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(const 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;\
274 /* Determine object index from a given position */
275 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
277 return (p - addr) / s->size;
280 static inline size_t slab_ksize(const struct kmem_cache *s)
282 #ifdef CONFIG_SLUB_DEBUG
284 * Debugging requires use of the padding between object
285 * and whatever may come after it.
287 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
292 * If we have the need to store the freelist pointer
293 * back there or track user information then we can
294 * only use the space before that information.
296 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
299 * Else we can use all the padding etc for the allocation
304 static inline int order_objects(int order, unsigned long size, int reserved)
306 return ((PAGE_SIZE << order) - reserved) / size;
309 static inline struct kmem_cache_order_objects oo_make(int order,
310 unsigned long size, int reserved)
312 struct kmem_cache_order_objects x = {
313 (order << OO_SHIFT) + order_objects(order, size, reserved)
319 static inline int oo_order(struct kmem_cache_order_objects x)
321 return x.x >> OO_SHIFT;
324 static inline int oo_objects(struct kmem_cache_order_objects x)
326 return x.x & OO_MASK;
330 * Determine a map of object in use on a page.
332 * Slab lock or node listlock must be held to guarantee that the page does
333 * not vanish from under us.
335 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
338 void *addr = page_address(page);
340 for (p = page->freelist; p; p = get_freepointer(s, p))
341 set_bit(slab_index(p, s, addr), map);
344 #ifdef CONFIG_SLUB_DEBUG
348 #ifdef CONFIG_SLUB_DEBUG_ON
349 static int slub_debug = DEBUG_DEFAULT_FLAGS;
351 static int slub_debug;
354 static char *slub_debug_slabs;
355 static int disable_higher_order_debug;
360 static void print_section(char *text, u8 *addr, unsigned int length)
368 for (i = 0; i < length; i++) {
370 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
373 printk(KERN_CONT " %02x", addr[i]);
375 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
377 printk(KERN_CONT " %s\n", ascii);
384 printk(KERN_CONT " ");
388 printk(KERN_CONT " %s\n", ascii);
392 static struct track *get_track(struct kmem_cache *s, void *object,
393 enum track_item alloc)
398 p = object + s->offset + sizeof(void *);
400 p = object + s->inuse;
405 static void set_track(struct kmem_cache *s, void *object,
406 enum track_item alloc, unsigned long addr)
408 struct track *p = get_track(s, object, alloc);
412 p->cpu = smp_processor_id();
413 p->pid = current->pid;
416 memset(p, 0, sizeof(struct track));
419 static void init_tracking(struct kmem_cache *s, void *object)
421 if (!(s->flags & SLAB_STORE_USER))
424 set_track(s, object, TRACK_FREE, 0UL);
425 set_track(s, object, TRACK_ALLOC, 0UL);
428 static void print_track(const char *s, struct track *t)
433 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
434 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
437 static void print_tracking(struct kmem_cache *s, void *object)
439 if (!(s->flags & SLAB_STORE_USER))
442 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
443 print_track("Freed", get_track(s, object, TRACK_FREE));
446 static void print_page_info(struct page *page)
448 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
449 page, page->objects, page->inuse, page->freelist, page->flags);
453 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
459 vsnprintf(buf, sizeof(buf), fmt, args);
461 printk(KERN_ERR "========================================"
462 "=====================================\n");
463 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
464 printk(KERN_ERR "----------------------------------------"
465 "-------------------------------------\n\n");
468 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
474 vsnprintf(buf, sizeof(buf), fmt, args);
476 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
479 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
481 unsigned int off; /* Offset of last byte */
482 u8 *addr = page_address(page);
484 print_tracking(s, p);
486 print_page_info(page);
488 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
489 p, p - addr, get_freepointer(s, p));
492 print_section("Bytes b4", p - 16, 16);
494 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
496 if (s->flags & SLAB_RED_ZONE)
497 print_section("Redzone", p + s->objsize,
498 s->inuse - s->objsize);
501 off = s->offset + sizeof(void *);
505 if (s->flags & SLAB_STORE_USER)
506 off += 2 * sizeof(struct track);
509 /* Beginning of the filler is the free pointer */
510 print_section("Padding", p + off, s->size - off);
515 static void object_err(struct kmem_cache *s, struct page *page,
516 u8 *object, char *reason)
518 slab_bug(s, "%s", reason);
519 print_trailer(s, page, object);
522 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
528 vsnprintf(buf, sizeof(buf), fmt, args);
530 slab_bug(s, "%s", buf);
531 print_page_info(page);
535 static void init_object(struct kmem_cache *s, void *object, u8 val)
539 if (s->flags & __OBJECT_POISON) {
540 memset(p, POISON_FREE, s->objsize - 1);
541 p[s->objsize - 1] = POISON_END;
544 if (s->flags & SLAB_RED_ZONE)
545 memset(p + s->objsize, val, s->inuse - s->objsize);
548 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
551 if (*start != (u8)value)
559 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
560 void *from, void *to)
562 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
563 memset(from, data, to - from);
566 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
567 u8 *object, char *what,
568 u8 *start, unsigned int value, unsigned int bytes)
573 fault = check_bytes(start, value, bytes);
578 while (end > fault && end[-1] == value)
581 slab_bug(s, "%s overwritten", what);
582 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
583 fault, end - 1, fault[0], value);
584 print_trailer(s, page, object);
586 restore_bytes(s, what, value, fault, end);
594 * Bytes of the object to be managed.
595 * If the freepointer may overlay the object then the free
596 * pointer is the first word of the object.
598 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
601 * object + s->objsize
602 * Padding to reach word boundary. This is also used for Redzoning.
603 * Padding is extended by another word if Redzoning is enabled and
606 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
607 * 0xcc (RED_ACTIVE) for objects in use.
610 * Meta data starts here.
612 * A. Free pointer (if we cannot overwrite object on free)
613 * B. Tracking data for SLAB_STORE_USER
614 * C. Padding to reach required alignment boundary or at mininum
615 * one word if debugging is on to be able to detect writes
616 * before the word boundary.
618 * Padding is done using 0x5a (POISON_INUSE)
621 * Nothing is used beyond s->size.
623 * If slabcaches are merged then the objsize and inuse boundaries are mostly
624 * ignored. And therefore no slab options that rely on these boundaries
625 * may be used with merged slabcaches.
628 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
630 unsigned long off = s->inuse; /* The end of info */
633 /* Freepointer is placed after the object. */
634 off += sizeof(void *);
636 if (s->flags & SLAB_STORE_USER)
637 /* We also have user information there */
638 off += 2 * sizeof(struct track);
643 return check_bytes_and_report(s, page, p, "Object padding",
644 p + off, POISON_INUSE, s->size - off);
647 /* Check the pad bytes at the end of a slab page */
648 static int slab_pad_check(struct kmem_cache *s, struct page *page)
656 if (!(s->flags & SLAB_POISON))
659 start = page_address(page);
660 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
661 end = start + length;
662 remainder = length % s->size;
666 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
669 while (end > fault && end[-1] == POISON_INUSE)
672 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
673 print_section("Padding", end - remainder, remainder);
675 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
679 static int check_object(struct kmem_cache *s, struct page *page,
680 void *object, u8 val)
683 u8 *endobject = object + s->objsize;
685 if (s->flags & SLAB_RED_ZONE) {
686 if (!check_bytes_and_report(s, page, object, "Redzone",
687 endobject, val, s->inuse - s->objsize))
690 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
691 check_bytes_and_report(s, page, p, "Alignment padding",
692 endobject, POISON_INUSE, s->inuse - s->objsize);
696 if (s->flags & SLAB_POISON) {
697 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
698 (!check_bytes_and_report(s, page, p, "Poison", p,
699 POISON_FREE, s->objsize - 1) ||
700 !check_bytes_and_report(s, page, p, "Poison",
701 p + s->objsize - 1, POISON_END, 1)))
704 * check_pad_bytes cleans up on its own.
706 check_pad_bytes(s, page, p);
709 if (!s->offset && val == SLUB_RED_ACTIVE)
711 * Object and freepointer overlap. Cannot check
712 * freepointer while object is allocated.
716 /* Check free pointer validity */
717 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
718 object_err(s, page, p, "Freepointer corrupt");
720 * No choice but to zap it and thus lose the remainder
721 * of the free objects in this slab. May cause
722 * another error because the object count is now wrong.
724 set_freepointer(s, p, NULL);
730 static int check_slab(struct kmem_cache *s, struct page *page)
734 VM_BUG_ON(!irqs_disabled());
736 if (!PageSlab(page)) {
737 slab_err(s, page, "Not a valid slab page");
741 maxobj = order_objects(compound_order(page), s->size, s->reserved);
742 if (page->objects > maxobj) {
743 slab_err(s, page, "objects %u > max %u",
744 s->name, page->objects, maxobj);
747 if (page->inuse > page->objects) {
748 slab_err(s, page, "inuse %u > max %u",
749 s->name, page->inuse, page->objects);
752 /* Slab_pad_check fixes things up after itself */
753 slab_pad_check(s, page);
758 * Determine if a certain object on a page is on the freelist. Must hold the
759 * slab lock to guarantee that the chains are in a consistent state.
761 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
764 void *fp = page->freelist;
766 unsigned long max_objects;
768 while (fp && nr <= page->objects) {
771 if (!check_valid_pointer(s, page, fp)) {
773 object_err(s, page, object,
774 "Freechain corrupt");
775 set_freepointer(s, object, NULL);
778 slab_err(s, page, "Freepointer corrupt");
779 page->freelist = NULL;
780 page->inuse = page->objects;
781 slab_fix(s, "Freelist cleared");
787 fp = get_freepointer(s, object);
791 max_objects = order_objects(compound_order(page), s->size, s->reserved);
792 if (max_objects > MAX_OBJS_PER_PAGE)
793 max_objects = MAX_OBJS_PER_PAGE;
795 if (page->objects != max_objects) {
796 slab_err(s, page, "Wrong number of objects. Found %d but "
797 "should be %d", page->objects, max_objects);
798 page->objects = max_objects;
799 slab_fix(s, "Number of objects adjusted.");
801 if (page->inuse != page->objects - nr) {
802 slab_err(s, page, "Wrong object count. Counter is %d but "
803 "counted were %d", page->inuse, page->objects - nr);
804 page->inuse = page->objects - nr;
805 slab_fix(s, "Object count adjusted.");
807 return search == NULL;
810 static void trace(struct kmem_cache *s, struct page *page, void *object,
813 if (s->flags & SLAB_TRACE) {
814 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
816 alloc ? "alloc" : "free",
821 print_section("Object", (void *)object, s->objsize);
828 * Hooks for other subsystems that check memory allocations. In a typical
829 * production configuration these hooks all should produce no code at all.
831 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
833 flags &= gfp_allowed_mask;
834 lockdep_trace_alloc(flags);
835 might_sleep_if(flags & __GFP_WAIT);
837 return should_failslab(s->objsize, flags, s->flags);
840 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
842 flags &= gfp_allowed_mask;
843 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
844 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
847 static inline void slab_free_hook(struct kmem_cache *s, void *x)
849 kmemleak_free_recursive(x, s->flags);
852 * Trouble is that we may no longer disable interupts in the fast path
853 * So in order to make the debug calls that expect irqs to be
854 * disabled we need to disable interrupts temporarily.
856 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
860 local_irq_save(flags);
861 kmemcheck_slab_free(s, x, s->objsize);
862 debug_check_no_locks_freed(x, s->objsize);
863 local_irq_restore(flags);
866 if (!(s->flags & SLAB_DEBUG_OBJECTS))
867 debug_check_no_obj_freed(x, s->objsize);
871 * Tracking of fully allocated slabs for debugging purposes.
873 static void add_full(struct kmem_cache_node *n, struct page *page)
875 spin_lock(&n->list_lock);
876 list_add(&page->lru, &n->full);
877 spin_unlock(&n->list_lock);
880 static void remove_full(struct kmem_cache *s, struct page *page)
882 struct kmem_cache_node *n;
884 if (!(s->flags & SLAB_STORE_USER))
887 n = get_node(s, page_to_nid(page));
889 spin_lock(&n->list_lock);
890 list_del(&page->lru);
891 spin_unlock(&n->list_lock);
894 /* Tracking of the number of slabs for debugging purposes */
895 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
897 struct kmem_cache_node *n = get_node(s, node);
899 return atomic_long_read(&n->nr_slabs);
902 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
904 return atomic_long_read(&n->nr_slabs);
907 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
909 struct kmem_cache_node *n = get_node(s, node);
912 * May be called early in order to allocate a slab for the
913 * kmem_cache_node structure. Solve the chicken-egg
914 * dilemma by deferring the increment of the count during
915 * bootstrap (see early_kmem_cache_node_alloc).
918 atomic_long_inc(&n->nr_slabs);
919 atomic_long_add(objects, &n->total_objects);
922 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
924 struct kmem_cache_node *n = get_node(s, node);
926 atomic_long_dec(&n->nr_slabs);
927 atomic_long_sub(objects, &n->total_objects);
930 /* Object debug checks for alloc/free paths */
931 static void setup_object_debug(struct kmem_cache *s, struct page *page,
934 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
937 init_object(s, object, SLUB_RED_INACTIVE);
938 init_tracking(s, object);
941 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
942 void *object, unsigned long addr)
944 if (!check_slab(s, page))
947 if (!on_freelist(s, page, object)) {
948 object_err(s, page, object, "Object already allocated");
952 if (!check_valid_pointer(s, page, object)) {
953 object_err(s, page, object, "Freelist Pointer check fails");
957 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
960 /* Success perform special debug activities for allocs */
961 if (s->flags & SLAB_STORE_USER)
962 set_track(s, object, TRACK_ALLOC, addr);
963 trace(s, page, object, 1);
964 init_object(s, object, SLUB_RED_ACTIVE);
968 if (PageSlab(page)) {
970 * If this is a slab page then lets do the best we can
971 * to avoid issues in the future. Marking all objects
972 * as used avoids touching the remaining objects.
974 slab_fix(s, "Marking all objects used");
975 page->inuse = page->objects;
976 page->freelist = NULL;
981 static noinline int free_debug_processing(struct kmem_cache *s,
982 struct page *page, void *object, unsigned long addr)
984 if (!check_slab(s, page))
987 if (!check_valid_pointer(s, page, object)) {
988 slab_err(s, page, "Invalid object pointer 0x%p", object);
992 if (on_freelist(s, page, object)) {
993 object_err(s, page, object, "Object already free");
997 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1000 if (unlikely(s != page->slab)) {
1001 if (!PageSlab(page)) {
1002 slab_err(s, page, "Attempt to free object(0x%p) "
1003 "outside of slab", object);
1004 } else if (!page->slab) {
1006 "SLUB <none>: no slab for object 0x%p.\n",
1010 object_err(s, page, object,
1011 "page slab pointer corrupt.");
1015 /* Special debug activities for freeing objects */
1016 if (!PageSlubFrozen(page) && !page->freelist)
1017 remove_full(s, page);
1018 if (s->flags & SLAB_STORE_USER)
1019 set_track(s, object, TRACK_FREE, addr);
1020 trace(s, page, object, 0);
1021 init_object(s, object, SLUB_RED_INACTIVE);
1025 slab_fix(s, "Object at 0x%p not freed", object);
1029 static int __init setup_slub_debug(char *str)
1031 slub_debug = DEBUG_DEFAULT_FLAGS;
1032 if (*str++ != '=' || !*str)
1034 * No options specified. Switch on full debugging.
1040 * No options but restriction on slabs. This means full
1041 * debugging for slabs matching a pattern.
1045 if (tolower(*str) == 'o') {
1047 * Avoid enabling debugging on caches if its minimum order
1048 * would increase as a result.
1050 disable_higher_order_debug = 1;
1057 * Switch off all debugging measures.
1062 * Determine which debug features should be switched on
1064 for (; *str && *str != ','; str++) {
1065 switch (tolower(*str)) {
1067 slub_debug |= SLAB_DEBUG_FREE;
1070 slub_debug |= SLAB_RED_ZONE;
1073 slub_debug |= SLAB_POISON;
1076 slub_debug |= SLAB_STORE_USER;
1079 slub_debug |= SLAB_TRACE;
1082 slub_debug |= SLAB_FAILSLAB;
1085 printk(KERN_ERR "slub_debug option '%c' "
1086 "unknown. skipped\n", *str);
1092 slub_debug_slabs = str + 1;
1097 __setup("slub_debug", setup_slub_debug);
1099 static unsigned long kmem_cache_flags(unsigned long objsize,
1100 unsigned long flags, const char *name,
1101 void (*ctor)(void *))
1104 * Enable debugging if selected on the kernel commandline.
1106 if (slub_debug && (!slub_debug_slabs ||
1107 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1108 flags |= slub_debug;
1113 static inline void setup_object_debug(struct kmem_cache *s,
1114 struct page *page, void *object) {}
1116 static inline int alloc_debug_processing(struct kmem_cache *s,
1117 struct page *page, void *object, unsigned long addr) { return 0; }
1119 static inline int free_debug_processing(struct kmem_cache *s,
1120 struct page *page, void *object, unsigned long addr) { return 0; }
1122 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1124 static inline int check_object(struct kmem_cache *s, struct page *page,
1125 void *object, u8 val) { return 1; }
1126 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1127 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1128 unsigned long flags, const char *name,
1129 void (*ctor)(void *))
1133 #define slub_debug 0
1135 #define disable_higher_order_debug 0
1137 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1139 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1141 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1143 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1146 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1149 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1152 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1154 #endif /* CONFIG_SLUB_DEBUG */
1157 * Slab allocation and freeing
1159 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1160 struct kmem_cache_order_objects oo)
1162 int order = oo_order(oo);
1164 flags |= __GFP_NOTRACK;
1166 if (node == NUMA_NO_NODE)
1167 return alloc_pages(flags, order);
1169 return alloc_pages_exact_node(node, flags, order);
1172 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1175 struct kmem_cache_order_objects oo = s->oo;
1178 flags |= s->allocflags;
1181 * Let the initial higher-order allocation fail under memory pressure
1182 * so we fall-back to the minimum order allocation.
1184 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1186 page = alloc_slab_page(alloc_gfp, node, oo);
1187 if (unlikely(!page)) {
1190 * Allocation may have failed due to fragmentation.
1191 * Try a lower order alloc if possible
1193 page = alloc_slab_page(flags, node, oo);
1197 stat(s, ORDER_FALLBACK);
1200 if (kmemcheck_enabled
1201 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1202 int pages = 1 << oo_order(oo);
1204 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1207 * Objects from caches that have a constructor don't get
1208 * cleared when they're allocated, so we need to do it here.
1211 kmemcheck_mark_uninitialized_pages(page, pages);
1213 kmemcheck_mark_unallocated_pages(page, pages);
1216 page->objects = oo_objects(oo);
1217 mod_zone_page_state(page_zone(page),
1218 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1219 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1225 static void setup_object(struct kmem_cache *s, struct page *page,
1228 setup_object_debug(s, page, object);
1229 if (unlikely(s->ctor))
1233 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1240 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1242 page = allocate_slab(s,
1243 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1247 inc_slabs_node(s, page_to_nid(page), page->objects);
1249 page->flags |= 1 << PG_slab;
1251 start = page_address(page);
1253 if (unlikely(s->flags & SLAB_POISON))
1254 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1257 for_each_object(p, s, start, page->objects) {
1258 setup_object(s, page, last);
1259 set_freepointer(s, last, p);
1262 setup_object(s, page, last);
1263 set_freepointer(s, last, NULL);
1265 page->freelist = start;
1271 static void __free_slab(struct kmem_cache *s, struct page *page)
1273 int order = compound_order(page);
1274 int pages = 1 << order;
1276 if (kmem_cache_debug(s)) {
1279 slab_pad_check(s, page);
1280 for_each_object(p, s, page_address(page),
1282 check_object(s, page, p, SLUB_RED_INACTIVE);
1285 kmemcheck_free_shadow(page, compound_order(page));
1287 mod_zone_page_state(page_zone(page),
1288 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1289 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1292 __ClearPageSlab(page);
1293 reset_page_mapcount(page);
1294 if (current->reclaim_state)
1295 current->reclaim_state->reclaimed_slab += pages;
1296 __free_pages(page, order);
1299 #define need_reserve_slab_rcu \
1300 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1302 static void rcu_free_slab(struct rcu_head *h)
1306 if (need_reserve_slab_rcu)
1307 page = virt_to_head_page(h);
1309 page = container_of((struct list_head *)h, struct page, lru);
1311 __free_slab(page->slab, page);
1314 static void free_slab(struct kmem_cache *s, struct page *page)
1316 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1317 struct rcu_head *head;
1319 if (need_reserve_slab_rcu) {
1320 int order = compound_order(page);
1321 int offset = (PAGE_SIZE << order) - s->reserved;
1323 VM_BUG_ON(s->reserved != sizeof(*head));
1324 head = page_address(page) + offset;
1327 * RCU free overloads the RCU head over the LRU
1329 head = (void *)&page->lru;
1332 call_rcu(head, rcu_free_slab);
1334 __free_slab(s, page);
1337 static void discard_slab(struct kmem_cache *s, struct page *page)
1339 dec_slabs_node(s, page_to_nid(page), page->objects);
1344 * Per slab locking using the pagelock
1346 static __always_inline void slab_lock(struct page *page)
1348 bit_spin_lock(PG_locked, &page->flags);
1351 static __always_inline void slab_unlock(struct page *page)
1353 __bit_spin_unlock(PG_locked, &page->flags);
1356 static __always_inline int slab_trylock(struct page *page)
1360 rc = bit_spin_trylock(PG_locked, &page->flags);
1365 * Management of partially allocated slabs
1367 static void add_partial(struct kmem_cache_node *n,
1368 struct page *page, int tail)
1370 spin_lock(&n->list_lock);
1373 list_add_tail(&page->lru, &n->partial);
1375 list_add(&page->lru, &n->partial);
1376 spin_unlock(&n->list_lock);
1379 static inline void __remove_partial(struct kmem_cache_node *n,
1382 list_del(&page->lru);
1386 static void remove_partial(struct kmem_cache *s, struct page *page)
1388 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1390 spin_lock(&n->list_lock);
1391 __remove_partial(n, page);
1392 spin_unlock(&n->list_lock);
1396 * Lock slab and remove from the partial list.
1398 * Must hold list_lock.
1400 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1403 if (slab_trylock(page)) {
1404 __remove_partial(n, page);
1405 __SetPageSlubFrozen(page);
1412 * Try to allocate a partial slab from a specific node.
1414 static struct page *get_partial_node(struct kmem_cache_node *n)
1419 * Racy check. If we mistakenly see no partial slabs then we
1420 * just allocate an empty slab. If we mistakenly try to get a
1421 * partial slab and there is none available then get_partials()
1424 if (!n || !n->nr_partial)
1427 spin_lock(&n->list_lock);
1428 list_for_each_entry(page, &n->partial, lru)
1429 if (lock_and_freeze_slab(n, page))
1433 spin_unlock(&n->list_lock);
1438 * Get a page from somewhere. Search in increasing NUMA distances.
1440 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1443 struct zonelist *zonelist;
1446 enum zone_type high_zoneidx = gfp_zone(flags);
1450 * The defrag ratio allows a configuration of the tradeoffs between
1451 * inter node defragmentation and node local allocations. A lower
1452 * defrag_ratio increases the tendency to do local allocations
1453 * instead of attempting to obtain partial slabs from other nodes.
1455 * If the defrag_ratio is set to 0 then kmalloc() always
1456 * returns node local objects. If the ratio is higher then kmalloc()
1457 * may return off node objects because partial slabs are obtained
1458 * from other nodes and filled up.
1460 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1461 * defrag_ratio = 1000) then every (well almost) allocation will
1462 * first attempt to defrag slab caches on other nodes. This means
1463 * scanning over all nodes to look for partial slabs which may be
1464 * expensive if we do it every time we are trying to find a slab
1465 * with available objects.
1467 if (!s->remote_node_defrag_ratio ||
1468 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1472 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1473 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1474 struct kmem_cache_node *n;
1476 n = get_node(s, zone_to_nid(zone));
1478 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1479 n->nr_partial > s->min_partial) {
1480 page = get_partial_node(n);
1493 * Get a partial page, lock it and return it.
1495 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1498 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1500 page = get_partial_node(get_node(s, searchnode));
1501 if (page || node != NUMA_NO_NODE)
1504 return get_any_partial(s, flags);
1508 * Move a page back to the lists.
1510 * Must be called with the slab lock held.
1512 * On exit the slab lock will have been dropped.
1514 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1517 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1519 __ClearPageSlubFrozen(page);
1522 if (page->freelist) {
1523 add_partial(n, page, tail);
1524 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1526 stat(s, DEACTIVATE_FULL);
1527 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1532 stat(s, DEACTIVATE_EMPTY);
1533 if (n->nr_partial < s->min_partial) {
1535 * Adding an empty slab to the partial slabs in order
1536 * to avoid page allocator overhead. This slab needs
1537 * to come after the other slabs with objects in
1538 * so that the others get filled first. That way the
1539 * size of the partial list stays small.
1541 * kmem_cache_shrink can reclaim any empty slabs from
1544 add_partial(n, page, 1);
1549 discard_slab(s, page);
1554 #ifdef CONFIG_CMPXCHG_LOCAL
1555 #ifdef CONFIG_PREEMPT
1557 * Calculate the next globally unique transaction for disambiguiation
1558 * during cmpxchg. The transactions start with the cpu number and are then
1559 * incremented by CONFIG_NR_CPUS.
1561 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1564 * No preemption supported therefore also no need to check for
1570 static inline unsigned long next_tid(unsigned long tid)
1572 return tid + TID_STEP;
1575 static inline unsigned int tid_to_cpu(unsigned long tid)
1577 return tid % TID_STEP;
1580 static inline unsigned long tid_to_event(unsigned long tid)
1582 return tid / TID_STEP;
1585 static inline unsigned int init_tid(int cpu)
1590 static inline void note_cmpxchg_failure(const char *n,
1591 const struct kmem_cache *s, unsigned long tid)
1593 #ifdef SLUB_DEBUG_CMPXCHG
1594 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1596 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1598 #ifdef CONFIG_PREEMPT
1599 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1600 printk("due to cpu change %d -> %d\n",
1601 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1604 if (tid_to_event(tid) != tid_to_event(actual_tid))
1605 printk("due to cpu running other code. Event %ld->%ld\n",
1606 tid_to_event(tid), tid_to_event(actual_tid));
1608 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1609 actual_tid, tid, next_tid(tid));
1611 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1616 void init_kmem_cache_cpus(struct kmem_cache *s)
1618 #ifdef CONFIG_CMPXCHG_LOCAL
1621 for_each_possible_cpu(cpu)
1622 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1627 * Remove the cpu slab
1629 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1632 struct page *page = c->page;
1636 stat(s, DEACTIVATE_REMOTE_FREES);
1638 * Merge cpu freelist into slab freelist. Typically we get here
1639 * because both freelists are empty. So this is unlikely
1642 while (unlikely(c->freelist)) {
1645 tail = 0; /* Hot objects. Put the slab first */
1647 /* Retrieve object from cpu_freelist */
1648 object = c->freelist;
1649 c->freelist = get_freepointer(s, c->freelist);
1651 /* And put onto the regular freelist */
1652 set_freepointer(s, object, page->freelist);
1653 page->freelist = object;
1657 #ifdef CONFIG_CMPXCHG_LOCAL
1658 c->tid = next_tid(c->tid);
1660 unfreeze_slab(s, page, tail);
1663 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1665 stat(s, CPUSLAB_FLUSH);
1667 deactivate_slab(s, c);
1673 * Called from IPI handler with interrupts disabled.
1675 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1677 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1679 if (likely(c && c->page))
1683 static void flush_cpu_slab(void *d)
1685 struct kmem_cache *s = d;
1687 __flush_cpu_slab(s, smp_processor_id());
1690 static void flush_all(struct kmem_cache *s)
1692 on_each_cpu(flush_cpu_slab, s, 1);
1696 * Check if the objects in a per cpu structure fit numa
1697 * locality expectations.
1699 static inline int node_match(struct kmem_cache_cpu *c, int node)
1702 if (node != NUMA_NO_NODE && c->node != node)
1708 static int count_free(struct page *page)
1710 return page->objects - page->inuse;
1713 static unsigned long count_partial(struct kmem_cache_node *n,
1714 int (*get_count)(struct page *))
1716 unsigned long flags;
1717 unsigned long x = 0;
1720 spin_lock_irqsave(&n->list_lock, flags);
1721 list_for_each_entry(page, &n->partial, lru)
1722 x += get_count(page);
1723 spin_unlock_irqrestore(&n->list_lock, flags);
1727 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1729 #ifdef CONFIG_SLUB_DEBUG
1730 return atomic_long_read(&n->total_objects);
1736 static noinline void
1737 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1742 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1744 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1745 "default order: %d, min order: %d\n", s->name, s->objsize,
1746 s->size, oo_order(s->oo), oo_order(s->min));
1748 if (oo_order(s->min) > get_order(s->objsize))
1749 printk(KERN_WARNING " %s debugging increased min order, use "
1750 "slub_debug=O to disable.\n", s->name);
1752 for_each_online_node(node) {
1753 struct kmem_cache_node *n = get_node(s, node);
1754 unsigned long nr_slabs;
1755 unsigned long nr_objs;
1756 unsigned long nr_free;
1761 nr_free = count_partial(n, count_free);
1762 nr_slabs = node_nr_slabs(n);
1763 nr_objs = node_nr_objs(n);
1766 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1767 node, nr_slabs, nr_objs, nr_free);
1772 * Slow path. The lockless freelist is empty or we need to perform
1775 * Interrupts are disabled.
1777 * Processing is still very fast if new objects have been freed to the
1778 * regular freelist. In that case we simply take over the regular freelist
1779 * as the lockless freelist and zap the regular freelist.
1781 * If that is not working then we fall back to the partial lists. We take the
1782 * first element of the freelist as the object to allocate now and move the
1783 * rest of the freelist to the lockless freelist.
1785 * And if we were unable to get a new slab from the partial slab lists then
1786 * we need to allocate a new slab. This is the slowest path since it involves
1787 * a call to the page allocator and the setup of a new slab.
1789 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1790 unsigned long addr, struct kmem_cache_cpu *c)
1794 #ifdef CONFIG_CMPXCHG_LOCAL
1795 unsigned long flags;
1797 local_irq_save(flags);
1798 #ifdef CONFIG_PREEMPT
1800 * We may have been preempted and rescheduled on a different
1801 * cpu before disabling interrupts. Need to reload cpu area
1804 c = this_cpu_ptr(s->cpu_slab);
1808 /* We handle __GFP_ZERO in the caller */
1809 gfpflags &= ~__GFP_ZERO;
1815 if (unlikely(!node_match(c, node)))
1818 stat(s, ALLOC_REFILL);
1821 object = c->page->freelist;
1822 if (unlikely(!object))
1824 if (kmem_cache_debug(s))
1827 c->freelist = get_freepointer(s, object);
1828 c->page->inuse = c->page->objects;
1829 c->page->freelist = NULL;
1830 c->node = page_to_nid(c->page);
1832 slab_unlock(c->page);
1833 #ifdef CONFIG_CMPXCHG_LOCAL
1834 c->tid = next_tid(c->tid);
1835 local_irq_restore(flags);
1837 stat(s, ALLOC_SLOWPATH);
1841 deactivate_slab(s, c);
1844 new = get_partial(s, gfpflags, node);
1847 stat(s, ALLOC_FROM_PARTIAL);
1851 gfpflags &= gfp_allowed_mask;
1852 if (gfpflags & __GFP_WAIT)
1855 new = new_slab(s, gfpflags, node);
1857 if (gfpflags & __GFP_WAIT)
1858 local_irq_disable();
1861 c = __this_cpu_ptr(s->cpu_slab);
1862 stat(s, ALLOC_SLAB);
1866 __SetPageSlubFrozen(new);
1870 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1871 slab_out_of_memory(s, gfpflags, node);
1872 #ifdef CONFIG_CMPXCHG_LOCAL
1873 local_irq_restore(flags);
1877 if (!alloc_debug_processing(s, c->page, object, addr))
1881 c->page->freelist = get_freepointer(s, object);
1882 c->node = NUMA_NO_NODE;
1887 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1888 * have the fastpath folded into their functions. So no function call
1889 * overhead for requests that can be satisfied on the fastpath.
1891 * The fastpath works by first checking if the lockless freelist can be used.
1892 * If not then __slab_alloc is called for slow processing.
1894 * Otherwise we can simply pick the next object from the lockless free list.
1896 static __always_inline void *slab_alloc(struct kmem_cache *s,
1897 gfp_t gfpflags, int node, unsigned long addr)
1900 struct kmem_cache_cpu *c;
1901 #ifdef CONFIG_CMPXCHG_LOCAL
1904 unsigned long flags;
1907 if (slab_pre_alloc_hook(s, gfpflags))
1910 #ifndef CONFIG_CMPXCHG_LOCAL
1911 local_irq_save(flags);
1917 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1918 * enabled. We may switch back and forth between cpus while
1919 * reading from one cpu area. That does not matter as long
1920 * as we end up on the original cpu again when doing the cmpxchg.
1922 c = __this_cpu_ptr(s->cpu_slab);
1924 #ifdef CONFIG_CMPXCHG_LOCAL
1926 * The transaction ids are globally unique per cpu and per operation on
1927 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1928 * occurs on the right processor and that there was no operation on the
1929 * linked list in between.
1935 object = c->freelist;
1936 if (unlikely(!object || !node_match(c, node)))
1938 object = __slab_alloc(s, gfpflags, node, addr, c);
1941 #ifdef CONFIG_CMPXCHG_LOCAL
1943 * The cmpxchg will only match if there was no additonal
1944 * operation and if we are on the right processor.
1946 * The cmpxchg does the following atomically (without lock semantics!)
1947 * 1. Relocate first pointer to the current per cpu area.
1948 * 2. Verify that tid and freelist have not been changed
1949 * 3. If they were not changed replace tid and freelist
1951 * Since this is without lock semantics the protection is only against
1952 * code executing on this cpu *not* from access by other cpus.
1954 if (unlikely(!this_cpu_cmpxchg_double(
1955 s->cpu_slab->freelist, s->cpu_slab->tid,
1957 get_freepointer(s, object), next_tid(tid)))) {
1959 note_cmpxchg_failure("slab_alloc", s, tid);
1963 c->freelist = get_freepointer(s, object);
1965 stat(s, ALLOC_FASTPATH);
1968 #ifndef CONFIG_CMPXCHG_LOCAL
1969 local_irq_restore(flags);
1972 if (unlikely(gfpflags & __GFP_ZERO) && object)
1973 memset(object, 0, s->objsize);
1975 slab_post_alloc_hook(s, gfpflags, object);
1980 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1982 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1984 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1988 EXPORT_SYMBOL(kmem_cache_alloc);
1990 #ifdef CONFIG_TRACING
1991 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1993 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1994 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
1997 EXPORT_SYMBOL(kmem_cache_alloc_trace);
1999 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2001 void *ret = kmalloc_order(size, flags, order);
2002 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2005 EXPORT_SYMBOL(kmalloc_order_trace);
2009 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2011 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2013 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2014 s->objsize, s->size, gfpflags, node);
2018 EXPORT_SYMBOL(kmem_cache_alloc_node);
2020 #ifdef CONFIG_TRACING
2021 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2023 int node, size_t size)
2025 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2027 trace_kmalloc_node(_RET_IP_, ret,
2028 size, s->size, gfpflags, node);
2031 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2036 * Slow patch handling. This may still be called frequently since objects
2037 * have a longer lifetime than the cpu slabs in most processing loads.
2039 * So we still attempt to reduce cache line usage. Just take the slab
2040 * lock and free the item. If there is no additional partial page
2041 * handling required then we can return immediately.
2043 static void __slab_free(struct kmem_cache *s, struct page *page,
2044 void *x, unsigned long addr)
2047 void **object = (void *)x;
2048 #ifdef CONFIG_CMPXCHG_LOCAL
2049 unsigned long flags;
2051 local_irq_save(flags);
2054 stat(s, FREE_SLOWPATH);
2056 if (kmem_cache_debug(s))
2060 prior = page->freelist;
2061 set_freepointer(s, object, prior);
2062 page->freelist = object;
2065 if (unlikely(PageSlubFrozen(page))) {
2066 stat(s, FREE_FROZEN);
2070 if (unlikely(!page->inuse))
2074 * Objects left in the slab. If it was not on the partial list before
2077 if (unlikely(!prior)) {
2078 add_partial(get_node(s, page_to_nid(page)), page, 1);
2079 stat(s, FREE_ADD_PARTIAL);
2084 #ifdef CONFIG_CMPXCHG_LOCAL
2085 local_irq_restore(flags);
2092 * Slab still on the partial list.
2094 remove_partial(s, page);
2095 stat(s, FREE_REMOVE_PARTIAL);
2098 #ifdef CONFIG_CMPXCHG_LOCAL
2099 local_irq_restore(flags);
2102 discard_slab(s, page);
2106 if (!free_debug_processing(s, page, x, addr))
2112 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2113 * can perform fastpath freeing without additional function calls.
2115 * The fastpath is only possible if we are freeing to the current cpu slab
2116 * of this processor. This typically the case if we have just allocated
2119 * If fastpath is not possible then fall back to __slab_free where we deal
2120 * with all sorts of special processing.
2122 static __always_inline void slab_free(struct kmem_cache *s,
2123 struct page *page, void *x, unsigned long addr)
2125 void **object = (void *)x;
2126 struct kmem_cache_cpu *c;
2127 #ifdef CONFIG_CMPXCHG_LOCAL
2130 unsigned long flags;
2133 slab_free_hook(s, x);
2135 #ifndef CONFIG_CMPXCHG_LOCAL
2136 local_irq_save(flags);
2143 * Determine the currently cpus per cpu slab.
2144 * The cpu may change afterward. However that does not matter since
2145 * data is retrieved via this pointer. If we are on the same cpu
2146 * during the cmpxchg then the free will succedd.
2148 c = __this_cpu_ptr(s->cpu_slab);
2150 #ifdef CONFIG_CMPXCHG_LOCAL
2155 if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
2156 set_freepointer(s, object, c->freelist);
2158 #ifdef CONFIG_CMPXCHG_LOCAL
2159 if (unlikely(!this_cpu_cmpxchg_double(
2160 s->cpu_slab->freelist, s->cpu_slab->tid,
2162 object, next_tid(tid)))) {
2164 note_cmpxchg_failure("slab_free", s, tid);
2168 c->freelist = object;
2170 stat(s, FREE_FASTPATH);
2172 __slab_free(s, page, x, addr);
2174 #ifndef CONFIG_CMPXCHG_LOCAL
2175 local_irq_restore(flags);
2179 void kmem_cache_free(struct kmem_cache *s, void *x)
2183 page = virt_to_head_page(x);
2185 slab_free(s, page, x, _RET_IP_);
2187 trace_kmem_cache_free(_RET_IP_, x);
2189 EXPORT_SYMBOL(kmem_cache_free);
2192 * Object placement in a slab is made very easy because we always start at
2193 * offset 0. If we tune the size of the object to the alignment then we can
2194 * get the required alignment by putting one properly sized object after
2197 * Notice that the allocation order determines the sizes of the per cpu
2198 * caches. Each processor has always one slab available for allocations.
2199 * Increasing the allocation order reduces the number of times that slabs
2200 * must be moved on and off the partial lists and is therefore a factor in
2205 * Mininum / Maximum order of slab pages. This influences locking overhead
2206 * and slab fragmentation. A higher order reduces the number of partial slabs
2207 * and increases the number of allocations possible without having to
2208 * take the list_lock.
2210 static int slub_min_order;
2211 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2212 static int slub_min_objects;
2215 * Merge control. If this is set then no merging of slab caches will occur.
2216 * (Could be removed. This was introduced to pacify the merge skeptics.)
2218 static int slub_nomerge;
2221 * Calculate the order of allocation given an slab object size.
2223 * The order of allocation has significant impact on performance and other
2224 * system components. Generally order 0 allocations should be preferred since
2225 * order 0 does not cause fragmentation in the page allocator. Larger objects
2226 * be problematic to put into order 0 slabs because there may be too much
2227 * unused space left. We go to a higher order if more than 1/16th of the slab
2230 * In order to reach satisfactory performance we must ensure that a minimum
2231 * number of objects is in one slab. Otherwise we may generate too much
2232 * activity on the partial lists which requires taking the list_lock. This is
2233 * less a concern for large slabs though which are rarely used.
2235 * slub_max_order specifies the order where we begin to stop considering the
2236 * number of objects in a slab as critical. If we reach slub_max_order then
2237 * we try to keep the page order as low as possible. So we accept more waste
2238 * of space in favor of a small page order.
2240 * Higher order allocations also allow the placement of more objects in a
2241 * slab and thereby reduce object handling overhead. If the user has
2242 * requested a higher mininum order then we start with that one instead of
2243 * the smallest order which will fit the object.
2245 static inline int slab_order(int size, int min_objects,
2246 int max_order, int fract_leftover, int reserved)
2250 int min_order = slub_min_order;
2252 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2253 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2255 for (order = max(min_order,
2256 fls(min_objects * size - 1) - PAGE_SHIFT);
2257 order <= max_order; order++) {
2259 unsigned long slab_size = PAGE_SIZE << order;
2261 if (slab_size < min_objects * size + reserved)
2264 rem = (slab_size - reserved) % size;
2266 if (rem <= slab_size / fract_leftover)
2274 static inline int calculate_order(int size, int reserved)
2282 * Attempt to find best configuration for a slab. This
2283 * works by first attempting to generate a layout with
2284 * the best configuration and backing off gradually.
2286 * First we reduce the acceptable waste in a slab. Then
2287 * we reduce the minimum objects required in a slab.
2289 min_objects = slub_min_objects;
2291 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2292 max_objects = order_objects(slub_max_order, size, reserved);
2293 min_objects = min(min_objects, max_objects);
2295 while (min_objects > 1) {
2297 while (fraction >= 4) {
2298 order = slab_order(size, min_objects,
2299 slub_max_order, fraction, reserved);
2300 if (order <= slub_max_order)
2308 * We were unable to place multiple objects in a slab. Now
2309 * lets see if we can place a single object there.
2311 order = slab_order(size, 1, slub_max_order, 1, reserved);
2312 if (order <= slub_max_order)
2316 * Doh this slab cannot be placed using slub_max_order.
2318 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2319 if (order < MAX_ORDER)
2325 * Figure out what the alignment of the objects will be.
2327 static unsigned long calculate_alignment(unsigned long flags,
2328 unsigned long align, unsigned long size)
2331 * If the user wants hardware cache aligned objects then follow that
2332 * suggestion if the object is sufficiently large.
2334 * The hardware cache alignment cannot override the specified
2335 * alignment though. If that is greater then use it.
2337 if (flags & SLAB_HWCACHE_ALIGN) {
2338 unsigned long ralign = cache_line_size();
2339 while (size <= ralign / 2)
2341 align = max(align, ralign);
2344 if (align < ARCH_SLAB_MINALIGN)
2345 align = ARCH_SLAB_MINALIGN;
2347 return ALIGN(align, sizeof(void *));
2351 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2354 spin_lock_init(&n->list_lock);
2355 INIT_LIST_HEAD(&n->partial);
2356 #ifdef CONFIG_SLUB_DEBUG
2357 atomic_long_set(&n->nr_slabs, 0);
2358 atomic_long_set(&n->total_objects, 0);
2359 INIT_LIST_HEAD(&n->full);
2363 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2365 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2366 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2368 #ifdef CONFIG_CMPXCHG_LOCAL
2370 * Must align to double word boundary for the double cmpxchg instructions
2373 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
2375 /* Regular alignment is sufficient */
2376 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2382 init_kmem_cache_cpus(s);
2387 static struct kmem_cache *kmem_cache_node;
2390 * No kmalloc_node yet so do it by hand. We know that this is the first
2391 * slab on the node for this slabcache. There are no concurrent accesses
2394 * Note that this function only works on the kmalloc_node_cache
2395 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2396 * memory on a fresh node that has no slab structures yet.
2398 static void early_kmem_cache_node_alloc(int node)
2401 struct kmem_cache_node *n;
2402 unsigned long flags;
2404 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2406 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2409 if (page_to_nid(page) != node) {
2410 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2412 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2413 "in order to be able to continue\n");
2418 page->freelist = get_freepointer(kmem_cache_node, n);
2420 kmem_cache_node->node[node] = n;
2421 #ifdef CONFIG_SLUB_DEBUG
2422 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2423 init_tracking(kmem_cache_node, n);
2425 init_kmem_cache_node(n, kmem_cache_node);
2426 inc_slabs_node(kmem_cache_node, node, page->objects);
2429 * lockdep requires consistent irq usage for each lock
2430 * so even though there cannot be a race this early in
2431 * the boot sequence, we still disable irqs.
2433 local_irq_save(flags);
2434 add_partial(n, page, 0);
2435 local_irq_restore(flags);
2438 static void free_kmem_cache_nodes(struct kmem_cache *s)
2442 for_each_node_state(node, N_NORMAL_MEMORY) {
2443 struct kmem_cache_node *n = s->node[node];
2446 kmem_cache_free(kmem_cache_node, n);
2448 s->node[node] = NULL;
2452 static int init_kmem_cache_nodes(struct kmem_cache *s)
2456 for_each_node_state(node, N_NORMAL_MEMORY) {
2457 struct kmem_cache_node *n;
2459 if (slab_state == DOWN) {
2460 early_kmem_cache_node_alloc(node);
2463 n = kmem_cache_alloc_node(kmem_cache_node,
2467 free_kmem_cache_nodes(s);
2472 init_kmem_cache_node(n, s);
2477 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2479 if (min < MIN_PARTIAL)
2481 else if (min > MAX_PARTIAL)
2483 s->min_partial = min;
2487 * calculate_sizes() determines the order and the distribution of data within
2490 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2492 unsigned long flags = s->flags;
2493 unsigned long size = s->objsize;
2494 unsigned long align = s->align;
2498 * Round up object size to the next word boundary. We can only
2499 * place the free pointer at word boundaries and this determines
2500 * the possible location of the free pointer.
2502 size = ALIGN(size, sizeof(void *));
2504 #ifdef CONFIG_SLUB_DEBUG
2506 * Determine if we can poison the object itself. If the user of
2507 * the slab may touch the object after free or before allocation
2508 * then we should never poison the object itself.
2510 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2512 s->flags |= __OBJECT_POISON;
2514 s->flags &= ~__OBJECT_POISON;
2518 * If we are Redzoning then check if there is some space between the
2519 * end of the object and the free pointer. If not then add an
2520 * additional word to have some bytes to store Redzone information.
2522 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2523 size += sizeof(void *);
2527 * With that we have determined the number of bytes in actual use
2528 * by the object. This is the potential offset to the free pointer.
2532 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2535 * Relocate free pointer after the object if it is not
2536 * permitted to overwrite the first word of the object on
2539 * This is the case if we do RCU, have a constructor or
2540 * destructor or are poisoning the objects.
2543 size += sizeof(void *);
2546 #ifdef CONFIG_SLUB_DEBUG
2547 if (flags & SLAB_STORE_USER)
2549 * Need to store information about allocs and frees after
2552 size += 2 * sizeof(struct track);
2554 if (flags & SLAB_RED_ZONE)
2556 * Add some empty padding so that we can catch
2557 * overwrites from earlier objects rather than let
2558 * tracking information or the free pointer be
2559 * corrupted if a user writes before the start
2562 size += sizeof(void *);
2566 * Determine the alignment based on various parameters that the
2567 * user specified and the dynamic determination of cache line size
2570 align = calculate_alignment(flags, align, s->objsize);
2574 * SLUB stores one object immediately after another beginning from
2575 * offset 0. In order to align the objects we have to simply size
2576 * each object to conform to the alignment.
2578 size = ALIGN(size, align);
2580 if (forced_order >= 0)
2581 order = forced_order;
2583 order = calculate_order(size, s->reserved);
2590 s->allocflags |= __GFP_COMP;
2592 if (s->flags & SLAB_CACHE_DMA)
2593 s->allocflags |= SLUB_DMA;
2595 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2596 s->allocflags |= __GFP_RECLAIMABLE;
2599 * Determine the number of objects per slab
2601 s->oo = oo_make(order, size, s->reserved);
2602 s->min = oo_make(get_order(size), size, s->reserved);
2603 if (oo_objects(s->oo) > oo_objects(s->max))
2606 return !!oo_objects(s->oo);
2610 static int kmem_cache_open(struct kmem_cache *s,
2611 const char *name, size_t size,
2612 size_t align, unsigned long flags,
2613 void (*ctor)(void *))
2615 memset(s, 0, kmem_size);
2620 s->flags = kmem_cache_flags(size, flags, name, ctor);
2623 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2624 s->reserved = sizeof(struct rcu_head);
2626 if (!calculate_sizes(s, -1))
2628 if (disable_higher_order_debug) {
2630 * Disable debugging flags that store metadata if the min slab
2633 if (get_order(s->size) > get_order(s->objsize)) {
2634 s->flags &= ~DEBUG_METADATA_FLAGS;
2636 if (!calculate_sizes(s, -1))
2642 * The larger the object size is, the more pages we want on the partial
2643 * list to avoid pounding the page allocator excessively.
2645 set_min_partial(s, ilog2(s->size));
2648 s->remote_node_defrag_ratio = 1000;
2650 if (!init_kmem_cache_nodes(s))
2653 if (alloc_kmem_cache_cpus(s))
2656 free_kmem_cache_nodes(s);
2658 if (flags & SLAB_PANIC)
2659 panic("Cannot create slab %s size=%lu realsize=%u "
2660 "order=%u offset=%u flags=%lx\n",
2661 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2667 * Determine the size of a slab object
2669 unsigned int kmem_cache_size(struct kmem_cache *s)
2673 EXPORT_SYMBOL(kmem_cache_size);
2675 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2678 #ifdef CONFIG_SLUB_DEBUG
2679 void *addr = page_address(page);
2681 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2682 sizeof(long), GFP_ATOMIC);
2685 slab_err(s, page, "%s", text);
2688 get_map(s, page, map);
2689 for_each_object(p, s, addr, page->objects) {
2691 if (!test_bit(slab_index(p, s, addr), map)) {
2692 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2694 print_tracking(s, p);
2703 * Attempt to free all partial slabs on a node.
2705 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2707 unsigned long flags;
2708 struct page *page, *h;
2710 spin_lock_irqsave(&n->list_lock, flags);
2711 list_for_each_entry_safe(page, h, &n->partial, lru) {
2713 __remove_partial(n, page);
2714 discard_slab(s, page);
2716 list_slab_objects(s, page,
2717 "Objects remaining on kmem_cache_close()");
2720 spin_unlock_irqrestore(&n->list_lock, flags);
2724 * Release all resources used by a slab cache.
2726 static inline int kmem_cache_close(struct kmem_cache *s)
2731 free_percpu(s->cpu_slab);
2732 /* Attempt to free all objects */
2733 for_each_node_state(node, N_NORMAL_MEMORY) {
2734 struct kmem_cache_node *n = get_node(s, node);
2737 if (n->nr_partial || slabs_node(s, node))
2740 free_kmem_cache_nodes(s);
2745 * Close a cache and release the kmem_cache structure
2746 * (must be used for caches created using kmem_cache_create)
2748 void kmem_cache_destroy(struct kmem_cache *s)
2750 down_write(&slub_lock);
2754 if (kmem_cache_close(s)) {
2755 printk(KERN_ERR "SLUB %s: %s called for cache that "
2756 "still has objects.\n", s->name, __func__);
2759 if (s->flags & SLAB_DESTROY_BY_RCU)
2761 sysfs_slab_remove(s);
2763 up_write(&slub_lock);
2765 EXPORT_SYMBOL(kmem_cache_destroy);
2767 /********************************************************************
2769 *******************************************************************/
2771 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2772 EXPORT_SYMBOL(kmalloc_caches);
2774 static struct kmem_cache *kmem_cache;
2776 #ifdef CONFIG_ZONE_DMA
2777 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2780 static int __init setup_slub_min_order(char *str)
2782 get_option(&str, &slub_min_order);
2787 __setup("slub_min_order=", setup_slub_min_order);
2789 static int __init setup_slub_max_order(char *str)
2791 get_option(&str, &slub_max_order);
2792 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2797 __setup("slub_max_order=", setup_slub_max_order);
2799 static int __init setup_slub_min_objects(char *str)
2801 get_option(&str, &slub_min_objects);
2806 __setup("slub_min_objects=", setup_slub_min_objects);
2808 static int __init setup_slub_nomerge(char *str)
2814 __setup("slub_nomerge", setup_slub_nomerge);
2816 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2817 int size, unsigned int flags)
2819 struct kmem_cache *s;
2821 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2824 * This function is called with IRQs disabled during early-boot on
2825 * single CPU so there's no need to take slub_lock here.
2827 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2831 list_add(&s->list, &slab_caches);
2835 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2840 * Conversion table for small slabs sizes / 8 to the index in the
2841 * kmalloc array. This is necessary for slabs < 192 since we have non power
2842 * of two cache sizes there. The size of larger slabs can be determined using
2845 static s8 size_index[24] = {
2872 static inline int size_index_elem(size_t bytes)
2874 return (bytes - 1) / 8;
2877 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2883 return ZERO_SIZE_PTR;
2885 index = size_index[size_index_elem(size)];
2887 index = fls(size - 1);
2889 #ifdef CONFIG_ZONE_DMA
2890 if (unlikely((flags & SLUB_DMA)))
2891 return kmalloc_dma_caches[index];
2894 return kmalloc_caches[index];
2897 void *__kmalloc(size_t size, gfp_t flags)
2899 struct kmem_cache *s;
2902 if (unlikely(size > SLUB_MAX_SIZE))
2903 return kmalloc_large(size, flags);
2905 s = get_slab(size, flags);
2907 if (unlikely(ZERO_OR_NULL_PTR(s)))
2910 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2912 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2916 EXPORT_SYMBOL(__kmalloc);
2919 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2924 flags |= __GFP_COMP | __GFP_NOTRACK;
2925 page = alloc_pages_node(node, flags, get_order(size));
2927 ptr = page_address(page);
2929 kmemleak_alloc(ptr, size, 1, flags);
2933 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2935 struct kmem_cache *s;
2938 if (unlikely(size > SLUB_MAX_SIZE)) {
2939 ret = kmalloc_large_node(size, flags, node);
2941 trace_kmalloc_node(_RET_IP_, ret,
2942 size, PAGE_SIZE << get_order(size),
2948 s = get_slab(size, flags);
2950 if (unlikely(ZERO_OR_NULL_PTR(s)))
2953 ret = slab_alloc(s, flags, node, _RET_IP_);
2955 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2959 EXPORT_SYMBOL(__kmalloc_node);
2962 size_t ksize(const void *object)
2966 if (unlikely(object == ZERO_SIZE_PTR))
2969 page = virt_to_head_page(object);
2971 if (unlikely(!PageSlab(page))) {
2972 WARN_ON(!PageCompound(page));
2973 return PAGE_SIZE << compound_order(page);
2976 return slab_ksize(page->slab);
2978 EXPORT_SYMBOL(ksize);
2980 void kfree(const void *x)
2983 void *object = (void *)x;
2985 trace_kfree(_RET_IP_, x);
2987 if (unlikely(ZERO_OR_NULL_PTR(x)))
2990 page = virt_to_head_page(x);
2991 if (unlikely(!PageSlab(page))) {
2992 BUG_ON(!PageCompound(page));
2997 slab_free(page->slab, page, object, _RET_IP_);
2999 EXPORT_SYMBOL(kfree);
3002 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3003 * the remaining slabs by the number of items in use. The slabs with the
3004 * most items in use come first. New allocations will then fill those up
3005 * and thus they can be removed from the partial lists.
3007 * The slabs with the least items are placed last. This results in them
3008 * being allocated from last increasing the chance that the last objects
3009 * are freed in them.
3011 int kmem_cache_shrink(struct kmem_cache *s)
3015 struct kmem_cache_node *n;
3018 int objects = oo_objects(s->max);
3019 struct list_head *slabs_by_inuse =
3020 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3021 unsigned long flags;
3023 if (!slabs_by_inuse)
3027 for_each_node_state(node, N_NORMAL_MEMORY) {
3028 n = get_node(s, node);
3033 for (i = 0; i < objects; i++)
3034 INIT_LIST_HEAD(slabs_by_inuse + i);
3036 spin_lock_irqsave(&n->list_lock, flags);
3039 * Build lists indexed by the items in use in each slab.
3041 * Note that concurrent frees may occur while we hold the
3042 * list_lock. page->inuse here is the upper limit.
3044 list_for_each_entry_safe(page, t, &n->partial, lru) {
3045 if (!page->inuse && slab_trylock(page)) {
3047 * Must hold slab lock here because slab_free
3048 * may have freed the last object and be
3049 * waiting to release the slab.
3051 __remove_partial(n, page);
3053 discard_slab(s, page);
3055 list_move(&page->lru,
3056 slabs_by_inuse + page->inuse);
3061 * Rebuild the partial list with the slabs filled up most
3062 * first and the least used slabs at the end.
3064 for (i = objects - 1; i >= 0; i--)
3065 list_splice(slabs_by_inuse + i, n->partial.prev);
3067 spin_unlock_irqrestore(&n->list_lock, flags);
3070 kfree(slabs_by_inuse);
3073 EXPORT_SYMBOL(kmem_cache_shrink);
3075 #if defined(CONFIG_MEMORY_HOTPLUG)
3076 static int slab_mem_going_offline_callback(void *arg)
3078 struct kmem_cache *s;
3080 down_read(&slub_lock);
3081 list_for_each_entry(s, &slab_caches, list)
3082 kmem_cache_shrink(s);
3083 up_read(&slub_lock);
3088 static void slab_mem_offline_callback(void *arg)
3090 struct kmem_cache_node *n;
3091 struct kmem_cache *s;
3092 struct memory_notify *marg = arg;
3095 offline_node = marg->status_change_nid;
3098 * If the node still has available memory. we need kmem_cache_node
3101 if (offline_node < 0)
3104 down_read(&slub_lock);
3105 list_for_each_entry(s, &slab_caches, list) {
3106 n = get_node(s, offline_node);
3109 * if n->nr_slabs > 0, slabs still exist on the node
3110 * that is going down. We were unable to free them,
3111 * and offline_pages() function shouldn't call this
3112 * callback. So, we must fail.
3114 BUG_ON(slabs_node(s, offline_node));
3116 s->node[offline_node] = NULL;
3117 kmem_cache_free(kmem_cache_node, n);
3120 up_read(&slub_lock);
3123 static int slab_mem_going_online_callback(void *arg)
3125 struct kmem_cache_node *n;
3126 struct kmem_cache *s;
3127 struct memory_notify *marg = arg;
3128 int nid = marg->status_change_nid;
3132 * If the node's memory is already available, then kmem_cache_node is
3133 * already created. Nothing to do.
3139 * We are bringing a node online. No memory is available yet. We must
3140 * allocate a kmem_cache_node structure in order to bring the node
3143 down_read(&slub_lock);
3144 list_for_each_entry(s, &slab_caches, list) {
3146 * XXX: kmem_cache_alloc_node will fallback to other nodes
3147 * since memory is not yet available from the node that
3150 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3155 init_kmem_cache_node(n, s);
3159 up_read(&slub_lock);
3163 static int slab_memory_callback(struct notifier_block *self,
3164 unsigned long action, void *arg)
3169 case MEM_GOING_ONLINE:
3170 ret = slab_mem_going_online_callback(arg);
3172 case MEM_GOING_OFFLINE:
3173 ret = slab_mem_going_offline_callback(arg);
3176 case MEM_CANCEL_ONLINE:
3177 slab_mem_offline_callback(arg);
3180 case MEM_CANCEL_OFFLINE:
3184 ret = notifier_from_errno(ret);
3190 #endif /* CONFIG_MEMORY_HOTPLUG */
3192 /********************************************************************
3193 * Basic setup of slabs
3194 *******************************************************************/
3197 * Used for early kmem_cache structures that were allocated using
3198 * the page allocator
3201 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3205 list_add(&s->list, &slab_caches);
3208 for_each_node_state(node, N_NORMAL_MEMORY) {
3209 struct kmem_cache_node *n = get_node(s, node);
3213 list_for_each_entry(p, &n->partial, lru)
3216 #ifdef CONFIG_SLUB_DEBUG
3217 list_for_each_entry(p, &n->full, lru)
3224 void __init kmem_cache_init(void)
3228 struct kmem_cache *temp_kmem_cache;
3230 struct kmem_cache *temp_kmem_cache_node;
3231 unsigned long kmalloc_size;
3233 kmem_size = offsetof(struct kmem_cache, node) +
3234 nr_node_ids * sizeof(struct kmem_cache_node *);
3236 /* Allocate two kmem_caches from the page allocator */
3237 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3238 order = get_order(2 * kmalloc_size);
3239 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3242 * Must first have the slab cache available for the allocations of the
3243 * struct kmem_cache_node's. There is special bootstrap code in
3244 * kmem_cache_open for slab_state == DOWN.
3246 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3248 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3249 sizeof(struct kmem_cache_node),
3250 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3252 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3254 /* Able to allocate the per node structures */
3255 slab_state = PARTIAL;
3257 temp_kmem_cache = kmem_cache;
3258 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3259 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3260 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3261 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3264 * Allocate kmem_cache_node properly from the kmem_cache slab.
3265 * kmem_cache_node is separately allocated so no need to
3266 * update any list pointers.
3268 temp_kmem_cache_node = kmem_cache_node;
3270 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3271 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3273 kmem_cache_bootstrap_fixup(kmem_cache_node);
3276 kmem_cache_bootstrap_fixup(kmem_cache);
3278 /* Free temporary boot structure */
3279 free_pages((unsigned long)temp_kmem_cache, order);
3281 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3284 * Patch up the size_index table if we have strange large alignment
3285 * requirements for the kmalloc array. This is only the case for
3286 * MIPS it seems. The standard arches will not generate any code here.
3288 * Largest permitted alignment is 256 bytes due to the way we
3289 * handle the index determination for the smaller caches.
3291 * Make sure that nothing crazy happens if someone starts tinkering
3292 * around with ARCH_KMALLOC_MINALIGN
3294 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3295 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3297 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3298 int elem = size_index_elem(i);
3299 if (elem >= ARRAY_SIZE(size_index))
3301 size_index[elem] = KMALLOC_SHIFT_LOW;
3304 if (KMALLOC_MIN_SIZE == 64) {
3306 * The 96 byte size cache is not used if the alignment
3309 for (i = 64 + 8; i <= 96; i += 8)
3310 size_index[size_index_elem(i)] = 7;
3311 } else if (KMALLOC_MIN_SIZE == 128) {
3313 * The 192 byte sized cache is not used if the alignment
3314 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3317 for (i = 128 + 8; i <= 192; i += 8)
3318 size_index[size_index_elem(i)] = 8;
3321 /* Caches that are not of the two-to-the-power-of size */
3322 if (KMALLOC_MIN_SIZE <= 32) {
3323 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3327 if (KMALLOC_MIN_SIZE <= 64) {
3328 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3332 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3333 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3339 /* Provide the correct kmalloc names now that the caches are up */
3340 if (KMALLOC_MIN_SIZE <= 32) {
3341 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3342 BUG_ON(!kmalloc_caches[1]->name);
3345 if (KMALLOC_MIN_SIZE <= 64) {
3346 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3347 BUG_ON(!kmalloc_caches[2]->name);
3350 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3351 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3354 kmalloc_caches[i]->name = s;
3358 register_cpu_notifier(&slab_notifier);
3361 #ifdef CONFIG_ZONE_DMA
3362 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3363 struct kmem_cache *s = kmalloc_caches[i];
3366 char *name = kasprintf(GFP_NOWAIT,
3367 "dma-kmalloc-%d", s->objsize);
3370 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3371 s->objsize, SLAB_CACHE_DMA);
3376 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3377 " CPUs=%d, Nodes=%d\n",
3378 caches, cache_line_size(),
3379 slub_min_order, slub_max_order, slub_min_objects,
3380 nr_cpu_ids, nr_node_ids);
3383 void __init kmem_cache_init_late(void)
3388 * Find a mergeable slab cache
3390 static int slab_unmergeable(struct kmem_cache *s)
3392 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3399 * We may have set a slab to be unmergeable during bootstrap.
3401 if (s->refcount < 0)
3407 static struct kmem_cache *find_mergeable(size_t size,
3408 size_t align, unsigned long flags, const char *name,
3409 void (*ctor)(void *))
3411 struct kmem_cache *s;
3413 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3419 size = ALIGN(size, sizeof(void *));
3420 align = calculate_alignment(flags, align, size);
3421 size = ALIGN(size, align);
3422 flags = kmem_cache_flags(size, flags, name, NULL);
3424 list_for_each_entry(s, &slab_caches, list) {
3425 if (slab_unmergeable(s))
3431 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3434 * Check if alignment is compatible.
3435 * Courtesy of Adrian Drzewiecki
3437 if ((s->size & ~(align - 1)) != s->size)
3440 if (s->size - size >= sizeof(void *))
3448 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3449 size_t align, unsigned long flags, void (*ctor)(void *))
3451 struct kmem_cache *s;
3457 down_write(&slub_lock);
3458 s = find_mergeable(size, align, flags, name, ctor);
3462 * Adjust the object sizes so that we clear
3463 * the complete object on kzalloc.
3465 s->objsize = max(s->objsize, (int)size);
3466 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3468 if (sysfs_slab_alias(s, name)) {
3472 up_write(&slub_lock);
3476 n = kstrdup(name, GFP_KERNEL);
3480 s = kmalloc(kmem_size, GFP_KERNEL);
3482 if (kmem_cache_open(s, n,
3483 size, align, flags, ctor)) {
3484 list_add(&s->list, &slab_caches);
3485 if (sysfs_slab_add(s)) {
3491 up_write(&slub_lock);
3498 up_write(&slub_lock);
3500 if (flags & SLAB_PANIC)
3501 panic("Cannot create slabcache %s\n", name);
3506 EXPORT_SYMBOL(kmem_cache_create);
3510 * Use the cpu notifier to insure that the cpu slabs are flushed when
3513 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3514 unsigned long action, void *hcpu)
3516 long cpu = (long)hcpu;
3517 struct kmem_cache *s;
3518 unsigned long flags;
3521 case CPU_UP_CANCELED:
3522 case CPU_UP_CANCELED_FROZEN:
3524 case CPU_DEAD_FROZEN:
3525 down_read(&slub_lock);
3526 list_for_each_entry(s, &slab_caches, list) {
3527 local_irq_save(flags);
3528 __flush_cpu_slab(s, cpu);
3529 local_irq_restore(flags);
3531 up_read(&slub_lock);
3539 static struct notifier_block __cpuinitdata slab_notifier = {
3540 .notifier_call = slab_cpuup_callback
3545 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3547 struct kmem_cache *s;
3550 if (unlikely(size > SLUB_MAX_SIZE))
3551 return kmalloc_large(size, gfpflags);
3553 s = get_slab(size, gfpflags);
3555 if (unlikely(ZERO_OR_NULL_PTR(s)))
3558 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3560 /* Honor the call site pointer we recieved. */
3561 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3567 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3568 int node, unsigned long caller)
3570 struct kmem_cache *s;
3573 if (unlikely(size > SLUB_MAX_SIZE)) {
3574 ret = kmalloc_large_node(size, gfpflags, node);
3576 trace_kmalloc_node(caller, ret,
3577 size, PAGE_SIZE << get_order(size),
3583 s = get_slab(size, gfpflags);
3585 if (unlikely(ZERO_OR_NULL_PTR(s)))
3588 ret = slab_alloc(s, gfpflags, node, caller);
3590 /* Honor the call site pointer we recieved. */
3591 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3598 static int count_inuse(struct page *page)
3603 static int count_total(struct page *page)
3605 return page->objects;
3609 #ifdef CONFIG_SLUB_DEBUG
3610 static int validate_slab(struct kmem_cache *s, struct page *page,
3614 void *addr = page_address(page);
3616 if (!check_slab(s, page) ||
3617 !on_freelist(s, page, NULL))
3620 /* Now we know that a valid freelist exists */
3621 bitmap_zero(map, page->objects);
3623 get_map(s, page, map);
3624 for_each_object(p, s, addr, page->objects) {
3625 if (test_bit(slab_index(p, s, addr), map))
3626 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3630 for_each_object(p, s, addr, page->objects)
3631 if (!test_bit(slab_index(p, s, addr), map))
3632 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3637 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3640 if (slab_trylock(page)) {
3641 validate_slab(s, page, map);
3644 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3648 static int validate_slab_node(struct kmem_cache *s,
3649 struct kmem_cache_node *n, unsigned long *map)
3651 unsigned long count = 0;
3653 unsigned long flags;
3655 spin_lock_irqsave(&n->list_lock, flags);
3657 list_for_each_entry(page, &n->partial, lru) {
3658 validate_slab_slab(s, page, map);
3661 if (count != n->nr_partial)
3662 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3663 "counter=%ld\n", s->name, count, n->nr_partial);
3665 if (!(s->flags & SLAB_STORE_USER))
3668 list_for_each_entry(page, &n->full, lru) {
3669 validate_slab_slab(s, page, map);
3672 if (count != atomic_long_read(&n->nr_slabs))
3673 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3674 "counter=%ld\n", s->name, count,
3675 atomic_long_read(&n->nr_slabs));
3678 spin_unlock_irqrestore(&n->list_lock, flags);
3682 static long validate_slab_cache(struct kmem_cache *s)
3685 unsigned long count = 0;
3686 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3687 sizeof(unsigned long), GFP_KERNEL);
3693 for_each_node_state(node, N_NORMAL_MEMORY) {
3694 struct kmem_cache_node *n = get_node(s, node);
3696 count += validate_slab_node(s, n, map);
3702 * Generate lists of code addresses where slabcache objects are allocated
3707 unsigned long count;
3714 DECLARE_BITMAP(cpus, NR_CPUS);
3720 unsigned long count;
3721 struct location *loc;
3724 static void free_loc_track(struct loc_track *t)
3727 free_pages((unsigned long)t->loc,
3728 get_order(sizeof(struct location) * t->max));
3731 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3736 order = get_order(sizeof(struct location) * max);
3738 l = (void *)__get_free_pages(flags, order);
3743 memcpy(l, t->loc, sizeof(struct location) * t->count);
3751 static int add_location(struct loc_track *t, struct kmem_cache *s,
3752 const struct track *track)
3754 long start, end, pos;
3756 unsigned long caddr;
3757 unsigned long age = jiffies - track->when;
3763 pos = start + (end - start + 1) / 2;
3766 * There is nothing at "end". If we end up there
3767 * we need to add something to before end.
3772 caddr = t->loc[pos].addr;
3773 if (track->addr == caddr) {
3779 if (age < l->min_time)
3781 if (age > l->max_time)
3784 if (track->pid < l->min_pid)
3785 l->min_pid = track->pid;
3786 if (track->pid > l->max_pid)
3787 l->max_pid = track->pid;
3789 cpumask_set_cpu(track->cpu,
3790 to_cpumask(l->cpus));
3792 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3796 if (track->addr < caddr)
3803 * Not found. Insert new tracking element.
3805 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3811 (t->count - pos) * sizeof(struct location));
3814 l->addr = track->addr;
3818 l->min_pid = track->pid;
3819 l->max_pid = track->pid;
3820 cpumask_clear(to_cpumask(l->cpus));
3821 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3822 nodes_clear(l->nodes);
3823 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3827 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3828 struct page *page, enum track_item alloc,
3831 void *addr = page_address(page);
3834 bitmap_zero(map, page->objects);
3835 get_map(s, page, map);
3837 for_each_object(p, s, addr, page->objects)
3838 if (!test_bit(slab_index(p, s, addr), map))
3839 add_location(t, s, get_track(s, p, alloc));
3842 static int list_locations(struct kmem_cache *s, char *buf,
3843 enum track_item alloc)
3847 struct loc_track t = { 0, 0, NULL };
3849 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3850 sizeof(unsigned long), GFP_KERNEL);
3852 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3855 return sprintf(buf, "Out of memory\n");
3857 /* Push back cpu slabs */
3860 for_each_node_state(node, N_NORMAL_MEMORY) {
3861 struct kmem_cache_node *n = get_node(s, node);
3862 unsigned long flags;
3865 if (!atomic_long_read(&n->nr_slabs))
3868 spin_lock_irqsave(&n->list_lock, flags);
3869 list_for_each_entry(page, &n->partial, lru)
3870 process_slab(&t, s, page, alloc, map);
3871 list_for_each_entry(page, &n->full, lru)
3872 process_slab(&t, s, page, alloc, map);
3873 spin_unlock_irqrestore(&n->list_lock, flags);
3876 for (i = 0; i < t.count; i++) {
3877 struct location *l = &t.loc[i];
3879 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3881 len += sprintf(buf + len, "%7ld ", l->count);
3884 len += sprintf(buf + len, "%pS", (void *)l->addr);
3886 len += sprintf(buf + len, "<not-available>");
3888 if (l->sum_time != l->min_time) {
3889 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3891 (long)div_u64(l->sum_time, l->count),
3894 len += sprintf(buf + len, " age=%ld",
3897 if (l->min_pid != l->max_pid)
3898 len += sprintf(buf + len, " pid=%ld-%ld",
3899 l->min_pid, l->max_pid);
3901 len += sprintf(buf + len, " pid=%ld",
3904 if (num_online_cpus() > 1 &&
3905 !cpumask_empty(to_cpumask(l->cpus)) &&
3906 len < PAGE_SIZE - 60) {
3907 len += sprintf(buf + len, " cpus=");
3908 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3909 to_cpumask(l->cpus));
3912 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3913 len < PAGE_SIZE - 60) {
3914 len += sprintf(buf + len, " nodes=");
3915 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3919 len += sprintf(buf + len, "\n");
3925 len += sprintf(buf, "No data\n");
3930 #ifdef SLUB_RESILIENCY_TEST
3931 static void resiliency_test(void)
3935 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3937 printk(KERN_ERR "SLUB resiliency testing\n");
3938 printk(KERN_ERR "-----------------------\n");
3939 printk(KERN_ERR "A. Corruption after allocation\n");
3941 p = kzalloc(16, GFP_KERNEL);
3943 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3944 " 0x12->0x%p\n\n", p + 16);
3946 validate_slab_cache(kmalloc_caches[4]);
3948 /* Hmmm... The next two are dangerous */
3949 p = kzalloc(32, GFP_KERNEL);
3950 p[32 + sizeof(void *)] = 0x34;
3951 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3952 " 0x34 -> -0x%p\n", p);
3954 "If allocated object is overwritten then not detectable\n\n");
3956 validate_slab_cache(kmalloc_caches[5]);
3957 p = kzalloc(64, GFP_KERNEL);
3958 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3960 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3963 "If allocated object is overwritten then not detectable\n\n");
3964 validate_slab_cache(kmalloc_caches[6]);
3966 printk(KERN_ERR "\nB. Corruption after free\n");
3967 p = kzalloc(128, GFP_KERNEL);
3970 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3971 validate_slab_cache(kmalloc_caches[7]);
3973 p = kzalloc(256, GFP_KERNEL);
3976 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3978 validate_slab_cache(kmalloc_caches[8]);
3980 p = kzalloc(512, GFP_KERNEL);
3983 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3984 validate_slab_cache(kmalloc_caches[9]);
3988 static void resiliency_test(void) {};
3993 enum slab_stat_type {
3994 SL_ALL, /* All slabs */
3995 SL_PARTIAL, /* Only partially allocated slabs */
3996 SL_CPU, /* Only slabs used for cpu caches */
3997 SL_OBJECTS, /* Determine allocated objects not slabs */
3998 SL_TOTAL /* Determine object capacity not slabs */
4001 #define SO_ALL (1 << SL_ALL)
4002 #define SO_PARTIAL (1 << SL_PARTIAL)
4003 #define SO_CPU (1 << SL_CPU)
4004 #define SO_OBJECTS (1 << SL_OBJECTS)
4005 #define SO_TOTAL (1 << SL_TOTAL)
4007 static ssize_t show_slab_objects(struct kmem_cache *s,
4008 char *buf, unsigned long flags)
4010 unsigned long total = 0;
4013 unsigned long *nodes;
4014 unsigned long *per_cpu;
4016 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4019 per_cpu = nodes + nr_node_ids;
4021 if (flags & SO_CPU) {
4024 for_each_possible_cpu(cpu) {
4025 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4027 if (!c || c->node < 0)
4031 if (flags & SO_TOTAL)
4032 x = c->page->objects;
4033 else if (flags & SO_OBJECTS)
4039 nodes[c->node] += x;
4045 lock_memory_hotplug();
4046 #ifdef CONFIG_SLUB_DEBUG
4047 if (flags & SO_ALL) {
4048 for_each_node_state(node, N_NORMAL_MEMORY) {
4049 struct kmem_cache_node *n = get_node(s, node);
4051 if (flags & SO_TOTAL)
4052 x = atomic_long_read(&n->total_objects);
4053 else if (flags & SO_OBJECTS)
4054 x = atomic_long_read(&n->total_objects) -
4055 count_partial(n, count_free);
4058 x = atomic_long_read(&n->nr_slabs);
4065 if (flags & SO_PARTIAL) {
4066 for_each_node_state(node, N_NORMAL_MEMORY) {
4067 struct kmem_cache_node *n = get_node(s, node);
4069 if (flags & SO_TOTAL)
4070 x = count_partial(n, count_total);
4071 else if (flags & SO_OBJECTS)
4072 x = count_partial(n, count_inuse);
4079 x = sprintf(buf, "%lu", total);
4081 for_each_node_state(node, N_NORMAL_MEMORY)
4083 x += sprintf(buf + x, " N%d=%lu",
4086 unlock_memory_hotplug();
4088 return x + sprintf(buf + x, "\n");
4091 #ifdef CONFIG_SLUB_DEBUG
4092 static int any_slab_objects(struct kmem_cache *s)
4096 for_each_online_node(node) {
4097 struct kmem_cache_node *n = get_node(s, node);
4102 if (atomic_long_read(&n->total_objects))
4109 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4110 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4112 struct slab_attribute {
4113 struct attribute attr;
4114 ssize_t (*show)(struct kmem_cache *s, char *buf);
4115 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4118 #define SLAB_ATTR_RO(_name) \
4119 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4121 #define SLAB_ATTR(_name) \
4122 static struct slab_attribute _name##_attr = \
4123 __ATTR(_name, 0644, _name##_show, _name##_store)
4125 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4127 return sprintf(buf, "%d\n", s->size);
4129 SLAB_ATTR_RO(slab_size);
4131 static ssize_t align_show(struct kmem_cache *s, char *buf)
4133 return sprintf(buf, "%d\n", s->align);
4135 SLAB_ATTR_RO(align);
4137 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4139 return sprintf(buf, "%d\n", s->objsize);
4141 SLAB_ATTR_RO(object_size);
4143 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4145 return sprintf(buf, "%d\n", oo_objects(s->oo));
4147 SLAB_ATTR_RO(objs_per_slab);
4149 static ssize_t order_store(struct kmem_cache *s,
4150 const char *buf, size_t length)
4152 unsigned long order;
4155 err = strict_strtoul(buf, 10, &order);
4159 if (order > slub_max_order || order < slub_min_order)
4162 calculate_sizes(s, order);
4166 static ssize_t order_show(struct kmem_cache *s, char *buf)
4168 return sprintf(buf, "%d\n", oo_order(s->oo));
4172 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4174 return sprintf(buf, "%lu\n", s->min_partial);
4177 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4183 err = strict_strtoul(buf, 10, &min);
4187 set_min_partial(s, min);
4190 SLAB_ATTR(min_partial);
4192 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4196 return sprintf(buf, "%pS\n", s->ctor);
4200 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4202 return sprintf(buf, "%d\n", s->refcount - 1);
4204 SLAB_ATTR_RO(aliases);
4206 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4208 return show_slab_objects(s, buf, SO_PARTIAL);
4210 SLAB_ATTR_RO(partial);
4212 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4214 return show_slab_objects(s, buf, SO_CPU);
4216 SLAB_ATTR_RO(cpu_slabs);
4218 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4220 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4222 SLAB_ATTR_RO(objects);
4224 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4226 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4228 SLAB_ATTR_RO(objects_partial);
4230 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4232 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4235 static ssize_t reclaim_account_store(struct kmem_cache *s,
4236 const char *buf, size_t length)
4238 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4240 s->flags |= SLAB_RECLAIM_ACCOUNT;
4243 SLAB_ATTR(reclaim_account);
4245 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4247 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4249 SLAB_ATTR_RO(hwcache_align);
4251 #ifdef CONFIG_ZONE_DMA
4252 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4254 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4256 SLAB_ATTR_RO(cache_dma);
4259 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4261 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4263 SLAB_ATTR_RO(destroy_by_rcu);
4265 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4267 return sprintf(buf, "%d\n", s->reserved);
4269 SLAB_ATTR_RO(reserved);
4271 #ifdef CONFIG_SLUB_DEBUG
4272 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4274 return show_slab_objects(s, buf, SO_ALL);
4276 SLAB_ATTR_RO(slabs);
4278 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4280 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4282 SLAB_ATTR_RO(total_objects);
4284 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4286 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4289 static ssize_t sanity_checks_store(struct kmem_cache *s,
4290 const char *buf, size_t length)
4292 s->flags &= ~SLAB_DEBUG_FREE;
4294 s->flags |= SLAB_DEBUG_FREE;
4297 SLAB_ATTR(sanity_checks);
4299 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4301 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4304 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4307 s->flags &= ~SLAB_TRACE;
4309 s->flags |= SLAB_TRACE;
4314 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4316 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4319 static ssize_t red_zone_store(struct kmem_cache *s,
4320 const char *buf, size_t length)
4322 if (any_slab_objects(s))
4325 s->flags &= ~SLAB_RED_ZONE;
4327 s->flags |= SLAB_RED_ZONE;
4328 calculate_sizes(s, -1);
4331 SLAB_ATTR(red_zone);
4333 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4335 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4338 static ssize_t poison_store(struct kmem_cache *s,
4339 const char *buf, size_t length)
4341 if (any_slab_objects(s))
4344 s->flags &= ~SLAB_POISON;
4346 s->flags |= SLAB_POISON;
4347 calculate_sizes(s, -1);
4352 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4354 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4357 static ssize_t store_user_store(struct kmem_cache *s,
4358 const char *buf, size_t length)
4360 if (any_slab_objects(s))
4363 s->flags &= ~SLAB_STORE_USER;
4365 s->flags |= SLAB_STORE_USER;
4366 calculate_sizes(s, -1);
4369 SLAB_ATTR(store_user);
4371 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4376 static ssize_t validate_store(struct kmem_cache *s,
4377 const char *buf, size_t length)
4381 if (buf[0] == '1') {
4382 ret = validate_slab_cache(s);
4388 SLAB_ATTR(validate);
4390 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4392 if (!(s->flags & SLAB_STORE_USER))
4394 return list_locations(s, buf, TRACK_ALLOC);
4396 SLAB_ATTR_RO(alloc_calls);
4398 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4400 if (!(s->flags & SLAB_STORE_USER))
4402 return list_locations(s, buf, TRACK_FREE);
4404 SLAB_ATTR_RO(free_calls);
4405 #endif /* CONFIG_SLUB_DEBUG */
4407 #ifdef CONFIG_FAILSLAB
4408 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4410 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4413 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4416 s->flags &= ~SLAB_FAILSLAB;
4418 s->flags |= SLAB_FAILSLAB;
4421 SLAB_ATTR(failslab);
4424 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4429 static ssize_t shrink_store(struct kmem_cache *s,
4430 const char *buf, size_t length)
4432 if (buf[0] == '1') {
4433 int rc = kmem_cache_shrink(s);
4444 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4446 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4449 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4450 const char *buf, size_t length)
4452 unsigned long ratio;
4455 err = strict_strtoul(buf, 10, &ratio);
4460 s->remote_node_defrag_ratio = ratio * 10;
4464 SLAB_ATTR(remote_node_defrag_ratio);
4467 #ifdef CONFIG_SLUB_STATS
4468 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4470 unsigned long sum = 0;
4473 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4478 for_each_online_cpu(cpu) {
4479 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4485 len = sprintf(buf, "%lu", sum);
4488 for_each_online_cpu(cpu) {
4489 if (data[cpu] && len < PAGE_SIZE - 20)
4490 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4494 return len + sprintf(buf + len, "\n");
4497 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4501 for_each_online_cpu(cpu)
4502 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4505 #define STAT_ATTR(si, text) \
4506 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4508 return show_stat(s, buf, si); \
4510 static ssize_t text##_store(struct kmem_cache *s, \
4511 const char *buf, size_t length) \
4513 if (buf[0] != '0') \
4515 clear_stat(s, si); \
4520 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4521 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4522 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4523 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4524 STAT_ATTR(FREE_FROZEN, free_frozen);
4525 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4526 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4527 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4528 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4529 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4530 STAT_ATTR(FREE_SLAB, free_slab);
4531 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4532 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4533 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4534 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4535 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4536 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4537 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4540 static struct attribute *slab_attrs[] = {
4541 &slab_size_attr.attr,
4542 &object_size_attr.attr,
4543 &objs_per_slab_attr.attr,
4545 &min_partial_attr.attr,
4547 &objects_partial_attr.attr,
4549 &cpu_slabs_attr.attr,
4553 &hwcache_align_attr.attr,
4554 &reclaim_account_attr.attr,
4555 &destroy_by_rcu_attr.attr,
4557 &reserved_attr.attr,
4558 #ifdef CONFIG_SLUB_DEBUG
4559 &total_objects_attr.attr,
4561 &sanity_checks_attr.attr,
4563 &red_zone_attr.attr,
4565 &store_user_attr.attr,
4566 &validate_attr.attr,
4567 &alloc_calls_attr.attr,
4568 &free_calls_attr.attr,
4570 #ifdef CONFIG_ZONE_DMA
4571 &cache_dma_attr.attr,
4574 &remote_node_defrag_ratio_attr.attr,
4576 #ifdef CONFIG_SLUB_STATS
4577 &alloc_fastpath_attr.attr,
4578 &alloc_slowpath_attr.attr,
4579 &free_fastpath_attr.attr,
4580 &free_slowpath_attr.attr,
4581 &free_frozen_attr.attr,
4582 &free_add_partial_attr.attr,
4583 &free_remove_partial_attr.attr,
4584 &alloc_from_partial_attr.attr,
4585 &alloc_slab_attr.attr,
4586 &alloc_refill_attr.attr,
4587 &free_slab_attr.attr,
4588 &cpuslab_flush_attr.attr,
4589 &deactivate_full_attr.attr,
4590 &deactivate_empty_attr.attr,
4591 &deactivate_to_head_attr.attr,
4592 &deactivate_to_tail_attr.attr,
4593 &deactivate_remote_frees_attr.attr,
4594 &order_fallback_attr.attr,
4596 #ifdef CONFIG_FAILSLAB
4597 &failslab_attr.attr,
4603 static struct attribute_group slab_attr_group = {
4604 .attrs = slab_attrs,
4607 static ssize_t slab_attr_show(struct kobject *kobj,
4608 struct attribute *attr,
4611 struct slab_attribute *attribute;
4612 struct kmem_cache *s;
4615 attribute = to_slab_attr(attr);
4618 if (!attribute->show)
4621 err = attribute->show(s, buf);
4626 static ssize_t slab_attr_store(struct kobject *kobj,
4627 struct attribute *attr,
4628 const char *buf, size_t len)
4630 struct slab_attribute *attribute;
4631 struct kmem_cache *s;
4634 attribute = to_slab_attr(attr);
4637 if (!attribute->store)
4640 err = attribute->store(s, buf, len);
4645 static void kmem_cache_release(struct kobject *kobj)
4647 struct kmem_cache *s = to_slab(kobj);
4653 static const struct sysfs_ops slab_sysfs_ops = {
4654 .show = slab_attr_show,
4655 .store = slab_attr_store,
4658 static struct kobj_type slab_ktype = {
4659 .sysfs_ops = &slab_sysfs_ops,
4660 .release = kmem_cache_release
4663 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4665 struct kobj_type *ktype = get_ktype(kobj);
4667 if (ktype == &slab_ktype)
4672 static const struct kset_uevent_ops slab_uevent_ops = {
4673 .filter = uevent_filter,
4676 static struct kset *slab_kset;
4678 #define ID_STR_LENGTH 64
4680 /* Create a unique string id for a slab cache:
4682 * Format :[flags-]size
4684 static char *create_unique_id(struct kmem_cache *s)
4686 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4693 * First flags affecting slabcache operations. We will only
4694 * get here for aliasable slabs so we do not need to support
4695 * too many flags. The flags here must cover all flags that
4696 * are matched during merging to guarantee that the id is
4699 if (s->flags & SLAB_CACHE_DMA)
4701 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4703 if (s->flags & SLAB_DEBUG_FREE)
4705 if (!(s->flags & SLAB_NOTRACK))
4709 p += sprintf(p, "%07d", s->size);
4710 BUG_ON(p > name + ID_STR_LENGTH - 1);
4714 static int sysfs_slab_add(struct kmem_cache *s)
4720 if (slab_state < SYSFS)
4721 /* Defer until later */
4724 unmergeable = slab_unmergeable(s);
4727 * Slabcache can never be merged so we can use the name proper.
4728 * This is typically the case for debug situations. In that
4729 * case we can catch duplicate names easily.
4731 sysfs_remove_link(&slab_kset->kobj, s->name);
4735 * Create a unique name for the slab as a target
4738 name = create_unique_id(s);
4741 s->kobj.kset = slab_kset;
4742 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4744 kobject_put(&s->kobj);
4748 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4750 kobject_del(&s->kobj);
4751 kobject_put(&s->kobj);
4754 kobject_uevent(&s->kobj, KOBJ_ADD);
4756 /* Setup first alias */
4757 sysfs_slab_alias(s, s->name);
4763 static void sysfs_slab_remove(struct kmem_cache *s)
4765 if (slab_state < SYSFS)
4767 * Sysfs has not been setup yet so no need to remove the
4772 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4773 kobject_del(&s->kobj);
4774 kobject_put(&s->kobj);
4778 * Need to buffer aliases during bootup until sysfs becomes
4779 * available lest we lose that information.
4781 struct saved_alias {
4782 struct kmem_cache *s;
4784 struct saved_alias *next;
4787 static struct saved_alias *alias_list;
4789 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4791 struct saved_alias *al;
4793 if (slab_state == SYSFS) {
4795 * If we have a leftover link then remove it.
4797 sysfs_remove_link(&slab_kset->kobj, name);
4798 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4801 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4807 al->next = alias_list;
4812 static int __init slab_sysfs_init(void)
4814 struct kmem_cache *s;
4817 down_write(&slub_lock);
4819 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4821 up_write(&slub_lock);
4822 printk(KERN_ERR "Cannot register slab subsystem.\n");
4828 list_for_each_entry(s, &slab_caches, list) {
4829 err = sysfs_slab_add(s);
4831 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4832 " to sysfs\n", s->name);
4835 while (alias_list) {
4836 struct saved_alias *al = alias_list;
4838 alias_list = alias_list->next;
4839 err = sysfs_slab_alias(al->s, al->name);
4841 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4842 " %s to sysfs\n", s->name);
4846 up_write(&slub_lock);
4851 __initcall(slab_sysfs_init);
4852 #endif /* CONFIG_SYSFS */
4855 * The /proc/slabinfo ABI
4857 #ifdef CONFIG_SLABINFO
4858 static void print_slabinfo_header(struct seq_file *m)
4860 seq_puts(m, "slabinfo - version: 2.1\n");
4861 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4862 "<objperslab> <pagesperslab>");
4863 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4864 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4868 static void *s_start(struct seq_file *m, loff_t *pos)
4872 down_read(&slub_lock);
4874 print_slabinfo_header(m);
4876 return seq_list_start(&slab_caches, *pos);
4879 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4881 return seq_list_next(p, &slab_caches, pos);
4884 static void s_stop(struct seq_file *m, void *p)
4886 up_read(&slub_lock);
4889 static int s_show(struct seq_file *m, void *p)
4891 unsigned long nr_partials = 0;
4892 unsigned long nr_slabs = 0;
4893 unsigned long nr_inuse = 0;
4894 unsigned long nr_objs = 0;
4895 unsigned long nr_free = 0;
4896 struct kmem_cache *s;
4899 s = list_entry(p, struct kmem_cache, list);
4901 for_each_online_node(node) {
4902 struct kmem_cache_node *n = get_node(s, node);
4907 nr_partials += n->nr_partial;
4908 nr_slabs += atomic_long_read(&n->nr_slabs);
4909 nr_objs += atomic_long_read(&n->total_objects);
4910 nr_free += count_partial(n, count_free);
4913 nr_inuse = nr_objs - nr_free;
4915 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4916 nr_objs, s->size, oo_objects(s->oo),
4917 (1 << oo_order(s->oo)));
4918 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4919 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4925 static const struct seq_operations slabinfo_op = {
4932 static int slabinfo_open(struct inode *inode, struct file *file)
4934 return seq_open(file, &slabinfo_op);
4937 static const struct file_operations proc_slabinfo_operations = {
4938 .open = slabinfo_open,
4940 .llseek = seq_lseek,
4941 .release = seq_release,
4944 static int __init slab_proc_init(void)
4946 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4949 module_init(slab_proc_init);
4950 #endif /* CONFIG_SLABINFO */