2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemcheck.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/mempolicy.h>
24 #include <linux/ctype.h>
25 #include <linux/debugobjects.h>
26 #include <linux/kallsyms.h>
27 #include <linux/memory.h>
28 #include <linux/math64.h>
29 #include <linux/fault-inject.h>
31 #include <trace/events/kmem.h>
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has noone operating on it and thus there is
68 * no danger of cacheline contention.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
135 * Mininum number of partial slabs. These will be left on the partial
136 * lists even if they are empty. kmem_cache_shrink may reclaim them.
138 #define MIN_PARTIAL 5
141 * Maximum number of desirable partial slabs.
142 * The existence of more partial slabs makes kmem_cache_shrink
143 * sort the partial list by the number of objects in the.
145 #define MAX_PARTIAL 10
147 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
148 SLAB_POISON | SLAB_STORE_USER)
151 * Debugging flags that require metadata to be stored in the slab. These get
152 * disabled when slub_debug=O is used and a cache's min order increases with
155 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
158 * Set of flags that will prevent slab merging
160 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
164 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
165 SLAB_CACHE_DMA | SLAB_NOTRACK)
168 #define OO_MASK ((1 << OO_SHIFT) - 1)
169 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
171 /* Internal SLUB flags */
172 #define __OBJECT_POISON 0x80000000UL /* Poison object */
174 static int kmem_size = sizeof(struct kmem_cache);
177 static struct notifier_block slab_notifier;
181 DOWN, /* No slab functionality available */
182 PARTIAL, /* Kmem_cache_node works */
183 UP, /* Everything works but does not show up in sysfs */
187 /* A list of all slab caches on the system */
188 static DECLARE_RWSEM(slub_lock);
189 static LIST_HEAD(slab_caches);
192 * Tracking user of a slab.
195 unsigned long addr; /* Called from address */
196 int cpu; /* Was running on cpu */
197 int pid; /* Pid context */
198 unsigned long when; /* When did the operation occur */
201 enum track_item { TRACK_ALLOC, TRACK_FREE };
204 static int sysfs_slab_add(struct kmem_cache *);
205 static int sysfs_slab_alias(struct kmem_cache *, const char *);
206 static void sysfs_slab_remove(struct kmem_cache *);
209 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
210 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
212 static inline void sysfs_slab_remove(struct kmem_cache *s)
220 static inline void stat(struct kmem_cache *s, enum stat_item si)
222 #ifdef CONFIG_SLUB_STATS
223 __this_cpu_inc(s->cpu_slab->stat[si]);
227 /********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
231 int slab_is_available(void)
233 return slab_state >= UP;
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
238 return s->node[node];
241 /* Verify that a pointer has an address that is valid within a slab page */
242 static inline int check_valid_pointer(struct kmem_cache *s,
243 struct page *page, const void *object)
250 base = page_address(page);
251 if (object < base || object >= base + page->objects * s->size ||
252 (object - base) % s->size) {
259 static inline void *get_freepointer(struct kmem_cache *s, void *object)
261 return *(void **)(object + s->offset);
264 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
266 *(void **)(object + s->offset) = fp;
269 /* Loop over all objects in a slab */
270 #define for_each_object(__p, __s, __addr, __objects) \
271 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
275 #define for_each_free_object(__p, __s, __free) \
276 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
278 /* Determine object index from a given position */
279 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
281 return (p - addr) / s->size;
284 static inline size_t slab_ksize(const struct kmem_cache *s)
286 #ifdef CONFIG_SLUB_DEBUG
288 * Debugging requires use of the padding between object
289 * and whatever may come after it.
291 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
296 * If we have the need to store the freelist pointer
297 * back there or track user information then we can
298 * only use the space before that information.
300 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
303 * Else we can use all the padding etc for the allocation
308 static inline struct kmem_cache_order_objects oo_make(int order,
311 struct kmem_cache_order_objects x = {
312 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
318 static inline int oo_order(struct kmem_cache_order_objects x)
320 return x.x >> OO_SHIFT;
323 static inline int oo_objects(struct kmem_cache_order_objects x)
325 return x.x & OO_MASK;
328 #ifdef CONFIG_SLUB_DEBUG
332 #ifdef CONFIG_SLUB_DEBUG_ON
333 static int slub_debug = DEBUG_DEFAULT_FLAGS;
335 static int slub_debug;
338 static char *slub_debug_slabs;
339 static int disable_higher_order_debug;
344 static void print_section(char *text, u8 *addr, unsigned int length)
352 for (i = 0; i < length; i++) {
354 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
357 printk(KERN_CONT " %02x", addr[i]);
359 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
361 printk(KERN_CONT " %s\n", ascii);
368 printk(KERN_CONT " ");
372 printk(KERN_CONT " %s\n", ascii);
376 static struct track *get_track(struct kmem_cache *s, void *object,
377 enum track_item alloc)
382 p = object + s->offset + sizeof(void *);
384 p = object + s->inuse;
389 static void set_track(struct kmem_cache *s, void *object,
390 enum track_item alloc, unsigned long addr)
392 struct track *p = get_track(s, object, alloc);
396 p->cpu = smp_processor_id();
397 p->pid = current->pid;
400 memset(p, 0, sizeof(struct track));
403 static void init_tracking(struct kmem_cache *s, void *object)
405 if (!(s->flags & SLAB_STORE_USER))
408 set_track(s, object, TRACK_FREE, 0UL);
409 set_track(s, object, TRACK_ALLOC, 0UL);
412 static void print_track(const char *s, struct track *t)
417 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
418 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
421 static void print_tracking(struct kmem_cache *s, void *object)
423 if (!(s->flags & SLAB_STORE_USER))
426 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
427 print_track("Freed", get_track(s, object, TRACK_FREE));
430 static void print_page_info(struct page *page)
432 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
433 page, page->objects, page->inuse, page->freelist, page->flags);
437 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
443 vsnprintf(buf, sizeof(buf), fmt, args);
445 printk(KERN_ERR "========================================"
446 "=====================================\n");
447 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
448 printk(KERN_ERR "----------------------------------------"
449 "-------------------------------------\n\n");
452 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
458 vsnprintf(buf, sizeof(buf), fmt, args);
460 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
463 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
465 unsigned int off; /* Offset of last byte */
466 u8 *addr = page_address(page);
468 print_tracking(s, p);
470 print_page_info(page);
472 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
473 p, p - addr, get_freepointer(s, p));
476 print_section("Bytes b4", p - 16, 16);
478 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
480 if (s->flags & SLAB_RED_ZONE)
481 print_section("Redzone", p + s->objsize,
482 s->inuse - s->objsize);
485 off = s->offset + sizeof(void *);
489 if (s->flags & SLAB_STORE_USER)
490 off += 2 * sizeof(struct track);
493 /* Beginning of the filler is the free pointer */
494 print_section("Padding", p + off, s->size - off);
499 static void object_err(struct kmem_cache *s, struct page *page,
500 u8 *object, char *reason)
502 slab_bug(s, "%s", reason);
503 print_trailer(s, page, object);
506 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
512 vsnprintf(buf, sizeof(buf), fmt, args);
514 slab_bug(s, "%s", buf);
515 print_page_info(page);
519 static void init_object(struct kmem_cache *s, void *object, u8 val)
523 if (s->flags & __OBJECT_POISON) {
524 memset(p, POISON_FREE, s->objsize - 1);
525 p[s->objsize - 1] = POISON_END;
528 if (s->flags & SLAB_RED_ZONE)
529 memset(p + s->objsize, val, s->inuse - s->objsize);
532 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
535 if (*start != (u8)value)
543 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
544 void *from, void *to)
546 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
547 memset(from, data, to - from);
550 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
551 u8 *object, char *what,
552 u8 *start, unsigned int value, unsigned int bytes)
557 fault = check_bytes(start, value, bytes);
562 while (end > fault && end[-1] == value)
565 slab_bug(s, "%s overwritten", what);
566 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
567 fault, end - 1, fault[0], value);
568 print_trailer(s, page, object);
570 restore_bytes(s, what, value, fault, end);
578 * Bytes of the object to be managed.
579 * If the freepointer may overlay the object then the free
580 * pointer is the first word of the object.
582 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
585 * object + s->objsize
586 * Padding to reach word boundary. This is also used for Redzoning.
587 * Padding is extended by another word if Redzoning is enabled and
590 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
591 * 0xcc (RED_ACTIVE) for objects in use.
594 * Meta data starts here.
596 * A. Free pointer (if we cannot overwrite object on free)
597 * B. Tracking data for SLAB_STORE_USER
598 * C. Padding to reach required alignment boundary or at mininum
599 * one word if debugging is on to be able to detect writes
600 * before the word boundary.
602 * Padding is done using 0x5a (POISON_INUSE)
605 * Nothing is used beyond s->size.
607 * If slabcaches are merged then the objsize and inuse boundaries are mostly
608 * ignored. And therefore no slab options that rely on these boundaries
609 * may be used with merged slabcaches.
612 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
614 unsigned long off = s->inuse; /* The end of info */
617 /* Freepointer is placed after the object. */
618 off += sizeof(void *);
620 if (s->flags & SLAB_STORE_USER)
621 /* We also have user information there */
622 off += 2 * sizeof(struct track);
627 return check_bytes_and_report(s, page, p, "Object padding",
628 p + off, POISON_INUSE, s->size - off);
631 /* Check the pad bytes at the end of a slab page */
632 static int slab_pad_check(struct kmem_cache *s, struct page *page)
640 if (!(s->flags & SLAB_POISON))
643 start = page_address(page);
644 length = (PAGE_SIZE << compound_order(page));
645 end = start + length;
646 remainder = length % s->size;
650 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
653 while (end > fault && end[-1] == POISON_INUSE)
656 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
657 print_section("Padding", end - remainder, remainder);
659 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
663 static int check_object(struct kmem_cache *s, struct page *page,
664 void *object, u8 val)
667 u8 *endobject = object + s->objsize;
669 if (s->flags & SLAB_RED_ZONE) {
670 if (!check_bytes_and_report(s, page, object, "Redzone",
671 endobject, val, s->inuse - s->objsize))
674 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
675 check_bytes_and_report(s, page, p, "Alignment padding",
676 endobject, POISON_INUSE, s->inuse - s->objsize);
680 if (s->flags & SLAB_POISON) {
681 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
682 (!check_bytes_and_report(s, page, p, "Poison", p,
683 POISON_FREE, s->objsize - 1) ||
684 !check_bytes_and_report(s, page, p, "Poison",
685 p + s->objsize - 1, POISON_END, 1)))
688 * check_pad_bytes cleans up on its own.
690 check_pad_bytes(s, page, p);
693 if (!s->offset && val == SLUB_RED_ACTIVE)
695 * Object and freepointer overlap. Cannot check
696 * freepointer while object is allocated.
700 /* Check free pointer validity */
701 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
702 object_err(s, page, p, "Freepointer corrupt");
704 * No choice but to zap it and thus lose the remainder
705 * of the free objects in this slab. May cause
706 * another error because the object count is now wrong.
708 set_freepointer(s, p, NULL);
714 static int check_slab(struct kmem_cache *s, struct page *page)
718 VM_BUG_ON(!irqs_disabled());
720 if (!PageSlab(page)) {
721 slab_err(s, page, "Not a valid slab page");
725 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
726 if (page->objects > maxobj) {
727 slab_err(s, page, "objects %u > max %u",
728 s->name, page->objects, maxobj);
731 if (page->inuse > page->objects) {
732 slab_err(s, page, "inuse %u > max %u",
733 s->name, page->inuse, page->objects);
736 /* Slab_pad_check fixes things up after itself */
737 slab_pad_check(s, page);
742 * Determine if a certain object on a page is on the freelist. Must hold the
743 * slab lock to guarantee that the chains are in a consistent state.
745 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
748 void *fp = page->freelist;
750 unsigned long max_objects;
752 while (fp && nr <= page->objects) {
755 if (!check_valid_pointer(s, page, fp)) {
757 object_err(s, page, object,
758 "Freechain corrupt");
759 set_freepointer(s, object, NULL);
762 slab_err(s, page, "Freepointer corrupt");
763 page->freelist = NULL;
764 page->inuse = page->objects;
765 slab_fix(s, "Freelist cleared");
771 fp = get_freepointer(s, object);
775 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
776 if (max_objects > MAX_OBJS_PER_PAGE)
777 max_objects = MAX_OBJS_PER_PAGE;
779 if (page->objects != max_objects) {
780 slab_err(s, page, "Wrong number of objects. Found %d but "
781 "should be %d", page->objects, max_objects);
782 page->objects = max_objects;
783 slab_fix(s, "Number of objects adjusted.");
785 if (page->inuse != page->objects - nr) {
786 slab_err(s, page, "Wrong object count. Counter is %d but "
787 "counted were %d", page->inuse, page->objects - nr);
788 page->inuse = page->objects - nr;
789 slab_fix(s, "Object count adjusted.");
791 return search == NULL;
794 static void trace(struct kmem_cache *s, struct page *page, void *object,
797 if (s->flags & SLAB_TRACE) {
798 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
800 alloc ? "alloc" : "free",
805 print_section("Object", (void *)object, s->objsize);
812 * Hooks for other subsystems that check memory allocations. In a typical
813 * production configuration these hooks all should produce no code at all.
815 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
817 flags &= gfp_allowed_mask;
818 lockdep_trace_alloc(flags);
819 might_sleep_if(flags & __GFP_WAIT);
821 return should_failslab(s->objsize, flags, s->flags);
824 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
826 flags &= gfp_allowed_mask;
827 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
828 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
831 static inline void slab_free_hook(struct kmem_cache *s, void *x)
833 kmemleak_free_recursive(x, s->flags);
836 static inline void slab_free_hook_irq(struct kmem_cache *s, void *object)
838 kmemcheck_slab_free(s, object, s->objsize);
839 debug_check_no_locks_freed(object, s->objsize);
840 if (!(s->flags & SLAB_DEBUG_OBJECTS))
841 debug_check_no_obj_freed(object, s->objsize);
845 * Tracking of fully allocated slabs for debugging purposes.
847 static void add_full(struct kmem_cache_node *n, struct page *page)
849 spin_lock(&n->list_lock);
850 list_add(&page->lru, &n->full);
851 spin_unlock(&n->list_lock);
854 static void remove_full(struct kmem_cache *s, struct page *page)
856 struct kmem_cache_node *n;
858 if (!(s->flags & SLAB_STORE_USER))
861 n = get_node(s, page_to_nid(page));
863 spin_lock(&n->list_lock);
864 list_del(&page->lru);
865 spin_unlock(&n->list_lock);
868 /* Tracking of the number of slabs for debugging purposes */
869 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
871 struct kmem_cache_node *n = get_node(s, node);
873 return atomic_long_read(&n->nr_slabs);
876 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
878 return atomic_long_read(&n->nr_slabs);
881 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
883 struct kmem_cache_node *n = get_node(s, node);
886 * May be called early in order to allocate a slab for the
887 * kmem_cache_node structure. Solve the chicken-egg
888 * dilemma by deferring the increment of the count during
889 * bootstrap (see early_kmem_cache_node_alloc).
892 atomic_long_inc(&n->nr_slabs);
893 atomic_long_add(objects, &n->total_objects);
896 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
898 struct kmem_cache_node *n = get_node(s, node);
900 atomic_long_dec(&n->nr_slabs);
901 atomic_long_sub(objects, &n->total_objects);
904 /* Object debug checks for alloc/free paths */
905 static void setup_object_debug(struct kmem_cache *s, struct page *page,
908 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
911 init_object(s, object, SLUB_RED_INACTIVE);
912 init_tracking(s, object);
915 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
916 void *object, unsigned long addr)
918 if (!check_slab(s, page))
921 if (!on_freelist(s, page, object)) {
922 object_err(s, page, object, "Object already allocated");
926 if (!check_valid_pointer(s, page, object)) {
927 object_err(s, page, object, "Freelist Pointer check fails");
931 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
934 /* Success perform special debug activities for allocs */
935 if (s->flags & SLAB_STORE_USER)
936 set_track(s, object, TRACK_ALLOC, addr);
937 trace(s, page, object, 1);
938 init_object(s, object, SLUB_RED_ACTIVE);
942 if (PageSlab(page)) {
944 * If this is a slab page then lets do the best we can
945 * to avoid issues in the future. Marking all objects
946 * as used avoids touching the remaining objects.
948 slab_fix(s, "Marking all objects used");
949 page->inuse = page->objects;
950 page->freelist = NULL;
955 static noinline int free_debug_processing(struct kmem_cache *s,
956 struct page *page, void *object, unsigned long addr)
958 if (!check_slab(s, page))
961 if (!check_valid_pointer(s, page, object)) {
962 slab_err(s, page, "Invalid object pointer 0x%p", object);
966 if (on_freelist(s, page, object)) {
967 object_err(s, page, object, "Object already free");
971 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
974 if (unlikely(s != page->slab)) {
975 if (!PageSlab(page)) {
976 slab_err(s, page, "Attempt to free object(0x%p) "
977 "outside of slab", object);
978 } else if (!page->slab) {
980 "SLUB <none>: no slab for object 0x%p.\n",
984 object_err(s, page, object,
985 "page slab pointer corrupt.");
989 /* Special debug activities for freeing objects */
990 if (!PageSlubFrozen(page) && !page->freelist)
991 remove_full(s, page);
992 if (s->flags & SLAB_STORE_USER)
993 set_track(s, object, TRACK_FREE, addr);
994 trace(s, page, object, 0);
995 init_object(s, object, SLUB_RED_INACTIVE);
999 slab_fix(s, "Object at 0x%p not freed", object);
1003 static int __init setup_slub_debug(char *str)
1005 slub_debug = DEBUG_DEFAULT_FLAGS;
1006 if (*str++ != '=' || !*str)
1008 * No options specified. Switch on full debugging.
1014 * No options but restriction on slabs. This means full
1015 * debugging for slabs matching a pattern.
1019 if (tolower(*str) == 'o') {
1021 * Avoid enabling debugging on caches if its minimum order
1022 * would increase as a result.
1024 disable_higher_order_debug = 1;
1031 * Switch off all debugging measures.
1036 * Determine which debug features should be switched on
1038 for (; *str && *str != ','; str++) {
1039 switch (tolower(*str)) {
1041 slub_debug |= SLAB_DEBUG_FREE;
1044 slub_debug |= SLAB_RED_ZONE;
1047 slub_debug |= SLAB_POISON;
1050 slub_debug |= SLAB_STORE_USER;
1053 slub_debug |= SLAB_TRACE;
1056 slub_debug |= SLAB_FAILSLAB;
1059 printk(KERN_ERR "slub_debug option '%c' "
1060 "unknown. skipped\n", *str);
1066 slub_debug_slabs = str + 1;
1071 __setup("slub_debug", setup_slub_debug);
1073 static unsigned long kmem_cache_flags(unsigned long objsize,
1074 unsigned long flags, const char *name,
1075 void (*ctor)(void *))
1078 * Enable debugging if selected on the kernel commandline.
1080 if (slub_debug && (!slub_debug_slabs ||
1081 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1082 flags |= slub_debug;
1087 static inline void setup_object_debug(struct kmem_cache *s,
1088 struct page *page, void *object) {}
1090 static inline int alloc_debug_processing(struct kmem_cache *s,
1091 struct page *page, void *object, unsigned long addr) { return 0; }
1093 static inline int free_debug_processing(struct kmem_cache *s,
1094 struct page *page, void *object, unsigned long addr) { return 0; }
1096 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1098 static inline int check_object(struct kmem_cache *s, struct page *page,
1099 void *object, u8 val) { return 1; }
1100 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1101 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1102 unsigned long flags, const char *name,
1103 void (*ctor)(void *))
1107 #define slub_debug 0
1109 #define disable_higher_order_debug 0
1111 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1113 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1115 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1117 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1120 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1123 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1126 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1128 static inline void slab_free_hook_irq(struct kmem_cache *s,
1131 #endif /* CONFIG_SLUB_DEBUG */
1134 * Slab allocation and freeing
1136 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1137 struct kmem_cache_order_objects oo)
1139 int order = oo_order(oo);
1141 flags |= __GFP_NOTRACK;
1143 if (node == NUMA_NO_NODE)
1144 return alloc_pages(flags, order);
1146 return alloc_pages_exact_node(node, flags, order);
1149 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1152 struct kmem_cache_order_objects oo = s->oo;
1155 flags |= s->allocflags;
1158 * Let the initial higher-order allocation fail under memory pressure
1159 * so we fall-back to the minimum order allocation.
1161 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1163 page = alloc_slab_page(alloc_gfp, node, oo);
1164 if (unlikely(!page)) {
1167 * Allocation may have failed due to fragmentation.
1168 * Try a lower order alloc if possible
1170 page = alloc_slab_page(flags, node, oo);
1174 stat(s, ORDER_FALLBACK);
1177 if (kmemcheck_enabled
1178 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1179 int pages = 1 << oo_order(oo);
1181 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1184 * Objects from caches that have a constructor don't get
1185 * cleared when they're allocated, so we need to do it here.
1188 kmemcheck_mark_uninitialized_pages(page, pages);
1190 kmemcheck_mark_unallocated_pages(page, pages);
1193 page->objects = oo_objects(oo);
1194 mod_zone_page_state(page_zone(page),
1195 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1196 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1202 static void setup_object(struct kmem_cache *s, struct page *page,
1205 setup_object_debug(s, page, object);
1206 if (unlikely(s->ctor))
1210 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1217 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1219 page = allocate_slab(s,
1220 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1224 inc_slabs_node(s, page_to_nid(page), page->objects);
1226 page->flags |= 1 << PG_slab;
1228 start = page_address(page);
1230 if (unlikely(s->flags & SLAB_POISON))
1231 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1234 for_each_object(p, s, start, page->objects) {
1235 setup_object(s, page, last);
1236 set_freepointer(s, last, p);
1239 setup_object(s, page, last);
1240 set_freepointer(s, last, NULL);
1242 page->freelist = start;
1248 static void __free_slab(struct kmem_cache *s, struct page *page)
1250 int order = compound_order(page);
1251 int pages = 1 << order;
1253 if (kmem_cache_debug(s)) {
1256 slab_pad_check(s, page);
1257 for_each_object(p, s, page_address(page),
1259 check_object(s, page, p, SLUB_RED_INACTIVE);
1262 kmemcheck_free_shadow(page, compound_order(page));
1264 mod_zone_page_state(page_zone(page),
1265 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1266 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1269 __ClearPageSlab(page);
1270 reset_page_mapcount(page);
1271 if (current->reclaim_state)
1272 current->reclaim_state->reclaimed_slab += pages;
1273 __free_pages(page, order);
1276 static void rcu_free_slab(struct rcu_head *h)
1280 page = container_of((struct list_head *)h, struct page, lru);
1281 __free_slab(page->slab, page);
1284 static void free_slab(struct kmem_cache *s, struct page *page)
1286 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1288 * RCU free overloads the RCU head over the LRU
1290 struct rcu_head *head = (void *)&page->lru;
1292 call_rcu(head, rcu_free_slab);
1294 __free_slab(s, page);
1297 static void discard_slab(struct kmem_cache *s, struct page *page)
1299 dec_slabs_node(s, page_to_nid(page), page->objects);
1304 * Per slab locking using the pagelock
1306 static __always_inline void slab_lock(struct page *page)
1308 bit_spin_lock(PG_locked, &page->flags);
1311 static __always_inline void slab_unlock(struct page *page)
1313 __bit_spin_unlock(PG_locked, &page->flags);
1316 static __always_inline int slab_trylock(struct page *page)
1320 rc = bit_spin_trylock(PG_locked, &page->flags);
1325 * Management of partially allocated slabs
1327 static void add_partial(struct kmem_cache_node *n,
1328 struct page *page, int tail)
1330 spin_lock(&n->list_lock);
1333 list_add_tail(&page->lru, &n->partial);
1335 list_add(&page->lru, &n->partial);
1336 spin_unlock(&n->list_lock);
1339 static inline void __remove_partial(struct kmem_cache_node *n,
1342 list_del(&page->lru);
1346 static void remove_partial(struct kmem_cache *s, struct page *page)
1348 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1350 spin_lock(&n->list_lock);
1351 __remove_partial(n, page);
1352 spin_unlock(&n->list_lock);
1356 * Lock slab and remove from the partial list.
1358 * Must hold list_lock.
1360 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1363 if (slab_trylock(page)) {
1364 __remove_partial(n, page);
1365 __SetPageSlubFrozen(page);
1372 * Try to allocate a partial slab from a specific node.
1374 static struct page *get_partial_node(struct kmem_cache_node *n)
1379 * Racy check. If we mistakenly see no partial slabs then we
1380 * just allocate an empty slab. If we mistakenly try to get a
1381 * partial slab and there is none available then get_partials()
1384 if (!n || !n->nr_partial)
1387 spin_lock(&n->list_lock);
1388 list_for_each_entry(page, &n->partial, lru)
1389 if (lock_and_freeze_slab(n, page))
1393 spin_unlock(&n->list_lock);
1398 * Get a page from somewhere. Search in increasing NUMA distances.
1400 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1403 struct zonelist *zonelist;
1406 enum zone_type high_zoneidx = gfp_zone(flags);
1410 * The defrag ratio allows a configuration of the tradeoffs between
1411 * inter node defragmentation and node local allocations. A lower
1412 * defrag_ratio increases the tendency to do local allocations
1413 * instead of attempting to obtain partial slabs from other nodes.
1415 * If the defrag_ratio is set to 0 then kmalloc() always
1416 * returns node local objects. If the ratio is higher then kmalloc()
1417 * may return off node objects because partial slabs are obtained
1418 * from other nodes and filled up.
1420 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1421 * defrag_ratio = 1000) then every (well almost) allocation will
1422 * first attempt to defrag slab caches on other nodes. This means
1423 * scanning over all nodes to look for partial slabs which may be
1424 * expensive if we do it every time we are trying to find a slab
1425 * with available objects.
1427 if (!s->remote_node_defrag_ratio ||
1428 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1432 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1433 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1434 struct kmem_cache_node *n;
1436 n = get_node(s, zone_to_nid(zone));
1438 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1439 n->nr_partial > s->min_partial) {
1440 page = get_partial_node(n);
1453 * Get a partial page, lock it and return it.
1455 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1458 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1460 page = get_partial_node(get_node(s, searchnode));
1461 if (page || node != -1)
1464 return get_any_partial(s, flags);
1468 * Move a page back to the lists.
1470 * Must be called with the slab lock held.
1472 * On exit the slab lock will have been dropped.
1474 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1477 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1479 __ClearPageSlubFrozen(page);
1482 if (page->freelist) {
1483 add_partial(n, page, tail);
1484 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1486 stat(s, DEACTIVATE_FULL);
1487 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1492 stat(s, DEACTIVATE_EMPTY);
1493 if (n->nr_partial < s->min_partial) {
1495 * Adding an empty slab to the partial slabs in order
1496 * to avoid page allocator overhead. This slab needs
1497 * to come after the other slabs with objects in
1498 * so that the others get filled first. That way the
1499 * size of the partial list stays small.
1501 * kmem_cache_shrink can reclaim any empty slabs from
1504 add_partial(n, page, 1);
1509 discard_slab(s, page);
1515 * Remove the cpu slab
1517 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1520 struct page *page = c->page;
1524 stat(s, DEACTIVATE_REMOTE_FREES);
1526 * Merge cpu freelist into slab freelist. Typically we get here
1527 * because both freelists are empty. So this is unlikely
1530 while (unlikely(c->freelist)) {
1533 tail = 0; /* Hot objects. Put the slab first */
1535 /* Retrieve object from cpu_freelist */
1536 object = c->freelist;
1537 c->freelist = get_freepointer(s, c->freelist);
1539 /* And put onto the regular freelist */
1540 set_freepointer(s, object, page->freelist);
1541 page->freelist = object;
1545 unfreeze_slab(s, page, tail);
1548 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1550 stat(s, CPUSLAB_FLUSH);
1552 deactivate_slab(s, c);
1558 * Called from IPI handler with interrupts disabled.
1560 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1562 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1564 if (likely(c && c->page))
1568 static void flush_cpu_slab(void *d)
1570 struct kmem_cache *s = d;
1572 __flush_cpu_slab(s, smp_processor_id());
1575 static void flush_all(struct kmem_cache *s)
1577 on_each_cpu(flush_cpu_slab, s, 1);
1581 * Check if the objects in a per cpu structure fit numa
1582 * locality expectations.
1584 static inline int node_match(struct kmem_cache_cpu *c, int node)
1587 if (node != NUMA_NO_NODE && c->node != node)
1593 static int count_free(struct page *page)
1595 return page->objects - page->inuse;
1598 static unsigned long count_partial(struct kmem_cache_node *n,
1599 int (*get_count)(struct page *))
1601 unsigned long flags;
1602 unsigned long x = 0;
1605 spin_lock_irqsave(&n->list_lock, flags);
1606 list_for_each_entry(page, &n->partial, lru)
1607 x += get_count(page);
1608 spin_unlock_irqrestore(&n->list_lock, flags);
1612 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1614 #ifdef CONFIG_SLUB_DEBUG
1615 return atomic_long_read(&n->total_objects);
1621 static noinline void
1622 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1627 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1629 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1630 "default order: %d, min order: %d\n", s->name, s->objsize,
1631 s->size, oo_order(s->oo), oo_order(s->min));
1633 if (oo_order(s->min) > get_order(s->objsize))
1634 printk(KERN_WARNING " %s debugging increased min order, use "
1635 "slub_debug=O to disable.\n", s->name);
1637 for_each_online_node(node) {
1638 struct kmem_cache_node *n = get_node(s, node);
1639 unsigned long nr_slabs;
1640 unsigned long nr_objs;
1641 unsigned long nr_free;
1646 nr_free = count_partial(n, count_free);
1647 nr_slabs = node_nr_slabs(n);
1648 nr_objs = node_nr_objs(n);
1651 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1652 node, nr_slabs, nr_objs, nr_free);
1657 * Slow path. The lockless freelist is empty or we need to perform
1660 * Interrupts are disabled.
1662 * Processing is still very fast if new objects have been freed to the
1663 * regular freelist. In that case we simply take over the regular freelist
1664 * as the lockless freelist and zap the regular freelist.
1666 * If that is not working then we fall back to the partial lists. We take the
1667 * first element of the freelist as the object to allocate now and move the
1668 * rest of the freelist to the lockless freelist.
1670 * And if we were unable to get a new slab from the partial slab lists then
1671 * we need to allocate a new slab. This is the slowest path since it involves
1672 * a call to the page allocator and the setup of a new slab.
1674 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1675 unsigned long addr, struct kmem_cache_cpu *c)
1680 /* We handle __GFP_ZERO in the caller */
1681 gfpflags &= ~__GFP_ZERO;
1687 if (unlikely(!node_match(c, node)))
1690 stat(s, ALLOC_REFILL);
1693 object = c->page->freelist;
1694 if (unlikely(!object))
1696 if (kmem_cache_debug(s))
1699 c->freelist = get_freepointer(s, object);
1700 c->page->inuse = c->page->objects;
1701 c->page->freelist = NULL;
1702 c->node = page_to_nid(c->page);
1704 slab_unlock(c->page);
1705 stat(s, ALLOC_SLOWPATH);
1709 deactivate_slab(s, c);
1712 new = get_partial(s, gfpflags, node);
1715 stat(s, ALLOC_FROM_PARTIAL);
1719 gfpflags &= gfp_allowed_mask;
1720 if (gfpflags & __GFP_WAIT)
1723 new = new_slab(s, gfpflags, node);
1725 if (gfpflags & __GFP_WAIT)
1726 local_irq_disable();
1729 c = __this_cpu_ptr(s->cpu_slab);
1730 stat(s, ALLOC_SLAB);
1734 __SetPageSlubFrozen(new);
1738 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1739 slab_out_of_memory(s, gfpflags, node);
1742 if (!alloc_debug_processing(s, c->page, object, addr))
1746 c->page->freelist = get_freepointer(s, object);
1747 c->node = NUMA_NO_NODE;
1752 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1753 * have the fastpath folded into their functions. So no function call
1754 * overhead for requests that can be satisfied on the fastpath.
1756 * The fastpath works by first checking if the lockless freelist can be used.
1757 * If not then __slab_alloc is called for slow processing.
1759 * Otherwise we can simply pick the next object from the lockless free list.
1761 static __always_inline void *slab_alloc(struct kmem_cache *s,
1762 gfp_t gfpflags, int node, unsigned long addr)
1765 struct kmem_cache_cpu *c;
1766 unsigned long flags;
1768 if (slab_pre_alloc_hook(s, gfpflags))
1771 local_irq_save(flags);
1772 c = __this_cpu_ptr(s->cpu_slab);
1773 object = c->freelist;
1774 if (unlikely(!object || !node_match(c, node)))
1776 object = __slab_alloc(s, gfpflags, node, addr, c);
1779 c->freelist = get_freepointer(s, object);
1780 stat(s, ALLOC_FASTPATH);
1782 local_irq_restore(flags);
1784 if (unlikely(gfpflags & __GFP_ZERO) && object)
1785 memset(object, 0, s->objsize);
1787 slab_post_alloc_hook(s, gfpflags, object);
1792 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1794 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1796 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1800 EXPORT_SYMBOL(kmem_cache_alloc);
1802 #ifdef CONFIG_TRACING
1803 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1805 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1806 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
1809 EXPORT_SYMBOL(kmem_cache_alloc_trace);
1811 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1813 void *ret = kmalloc_order(size, flags, order);
1814 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1817 EXPORT_SYMBOL(kmalloc_order_trace);
1821 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1823 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1825 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1826 s->objsize, s->size, gfpflags, node);
1830 EXPORT_SYMBOL(kmem_cache_alloc_node);
1832 #ifdef CONFIG_TRACING
1833 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
1835 int node, size_t size)
1837 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1839 trace_kmalloc_node(_RET_IP_, ret,
1840 size, s->size, gfpflags, node);
1843 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
1848 * Slow patch handling. This may still be called frequently since objects
1849 * have a longer lifetime than the cpu slabs in most processing loads.
1851 * So we still attempt to reduce cache line usage. Just take the slab
1852 * lock and free the item. If there is no additional partial page
1853 * handling required then we can return immediately.
1855 static void __slab_free(struct kmem_cache *s, struct page *page,
1856 void *x, unsigned long addr)
1859 void **object = (void *)x;
1861 stat(s, FREE_SLOWPATH);
1864 if (kmem_cache_debug(s))
1868 prior = page->freelist;
1869 set_freepointer(s, object, prior);
1870 page->freelist = object;
1873 if (unlikely(PageSlubFrozen(page))) {
1874 stat(s, FREE_FROZEN);
1878 if (unlikely(!page->inuse))
1882 * Objects left in the slab. If it was not on the partial list before
1885 if (unlikely(!prior)) {
1886 add_partial(get_node(s, page_to_nid(page)), page, 1);
1887 stat(s, FREE_ADD_PARTIAL);
1897 * Slab still on the partial list.
1899 remove_partial(s, page);
1900 stat(s, FREE_REMOVE_PARTIAL);
1904 discard_slab(s, page);
1908 if (!free_debug_processing(s, page, x, addr))
1914 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1915 * can perform fastpath freeing without additional function calls.
1917 * The fastpath is only possible if we are freeing to the current cpu slab
1918 * of this processor. This typically the case if we have just allocated
1921 * If fastpath is not possible then fall back to __slab_free where we deal
1922 * with all sorts of special processing.
1924 static __always_inline void slab_free(struct kmem_cache *s,
1925 struct page *page, void *x, unsigned long addr)
1927 void **object = (void *)x;
1928 struct kmem_cache_cpu *c;
1929 unsigned long flags;
1931 slab_free_hook(s, x);
1933 local_irq_save(flags);
1934 c = __this_cpu_ptr(s->cpu_slab);
1936 slab_free_hook_irq(s, x);
1938 if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
1939 set_freepointer(s, object, c->freelist);
1940 c->freelist = object;
1941 stat(s, FREE_FASTPATH);
1943 __slab_free(s, page, x, addr);
1945 local_irq_restore(flags);
1948 void kmem_cache_free(struct kmem_cache *s, void *x)
1952 page = virt_to_head_page(x);
1954 slab_free(s, page, x, _RET_IP_);
1956 trace_kmem_cache_free(_RET_IP_, x);
1958 EXPORT_SYMBOL(kmem_cache_free);
1961 * Object placement in a slab is made very easy because we always start at
1962 * offset 0. If we tune the size of the object to the alignment then we can
1963 * get the required alignment by putting one properly sized object after
1966 * Notice that the allocation order determines the sizes of the per cpu
1967 * caches. Each processor has always one slab available for allocations.
1968 * Increasing the allocation order reduces the number of times that slabs
1969 * must be moved on and off the partial lists and is therefore a factor in
1974 * Mininum / Maximum order of slab pages. This influences locking overhead
1975 * and slab fragmentation. A higher order reduces the number of partial slabs
1976 * and increases the number of allocations possible without having to
1977 * take the list_lock.
1979 static int slub_min_order;
1980 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1981 static int slub_min_objects;
1984 * Merge control. If this is set then no merging of slab caches will occur.
1985 * (Could be removed. This was introduced to pacify the merge skeptics.)
1987 static int slub_nomerge;
1990 * Calculate the order of allocation given an slab object size.
1992 * The order of allocation has significant impact on performance and other
1993 * system components. Generally order 0 allocations should be preferred since
1994 * order 0 does not cause fragmentation in the page allocator. Larger objects
1995 * be problematic to put into order 0 slabs because there may be too much
1996 * unused space left. We go to a higher order if more than 1/16th of the slab
1999 * In order to reach satisfactory performance we must ensure that a minimum
2000 * number of objects is in one slab. Otherwise we may generate too much
2001 * activity on the partial lists which requires taking the list_lock. This is
2002 * less a concern for large slabs though which are rarely used.
2004 * slub_max_order specifies the order where we begin to stop considering the
2005 * number of objects in a slab as critical. If we reach slub_max_order then
2006 * we try to keep the page order as low as possible. So we accept more waste
2007 * of space in favor of a small page order.
2009 * Higher order allocations also allow the placement of more objects in a
2010 * slab and thereby reduce object handling overhead. If the user has
2011 * requested a higher mininum order then we start with that one instead of
2012 * the smallest order which will fit the object.
2014 static inline int slab_order(int size, int min_objects,
2015 int max_order, int fract_leftover)
2019 int min_order = slub_min_order;
2021 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
2022 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2024 for (order = max(min_order,
2025 fls(min_objects * size - 1) - PAGE_SHIFT);
2026 order <= max_order; order++) {
2028 unsigned long slab_size = PAGE_SIZE << order;
2030 if (slab_size < min_objects * size)
2033 rem = slab_size % size;
2035 if (rem <= slab_size / fract_leftover)
2043 static inline int calculate_order(int size)
2051 * Attempt to find best configuration for a slab. This
2052 * works by first attempting to generate a layout with
2053 * the best configuration and backing off gradually.
2055 * First we reduce the acceptable waste in a slab. Then
2056 * we reduce the minimum objects required in a slab.
2058 min_objects = slub_min_objects;
2060 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2061 max_objects = (PAGE_SIZE << slub_max_order)/size;
2062 min_objects = min(min_objects, max_objects);
2064 while (min_objects > 1) {
2066 while (fraction >= 4) {
2067 order = slab_order(size, min_objects,
2068 slub_max_order, fraction);
2069 if (order <= slub_max_order)
2077 * We were unable to place multiple objects in a slab. Now
2078 * lets see if we can place a single object there.
2080 order = slab_order(size, 1, slub_max_order, 1);
2081 if (order <= slub_max_order)
2085 * Doh this slab cannot be placed using slub_max_order.
2087 order = slab_order(size, 1, MAX_ORDER, 1);
2088 if (order < MAX_ORDER)
2094 * Figure out what the alignment of the objects will be.
2096 static unsigned long calculate_alignment(unsigned long flags,
2097 unsigned long align, unsigned long size)
2100 * If the user wants hardware cache aligned objects then follow that
2101 * suggestion if the object is sufficiently large.
2103 * The hardware cache alignment cannot override the specified
2104 * alignment though. If that is greater then use it.
2106 if (flags & SLAB_HWCACHE_ALIGN) {
2107 unsigned long ralign = cache_line_size();
2108 while (size <= ralign / 2)
2110 align = max(align, ralign);
2113 if (align < ARCH_SLAB_MINALIGN)
2114 align = ARCH_SLAB_MINALIGN;
2116 return ALIGN(align, sizeof(void *));
2120 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2123 spin_lock_init(&n->list_lock);
2124 INIT_LIST_HEAD(&n->partial);
2125 #ifdef CONFIG_SLUB_DEBUG
2126 atomic_long_set(&n->nr_slabs, 0);
2127 atomic_long_set(&n->total_objects, 0);
2128 INIT_LIST_HEAD(&n->full);
2132 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2134 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2135 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2137 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2139 return s->cpu_slab != NULL;
2142 static struct kmem_cache *kmem_cache_node;
2145 * No kmalloc_node yet so do it by hand. We know that this is the first
2146 * slab on the node for this slabcache. There are no concurrent accesses
2149 * Note that this function only works on the kmalloc_node_cache
2150 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2151 * memory on a fresh node that has no slab structures yet.
2153 static void early_kmem_cache_node_alloc(int node)
2156 struct kmem_cache_node *n;
2157 unsigned long flags;
2159 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2161 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2164 if (page_to_nid(page) != node) {
2165 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2167 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2168 "in order to be able to continue\n");
2173 page->freelist = get_freepointer(kmem_cache_node, n);
2175 kmem_cache_node->node[node] = n;
2176 #ifdef CONFIG_SLUB_DEBUG
2177 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2178 init_tracking(kmem_cache_node, n);
2180 init_kmem_cache_node(n, kmem_cache_node);
2181 inc_slabs_node(kmem_cache_node, node, page->objects);
2184 * lockdep requires consistent irq usage for each lock
2185 * so even though there cannot be a race this early in
2186 * the boot sequence, we still disable irqs.
2188 local_irq_save(flags);
2189 add_partial(n, page, 0);
2190 local_irq_restore(flags);
2193 static void free_kmem_cache_nodes(struct kmem_cache *s)
2197 for_each_node_state(node, N_NORMAL_MEMORY) {
2198 struct kmem_cache_node *n = s->node[node];
2201 kmem_cache_free(kmem_cache_node, n);
2203 s->node[node] = NULL;
2207 static int init_kmem_cache_nodes(struct kmem_cache *s)
2211 for_each_node_state(node, N_NORMAL_MEMORY) {
2212 struct kmem_cache_node *n;
2214 if (slab_state == DOWN) {
2215 early_kmem_cache_node_alloc(node);
2218 n = kmem_cache_alloc_node(kmem_cache_node,
2222 free_kmem_cache_nodes(s);
2227 init_kmem_cache_node(n, s);
2232 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2234 if (min < MIN_PARTIAL)
2236 else if (min > MAX_PARTIAL)
2238 s->min_partial = min;
2242 * calculate_sizes() determines the order and the distribution of data within
2245 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2247 unsigned long flags = s->flags;
2248 unsigned long size = s->objsize;
2249 unsigned long align = s->align;
2253 * Round up object size to the next word boundary. We can only
2254 * place the free pointer at word boundaries and this determines
2255 * the possible location of the free pointer.
2257 size = ALIGN(size, sizeof(void *));
2259 #ifdef CONFIG_SLUB_DEBUG
2261 * Determine if we can poison the object itself. If the user of
2262 * the slab may touch the object after free or before allocation
2263 * then we should never poison the object itself.
2265 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2267 s->flags |= __OBJECT_POISON;
2269 s->flags &= ~__OBJECT_POISON;
2273 * If we are Redzoning then check if there is some space between the
2274 * end of the object and the free pointer. If not then add an
2275 * additional word to have some bytes to store Redzone information.
2277 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2278 size += sizeof(void *);
2282 * With that we have determined the number of bytes in actual use
2283 * by the object. This is the potential offset to the free pointer.
2287 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2290 * Relocate free pointer after the object if it is not
2291 * permitted to overwrite the first word of the object on
2294 * This is the case if we do RCU, have a constructor or
2295 * destructor or are poisoning the objects.
2298 size += sizeof(void *);
2301 #ifdef CONFIG_SLUB_DEBUG
2302 if (flags & SLAB_STORE_USER)
2304 * Need to store information about allocs and frees after
2307 size += 2 * sizeof(struct track);
2309 if (flags & SLAB_RED_ZONE)
2311 * Add some empty padding so that we can catch
2312 * overwrites from earlier objects rather than let
2313 * tracking information or the free pointer be
2314 * corrupted if a user writes before the start
2317 size += sizeof(void *);
2321 * Determine the alignment based on various parameters that the
2322 * user specified and the dynamic determination of cache line size
2325 align = calculate_alignment(flags, align, s->objsize);
2329 * SLUB stores one object immediately after another beginning from
2330 * offset 0. In order to align the objects we have to simply size
2331 * each object to conform to the alignment.
2333 size = ALIGN(size, align);
2335 if (forced_order >= 0)
2336 order = forced_order;
2338 order = calculate_order(size);
2345 s->allocflags |= __GFP_COMP;
2347 if (s->flags & SLAB_CACHE_DMA)
2348 s->allocflags |= SLUB_DMA;
2350 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2351 s->allocflags |= __GFP_RECLAIMABLE;
2354 * Determine the number of objects per slab
2356 s->oo = oo_make(order, size);
2357 s->min = oo_make(get_order(size), size);
2358 if (oo_objects(s->oo) > oo_objects(s->max))
2361 return !!oo_objects(s->oo);
2365 static int kmem_cache_open(struct kmem_cache *s,
2366 const char *name, size_t size,
2367 size_t align, unsigned long flags,
2368 void (*ctor)(void *))
2370 memset(s, 0, kmem_size);
2375 s->flags = kmem_cache_flags(size, flags, name, ctor);
2377 if (!calculate_sizes(s, -1))
2379 if (disable_higher_order_debug) {
2381 * Disable debugging flags that store metadata if the min slab
2384 if (get_order(s->size) > get_order(s->objsize)) {
2385 s->flags &= ~DEBUG_METADATA_FLAGS;
2387 if (!calculate_sizes(s, -1))
2393 * The larger the object size is, the more pages we want on the partial
2394 * list to avoid pounding the page allocator excessively.
2396 set_min_partial(s, ilog2(s->size));
2399 s->remote_node_defrag_ratio = 1000;
2401 if (!init_kmem_cache_nodes(s))
2404 if (alloc_kmem_cache_cpus(s))
2407 free_kmem_cache_nodes(s);
2409 if (flags & SLAB_PANIC)
2410 panic("Cannot create slab %s size=%lu realsize=%u "
2411 "order=%u offset=%u flags=%lx\n",
2412 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2418 * Determine the size of a slab object
2420 unsigned int kmem_cache_size(struct kmem_cache *s)
2424 EXPORT_SYMBOL(kmem_cache_size);
2426 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2429 #ifdef CONFIG_SLUB_DEBUG
2430 void *addr = page_address(page);
2432 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2433 sizeof(long), GFP_ATOMIC);
2436 slab_err(s, page, "%s", text);
2438 for_each_free_object(p, s, page->freelist)
2439 set_bit(slab_index(p, s, addr), map);
2441 for_each_object(p, s, addr, page->objects) {
2443 if (!test_bit(slab_index(p, s, addr), map)) {
2444 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2446 print_tracking(s, p);
2455 * Attempt to free all partial slabs on a node.
2457 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2459 unsigned long flags;
2460 struct page *page, *h;
2462 spin_lock_irqsave(&n->list_lock, flags);
2463 list_for_each_entry_safe(page, h, &n->partial, lru) {
2465 __remove_partial(n, page);
2466 discard_slab(s, page);
2468 list_slab_objects(s, page,
2469 "Objects remaining on kmem_cache_close()");
2472 spin_unlock_irqrestore(&n->list_lock, flags);
2476 * Release all resources used by a slab cache.
2478 static inline int kmem_cache_close(struct kmem_cache *s)
2483 free_percpu(s->cpu_slab);
2484 /* Attempt to free all objects */
2485 for_each_node_state(node, N_NORMAL_MEMORY) {
2486 struct kmem_cache_node *n = get_node(s, node);
2489 if (n->nr_partial || slabs_node(s, node))
2492 free_kmem_cache_nodes(s);
2497 * Close a cache and release the kmem_cache structure
2498 * (must be used for caches created using kmem_cache_create)
2500 void kmem_cache_destroy(struct kmem_cache *s)
2502 down_write(&slub_lock);
2506 if (kmem_cache_close(s)) {
2507 printk(KERN_ERR "SLUB %s: %s called for cache that "
2508 "still has objects.\n", s->name, __func__);
2511 if (s->flags & SLAB_DESTROY_BY_RCU)
2513 sysfs_slab_remove(s);
2515 up_write(&slub_lock);
2517 EXPORT_SYMBOL(kmem_cache_destroy);
2519 /********************************************************************
2521 *******************************************************************/
2523 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2524 EXPORT_SYMBOL(kmalloc_caches);
2526 static struct kmem_cache *kmem_cache;
2528 #ifdef CONFIG_ZONE_DMA
2529 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2532 static int __init setup_slub_min_order(char *str)
2534 get_option(&str, &slub_min_order);
2539 __setup("slub_min_order=", setup_slub_min_order);
2541 static int __init setup_slub_max_order(char *str)
2543 get_option(&str, &slub_max_order);
2544 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2549 __setup("slub_max_order=", setup_slub_max_order);
2551 static int __init setup_slub_min_objects(char *str)
2553 get_option(&str, &slub_min_objects);
2558 __setup("slub_min_objects=", setup_slub_min_objects);
2560 static int __init setup_slub_nomerge(char *str)
2566 __setup("slub_nomerge", setup_slub_nomerge);
2568 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2569 int size, unsigned int flags)
2571 struct kmem_cache *s;
2573 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2576 * This function is called with IRQs disabled during early-boot on
2577 * single CPU so there's no need to take slub_lock here.
2579 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2583 list_add(&s->list, &slab_caches);
2587 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2592 * Conversion table for small slabs sizes / 8 to the index in the
2593 * kmalloc array. This is necessary for slabs < 192 since we have non power
2594 * of two cache sizes there. The size of larger slabs can be determined using
2597 static s8 size_index[24] = {
2624 static inline int size_index_elem(size_t bytes)
2626 return (bytes - 1) / 8;
2629 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2635 return ZERO_SIZE_PTR;
2637 index = size_index[size_index_elem(size)];
2639 index = fls(size - 1);
2641 #ifdef CONFIG_ZONE_DMA
2642 if (unlikely((flags & SLUB_DMA)))
2643 return kmalloc_dma_caches[index];
2646 return kmalloc_caches[index];
2649 void *__kmalloc(size_t size, gfp_t flags)
2651 struct kmem_cache *s;
2654 if (unlikely(size > SLUB_MAX_SIZE))
2655 return kmalloc_large(size, flags);
2657 s = get_slab(size, flags);
2659 if (unlikely(ZERO_OR_NULL_PTR(s)))
2662 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2664 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2668 EXPORT_SYMBOL(__kmalloc);
2671 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2676 flags |= __GFP_COMP | __GFP_NOTRACK;
2677 page = alloc_pages_node(node, flags, get_order(size));
2679 ptr = page_address(page);
2681 kmemleak_alloc(ptr, size, 1, flags);
2685 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2687 struct kmem_cache *s;
2690 if (unlikely(size > SLUB_MAX_SIZE)) {
2691 ret = kmalloc_large_node(size, flags, node);
2693 trace_kmalloc_node(_RET_IP_, ret,
2694 size, PAGE_SIZE << get_order(size),
2700 s = get_slab(size, flags);
2702 if (unlikely(ZERO_OR_NULL_PTR(s)))
2705 ret = slab_alloc(s, flags, node, _RET_IP_);
2707 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2711 EXPORT_SYMBOL(__kmalloc_node);
2714 size_t ksize(const void *object)
2718 if (unlikely(object == ZERO_SIZE_PTR))
2721 page = virt_to_head_page(object);
2723 if (unlikely(!PageSlab(page))) {
2724 WARN_ON(!PageCompound(page));
2725 return PAGE_SIZE << compound_order(page);
2728 return slab_ksize(page->slab);
2730 EXPORT_SYMBOL(ksize);
2732 void kfree(const void *x)
2735 void *object = (void *)x;
2737 trace_kfree(_RET_IP_, x);
2739 if (unlikely(ZERO_OR_NULL_PTR(x)))
2742 page = virt_to_head_page(x);
2743 if (unlikely(!PageSlab(page))) {
2744 BUG_ON(!PageCompound(page));
2749 slab_free(page->slab, page, object, _RET_IP_);
2751 EXPORT_SYMBOL(kfree);
2754 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2755 * the remaining slabs by the number of items in use. The slabs with the
2756 * most items in use come first. New allocations will then fill those up
2757 * and thus they can be removed from the partial lists.
2759 * The slabs with the least items are placed last. This results in them
2760 * being allocated from last increasing the chance that the last objects
2761 * are freed in them.
2763 int kmem_cache_shrink(struct kmem_cache *s)
2767 struct kmem_cache_node *n;
2770 int objects = oo_objects(s->max);
2771 struct list_head *slabs_by_inuse =
2772 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2773 unsigned long flags;
2775 if (!slabs_by_inuse)
2779 for_each_node_state(node, N_NORMAL_MEMORY) {
2780 n = get_node(s, node);
2785 for (i = 0; i < objects; i++)
2786 INIT_LIST_HEAD(slabs_by_inuse + i);
2788 spin_lock_irqsave(&n->list_lock, flags);
2791 * Build lists indexed by the items in use in each slab.
2793 * Note that concurrent frees may occur while we hold the
2794 * list_lock. page->inuse here is the upper limit.
2796 list_for_each_entry_safe(page, t, &n->partial, lru) {
2797 if (!page->inuse && slab_trylock(page)) {
2799 * Must hold slab lock here because slab_free
2800 * may have freed the last object and be
2801 * waiting to release the slab.
2803 __remove_partial(n, page);
2805 discard_slab(s, page);
2807 list_move(&page->lru,
2808 slabs_by_inuse + page->inuse);
2813 * Rebuild the partial list with the slabs filled up most
2814 * first and the least used slabs at the end.
2816 for (i = objects - 1; i >= 0; i--)
2817 list_splice(slabs_by_inuse + i, n->partial.prev);
2819 spin_unlock_irqrestore(&n->list_lock, flags);
2822 kfree(slabs_by_inuse);
2825 EXPORT_SYMBOL(kmem_cache_shrink);
2827 #if defined(CONFIG_MEMORY_HOTPLUG)
2828 static int slab_mem_going_offline_callback(void *arg)
2830 struct kmem_cache *s;
2832 down_read(&slub_lock);
2833 list_for_each_entry(s, &slab_caches, list)
2834 kmem_cache_shrink(s);
2835 up_read(&slub_lock);
2840 static void slab_mem_offline_callback(void *arg)
2842 struct kmem_cache_node *n;
2843 struct kmem_cache *s;
2844 struct memory_notify *marg = arg;
2847 offline_node = marg->status_change_nid;
2850 * If the node still has available memory. we need kmem_cache_node
2853 if (offline_node < 0)
2856 down_read(&slub_lock);
2857 list_for_each_entry(s, &slab_caches, list) {
2858 n = get_node(s, offline_node);
2861 * if n->nr_slabs > 0, slabs still exist on the node
2862 * that is going down. We were unable to free them,
2863 * and offline_pages() function shouldn't call this
2864 * callback. So, we must fail.
2866 BUG_ON(slabs_node(s, offline_node));
2868 s->node[offline_node] = NULL;
2869 kmem_cache_free(kmem_cache_node, n);
2872 up_read(&slub_lock);
2875 static int slab_mem_going_online_callback(void *arg)
2877 struct kmem_cache_node *n;
2878 struct kmem_cache *s;
2879 struct memory_notify *marg = arg;
2880 int nid = marg->status_change_nid;
2884 * If the node's memory is already available, then kmem_cache_node is
2885 * already created. Nothing to do.
2891 * We are bringing a node online. No memory is available yet. We must
2892 * allocate a kmem_cache_node structure in order to bring the node
2895 down_read(&slub_lock);
2896 list_for_each_entry(s, &slab_caches, list) {
2898 * XXX: kmem_cache_alloc_node will fallback to other nodes
2899 * since memory is not yet available from the node that
2902 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
2907 init_kmem_cache_node(n, s);
2911 up_read(&slub_lock);
2915 static int slab_memory_callback(struct notifier_block *self,
2916 unsigned long action, void *arg)
2921 case MEM_GOING_ONLINE:
2922 ret = slab_mem_going_online_callback(arg);
2924 case MEM_GOING_OFFLINE:
2925 ret = slab_mem_going_offline_callback(arg);
2928 case MEM_CANCEL_ONLINE:
2929 slab_mem_offline_callback(arg);
2932 case MEM_CANCEL_OFFLINE:
2936 ret = notifier_from_errno(ret);
2942 #endif /* CONFIG_MEMORY_HOTPLUG */
2944 /********************************************************************
2945 * Basic setup of slabs
2946 *******************************************************************/
2949 * Used for early kmem_cache structures that were allocated using
2950 * the page allocator
2953 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
2957 list_add(&s->list, &slab_caches);
2960 for_each_node_state(node, N_NORMAL_MEMORY) {
2961 struct kmem_cache_node *n = get_node(s, node);
2965 list_for_each_entry(p, &n->partial, lru)
2968 #ifdef CONFIG_SLAB_DEBUG
2969 list_for_each_entry(p, &n->full, lru)
2976 void __init kmem_cache_init(void)
2980 struct kmem_cache *temp_kmem_cache;
2982 struct kmem_cache *temp_kmem_cache_node;
2983 unsigned long kmalloc_size;
2985 kmem_size = offsetof(struct kmem_cache, node) +
2986 nr_node_ids * sizeof(struct kmem_cache_node *);
2988 /* Allocate two kmem_caches from the page allocator */
2989 kmalloc_size = ALIGN(kmem_size, cache_line_size());
2990 order = get_order(2 * kmalloc_size);
2991 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
2994 * Must first have the slab cache available for the allocations of the
2995 * struct kmem_cache_node's. There is special bootstrap code in
2996 * kmem_cache_open for slab_state == DOWN.
2998 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3000 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3001 sizeof(struct kmem_cache_node),
3002 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3004 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3006 /* Able to allocate the per node structures */
3007 slab_state = PARTIAL;
3009 temp_kmem_cache = kmem_cache;
3010 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3011 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3012 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3013 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3016 * Allocate kmem_cache_node properly from the kmem_cache slab.
3017 * kmem_cache_node is separately allocated so no need to
3018 * update any list pointers.
3020 temp_kmem_cache_node = kmem_cache_node;
3022 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3023 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3025 kmem_cache_bootstrap_fixup(kmem_cache_node);
3028 kmem_cache_bootstrap_fixup(kmem_cache);
3030 /* Free temporary boot structure */
3031 free_pages((unsigned long)temp_kmem_cache, order);
3033 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3036 * Patch up the size_index table if we have strange large alignment
3037 * requirements for the kmalloc array. This is only the case for
3038 * MIPS it seems. The standard arches will not generate any code here.
3040 * Largest permitted alignment is 256 bytes due to the way we
3041 * handle the index determination for the smaller caches.
3043 * Make sure that nothing crazy happens if someone starts tinkering
3044 * around with ARCH_KMALLOC_MINALIGN
3046 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3047 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3049 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3050 int elem = size_index_elem(i);
3051 if (elem >= ARRAY_SIZE(size_index))
3053 size_index[elem] = KMALLOC_SHIFT_LOW;
3056 if (KMALLOC_MIN_SIZE == 64) {
3058 * The 96 byte size cache is not used if the alignment
3061 for (i = 64 + 8; i <= 96; i += 8)
3062 size_index[size_index_elem(i)] = 7;
3063 } else if (KMALLOC_MIN_SIZE == 128) {
3065 * The 192 byte sized cache is not used if the alignment
3066 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3069 for (i = 128 + 8; i <= 192; i += 8)
3070 size_index[size_index_elem(i)] = 8;
3073 /* Caches that are not of the two-to-the-power-of size */
3074 if (KMALLOC_MIN_SIZE <= 32) {
3075 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3079 if (KMALLOC_MIN_SIZE <= 64) {
3080 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3084 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3085 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3091 /* Provide the correct kmalloc names now that the caches are up */
3092 if (KMALLOC_MIN_SIZE <= 32) {
3093 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3094 BUG_ON(!kmalloc_caches[1]->name);
3097 if (KMALLOC_MIN_SIZE <= 64) {
3098 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3099 BUG_ON(!kmalloc_caches[2]->name);
3102 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3103 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3106 kmalloc_caches[i]->name = s;
3110 register_cpu_notifier(&slab_notifier);
3113 #ifdef CONFIG_ZONE_DMA
3114 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3115 struct kmem_cache *s = kmalloc_caches[i];
3118 char *name = kasprintf(GFP_NOWAIT,
3119 "dma-kmalloc-%d", s->objsize);
3122 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3123 s->objsize, SLAB_CACHE_DMA);
3128 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3129 " CPUs=%d, Nodes=%d\n",
3130 caches, cache_line_size(),
3131 slub_min_order, slub_max_order, slub_min_objects,
3132 nr_cpu_ids, nr_node_ids);
3135 void __init kmem_cache_init_late(void)
3140 * Find a mergeable slab cache
3142 static int slab_unmergeable(struct kmem_cache *s)
3144 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3151 * We may have set a slab to be unmergeable during bootstrap.
3153 if (s->refcount < 0)
3159 static struct kmem_cache *find_mergeable(size_t size,
3160 size_t align, unsigned long flags, const char *name,
3161 void (*ctor)(void *))
3163 struct kmem_cache *s;
3165 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3171 size = ALIGN(size, sizeof(void *));
3172 align = calculate_alignment(flags, align, size);
3173 size = ALIGN(size, align);
3174 flags = kmem_cache_flags(size, flags, name, NULL);
3176 list_for_each_entry(s, &slab_caches, list) {
3177 if (slab_unmergeable(s))
3183 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3186 * Check if alignment is compatible.
3187 * Courtesy of Adrian Drzewiecki
3189 if ((s->size & ~(align - 1)) != s->size)
3192 if (s->size - size >= sizeof(void *))
3200 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3201 size_t align, unsigned long flags, void (*ctor)(void *))
3203 struct kmem_cache *s;
3209 down_write(&slub_lock);
3210 s = find_mergeable(size, align, flags, name, ctor);
3214 * Adjust the object sizes so that we clear
3215 * the complete object on kzalloc.
3217 s->objsize = max(s->objsize, (int)size);
3218 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3220 if (sysfs_slab_alias(s, name)) {
3224 up_write(&slub_lock);
3228 n = kstrdup(name, GFP_KERNEL);
3232 s = kmalloc(kmem_size, GFP_KERNEL);
3234 if (kmem_cache_open(s, n,
3235 size, align, flags, ctor)) {
3236 list_add(&s->list, &slab_caches);
3237 if (sysfs_slab_add(s)) {
3243 up_write(&slub_lock);
3250 up_write(&slub_lock);
3252 if (flags & SLAB_PANIC)
3253 panic("Cannot create slabcache %s\n", name);
3258 EXPORT_SYMBOL(kmem_cache_create);
3262 * Use the cpu notifier to insure that the cpu slabs are flushed when
3265 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3266 unsigned long action, void *hcpu)
3268 long cpu = (long)hcpu;
3269 struct kmem_cache *s;
3270 unsigned long flags;
3273 case CPU_UP_CANCELED:
3274 case CPU_UP_CANCELED_FROZEN:
3276 case CPU_DEAD_FROZEN:
3277 down_read(&slub_lock);
3278 list_for_each_entry(s, &slab_caches, list) {
3279 local_irq_save(flags);
3280 __flush_cpu_slab(s, cpu);
3281 local_irq_restore(flags);
3283 up_read(&slub_lock);
3291 static struct notifier_block __cpuinitdata slab_notifier = {
3292 .notifier_call = slab_cpuup_callback
3297 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3299 struct kmem_cache *s;
3302 if (unlikely(size > SLUB_MAX_SIZE))
3303 return kmalloc_large(size, gfpflags);
3305 s = get_slab(size, gfpflags);
3307 if (unlikely(ZERO_OR_NULL_PTR(s)))
3310 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3312 /* Honor the call site pointer we recieved. */
3313 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3319 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3320 int node, unsigned long caller)
3322 struct kmem_cache *s;
3325 if (unlikely(size > SLUB_MAX_SIZE)) {
3326 ret = kmalloc_large_node(size, gfpflags, node);
3328 trace_kmalloc_node(caller, ret,
3329 size, PAGE_SIZE << get_order(size),
3335 s = get_slab(size, gfpflags);
3337 if (unlikely(ZERO_OR_NULL_PTR(s)))
3340 ret = slab_alloc(s, gfpflags, node, caller);
3342 /* Honor the call site pointer we recieved. */
3343 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3350 static int count_inuse(struct page *page)
3355 static int count_total(struct page *page)
3357 return page->objects;
3361 #ifdef CONFIG_SLUB_DEBUG
3362 static int validate_slab(struct kmem_cache *s, struct page *page,
3366 void *addr = page_address(page);
3368 if (!check_slab(s, page) ||
3369 !on_freelist(s, page, NULL))
3372 /* Now we know that a valid freelist exists */
3373 bitmap_zero(map, page->objects);
3375 for_each_free_object(p, s, page->freelist) {
3376 set_bit(slab_index(p, s, addr), map);
3377 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3381 for_each_object(p, s, addr, page->objects)
3382 if (!test_bit(slab_index(p, s, addr), map))
3383 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3388 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3391 if (slab_trylock(page)) {
3392 validate_slab(s, page, map);
3395 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3399 static int validate_slab_node(struct kmem_cache *s,
3400 struct kmem_cache_node *n, unsigned long *map)
3402 unsigned long count = 0;
3404 unsigned long flags;
3406 spin_lock_irqsave(&n->list_lock, flags);
3408 list_for_each_entry(page, &n->partial, lru) {
3409 validate_slab_slab(s, page, map);
3412 if (count != n->nr_partial)
3413 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3414 "counter=%ld\n", s->name, count, n->nr_partial);
3416 if (!(s->flags & SLAB_STORE_USER))
3419 list_for_each_entry(page, &n->full, lru) {
3420 validate_slab_slab(s, page, map);
3423 if (count != atomic_long_read(&n->nr_slabs))
3424 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3425 "counter=%ld\n", s->name, count,
3426 atomic_long_read(&n->nr_slabs));
3429 spin_unlock_irqrestore(&n->list_lock, flags);
3433 static long validate_slab_cache(struct kmem_cache *s)
3436 unsigned long count = 0;
3437 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3438 sizeof(unsigned long), GFP_KERNEL);
3444 for_each_node_state(node, N_NORMAL_MEMORY) {
3445 struct kmem_cache_node *n = get_node(s, node);
3447 count += validate_slab_node(s, n, map);
3453 * Generate lists of code addresses where slabcache objects are allocated
3458 unsigned long count;
3465 DECLARE_BITMAP(cpus, NR_CPUS);
3471 unsigned long count;
3472 struct location *loc;
3475 static void free_loc_track(struct loc_track *t)
3478 free_pages((unsigned long)t->loc,
3479 get_order(sizeof(struct location) * t->max));
3482 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3487 order = get_order(sizeof(struct location) * max);
3489 l = (void *)__get_free_pages(flags, order);
3494 memcpy(l, t->loc, sizeof(struct location) * t->count);
3502 static int add_location(struct loc_track *t, struct kmem_cache *s,
3503 const struct track *track)
3505 long start, end, pos;
3507 unsigned long caddr;
3508 unsigned long age = jiffies - track->when;
3514 pos = start + (end - start + 1) / 2;
3517 * There is nothing at "end". If we end up there
3518 * we need to add something to before end.
3523 caddr = t->loc[pos].addr;
3524 if (track->addr == caddr) {
3530 if (age < l->min_time)
3532 if (age > l->max_time)
3535 if (track->pid < l->min_pid)
3536 l->min_pid = track->pid;
3537 if (track->pid > l->max_pid)
3538 l->max_pid = track->pid;
3540 cpumask_set_cpu(track->cpu,
3541 to_cpumask(l->cpus));
3543 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3547 if (track->addr < caddr)
3554 * Not found. Insert new tracking element.
3556 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3562 (t->count - pos) * sizeof(struct location));
3565 l->addr = track->addr;
3569 l->min_pid = track->pid;
3570 l->max_pid = track->pid;
3571 cpumask_clear(to_cpumask(l->cpus));
3572 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3573 nodes_clear(l->nodes);
3574 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3578 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3579 struct page *page, enum track_item alloc,
3582 void *addr = page_address(page);
3585 bitmap_zero(map, page->objects);
3586 for_each_free_object(p, s, page->freelist)
3587 set_bit(slab_index(p, s, addr), map);
3589 for_each_object(p, s, addr, page->objects)
3590 if (!test_bit(slab_index(p, s, addr), map))
3591 add_location(t, s, get_track(s, p, alloc));
3594 static int list_locations(struct kmem_cache *s, char *buf,
3595 enum track_item alloc)
3599 struct loc_track t = { 0, 0, NULL };
3601 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3602 sizeof(unsigned long), GFP_KERNEL);
3604 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3607 return sprintf(buf, "Out of memory\n");
3609 /* Push back cpu slabs */
3612 for_each_node_state(node, N_NORMAL_MEMORY) {
3613 struct kmem_cache_node *n = get_node(s, node);
3614 unsigned long flags;
3617 if (!atomic_long_read(&n->nr_slabs))
3620 spin_lock_irqsave(&n->list_lock, flags);
3621 list_for_each_entry(page, &n->partial, lru)
3622 process_slab(&t, s, page, alloc, map);
3623 list_for_each_entry(page, &n->full, lru)
3624 process_slab(&t, s, page, alloc, map);
3625 spin_unlock_irqrestore(&n->list_lock, flags);
3628 for (i = 0; i < t.count; i++) {
3629 struct location *l = &t.loc[i];
3631 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3633 len += sprintf(buf + len, "%7ld ", l->count);
3636 len += sprintf(buf + len, "%pS", (void *)l->addr);
3638 len += sprintf(buf + len, "<not-available>");
3640 if (l->sum_time != l->min_time) {
3641 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3643 (long)div_u64(l->sum_time, l->count),
3646 len += sprintf(buf + len, " age=%ld",
3649 if (l->min_pid != l->max_pid)
3650 len += sprintf(buf + len, " pid=%ld-%ld",
3651 l->min_pid, l->max_pid);
3653 len += sprintf(buf + len, " pid=%ld",
3656 if (num_online_cpus() > 1 &&
3657 !cpumask_empty(to_cpumask(l->cpus)) &&
3658 len < PAGE_SIZE - 60) {
3659 len += sprintf(buf + len, " cpus=");
3660 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3661 to_cpumask(l->cpus));
3664 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3665 len < PAGE_SIZE - 60) {
3666 len += sprintf(buf + len, " nodes=");
3667 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3671 len += sprintf(buf + len, "\n");
3677 len += sprintf(buf, "No data\n");
3682 #ifdef SLUB_RESILIENCY_TEST
3683 static void resiliency_test(void)
3687 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3689 printk(KERN_ERR "SLUB resiliency testing\n");
3690 printk(KERN_ERR "-----------------------\n");
3691 printk(KERN_ERR "A. Corruption after allocation\n");
3693 p = kzalloc(16, GFP_KERNEL);
3695 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3696 " 0x12->0x%p\n\n", p + 16);
3698 validate_slab_cache(kmalloc_caches[4]);
3700 /* Hmmm... The next two are dangerous */
3701 p = kzalloc(32, GFP_KERNEL);
3702 p[32 + sizeof(void *)] = 0x34;
3703 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3704 " 0x34 -> -0x%p\n", p);
3706 "If allocated object is overwritten then not detectable\n\n");
3708 validate_slab_cache(kmalloc_caches[5]);
3709 p = kzalloc(64, GFP_KERNEL);
3710 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3712 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3715 "If allocated object is overwritten then not detectable\n\n");
3716 validate_slab_cache(kmalloc_caches[6]);
3718 printk(KERN_ERR "\nB. Corruption after free\n");
3719 p = kzalloc(128, GFP_KERNEL);
3722 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3723 validate_slab_cache(kmalloc_caches[7]);
3725 p = kzalloc(256, GFP_KERNEL);
3728 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3730 validate_slab_cache(kmalloc_caches[8]);
3732 p = kzalloc(512, GFP_KERNEL);
3735 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3736 validate_slab_cache(kmalloc_caches[9]);
3740 static void resiliency_test(void) {};
3745 enum slab_stat_type {
3746 SL_ALL, /* All slabs */
3747 SL_PARTIAL, /* Only partially allocated slabs */
3748 SL_CPU, /* Only slabs used for cpu caches */
3749 SL_OBJECTS, /* Determine allocated objects not slabs */
3750 SL_TOTAL /* Determine object capacity not slabs */
3753 #define SO_ALL (1 << SL_ALL)
3754 #define SO_PARTIAL (1 << SL_PARTIAL)
3755 #define SO_CPU (1 << SL_CPU)
3756 #define SO_OBJECTS (1 << SL_OBJECTS)
3757 #define SO_TOTAL (1 << SL_TOTAL)
3759 static ssize_t show_slab_objects(struct kmem_cache *s,
3760 char *buf, unsigned long flags)
3762 unsigned long total = 0;
3765 unsigned long *nodes;
3766 unsigned long *per_cpu;
3768 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3771 per_cpu = nodes + nr_node_ids;
3773 if (flags & SO_CPU) {
3776 for_each_possible_cpu(cpu) {
3777 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3779 if (!c || c->node < 0)
3783 if (flags & SO_TOTAL)
3784 x = c->page->objects;
3785 else if (flags & SO_OBJECTS)
3791 nodes[c->node] += x;
3797 lock_memory_hotplug();
3798 #ifdef CONFIG_SLUB_DEBUG
3799 if (flags & SO_ALL) {
3800 for_each_node_state(node, N_NORMAL_MEMORY) {
3801 struct kmem_cache_node *n = get_node(s, node);
3803 if (flags & SO_TOTAL)
3804 x = atomic_long_read(&n->total_objects);
3805 else if (flags & SO_OBJECTS)
3806 x = atomic_long_read(&n->total_objects) -
3807 count_partial(n, count_free);
3810 x = atomic_long_read(&n->nr_slabs);
3817 if (flags & SO_PARTIAL) {
3818 for_each_node_state(node, N_NORMAL_MEMORY) {
3819 struct kmem_cache_node *n = get_node(s, node);
3821 if (flags & SO_TOTAL)
3822 x = count_partial(n, count_total);
3823 else if (flags & SO_OBJECTS)
3824 x = count_partial(n, count_inuse);
3831 x = sprintf(buf, "%lu", total);
3833 for_each_node_state(node, N_NORMAL_MEMORY)
3835 x += sprintf(buf + x, " N%d=%lu",
3838 unlock_memory_hotplug();
3840 return x + sprintf(buf + x, "\n");
3843 #ifdef CONFIG_SLUB_DEBUG
3844 static int any_slab_objects(struct kmem_cache *s)
3848 for_each_online_node(node) {
3849 struct kmem_cache_node *n = get_node(s, node);
3854 if (atomic_long_read(&n->total_objects))
3861 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3862 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3864 struct slab_attribute {
3865 struct attribute attr;
3866 ssize_t (*show)(struct kmem_cache *s, char *buf);
3867 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3870 #define SLAB_ATTR_RO(_name) \
3871 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3873 #define SLAB_ATTR(_name) \
3874 static struct slab_attribute _name##_attr = \
3875 __ATTR(_name, 0644, _name##_show, _name##_store)
3877 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3879 return sprintf(buf, "%d\n", s->size);
3881 SLAB_ATTR_RO(slab_size);
3883 static ssize_t align_show(struct kmem_cache *s, char *buf)
3885 return sprintf(buf, "%d\n", s->align);
3887 SLAB_ATTR_RO(align);
3889 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3891 return sprintf(buf, "%d\n", s->objsize);
3893 SLAB_ATTR_RO(object_size);
3895 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3897 return sprintf(buf, "%d\n", oo_objects(s->oo));
3899 SLAB_ATTR_RO(objs_per_slab);
3901 static ssize_t order_store(struct kmem_cache *s,
3902 const char *buf, size_t length)
3904 unsigned long order;
3907 err = strict_strtoul(buf, 10, &order);
3911 if (order > slub_max_order || order < slub_min_order)
3914 calculate_sizes(s, order);
3918 static ssize_t order_show(struct kmem_cache *s, char *buf)
3920 return sprintf(buf, "%d\n", oo_order(s->oo));
3924 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3926 return sprintf(buf, "%lu\n", s->min_partial);
3929 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3935 err = strict_strtoul(buf, 10, &min);
3939 set_min_partial(s, min);
3942 SLAB_ATTR(min_partial);
3944 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3948 return sprintf(buf, "%pS\n", s->ctor);
3952 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3954 return sprintf(buf, "%d\n", s->refcount - 1);
3956 SLAB_ATTR_RO(aliases);
3958 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3960 return show_slab_objects(s, buf, SO_PARTIAL);
3962 SLAB_ATTR_RO(partial);
3964 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3966 return show_slab_objects(s, buf, SO_CPU);
3968 SLAB_ATTR_RO(cpu_slabs);
3970 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3972 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3974 SLAB_ATTR_RO(objects);
3976 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3978 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3980 SLAB_ATTR_RO(objects_partial);
3982 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3984 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3987 static ssize_t reclaim_account_store(struct kmem_cache *s,
3988 const char *buf, size_t length)
3990 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3992 s->flags |= SLAB_RECLAIM_ACCOUNT;
3995 SLAB_ATTR(reclaim_account);
3997 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3999 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4001 SLAB_ATTR_RO(hwcache_align);
4003 #ifdef CONFIG_ZONE_DMA
4004 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4006 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4008 SLAB_ATTR_RO(cache_dma);
4011 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4013 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4015 SLAB_ATTR_RO(destroy_by_rcu);
4017 #ifdef CONFIG_SLUB_DEBUG
4018 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4020 return show_slab_objects(s, buf, SO_ALL);
4022 SLAB_ATTR_RO(slabs);
4024 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4026 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4028 SLAB_ATTR_RO(total_objects);
4030 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4032 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4035 static ssize_t sanity_checks_store(struct kmem_cache *s,
4036 const char *buf, size_t length)
4038 s->flags &= ~SLAB_DEBUG_FREE;
4040 s->flags |= SLAB_DEBUG_FREE;
4043 SLAB_ATTR(sanity_checks);
4045 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4047 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4050 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4053 s->flags &= ~SLAB_TRACE;
4055 s->flags |= SLAB_TRACE;
4060 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4062 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4065 static ssize_t red_zone_store(struct kmem_cache *s,
4066 const char *buf, size_t length)
4068 if (any_slab_objects(s))
4071 s->flags &= ~SLAB_RED_ZONE;
4073 s->flags |= SLAB_RED_ZONE;
4074 calculate_sizes(s, -1);
4077 SLAB_ATTR(red_zone);
4079 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4081 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4084 static ssize_t poison_store(struct kmem_cache *s,
4085 const char *buf, size_t length)
4087 if (any_slab_objects(s))
4090 s->flags &= ~SLAB_POISON;
4092 s->flags |= SLAB_POISON;
4093 calculate_sizes(s, -1);
4098 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4100 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4103 static ssize_t store_user_store(struct kmem_cache *s,
4104 const char *buf, size_t length)
4106 if (any_slab_objects(s))
4109 s->flags &= ~SLAB_STORE_USER;
4111 s->flags |= SLAB_STORE_USER;
4112 calculate_sizes(s, -1);
4115 SLAB_ATTR(store_user);
4117 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4122 static ssize_t validate_store(struct kmem_cache *s,
4123 const char *buf, size_t length)
4127 if (buf[0] == '1') {
4128 ret = validate_slab_cache(s);
4134 SLAB_ATTR(validate);
4136 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4138 if (!(s->flags & SLAB_STORE_USER))
4140 return list_locations(s, buf, TRACK_ALLOC);
4142 SLAB_ATTR_RO(alloc_calls);
4144 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4146 if (!(s->flags & SLAB_STORE_USER))
4148 return list_locations(s, buf, TRACK_FREE);
4150 SLAB_ATTR_RO(free_calls);
4151 #endif /* CONFIG_SLUB_DEBUG */
4153 #ifdef CONFIG_FAILSLAB
4154 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4156 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4159 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4162 s->flags &= ~SLAB_FAILSLAB;
4164 s->flags |= SLAB_FAILSLAB;
4167 SLAB_ATTR(failslab);
4170 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4175 static ssize_t shrink_store(struct kmem_cache *s,
4176 const char *buf, size_t length)
4178 if (buf[0] == '1') {
4179 int rc = kmem_cache_shrink(s);
4190 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4192 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4195 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4196 const char *buf, size_t length)
4198 unsigned long ratio;
4201 err = strict_strtoul(buf, 10, &ratio);
4206 s->remote_node_defrag_ratio = ratio * 10;
4210 SLAB_ATTR(remote_node_defrag_ratio);
4213 #ifdef CONFIG_SLUB_STATS
4214 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4216 unsigned long sum = 0;
4219 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4224 for_each_online_cpu(cpu) {
4225 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4231 len = sprintf(buf, "%lu", sum);
4234 for_each_online_cpu(cpu) {
4235 if (data[cpu] && len < PAGE_SIZE - 20)
4236 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4240 return len + sprintf(buf + len, "\n");
4243 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4247 for_each_online_cpu(cpu)
4248 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4251 #define STAT_ATTR(si, text) \
4252 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4254 return show_stat(s, buf, si); \
4256 static ssize_t text##_store(struct kmem_cache *s, \
4257 const char *buf, size_t length) \
4259 if (buf[0] != '0') \
4261 clear_stat(s, si); \
4266 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4267 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4268 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4269 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4270 STAT_ATTR(FREE_FROZEN, free_frozen);
4271 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4272 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4273 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4274 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4275 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4276 STAT_ATTR(FREE_SLAB, free_slab);
4277 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4278 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4279 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4280 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4281 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4282 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4283 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4286 static struct attribute *slab_attrs[] = {
4287 &slab_size_attr.attr,
4288 &object_size_attr.attr,
4289 &objs_per_slab_attr.attr,
4291 &min_partial_attr.attr,
4293 &objects_partial_attr.attr,
4295 &cpu_slabs_attr.attr,
4299 &hwcache_align_attr.attr,
4300 &reclaim_account_attr.attr,
4301 &destroy_by_rcu_attr.attr,
4303 #ifdef CONFIG_SLUB_DEBUG
4304 &total_objects_attr.attr,
4306 &sanity_checks_attr.attr,
4308 &red_zone_attr.attr,
4310 &store_user_attr.attr,
4311 &validate_attr.attr,
4312 &alloc_calls_attr.attr,
4313 &free_calls_attr.attr,
4315 #ifdef CONFIG_ZONE_DMA
4316 &cache_dma_attr.attr,
4319 &remote_node_defrag_ratio_attr.attr,
4321 #ifdef CONFIG_SLUB_STATS
4322 &alloc_fastpath_attr.attr,
4323 &alloc_slowpath_attr.attr,
4324 &free_fastpath_attr.attr,
4325 &free_slowpath_attr.attr,
4326 &free_frozen_attr.attr,
4327 &free_add_partial_attr.attr,
4328 &free_remove_partial_attr.attr,
4329 &alloc_from_partial_attr.attr,
4330 &alloc_slab_attr.attr,
4331 &alloc_refill_attr.attr,
4332 &free_slab_attr.attr,
4333 &cpuslab_flush_attr.attr,
4334 &deactivate_full_attr.attr,
4335 &deactivate_empty_attr.attr,
4336 &deactivate_to_head_attr.attr,
4337 &deactivate_to_tail_attr.attr,
4338 &deactivate_remote_frees_attr.attr,
4339 &order_fallback_attr.attr,
4341 #ifdef CONFIG_FAILSLAB
4342 &failslab_attr.attr,
4348 static struct attribute_group slab_attr_group = {
4349 .attrs = slab_attrs,
4352 static ssize_t slab_attr_show(struct kobject *kobj,
4353 struct attribute *attr,
4356 struct slab_attribute *attribute;
4357 struct kmem_cache *s;
4360 attribute = to_slab_attr(attr);
4363 if (!attribute->show)
4366 err = attribute->show(s, buf);
4371 static ssize_t slab_attr_store(struct kobject *kobj,
4372 struct attribute *attr,
4373 const char *buf, size_t len)
4375 struct slab_attribute *attribute;
4376 struct kmem_cache *s;
4379 attribute = to_slab_attr(attr);
4382 if (!attribute->store)
4385 err = attribute->store(s, buf, len);
4390 static void kmem_cache_release(struct kobject *kobj)
4392 struct kmem_cache *s = to_slab(kobj);
4398 static const struct sysfs_ops slab_sysfs_ops = {
4399 .show = slab_attr_show,
4400 .store = slab_attr_store,
4403 static struct kobj_type slab_ktype = {
4404 .sysfs_ops = &slab_sysfs_ops,
4405 .release = kmem_cache_release
4408 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4410 struct kobj_type *ktype = get_ktype(kobj);
4412 if (ktype == &slab_ktype)
4417 static const struct kset_uevent_ops slab_uevent_ops = {
4418 .filter = uevent_filter,
4421 static struct kset *slab_kset;
4423 #define ID_STR_LENGTH 64
4425 /* Create a unique string id for a slab cache:
4427 * Format :[flags-]size
4429 static char *create_unique_id(struct kmem_cache *s)
4431 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4438 * First flags affecting slabcache operations. We will only
4439 * get here for aliasable slabs so we do not need to support
4440 * too many flags. The flags here must cover all flags that
4441 * are matched during merging to guarantee that the id is
4444 if (s->flags & SLAB_CACHE_DMA)
4446 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4448 if (s->flags & SLAB_DEBUG_FREE)
4450 if (!(s->flags & SLAB_NOTRACK))
4454 p += sprintf(p, "%07d", s->size);
4455 BUG_ON(p > name + ID_STR_LENGTH - 1);
4459 static int sysfs_slab_add(struct kmem_cache *s)
4465 if (slab_state < SYSFS)
4466 /* Defer until later */
4469 unmergeable = slab_unmergeable(s);
4472 * Slabcache can never be merged so we can use the name proper.
4473 * This is typically the case for debug situations. In that
4474 * case we can catch duplicate names easily.
4476 sysfs_remove_link(&slab_kset->kobj, s->name);
4480 * Create a unique name for the slab as a target
4483 name = create_unique_id(s);
4486 s->kobj.kset = slab_kset;
4487 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4489 kobject_put(&s->kobj);
4493 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4495 kobject_del(&s->kobj);
4496 kobject_put(&s->kobj);
4499 kobject_uevent(&s->kobj, KOBJ_ADD);
4501 /* Setup first alias */
4502 sysfs_slab_alias(s, s->name);
4508 static void sysfs_slab_remove(struct kmem_cache *s)
4510 if (slab_state < SYSFS)
4512 * Sysfs has not been setup yet so no need to remove the
4517 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4518 kobject_del(&s->kobj);
4519 kobject_put(&s->kobj);
4523 * Need to buffer aliases during bootup until sysfs becomes
4524 * available lest we lose that information.
4526 struct saved_alias {
4527 struct kmem_cache *s;
4529 struct saved_alias *next;
4532 static struct saved_alias *alias_list;
4534 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4536 struct saved_alias *al;
4538 if (slab_state == SYSFS) {
4540 * If we have a leftover link then remove it.
4542 sysfs_remove_link(&slab_kset->kobj, name);
4543 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4546 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4552 al->next = alias_list;
4557 static int __init slab_sysfs_init(void)
4559 struct kmem_cache *s;
4562 down_write(&slub_lock);
4564 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4566 up_write(&slub_lock);
4567 printk(KERN_ERR "Cannot register slab subsystem.\n");
4573 list_for_each_entry(s, &slab_caches, list) {
4574 err = sysfs_slab_add(s);
4576 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4577 " to sysfs\n", s->name);
4580 while (alias_list) {
4581 struct saved_alias *al = alias_list;
4583 alias_list = alias_list->next;
4584 err = sysfs_slab_alias(al->s, al->name);
4586 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4587 " %s to sysfs\n", s->name);
4591 up_write(&slub_lock);
4596 __initcall(slab_sysfs_init);
4597 #endif /* CONFIG_SYSFS */
4600 * The /proc/slabinfo ABI
4602 #ifdef CONFIG_SLABINFO
4603 static void print_slabinfo_header(struct seq_file *m)
4605 seq_puts(m, "slabinfo - version: 2.1\n");
4606 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4607 "<objperslab> <pagesperslab>");
4608 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4609 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4613 static void *s_start(struct seq_file *m, loff_t *pos)
4617 down_read(&slub_lock);
4619 print_slabinfo_header(m);
4621 return seq_list_start(&slab_caches, *pos);
4624 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4626 return seq_list_next(p, &slab_caches, pos);
4629 static void s_stop(struct seq_file *m, void *p)
4631 up_read(&slub_lock);
4634 static int s_show(struct seq_file *m, void *p)
4636 unsigned long nr_partials = 0;
4637 unsigned long nr_slabs = 0;
4638 unsigned long nr_inuse = 0;
4639 unsigned long nr_objs = 0;
4640 unsigned long nr_free = 0;
4641 struct kmem_cache *s;
4644 s = list_entry(p, struct kmem_cache, list);
4646 for_each_online_node(node) {
4647 struct kmem_cache_node *n = get_node(s, node);
4652 nr_partials += n->nr_partial;
4653 nr_slabs += atomic_long_read(&n->nr_slabs);
4654 nr_objs += atomic_long_read(&n->total_objects);
4655 nr_free += count_partial(n, count_free);
4658 nr_inuse = nr_objs - nr_free;
4660 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4661 nr_objs, s->size, oo_objects(s->oo),
4662 (1 << oo_order(s->oo)));
4663 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4664 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4670 static const struct seq_operations slabinfo_op = {
4677 static int slabinfo_open(struct inode *inode, struct file *file)
4679 return seq_open(file, &slabinfo_op);
4682 static const struct file_operations proc_slabinfo_operations = {
4683 .open = slabinfo_open,
4685 .llseek = seq_lseek,
4686 .release = seq_release,
4689 static int __init slab_proc_init(void)
4691 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4694 module_init(slab_proc_init);
4695 #endif /* CONFIG_SLABINFO */