);
5704 pr_err("Cannot register slab subsystem.\n");
5710 list_for_each_entry(s, &slab_caches, list) {
5711 err = sysfs_slab_add(s);
5713 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5717 while (alias_list) {
5718 struct saved_alias *al = alias_list;
5720 alias_list = alias_list->next;
5721 err = sysfs_slab_alias(al->s, al->name);
5723 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5728 mutex_unlock(&slab_mutex);
5733 __initcall(slab_sysfs_init);
5734 #endif /* CONFIG_SYSFS */
5737 * The /proc/slabinfo ABI
5739 #ifdef CONFIG_SLUB_DEBUG
5740 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5742 unsigned long nr_slabs = 0;
5743 unsigned long nr_objs = 0;
5744 unsigned long nr_free = 0;
5746 struct kmem_cache_node *n;
5748 for_each_kmem_cache_node(s, node, n) {
5749 nr_slabs += node_nr_slabs(n);
5750 nr_objs += node_nr_objs(n);
5751 nr_free += count_partial(n, count_free);
5754 sinfo->active_objs = nr_objs - nr_free;
5755 sinfo->num_objs = nr_objs;
5756 sinfo->active_slabs = nr_slabs;
5757 sinfo->num_slabs = nr_slabs;
5758 sinfo->objects_per_slab = oo_objects(s->oo);
5759 sinfo->cache_order = oo_order(s->oo);
5762 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5766 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5767 size_t count, loff_t *ppos)
5771 #endif /* CONFIG_SLUB_DEBUG */