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 no one 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 *get_freepointer_safe(struct kmem_cache *s, void *object)
268 #ifdef CONFIG_DEBUG_PAGEALLOC
269 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
271 p = get_freepointer(s, object);
276 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
278 *(void **)(object + s->offset) = fp;
281 /* Loop over all objects in a slab */
282 #define for_each_object(__p, __s, __addr, __objects) \
283 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
287 #define for_each_free_object(__p, __s, __free) \
288 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 return (p - addr) / s->size;
296 static inline size_t slab_ksize(const struct kmem_cache *s)
298 #ifdef CONFIG_SLUB_DEBUG
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
315 * Else we can use all the padding etc for the allocation
320 static inline int order_objects(int order, unsigned long size, int reserved)
322 return ((PAGE_SIZE << order) - reserved) / size;
325 static inline struct kmem_cache_order_objects oo_make(int order,
326 unsigned long size, int reserved)
328 struct kmem_cache_order_objects x = {
329 (order << OO_SHIFT) + order_objects(order, size, reserved)
335 static inline int oo_order(struct kmem_cache_order_objects x)
337 return x.x >> OO_SHIFT;
340 static inline int oo_objects(struct kmem_cache_order_objects x)
342 return x.x & OO_MASK;
345 #ifdef CONFIG_SLUB_DEBUG
349 #ifdef CONFIG_SLUB_DEBUG_ON
350 static int slub_debug = DEBUG_DEFAULT_FLAGS;
352 static int slub_debug;
355 static char *slub_debug_slabs;
356 static int disable_higher_order_debug;
361 static void print_section(char *text, u8 *addr, unsigned int length)
369 for (i = 0; i < length; i++) {
371 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
374 printk(KERN_CONT " %02x", addr[i]);
376 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
378 printk(KERN_CONT " %s\n", ascii);
385 printk(KERN_CONT " ");
389 printk(KERN_CONT " %s\n", ascii);
393 static struct track *get_track(struct kmem_cache *s, void *object,
394 enum track_item alloc)
399 p = object + s->offset + sizeof(void *);
401 p = object + s->inuse;
406 static void set_track(struct kmem_cache *s, void *object,
407 enum track_item alloc, unsigned long addr)
409 struct track *p = get_track(s, object, alloc);
413 p->cpu = smp_processor_id();
414 p->pid = current->pid;
417 memset(p, 0, sizeof(struct track));
420 static void init_tracking(struct kmem_cache *s, void *object)
422 if (!(s->flags & SLAB_STORE_USER))
425 set_track(s, object, TRACK_FREE, 0UL);
426 set_track(s, object, TRACK_ALLOC, 0UL);
429 static void print_track(const char *s, struct track *t)
434 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
435 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
438 static void print_tracking(struct kmem_cache *s, void *object)
440 if (!(s->flags & SLAB_STORE_USER))
443 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
444 print_track("Freed", get_track(s, object, TRACK_FREE));
447 static void print_page_info(struct page *page)
449 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
450 page, page->objects, page->inuse, page->freelist, page->flags);
454 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
460 vsnprintf(buf, sizeof(buf), fmt, args);
462 printk(KERN_ERR "========================================"
463 "=====================================\n");
464 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
465 printk(KERN_ERR "----------------------------------------"
466 "-------------------------------------\n\n");
469 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
475 vsnprintf(buf, sizeof(buf), fmt, args);
477 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
480 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
482 unsigned int off; /* Offset of last byte */
483 u8 *addr = page_address(page);
485 print_tracking(s, p);
487 print_page_info(page);
489 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
490 p, p - addr, get_freepointer(s, p));
493 print_section("Bytes b4", p - 16, 16);
495 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
497 if (s->flags & SLAB_RED_ZONE)
498 print_section("Redzone", p + s->objsize,
499 s->inuse - s->objsize);
502 off = s->offset + sizeof(void *);
506 if (s->flags & SLAB_STORE_USER)
507 off += 2 * sizeof(struct track);
510 /* Beginning of the filler is the free pointer */
511 print_section("Padding", p + off, s->size - off);
516 static void object_err(struct kmem_cache *s, struct page *page,
517 u8 *object, char *reason)
519 slab_bug(s, "%s", reason);
520 print_trailer(s, page, object);
523 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
529 vsnprintf(buf, sizeof(buf), fmt, args);
531 slab_bug(s, "%s", buf);
532 print_page_info(page);
536 static void init_object(struct kmem_cache *s, void *object, u8 val)
540 if (s->flags & __OBJECT_POISON) {
541 memset(p, POISON_FREE, s->objsize - 1);
542 p[s->objsize - 1] = POISON_END;
545 if (s->flags & SLAB_RED_ZONE)
546 memset(p + s->objsize, val, s->inuse - s->objsize);
549 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
552 if (*start != (u8)value)
560 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
561 void *from, void *to)
563 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
564 memset(from, data, to - from);
567 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
568 u8 *object, char *what,
569 u8 *start, unsigned int value, unsigned int bytes)
574 fault = check_bytes(start, value, bytes);
579 while (end > fault && end[-1] == value)
582 slab_bug(s, "%s overwritten", what);
583 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
584 fault, end - 1, fault[0], value);
585 print_trailer(s, page, object);
587 restore_bytes(s, what, value, fault, end);
595 * Bytes of the object to be managed.
596 * If the freepointer may overlay the object then the free
597 * pointer is the first word of the object.
599 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
602 * object + s->objsize
603 * Padding to reach word boundary. This is also used for Redzoning.
604 * Padding is extended by another word if Redzoning is enabled and
607 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
608 * 0xcc (RED_ACTIVE) for objects in use.
611 * Meta data starts here.
613 * A. Free pointer (if we cannot overwrite object on free)
614 * B. Tracking data for SLAB_STORE_USER
615 * C. Padding to reach required alignment boundary or at mininum
616 * one word if debugging is on to be able to detect writes
617 * before the word boundary.
619 * Padding is done using 0x5a (POISON_INUSE)
622 * Nothing is used beyond s->size.
624 * If slabcaches are merged then the objsize and inuse boundaries are mostly
625 * ignored. And therefore no slab options that rely on these boundaries
626 * may be used with merged slabcaches.
629 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
631 unsigned long off = s->inuse; /* The end of info */
634 /* Freepointer is placed after the object. */
635 off += sizeof(void *);
637 if (s->flags & SLAB_STORE_USER)
638 /* We also have user information there */
639 off += 2 * sizeof(struct track);
644 return check_bytes_and_report(s, page, p, "Object padding",
645 p + off, POISON_INUSE, s->size - off);
648 /* Check the pad bytes at the end of a slab page */
649 static int slab_pad_check(struct kmem_cache *s, struct page *page)
657 if (!(s->flags & SLAB_POISON))
660 start = page_address(page);
661 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
662 end = start + length;
663 remainder = length % s->size;
667 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
670 while (end > fault && end[-1] == POISON_INUSE)
673 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
674 print_section("Padding", end - remainder, remainder);
676 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
680 static int check_object(struct kmem_cache *s, struct page *page,
681 void *object, u8 val)
684 u8 *endobject = object + s->objsize;
686 if (s->flags & SLAB_RED_ZONE) {
687 if (!check_bytes_and_report(s, page, object, "Redzone",
688 endobject, val, s->inuse - s->objsize))
691 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
692 check_bytes_and_report(s, page, p, "Alignment padding",
693 endobject, POISON_INUSE, s->inuse - s->objsize);
697 if (s->flags & SLAB_POISON) {
698 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
699 (!check_bytes_and_report(s, page, p, "Poison", p,
700 POISON_FREE, s->objsize - 1) ||
701 !check_bytes_and_report(s, page, p, "Poison",
702 p + s->objsize - 1, POISON_END, 1)))
705 * check_pad_bytes cleans up on its own.
707 check_pad_bytes(s, page, p);
710 if (!s->offset && val == SLUB_RED_ACTIVE)
712 * Object and freepointer overlap. Cannot check
713 * freepointer while object is allocated.
717 /* Check free pointer validity */
718 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
719 object_err(s, page, p, "Freepointer corrupt");
721 * No choice but to zap it and thus lose the remainder
722 * of the free objects in this slab. May cause
723 * another error because the object count is now wrong.
725 set_freepointer(s, p, NULL);
731 static int check_slab(struct kmem_cache *s, struct page *page)
735 VM_BUG_ON(!irqs_disabled());
737 if (!PageSlab(page)) {
738 slab_err(s, page, "Not a valid slab page");
742 maxobj = order_objects(compound_order(page), s->size, s->reserved);
743 if (page->objects > maxobj) {
744 slab_err(s, page, "objects %u > max %u",
745 s->name, page->objects, maxobj);
748 if (page->inuse > page->objects) {
749 slab_err(s, page, "inuse %u > max %u",
750 s->name, page->inuse, page->objects);
753 /* Slab_pad_check fixes things up after itself */
754 slab_pad_check(s, page);
759 * Determine if a certain object on a page is on the freelist. Must hold the
760 * slab lock to guarantee that the chains are in a consistent state.
762 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
765 void *fp = page->freelist;
767 unsigned long max_objects;
769 while (fp && nr <= page->objects) {
772 if (!check_valid_pointer(s, page, fp)) {
774 object_err(s, page, object,
775 "Freechain corrupt");
776 set_freepointer(s, object, NULL);
779 slab_err(s, page, "Freepointer corrupt");
780 page->freelist = NULL;
781 page->inuse = page->objects;
782 slab_fix(s, "Freelist cleared");
788 fp = get_freepointer(s, object);
792 max_objects = order_objects(compound_order(page), s->size, s->reserved);
793 if (max_objects > MAX_OBJS_PER_PAGE)
794 max_objects = MAX_OBJS_PER_PAGE;
796 if (page->objects != max_objects) {
797 slab_err(s, page, "Wrong number of objects. Found %d but "
798 "should be %d", page->objects, max_objects);
799 page->objects = max_objects;
800 slab_fix(s, "Number of objects adjusted.");
802 if (page->inuse != page->objects - nr) {
803 slab_err(s, page, "Wrong object count. Counter is %d but "
804 "counted were %d", page->inuse, page->objects - nr);
805 page->inuse = page->objects - nr;
806 slab_fix(s, "Object count adjusted.");
808 return search == NULL;
811 static void trace(struct kmem_cache *s, struct page *page, void *object,
814 if (s->flags & SLAB_TRACE) {
815 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
817 alloc ? "alloc" : "free",
822 print_section("Object", (void *)object, s->objsize);
829 * Hooks for other subsystems that check memory allocations. In a typical
830 * production configuration these hooks all should produce no code at all.
832 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
834 flags &= gfp_allowed_mask;
835 lockdep_trace_alloc(flags);
836 might_sleep_if(flags & __GFP_WAIT);
838 return should_failslab(s->objsize, flags, s->flags);
841 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
843 flags &= gfp_allowed_mask;
844 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
845 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
848 static inline void slab_free_hook(struct kmem_cache *s, void *x)
850 kmemleak_free_recursive(x, s->flags);
853 * Trouble is that we may no longer disable interupts in the fast path
854 * So in order to make the debug calls that expect irqs to be
855 * disabled we need to disable interrupts temporarily.
857 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
861 local_irq_save(flags);
862 kmemcheck_slab_free(s, x, s->objsize);
863 debug_check_no_locks_freed(x, s->objsize);
864 local_irq_restore(flags);
867 if (!(s->flags & SLAB_DEBUG_OBJECTS))
868 debug_check_no_obj_freed(x, s->objsize);
872 * Tracking of fully allocated slabs for debugging purposes.
874 static void add_full(struct kmem_cache_node *n, struct page *page)
876 spin_lock(&n->list_lock);
877 list_add(&page->lru, &n->full);
878 spin_unlock(&n->list_lock);
881 static void remove_full(struct kmem_cache *s, struct page *page)
883 struct kmem_cache_node *n;
885 if (!(s->flags & SLAB_STORE_USER))
888 n = get_node(s, page_to_nid(page));
890 spin_lock(&n->list_lock);
891 list_del(&page->lru);
892 spin_unlock(&n->list_lock);
895 /* Tracking of the number of slabs for debugging purposes */
896 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
898 struct kmem_cache_node *n = get_node(s, node);
900 return atomic_long_read(&n->nr_slabs);
903 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
905 return atomic_long_read(&n->nr_slabs);
908 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
910 struct kmem_cache_node *n = get_node(s, node);
913 * May be called early in order to allocate a slab for the
914 * kmem_cache_node structure. Solve the chicken-egg
915 * dilemma by deferring the increment of the count during
916 * bootstrap (see early_kmem_cache_node_alloc).
919 atomic_long_inc(&n->nr_slabs);
920 atomic_long_add(objects, &n->total_objects);
923 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
925 struct kmem_cache_node *n = get_node(s, node);
927 atomic_long_dec(&n->nr_slabs);
928 atomic_long_sub(objects, &n->total_objects);
931 /* Object debug checks for alloc/free paths */
932 static void setup_object_debug(struct kmem_cache *s, struct page *page,
935 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
938 init_object(s, object, SLUB_RED_INACTIVE);
939 init_tracking(s, object);
942 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
943 void *object, unsigned long addr)
945 if (!check_slab(s, page))
948 if (!on_freelist(s, page, object)) {
949 object_err(s, page, object, "Object already allocated");
953 if (!check_valid_pointer(s, page, object)) {
954 object_err(s, page, object, "Freelist Pointer check fails");
958 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
961 /* Success perform special debug activities for allocs */
962 if (s->flags & SLAB_STORE_USER)
963 set_track(s, object, TRACK_ALLOC, addr);
964 trace(s, page, object, 1);
965 init_object(s, object, SLUB_RED_ACTIVE);
969 if (PageSlab(page)) {
971 * If this is a slab page then lets do the best we can
972 * to avoid issues in the future. Marking all objects
973 * as used avoids touching the remaining objects.
975 slab_fix(s, "Marking all objects used");
976 page->inuse = page->objects;
977 page->freelist = NULL;
982 static noinline int free_debug_processing(struct kmem_cache *s,
983 struct page *page, void *object, unsigned long addr)
985 if (!check_slab(s, page))
988 if (!check_valid_pointer(s, page, object)) {
989 slab_err(s, page, "Invalid object pointer 0x%p", object);
993 if (on_freelist(s, page, object)) {
994 object_err(s, page, object, "Object already free");
998 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1001 if (unlikely(s != page->slab)) {
1002 if (!PageSlab(page)) {
1003 slab_err(s, page, "Attempt to free object(0x%p) "
1004 "outside of slab", object);
1005 } else if (!page->slab) {
1007 "SLUB <none>: no slab for object 0x%p.\n",
1011 object_err(s, page, object,
1012 "page slab pointer corrupt.");
1016 /* Special debug activities for freeing objects */
1017 if (!PageSlubFrozen(page) && !page->freelist)
1018 remove_full(s, page);
1019 if (s->flags & SLAB_STORE_USER)
1020 set_track(s, object, TRACK_FREE, addr);
1021 trace(s, page, object, 0);
1022 init_object(s, object, SLUB_RED_INACTIVE);
1026 slab_fix(s, "Object at 0x%p not freed", object);
1030 static int __init setup_slub_debug(char *str)
1032 slub_debug = DEBUG_DEFAULT_FLAGS;
1033 if (*str++ != '=' || !*str)
1035 * No options specified. Switch on full debugging.
1041 * No options but restriction on slabs. This means full
1042 * debugging for slabs matching a pattern.
1046 if (tolower(*str) == 'o') {
1048 * Avoid enabling debugging on caches if its minimum order
1049 * would increase as a result.
1051 disable_higher_order_debug = 1;
1058 * Switch off all debugging measures.
1063 * Determine which debug features should be switched on
1065 for (; *str && *str != ','; str++) {
1066 switch (tolower(*str)) {
1068 slub_debug |= SLAB_DEBUG_FREE;
1071 slub_debug |= SLAB_RED_ZONE;
1074 slub_debug |= SLAB_POISON;
1077 slub_debug |= SLAB_STORE_USER;
1080 slub_debug |= SLAB_TRACE;
1083 slub_debug |= SLAB_FAILSLAB;
1086 printk(KERN_ERR "slub_debug option '%c' "
1087 "unknown. skipped\n", *str);
1093 slub_debug_slabs = str + 1;
1098 __setup("slub_debug", setup_slub_debug);
1100 static unsigned long kmem_cache_flags(unsigned long objsize,
1101 unsigned long flags, const char *name,
1102 void (*ctor)(void *))
1105 * Enable debugging if selected on the kernel commandline.
1107 if (slub_debug && (!slub_debug_slabs ||
1108 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1109 flags |= slub_debug;
1114 static inline void setup_object_debug(struct kmem_cache *s,
1115 struct page *page, void *object) {}
1117 static inline int alloc_debug_processing(struct kmem_cache *s,
1118 struct page *page, void *object, unsigned long addr) { return 0; }
1120 static inline int free_debug_processing(struct kmem_cache *s,
1121 struct page *page, void *object, unsigned long addr) { return 0; }
1123 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1125 static inline int check_object(struct kmem_cache *s, struct page *page,
1126 void *object, u8 val) { return 1; }
1127 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1128 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1129 unsigned long flags, const char *name,
1130 void (*ctor)(void *))
1134 #define slub_debug 0
1136 #define disable_higher_order_debug 0
1138 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1140 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1142 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1144 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1147 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1150 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1153 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1155 #endif /* CONFIG_SLUB_DEBUG */
1158 * Slab allocation and freeing
1160 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1161 struct kmem_cache_order_objects oo)
1163 int order = oo_order(oo);
1165 flags |= __GFP_NOTRACK;
1167 if (node == NUMA_NO_NODE)
1168 return alloc_pages(flags, order);
1170 return alloc_pages_exact_node(node, flags, order);
1173 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1176 struct kmem_cache_order_objects oo = s->oo;
1179 flags |= s->allocflags;
1182 * Let the initial higher-order allocation fail under memory pressure
1183 * so we fall-back to the minimum order allocation.
1185 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1187 page = alloc_slab_page(alloc_gfp, node, oo);
1188 if (unlikely(!page)) {
1191 * Allocation may have failed due to fragmentation.
1192 * Try a lower order alloc if possible
1194 page = alloc_slab_page(flags, node, oo);
1198 stat(s, ORDER_FALLBACK);
1201 if (kmemcheck_enabled
1202 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1203 int pages = 1 << oo_order(oo);
1205 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1208 * Objects from caches that have a constructor don't get
1209 * cleared when they're allocated, so we need to do it here.
1212 kmemcheck_mark_uninitialized_pages(page, pages);
1214 kmemcheck_mark_unallocated_pages(page, pages);
1217 page->objects = oo_objects(oo);
1218 mod_zone_page_state(page_zone(page),
1219 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1220 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1226 static void setup_object(struct kmem_cache *s, struct page *page,
1229 setup_object_debug(s, page, object);
1230 if (unlikely(s->ctor))
1234 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1241 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1243 page = allocate_slab(s,
1244 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1248 inc_slabs_node(s, page_to_nid(page), page->objects);
1250 page->flags |= 1 << PG_slab;
1252 start = page_address(page);
1254 if (unlikely(s->flags & SLAB_POISON))
1255 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1258 for_each_object(p, s, start, page->objects) {
1259 setup_object(s, page, last);
1260 set_freepointer(s, last, p);
1263 setup_object(s, page, last);
1264 set_freepointer(s, last, NULL);
1266 page->freelist = start;
1272 static void __free_slab(struct kmem_cache *s, struct page *page)
1274 int order = compound_order(page);
1275 int pages = 1 << order;
1277 if (kmem_cache_debug(s)) {
1280 slab_pad_check(s, page);
1281 for_each_object(p, s, page_address(page),
1283 check_object(s, page, p, SLUB_RED_INACTIVE);
1286 kmemcheck_free_shadow(page, compound_order(page));
1288 mod_zone_page_state(page_zone(page),
1289 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1290 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1293 __ClearPageSlab(page);
1294 reset_page_mapcount(page);
1295 if (current->reclaim_state)
1296 current->reclaim_state->reclaimed_slab += pages;
1297 __free_pages(page, order);
1300 #define need_reserve_slab_rcu \
1301 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1303 static void rcu_free_slab(struct rcu_head *h)
1307 if (need_reserve_slab_rcu)
1308 page = virt_to_head_page(h);
1310 page = container_of((struct list_head *)h, struct page, lru);
1312 __free_slab(page->slab, page);
1315 static void free_slab(struct kmem_cache *s, struct page *page)
1317 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1318 struct rcu_head *head;
1320 if (need_reserve_slab_rcu) {
1321 int order = compound_order(page);
1322 int offset = (PAGE_SIZE << order) - s->reserved;
1324 VM_BUG_ON(s->reserved != sizeof(*head));
1325 head = page_address(page) + offset;
1328 * RCU free overloads the RCU head over the LRU
1330 head = (void *)&page->lru;
1333 call_rcu(head, rcu_free_slab);
1335 __free_slab(s, page);
1338 static void discard_slab(struct kmem_cache *s, struct page *page)
1340 dec_slabs_node(s, page_to_nid(page), page->objects);
1345 * Per slab locking using the pagelock
1347 static __always_inline void slab_lock(struct page *page)
1349 bit_spin_lock(PG_locked, &page->flags);
1352 static __always_inline void slab_unlock(struct page *page)
1354 __bit_spin_unlock(PG_locked, &page->flags);
1357 static __always_inline int slab_trylock(struct page *page)
1361 rc = bit_spin_trylock(PG_locked, &page->flags);
1366 * Management of partially allocated slabs
1368 static void add_partial(struct kmem_cache_node *n,
1369 struct page *page, int tail)
1371 spin_lock(&n->list_lock);
1374 list_add_tail(&page->lru, &n->partial);
1376 list_add(&page->lru, &n->partial);
1377 spin_unlock(&n->list_lock);
1380 static inline void __remove_partial(struct kmem_cache_node *n,
1383 list_del(&page->lru);
1387 static void remove_partial(struct kmem_cache *s, struct page *page)
1389 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1391 spin_lock(&n->list_lock);
1392 __remove_partial(n, page);
1393 spin_unlock(&n->list_lock);
1397 * Lock slab and remove from the partial list.
1399 * Must hold list_lock.
1401 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1404 if (slab_trylock(page)) {
1405 __remove_partial(n, page);
1406 __SetPageSlubFrozen(page);
1413 * Try to allocate a partial slab from a specific node.
1415 static struct page *get_partial_node(struct kmem_cache_node *n)
1420 * Racy check. If we mistakenly see no partial slabs then we
1421 * just allocate an empty slab. If we mistakenly try to get a
1422 * partial slab and there is none available then get_partials()
1425 if (!n || !n->nr_partial)
1428 spin_lock(&n->list_lock);
1429 list_for_each_entry(page, &n->partial, lru)
1430 if (lock_and_freeze_slab(n, page))
1434 spin_unlock(&n->list_lock);
1439 * Get a page from somewhere. Search in increasing NUMA distances.
1441 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1444 struct zonelist *zonelist;
1447 enum zone_type high_zoneidx = gfp_zone(flags);
1451 * The defrag ratio allows a configuration of the tradeoffs between
1452 * inter node defragmentation and node local allocations. A lower
1453 * defrag_ratio increases the tendency to do local allocations
1454 * instead of attempting to obtain partial slabs from other nodes.
1456 * If the defrag_ratio is set to 0 then kmalloc() always
1457 * returns node local objects. If the ratio is higher then kmalloc()
1458 * may return off node objects because partial slabs are obtained
1459 * from other nodes and filled up.
1461 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1462 * defrag_ratio = 1000) then every (well almost) allocation will
1463 * first attempt to defrag slab caches on other nodes. This means
1464 * scanning over all nodes to look for partial slabs which may be
1465 * expensive if we do it every time we are trying to find a slab
1466 * with available objects.
1468 if (!s->remote_node_defrag_ratio ||
1469 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1473 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1474 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1475 struct kmem_cache_node *n;
1477 n = get_node(s, zone_to_nid(zone));
1479 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1480 n->nr_partial > s->min_partial) {
1481 page = get_partial_node(n);
1494 * Get a partial page, lock it and return it.
1496 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1499 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1501 page = get_partial_node(get_node(s, searchnode));
1502 if (page || node != -1)
1505 return get_any_partial(s, flags);
1509 * Move a page back to the lists.
1511 * Must be called with the slab lock held.
1513 * On exit the slab lock will have been dropped.
1515 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1518 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1520 __ClearPageSlubFrozen(page);
1523 if (page->freelist) {
1524 add_partial(n, page, tail);
1525 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1527 stat(s, DEACTIVATE_FULL);
1528 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1533 stat(s, DEACTIVATE_EMPTY);
1534 if (n->nr_partial < s->min_partial) {
1536 * Adding an empty slab to the partial slabs in order
1537 * to avoid page allocator overhead. This slab needs
1538 * to come after the other slabs with objects in
1539 * so that the others get filled first. That way the
1540 * size of the partial list stays small.
1542 * kmem_cache_shrink can reclaim any empty slabs from
1545 add_partial(n, page, 1);
1550 discard_slab(s, page);
1555 #ifdef CONFIG_CMPXCHG_LOCAL
1556 #ifdef CONFIG_PREEMPT
1558 * Calculate the next globally unique transaction for disambiguiation
1559 * during cmpxchg. The transactions start with the cpu number and are then
1560 * incremented by CONFIG_NR_CPUS.
1562 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1565 * No preemption supported therefore also no need to check for
1571 static inline unsigned long next_tid(unsigned long tid)
1573 return tid + TID_STEP;
1576 static inline unsigned int tid_to_cpu(unsigned long tid)
1578 return tid % TID_STEP;
1581 static inline unsigned long tid_to_event(unsigned long tid)
1583 return tid / TID_STEP;
1586 static inline unsigned int init_tid(int cpu)
1591 static inline void note_cmpxchg_failure(const char *n,
1592 const struct kmem_cache *s, unsigned long tid)
1594 #ifdef SLUB_DEBUG_CMPXCHG
1595 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1597 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1599 #ifdef CONFIG_PREEMPT
1600 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1601 printk("due to cpu change %d -> %d\n",
1602 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1605 if (tid_to_event(tid) != tid_to_event(actual_tid))
1606 printk("due to cpu running other code. Event %ld->%ld\n",
1607 tid_to_event(tid), tid_to_event(actual_tid));
1609 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1610 actual_tid, tid, next_tid(tid));
1612 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1617 void init_kmem_cache_cpus(struct kmem_cache *s)
1619 #ifdef CONFIG_CMPXCHG_LOCAL
1622 for_each_possible_cpu(cpu)
1623 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1628 * Remove the cpu slab
1630 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1633 struct page *page = c->page;
1637 stat(s, DEACTIVATE_REMOTE_FREES);
1639 * Merge cpu freelist into slab freelist. Typically we get here
1640 * because both freelists are empty. So this is unlikely
1643 while (unlikely(c->freelist)) {
1646 tail = 0; /* Hot objects. Put the slab first */
1648 /* Retrieve object from cpu_freelist */
1649 object = c->freelist;
1650 c->freelist = get_freepointer(s, c->freelist);
1652 /* And put onto the regular freelist */
1653 set_freepointer(s, object, page->freelist);
1654 page->freelist = object;
1658 #ifdef CONFIG_CMPXCHG_LOCAL
1659 c->tid = next_tid(c->tid);
1661 unfreeze_slab(s, page, tail);
1664 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1666 stat(s, CPUSLAB_FLUSH);
1668 deactivate_slab(s, c);
1674 * Called from IPI handler with interrupts disabled.
1676 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1678 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1680 if (likely(c && c->page))
1684 static void flush_cpu_slab(void *d)
1686 struct kmem_cache *s = d;
1688 __flush_cpu_slab(s, smp_processor_id());
1691 static void flush_all(struct kmem_cache *s)
1693 on_each_cpu(flush_cpu_slab, s, 1);
1697 * Check if the objects in a per cpu structure fit numa
1698 * locality expectations.
1700 static inline int node_match(struct kmem_cache_cpu *c, int node)
1703 if (node != NUMA_NO_NODE && c->node != node)
1709 static int count_free(struct page *page)
1711 return page->objects - page->inuse;
1714 static unsigned long count_partial(struct kmem_cache_node *n,
1715 int (*get_count)(struct page *))
1717 unsigned long flags;
1718 unsigned long x = 0;
1721 spin_lock_irqsave(&n->list_lock, flags);
1722 list_for_each_entry(page, &n->partial, lru)
1723 x += get_count(page);
1724 spin_unlock_irqrestore(&n->list_lock, flags);
1728 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1730 #ifdef CONFIG_SLUB_DEBUG
1731 return atomic_long_read(&n->total_objects);
1737 static noinline void
1738 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1743 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1745 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1746 "default order: %d, min order: %d\n", s->name, s->objsize,
1747 s->size, oo_order(s->oo), oo_order(s->min));
1749 if (oo_order(s->min) > get_order(s->objsize))
1750 printk(KERN_WARNING " %s debugging increased min order, use "
1751 "slub_debug=O to disable.\n", s->name);
1753 for_each_online_node(node) {
1754 struct kmem_cache_node *n = get_node(s, node);
1755 unsigned long nr_slabs;
1756 unsigned long nr_objs;
1757 unsigned long nr_free;
1762 nr_free = count_partial(n, count_free);
1763 nr_slabs = node_nr_slabs(n);
1764 nr_objs = node_nr_objs(n);
1767 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1768 node, nr_slabs, nr_objs, nr_free);
1773 * Slow path. The lockless freelist is empty or we need to perform
1776 * Interrupts are disabled.
1778 * Processing is still very fast if new objects have been freed to the
1779 * regular freelist. In that case we simply take over the regular freelist
1780 * as the lockless freelist and zap the regular freelist.
1782 * If that is not working then we fall back to the partial lists. We take the
1783 * first element of the freelist as the object to allocate now and move the
1784 * rest of the freelist to the lockless freelist.
1786 * And if we were unable to get a new slab from the partial slab lists then
1787 * we need to allocate a new slab. This is the slowest path since it involves
1788 * a call to the page allocator and the setup of a new slab.
1790 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1791 unsigned long addr, struct kmem_cache_cpu *c)
1795 #ifdef CONFIG_CMPXCHG_LOCAL
1796 unsigned long flags;
1798 local_irq_save(flags);
1799 #ifdef CONFIG_PREEMPT
1801 * We may have been preempted and rescheduled on a different
1802 * cpu before disabling interrupts. Need to reload cpu area
1805 c = this_cpu_ptr(s->cpu_slab);
1809 /* We handle __GFP_ZERO in the caller */
1810 gfpflags &= ~__GFP_ZERO;
1816 if (unlikely(!node_match(c, node)))
1819 stat(s, ALLOC_REFILL);
1822 object = c->page->freelist;
1823 if (unlikely(!object))
1825 if (kmem_cache_debug(s))
1828 c->freelist = get_freepointer(s, object);
1829 c->page->inuse = c->page->objects;
1830 c->page->freelist = NULL;
1831 c->node = page_to_nid(c->page);
1833 slab_unlock(c->page);
1834 #ifdef CONFIG_CMPXCHG_LOCAL
1835 c->tid = next_tid(c->tid);
1836 local_irq_restore(flags);
1838 stat(s, ALLOC_SLOWPATH);
1842 deactivate_slab(s, c);
1845 new = get_partial(s, gfpflags, node);
1848 stat(s, ALLOC_FROM_PARTIAL);
1852 gfpflags &= gfp_allowed_mask;
1853 if (gfpflags & __GFP_WAIT)
1856 new = new_slab(s, gfpflags, node);
1858 if (gfpflags & __GFP_WAIT)
1859 local_irq_disable();
1862 c = __this_cpu_ptr(s->cpu_slab);
1863 stat(s, ALLOC_SLAB);
1867 __SetPageSlubFrozen(new);
1871 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1872 slab_out_of_memory(s, gfpflags, node);
1873 #ifdef CONFIG_CMPXCHG_LOCAL
1874 local_irq_restore(flags);
1878 if (!alloc_debug_processing(s, c->page, object, addr))
1882 c->page->freelist = get_freepointer(s, object);
1883 c->node = NUMA_NO_NODE;
1888 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1889 * have the fastpath folded into their functions. So no function call
1890 * overhead for requests that can be satisfied on the fastpath.
1892 * The fastpath works by first checking if the lockless freelist can be used.
1893 * If not then __slab_alloc is called for slow processing.
1895 * Otherwise we can simply pick the next object from the lockless free list.
1897 static __always_inline void *slab_alloc(struct kmem_cache *s,
1898 gfp_t gfpflags, int node, unsigned long addr)
1901 struct kmem_cache_cpu *c;
1902 #ifdef CONFIG_CMPXCHG_LOCAL
1905 unsigned long flags;
1908 if (slab_pre_alloc_hook(s, gfpflags))
1911 #ifndef CONFIG_CMPXCHG_LOCAL
1912 local_irq_save(flags);
1918 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1919 * enabled. We may switch back and forth between cpus while
1920 * reading from one cpu area. That does not matter as long
1921 * as we end up on the original cpu again when doing the cmpxchg.
1923 c = __this_cpu_ptr(s->cpu_slab);
1925 #ifdef CONFIG_CMPXCHG_LOCAL
1927 * The transaction ids are globally unique per cpu and per operation on
1928 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1929 * occurs on the right processor and that there was no operation on the
1930 * linked list in between.
1936 object = c->freelist;
1937 if (unlikely(!object || !node_match(c, node)))
1939 object = __slab_alloc(s, gfpflags, node, addr, c);
1942 #ifdef CONFIG_CMPXCHG_LOCAL
1944 * The cmpxchg will only match if there was no additional
1945 * operation and if we are on the right processor.
1947 * The cmpxchg does the following atomically (without lock semantics!)
1948 * 1. Relocate first pointer to the current per cpu area.
1949 * 2. Verify that tid and freelist have not been changed
1950 * 3. If they were not changed replace tid and freelist
1952 * Since this is without lock semantics the protection is only against
1953 * code executing on this cpu *not* from access by other cpus.
1955 if (unlikely(!irqsafe_cpu_cmpxchg_double(
1956 s->cpu_slab->freelist, s->cpu_slab->tid,
1958 get_freepointer_safe(s, object), next_tid(tid)))) {
1960 note_cmpxchg_failure("slab_alloc", s, tid);
1964 c->freelist = get_freepointer(s, object);
1966 stat(s, ALLOC_FASTPATH);
1969 #ifndef CONFIG_CMPXCHG_LOCAL
1970 local_irq_restore(flags);
1973 if (unlikely(gfpflags & __GFP_ZERO) && object)
1974 memset(object, 0, s->objsize);
1976 slab_post_alloc_hook(s, gfpflags, object);
1981 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1983 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1985 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1989 EXPORT_SYMBOL(kmem_cache_alloc);
1991 #ifdef CONFIG_TRACING
1992 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1994 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1995 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
1998 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2000 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2002 void *ret = kmalloc_order(size, flags, order);
2003 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2006 EXPORT_SYMBOL(kmalloc_order_trace);
2010 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2012 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2014 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2015 s->objsize, s->size, gfpflags, node);
2019 EXPORT_SYMBOL(kmem_cache_alloc_node);
2021 #ifdef CONFIG_TRACING
2022 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2024 int node, size_t size)
2026 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2028 trace_kmalloc_node(_RET_IP_, ret,
2029 size, s->size, gfpflags, node);
2032 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2037 * Slow patch handling. This may still be called frequently since objects
2038 * have a longer lifetime than the cpu slabs in most processing loads.
2040 * So we still attempt to reduce cache line usage. Just take the slab
2041 * lock and free the item. If there is no additional partial page
2042 * handling required then we can return immediately.
2044 static void __slab_free(struct kmem_cache *s, struct page *page,
2045 void *x, unsigned long addr)
2048 void **object = (void *)x;
2049 #ifdef CONFIG_CMPXCHG_LOCAL
2050 unsigned long flags;
2052 local_irq_save(flags);
2055 stat(s, FREE_SLOWPATH);
2057 if (kmem_cache_debug(s))
2061 prior = page->freelist;
2062 set_freepointer(s, object, prior);
2063 page->freelist = object;
2066 if (unlikely(PageSlubFrozen(page))) {
2067 stat(s, FREE_FROZEN);
2071 if (unlikely(!page->inuse))
2075 * Objects left in the slab. If it was not on the partial list before
2078 if (unlikely(!prior)) {
2079 add_partial(get_node(s, page_to_nid(page)), page, 1);
2080 stat(s, FREE_ADD_PARTIAL);
2085 #ifdef CONFIG_CMPXCHG_LOCAL
2086 local_irq_restore(flags);
2093 * Slab still on the partial list.
2095 remove_partial(s, page);
2096 stat(s, FREE_REMOVE_PARTIAL);
2099 #ifdef CONFIG_CMPXCHG_LOCAL
2100 local_irq_restore(flags);
2103 discard_slab(s, page);
2107 if (!free_debug_processing(s, page, x, addr))
2113 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2114 * can perform fastpath freeing without additional function calls.
2116 * The fastpath is only possible if we are freeing to the current cpu slab
2117 * of this processor. This typically the case if we have just allocated
2120 * If fastpath is not possible then fall back to __slab_free where we deal
2121 * with all sorts of special processing.
2123 static __always_inline void slab_free(struct kmem_cache *s,
2124 struct page *page, void *x, unsigned long addr)
2126 void **object = (void *)x;
2127 struct kmem_cache_cpu *c;
2128 #ifdef CONFIG_CMPXCHG_LOCAL
2131 unsigned long flags;
2134 slab_free_hook(s, x);
2136 #ifndef CONFIG_CMPXCHG_LOCAL
2137 local_irq_save(flags);
2144 * Determine the currently cpus per cpu slab.
2145 * The cpu may change afterward. However that does not matter since
2146 * data is retrieved via this pointer. If we are on the same cpu
2147 * during the cmpxchg then the free will succedd.
2149 c = __this_cpu_ptr(s->cpu_slab);
2151 #ifdef CONFIG_CMPXCHG_LOCAL
2156 if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
2157 set_freepointer(s, object, c->freelist);
2159 #ifdef CONFIG_CMPXCHG_LOCAL
2160 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2161 s->cpu_slab->freelist, s->cpu_slab->tid,
2163 object, next_tid(tid)))) {
2165 note_cmpxchg_failure("slab_free", s, tid);
2169 c->freelist = object;
2171 stat(s, FREE_FASTPATH);
2173 __slab_free(s, page, x, addr);
2175 #ifndef CONFIG_CMPXCHG_LOCAL
2176 local_irq_restore(flags);
2180 void kmem_cache_free(struct kmem_cache *s, void *x)
2184 page = virt_to_head_page(x);
2186 slab_free(s, page, x, _RET_IP_);
2188 trace_kmem_cache_free(_RET_IP_, x);
2190 EXPORT_SYMBOL(kmem_cache_free);
2193 * Object placement in a slab is made very easy because we always start at
2194 * offset 0. If we tune the size of the object to the alignment then we can
2195 * get the required alignment by putting one properly sized object after
2198 * Notice that the allocation order determines the sizes of the per cpu
2199 * caches. Each processor has always one slab available for allocations.
2200 * Increasing the allocation order reduces the number of times that slabs
2201 * must be moved on and off the partial lists and is therefore a factor in
2206 * Mininum / Maximum order of slab pages. This influences locking overhead
2207 * and slab fragmentation. A higher order reduces the number of partial slabs
2208 * and increases the number of allocations possible without having to
2209 * take the list_lock.
2211 static int slub_min_order;
2212 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2213 static int slub_min_objects;
2216 * Merge control. If this is set then no merging of slab caches will occur.
2217 * (Could be removed. This was introduced to pacify the merge skeptics.)
2219 static int slub_nomerge;
2222 * Calculate the order of allocation given an slab object size.
2224 * The order of allocation has significant impact on performance and other
2225 * system components. Generally order 0 allocations should be preferred since
2226 * order 0 does not cause fragmentation in the page allocator. Larger objects
2227 * be problematic to put into order 0 slabs because there may be too much
2228 * unused space left. We go to a higher order if more than 1/16th of the slab
2231 * In order to reach satisfactory performance we must ensure that a minimum
2232 * number of objects is in one slab. Otherwise we may generate too much
2233 * activity on the partial lists which requires taking the list_lock. This is
2234 * less a concern for large slabs though which are rarely used.
2236 * slub_max_order specifies the order where we begin to stop considering the
2237 * number of objects in a slab as critical. If we reach slub_max_order then
2238 * we try to keep the page order as low as possible. So we accept more waste
2239 * of space in favor of a small page order.
2241 * Higher order allocations also allow the placement of more objects in a
2242 * slab and thereby reduce object handling overhead. If the user has
2243 * requested a higher mininum order then we start with that one instead of
2244 * the smallest order which will fit the object.
2246 static inline int slab_order(int size, int min_objects,
2247 int max_order, int fract_leftover, int reserved)
2251 int min_order = slub_min_order;
2253 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2254 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2256 for (order = max(min_order,
2257 fls(min_objects * size - 1) - PAGE_SHIFT);
2258 order <= max_order; order++) {
2260 unsigned long slab_size = PAGE_SIZE << order;
2262 if (slab_size < min_objects * size + reserved)
2265 rem = (slab_size - reserved) % size;
2267 if (rem <= slab_size / fract_leftover)
2275 static inline int calculate_order(int size, int reserved)
2283 * Attempt to find best configuration for a slab. This
2284 * works by first attempting to generate a layout with
2285 * the best configuration and backing off gradually.
2287 * First we reduce the acceptable waste in a slab. Then
2288 * we reduce the minimum objects required in a slab.
2290 min_objects = slub_min_objects;
2292 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2293 max_objects = order_objects(slub_max_order, size, reserved);
2294 min_objects = min(min_objects, max_objects);
2296 while (min_objects > 1) {
2298 while (fraction >= 4) {
2299 order = slab_order(size, min_objects,
2300 slub_max_order, fraction, reserved);
2301 if (order <= slub_max_order)
2309 * We were unable to place multiple objects in a slab. Now
2310 * lets see if we can place a single object there.
2312 order = slab_order(size, 1, slub_max_order, 1, reserved);
2313 if (order <= slub_max_order)
2317 * Doh this slab cannot be placed using slub_max_order.
2319 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2320 if (order < MAX_ORDER)
2326 * Figure out what the alignment of the objects will be.
2328 static unsigned long calculate_alignment(unsigned long flags,
2329 unsigned long align, unsigned long size)
2332 * If the user wants hardware cache aligned objects then follow that
2333 * suggestion if the object is sufficiently large.
2335 * The hardware cache alignment cannot override the specified
2336 * alignment though. If that is greater then use it.
2338 if (flags & SLAB_HWCACHE_ALIGN) {
2339 unsigned long ralign = cache_line_size();
2340 while (size <= ralign / 2)
2342 align = max(align, ralign);
2345 if (align < ARCH_SLAB_MINALIGN)
2346 align = ARCH_SLAB_MINALIGN;
2348 return ALIGN(align, sizeof(void *));
2352 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2355 spin_lock_init(&n->list_lock);
2356 INIT_LIST_HEAD(&n->partial);
2357 #ifdef CONFIG_SLUB_DEBUG
2358 atomic_long_set(&n->nr_slabs, 0);
2359 atomic_long_set(&n->total_objects, 0);
2360 INIT_LIST_HEAD(&n->full);
2364 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2366 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2367 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2369 #ifdef CONFIG_CMPXCHG_LOCAL
2371 * Must align to double word boundary for the double cmpxchg instructions
2374 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
2376 /* Regular alignment is sufficient */
2377 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2383 init_kmem_cache_cpus(s);
2388 static struct kmem_cache *kmem_cache_node;
2391 * No kmalloc_node yet so do it by hand. We know that this is the first
2392 * slab on the node for this slabcache. There are no concurrent accesses
2395 * Note that this function only works on the kmalloc_node_cache
2396 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2397 * memory on a fresh node that has no slab structures yet.
2399 static void early_kmem_cache_node_alloc(int node)
2402 struct kmem_cache_node *n;
2403 unsigned long flags;
2405 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2407 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2410 if (page_to_nid(page) != node) {
2411 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2413 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2414 "in order to be able to continue\n");
2419 page->freelist = get_freepointer(kmem_cache_node, n);
2421 kmem_cache_node->node[node] = n;
2422 #ifdef CONFIG_SLUB_DEBUG
2423 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2424 init_tracking(kmem_cache_node, n);
2426 init_kmem_cache_node(n, kmem_cache_node);
2427 inc_slabs_node(kmem_cache_node, node, page->objects);
2430 * lockdep requires consistent irq usage for each lock
2431 * so even though there cannot be a race this early in
2432 * the boot sequence, we still disable irqs.
2434 local_irq_save(flags);
2435 add_partial(n, page, 0);
2436 local_irq_restore(flags);
2439 static void free_kmem_cache_nodes(struct kmem_cache *s)
2443 for_each_node_state(node, N_NORMAL_MEMORY) {
2444 struct kmem_cache_node *n = s->node[node];
2447 kmem_cache_free(kmem_cache_node, n);
2449 s->node[node] = NULL;
2453 static int init_kmem_cache_nodes(struct kmem_cache *s)
2457 for_each_node_state(node, N_NORMAL_MEMORY) {
2458 struct kmem_cache_node *n;
2460 if (slab_state == DOWN) {
2461 early_kmem_cache_node_alloc(node);
2464 n = kmem_cache_alloc_node(kmem_cache_node,
2468 free_kmem_cache_nodes(s);
2473 init_kmem_cache_node(n, s);
2478 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2480 if (min < MIN_PARTIAL)
2482 else if (min > MAX_PARTIAL)
2484 s->min_partial = min;
2488 * calculate_sizes() determines the order and the distribution of data within
2491 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2493 unsigned long flags = s->flags;
2494 unsigned long size = s->objsize;
2495 unsigned long align = s->align;
2499 * Round up object size to the next word boundary. We can only
2500 * place the free pointer at word boundaries and this determines
2501 * the possible location of the free pointer.
2503 size = ALIGN(size, sizeof(void *));
2505 #ifdef CONFIG_SLUB_DEBUG
2507 * Determine if we can poison the object itself. If the user of
2508 * the slab may touch the object after free or before allocation
2509 * then we should never poison the object itself.
2511 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2513 s->flags |= __OBJECT_POISON;
2515 s->flags &= ~__OBJECT_POISON;
2519 * If we are Redzoning then check if there is some space between the
2520 * end of the object and the free pointer. If not then add an
2521 * additional word to have some bytes to store Redzone information.
2523 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2524 size += sizeof(void *);
2528 * With that we have determined the number of bytes in actual use
2529 * by the object. This is the potential offset to the free pointer.
2533 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2536 * Relocate free pointer after the object if it is not
2537 * permitted to overwrite the first word of the object on
2540 * This is the case if we do RCU, have a constructor or
2541 * destructor or are poisoning the objects.
2544 size += sizeof(void *);
2547 #ifdef CONFIG_SLUB_DEBUG
2548 if (flags & SLAB_STORE_USER)
2550 * Need to store information about allocs and frees after
2553 size += 2 * sizeof(struct track);
2555 if (flags & SLAB_RED_ZONE)
2557 * Add some empty padding so that we can catch
2558 * overwrites from earlier objects rather than let
2559 * tracking information or the free pointer be
2560 * corrupted if a user writes before the start
2563 size += sizeof(void *);
2567 * Determine the alignment based on various parameters that the
2568 * user specified and the dynamic determination of cache line size
2571 align = calculate_alignment(flags, align, s->objsize);
2575 * SLUB stores one object immediately after another beginning from
2576 * offset 0. In order to align the objects we have to simply size
2577 * each object to conform to the alignment.
2579 size = ALIGN(size, align);
2581 if (forced_order >= 0)
2582 order = forced_order;
2584 order = calculate_order(size, s->reserved);
2591 s->allocflags |= __GFP_COMP;
2593 if (s->flags & SLAB_CACHE_DMA)
2594 s->allocflags |= SLUB_DMA;
2596 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2597 s->allocflags |= __GFP_RECLAIMABLE;
2600 * Determine the number of objects per slab
2602 s->oo = oo_make(order, size, s->reserved);
2603 s->min = oo_make(get_order(size), size, s->reserved);
2604 if (oo_objects(s->oo) > oo_objects(s->max))
2607 return !!oo_objects(s->oo);
2611 static int kmem_cache_open(struct kmem_cache *s,
2612 const char *name, size_t size,
2613 size_t align, unsigned long flags,
2614 void (*ctor)(void *))
2616 memset(s, 0, kmem_size);
2621 s->flags = kmem_cache_flags(size, flags, name, ctor);
2624 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2625 s->reserved = sizeof(struct rcu_head);
2627 if (!calculate_sizes(s, -1))
2629 if (disable_higher_order_debug) {
2631 * Disable debugging flags that store metadata if the min slab
2634 if (get_order(s->size) > get_order(s->objsize)) {
2635 s->flags &= ~DEBUG_METADATA_FLAGS;
2637 if (!calculate_sizes(s, -1))
2643 * The larger the object size is, the more pages we want on the partial
2644 * list to avoid pounding the page allocator excessively.
2646 set_min_partial(s, ilog2(s->size));
2649 s->remote_node_defrag_ratio = 1000;
2651 if (!init_kmem_cache_nodes(s))
2654 if (alloc_kmem_cache_cpus(s))
2657 free_kmem_cache_nodes(s);
2659 if (flags & SLAB_PANIC)
2660 panic("Cannot create slab %s size=%lu realsize=%u "
2661 "order=%u offset=%u flags=%lx\n",
2662 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2668 * Determine the size of a slab object
2670 unsigned int kmem_cache_size(struct kmem_cache *s)
2674 EXPORT_SYMBOL(kmem_cache_size);
2676 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2679 #ifdef CONFIG_SLUB_DEBUG
2680 void *addr = page_address(page);
2682 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2683 sizeof(long), GFP_ATOMIC);
2686 slab_err(s, page, "%s", text);
2688 for_each_free_object(p, s, page->freelist)
2689 set_bit(slab_index(p, s, addr), map);
2691 for_each_object(p, s, addr, page->objects) {
2693 if (!test_bit(slab_index(p, s, addr), map)) {
2694 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2696 print_tracking(s, p);
2705 * Attempt to free all partial slabs on a node.
2707 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2709 unsigned long flags;
2710 struct page *page, *h;
2712 spin_lock_irqsave(&n->list_lock, flags);
2713 list_for_each_entry_safe(page, h, &n->partial, lru) {
2715 __remove_partial(n, page);
2716 discard_slab(s, page);
2718 list_slab_objects(s, page,
2719 "Objects remaining on kmem_cache_close()");
2722 spin_unlock_irqrestore(&n->list_lock, flags);
2726 * Release all resources used by a slab cache.
2728 static inline int kmem_cache_close(struct kmem_cache *s)
2733 free_percpu(s->cpu_slab);
2734 /* Attempt to free all objects */
2735 for_each_node_state(node, N_NORMAL_MEMORY) {
2736 struct kmem_cache_node *n = get_node(s, node);
2739 if (n->nr_partial || slabs_node(s, node))
2742 free_kmem_cache_nodes(s);
2747 * Close a cache and release the kmem_cache structure
2748 * (must be used for caches created using kmem_cache_create)
2750 void kmem_cache_destroy(struct kmem_cache *s)
2752 down_write(&slub_lock);
2756 if (kmem_cache_close(s)) {
2757 printk(KERN_ERR "SLUB %s: %s called for cache that "
2758 "still has objects.\n", s->name, __func__);
2761 if (s->flags & SLAB_DESTROY_BY_RCU)
2763 sysfs_slab_remove(s);
2765 up_write(&slub_lock);
2767 EXPORT_SYMBOL(kmem_cache_destroy);
2769 /********************************************************************
2771 *******************************************************************/
2773 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2774 EXPORT_SYMBOL(kmalloc_caches);
2776 static struct kmem_cache *kmem_cache;
2778 #ifdef CONFIG_ZONE_DMA
2779 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2782 static int __init setup_slub_min_order(char *str)
2784 get_option(&str, &slub_min_order);
2789 __setup("slub_min_order=", setup_slub_min_order);
2791 static int __init setup_slub_max_order(char *str)
2793 get_option(&str, &slub_max_order);
2794 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2799 __setup("slub_max_order=", setup_slub_max_order);
2801 static int __init setup_slub_min_objects(char *str)
2803 get_option(&str, &slub_min_objects);
2808 __setup("slub_min_objects=", setup_slub_min_objects);
2810 static int __init setup_slub_nomerge(char *str)
2816 __setup("slub_nomerge", setup_slub_nomerge);
2818 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2819 int size, unsigned int flags)
2821 struct kmem_cache *s;
2823 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2826 * This function is called with IRQs disabled during early-boot on
2827 * single CPU so there's no need to take slub_lock here.
2829 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2833 list_add(&s->list, &slab_caches);
2837 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2842 * Conversion table for small slabs sizes / 8 to the index in the
2843 * kmalloc array. This is necessary for slabs < 192 since we have non power
2844 * of two cache sizes there. The size of larger slabs can be determined using
2847 static s8 size_index[24] = {
2874 static inline int size_index_elem(size_t bytes)
2876 return (bytes - 1) / 8;
2879 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2885 return ZERO_SIZE_PTR;
2887 index = size_index[size_index_elem(size)];
2889 index = fls(size - 1);
2891 #ifdef CONFIG_ZONE_DMA
2892 if (unlikely((flags & SLUB_DMA)))
2893 return kmalloc_dma_caches[index];
2896 return kmalloc_caches[index];
2899 void *__kmalloc(size_t size, gfp_t flags)
2901 struct kmem_cache *s;
2904 if (unlikely(size > SLUB_MAX_SIZE))
2905 return kmalloc_large(size, flags);
2907 s = get_slab(size, flags);
2909 if (unlikely(ZERO_OR_NULL_PTR(s)))
2912 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2914 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2918 EXPORT_SYMBOL(__kmalloc);
2921 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2926 flags |= __GFP_COMP | __GFP_NOTRACK;
2927 page = alloc_pages_node(node, flags, get_order(size));
2929 ptr = page_address(page);
2931 kmemleak_alloc(ptr, size, 1, flags);
2935 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2937 struct kmem_cache *s;
2940 if (unlikely(size > SLUB_MAX_SIZE)) {
2941 ret = kmalloc_large_node(size, flags, node);
2943 trace_kmalloc_node(_RET_IP_, ret,
2944 size, PAGE_SIZE << get_order(size),
2950 s = get_slab(size, flags);
2952 if (unlikely(ZERO_OR_NULL_PTR(s)))
2955 ret = slab_alloc(s, flags, node, _RET_IP_);
2957 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2961 EXPORT_SYMBOL(__kmalloc_node);
2964 size_t ksize(const void *object)
2968 if (unlikely(object == ZERO_SIZE_PTR))
2971 page = virt_to_head_page(object);
2973 if (unlikely(!PageSlab(page))) {
2974 WARN_ON(!PageCompound(page));
2975 return PAGE_SIZE << compound_order(page);
2978 return slab_ksize(page->slab);
2980 EXPORT_SYMBOL(ksize);
2982 void kfree(const void *x)
2985 void *object = (void *)x;
2987 trace_kfree(_RET_IP_, x);
2989 if (unlikely(ZERO_OR_NULL_PTR(x)))
2992 page = virt_to_head_page(x);
2993 if (unlikely(!PageSlab(page))) {
2994 BUG_ON(!PageCompound(page));
2999 slab_free(page->slab, page, object, _RET_IP_);
3001 EXPORT_SYMBOL(kfree);
3004 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3005 * the remaining slabs by the number of items in use. The slabs with the
3006 * most items in use come first. New allocations will then fill those up
3007 * and thus they can be removed from the partial lists.
3009 * The slabs with the least items are placed last. This results in them
3010 * being allocated from last increasing the chance that the last objects
3011 * are freed in them.
3013 int kmem_cache_shrink(struct kmem_cache *s)
3017 struct kmem_cache_node *n;
3020 int objects = oo_objects(s->max);
3021 struct list_head *slabs_by_inuse =
3022 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3023 unsigned long flags;
3025 if (!slabs_by_inuse)
3029 for_each_node_state(node, N_NORMAL_MEMORY) {
3030 n = get_node(s, node);
3035 for (i = 0; i < objects; i++)
3036 INIT_LIST_HEAD(slabs_by_inuse + i);
3038 spin_lock_irqsave(&n->list_lock, flags);
3041 * Build lists indexed by the items in use in each slab.
3043 * Note that concurrent frees may occur while we hold the
3044 * list_lock. page->inuse here is the upper limit.
3046 list_for_each_entry_safe(page, t, &n->partial, lru) {
3047 if (!page->inuse && slab_trylock(page)) {
3049 * Must hold slab lock here because slab_free
3050 * may have freed the last object and be
3051 * waiting to release the slab.
3053 __remove_partial(n, page);
3055 discard_slab(s, page);
3057 list_move(&page->lru,
3058 slabs_by_inuse + page->inuse);
3063 * Rebuild the partial list with the slabs filled up most
3064 * first and the least used slabs at the end.
3066 for (i = objects - 1; i >= 0; i--)
3067 list_splice(slabs_by_inuse + i, n->partial.prev);
3069 spin_unlock_irqrestore(&n->list_lock, flags);
3072 kfree(slabs_by_inuse);
3075 EXPORT_SYMBOL(kmem_cache_shrink);
3077 #if defined(CONFIG_MEMORY_HOTPLUG)
3078 static int slab_mem_going_offline_callback(void *arg)
3080 struct kmem_cache *s;
3082 down_read(&slub_lock);
3083 list_for_each_entry(s, &slab_caches, list)
3084 kmem_cache_shrink(s);
3085 up_read(&slub_lock);
3090 static void slab_mem_offline_callback(void *arg)
3092 struct kmem_cache_node *n;
3093 struct kmem_cache *s;
3094 struct memory_notify *marg = arg;
3097 offline_node = marg->status_change_nid;
3100 * If the node still has available memory. we need kmem_cache_node
3103 if (offline_node < 0)
3106 down_read(&slub_lock);
3107 list_for_each_entry(s, &slab_caches, list) {
3108 n = get_node(s, offline_node);
3111 * if n->nr_slabs > 0, slabs still exist on the node
3112 * that is going down. We were unable to free them,
3113 * and offline_pages() function shouldn't call this
3114 * callback. So, we must fail.
3116 BUG_ON(slabs_node(s, offline_node));
3118 s->node[offline_node] = NULL;
3119 kmem_cache_free(kmem_cache_node, n);
3122 up_read(&slub_lock);
3125 static int slab_mem_going_online_callback(void *arg)
3127 struct kmem_cache_node *n;
3128 struct kmem_cache *s;
3129 struct memory_notify *marg = arg;
3130 int nid = marg->status_change_nid;
3134 * If the node's memory is already available, then kmem_cache_node is
3135 * already created. Nothing to do.
3141 * We are bringing a node online. No memory is available yet. We must
3142 * allocate a kmem_cache_node structure in order to bring the node
3145 down_read(&slub_lock);
3146 list_for_each_entry(s, &slab_caches, list) {
3148 * XXX: kmem_cache_alloc_node will fallback to other nodes
3149 * since memory is not yet available from the node that
3152 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3157 init_kmem_cache_node(n, s);
3161 up_read(&slub_lock);
3165 static int slab_memory_callback(struct notifier_block *self,
3166 unsigned long action, void *arg)
3171 case MEM_GOING_ONLINE:
3172 ret = slab_mem_going_online_callback(arg);
3174 case MEM_GOING_OFFLINE:
3175 ret = slab_mem_going_offline_callback(arg);
3178 case MEM_CANCEL_ONLINE:
3179 slab_mem_offline_callback(arg);
3182 case MEM_CANCEL_OFFLINE:
3186 ret = notifier_from_errno(ret);
3192 #endif /* CONFIG_MEMORY_HOTPLUG */
3194 /********************************************************************
3195 * Basic setup of slabs
3196 *******************************************************************/
3199 * Used for early kmem_cache structures that were allocated using
3200 * the page allocator
3203 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3207 list_add(&s->list, &slab_caches);
3210 for_each_node_state(node, N_NORMAL_MEMORY) {
3211 struct kmem_cache_node *n = get_node(s, node);
3215 list_for_each_entry(p, &n->partial, lru)
3218 #ifdef CONFIG_SLAB_DEBUG
3219 list_for_each_entry(p, &n->full, lru)
3226 void __init kmem_cache_init(void)
3230 struct kmem_cache *temp_kmem_cache;
3232 struct kmem_cache *temp_kmem_cache_node;
3233 unsigned long kmalloc_size;
3235 kmem_size = offsetof(struct kmem_cache, node) +
3236 nr_node_ids * sizeof(struct kmem_cache_node *);
3238 /* Allocate two kmem_caches from the page allocator */
3239 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3240 order = get_order(2 * kmalloc_size);
3241 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3244 * Must first have the slab cache available for the allocations of the
3245 * struct kmem_cache_node's. There is special bootstrap code in
3246 * kmem_cache_open for slab_state == DOWN.
3248 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3250 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3251 sizeof(struct kmem_cache_node),
3252 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3254 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3256 /* Able to allocate the per node structures */
3257 slab_state = PARTIAL;
3259 temp_kmem_cache = kmem_cache;
3260 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3261 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3262 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3263 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3266 * Allocate kmem_cache_node properly from the kmem_cache slab.
3267 * kmem_cache_node is separately allocated so no need to
3268 * update any list pointers.
3270 temp_kmem_cache_node = kmem_cache_node;
3272 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3273 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3275 kmem_cache_bootstrap_fixup(kmem_cache_node);
3278 kmem_cache_bootstrap_fixup(kmem_cache);
3280 /* Free temporary boot structure */
3281 free_pages((unsigned long)temp_kmem_cache, order);
3283 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3286 * Patch up the size_index table if we have strange large alignment
3287 * requirements for the kmalloc array. This is only the case for
3288 * MIPS it seems. The standard arches will not generate any code here.
3290 * Largest permitted alignment is 256 bytes due to the way we
3291 * handle the index determination for the smaller caches.
3293 * Make sure that nothing crazy happens if someone starts tinkering
3294 * around with ARCH_KMALLOC_MINALIGN
3296 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3297 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3299 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3300 int elem = size_index_elem(i);
3301 if (elem >= ARRAY_SIZE(size_index))
3303 size_index[elem] = KMALLOC_SHIFT_LOW;
3306 if (KMALLOC_MIN_SIZE == 64) {
3308 * The 96 byte size cache is not used if the alignment
3311 for (i = 64 + 8; i <= 96; i += 8)
3312 size_index[size_index_elem(i)] = 7;
3313 } else if (KMALLOC_MIN_SIZE == 128) {
3315 * The 192 byte sized cache is not used if the alignment
3316 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3319 for (i = 128 + 8; i <= 192; i += 8)
3320 size_index[size_index_elem(i)] = 8;
3323 /* Caches that are not of the two-to-the-power-of size */
3324 if (KMALLOC_MIN_SIZE <= 32) {
3325 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3329 if (KMALLOC_MIN_SIZE <= 64) {
3330 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3334 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3335 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3341 /* Provide the correct kmalloc names now that the caches are up */
3342 if (KMALLOC_MIN_SIZE <= 32) {
3343 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3344 BUG_ON(!kmalloc_caches[1]->name);
3347 if (KMALLOC_MIN_SIZE <= 64) {
3348 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3349 BUG_ON(!kmalloc_caches[2]->name);
3352 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3353 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3356 kmalloc_caches[i]->name = s;
3360 register_cpu_notifier(&slab_notifier);
3363 #ifdef CONFIG_ZONE_DMA
3364 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3365 struct kmem_cache *s = kmalloc_caches[i];
3368 char *name = kasprintf(GFP_NOWAIT,
3369 "dma-kmalloc-%d", s->objsize);
3372 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3373 s->objsize, SLAB_CACHE_DMA);
3378 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3379 " CPUs=%d, Nodes=%d\n",
3380 caches, cache_line_size(),
3381 slub_min_order, slub_max_order, slub_min_objects,
3382 nr_cpu_ids, nr_node_ids);
3385 void __init kmem_cache_init_late(void)
3390 * Find a mergeable slab cache
3392 static int slab_unmergeable(struct kmem_cache *s)
3394 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3401 * We may have set a slab to be unmergeable during bootstrap.
3403 if (s->refcount < 0)
3409 static struct kmem_cache *find_mergeable(size_t size,
3410 size_t align, unsigned long flags, const char *name,
3411 void (*ctor)(void *))
3413 struct kmem_cache *s;
3415 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3421 size = ALIGN(size, sizeof(void *));
3422 align = calculate_alignment(flags, align, size);
3423 size = ALIGN(size, align);
3424 flags = kmem_cache_flags(size, flags, name, NULL);
3426 list_for_each_entry(s, &slab_caches, list) {
3427 if (slab_unmergeable(s))
3433 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3436 * Check if alignment is compatible.
3437 * Courtesy of Adrian Drzewiecki
3439 if ((s->size & ~(align - 1)) != s->size)
3442 if (s->size - size >= sizeof(void *))
3450 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3451 size_t align, unsigned long flags, void (*ctor)(void *))
3453 struct kmem_cache *s;
3459 down_write(&slub_lock);
3460 s = find_mergeable(size, align, flags, name, ctor);
3464 * Adjust the object sizes so that we clear
3465 * the complete object on kzalloc.
3467 s->objsize = max(s->objsize, (int)size);
3468 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3470 if (sysfs_slab_alias(s, name)) {
3474 up_write(&slub_lock);
3478 n = kstrdup(name, GFP_KERNEL);
3482 s = kmalloc(kmem_size, GFP_KERNEL);
3484 if (kmem_cache_open(s, n,
3485 size, align, flags, ctor)) {
3486 list_add(&s->list, &slab_caches);
3487 if (sysfs_slab_add(s)) {
3493 up_write(&slub_lock);
3500 up_write(&slub_lock);
3502 if (flags & SLAB_PANIC)
3503 panic("Cannot create slabcache %s\n", name);
3508 EXPORT_SYMBOL(kmem_cache_create);
3512 * Use the cpu notifier to insure that the cpu slabs are flushed when
3515 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3516 unsigned long action, void *hcpu)
3518 long cpu = (long)hcpu;
3519 struct kmem_cache *s;
3520 unsigned long flags;
3523 case CPU_UP_CANCELED:
3524 case CPU_UP_CANCELED_FROZEN:
3526 case CPU_DEAD_FROZEN:
3527 down_read(&slub_lock);
3528 list_for_each_entry(s, &slab_caches, list) {
3529 local_irq_save(flags);
3530 __flush_cpu_slab(s, cpu);
3531 local_irq_restore(flags);
3533 up_read(&slub_lock);
3541 static struct notifier_block __cpuinitdata slab_notifier = {
3542 .notifier_call = slab_cpuup_callback
3547 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3549 struct kmem_cache *s;
3552 if (unlikely(size > SLUB_MAX_SIZE))
3553 return kmalloc_large(size, gfpflags);
3555 s = get_slab(size, gfpflags);
3557 if (unlikely(ZERO_OR_NULL_PTR(s)))
3560 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3562 /* Honor the call site pointer we received. */
3563 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3569 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3570 int node, unsigned long caller)
3572 struct kmem_cache *s;
3575 if (unlikely(size > SLUB_MAX_SIZE)) {
3576 ret = kmalloc_large_node(size, gfpflags, node);
3578 trace_kmalloc_node(caller, ret,
3579 size, PAGE_SIZE << get_order(size),
3585 s = get_slab(size, gfpflags);
3587 if (unlikely(ZERO_OR_NULL_PTR(s)))
3590 ret = slab_alloc(s, gfpflags, node, caller);
3592 /* Honor the call site pointer we received. */
3593 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3600 static int count_inuse(struct page *page)
3605 static int count_total(struct page *page)
3607 return page->objects;
3611 #ifdef CONFIG_SLUB_DEBUG
3612 static int validate_slab(struct kmem_cache *s, struct page *page,
3616 void *addr = page_address(page);
3618 if (!check_slab(s, page) ||
3619 !on_freelist(s, page, NULL))
3622 /* Now we know that a valid freelist exists */
3623 bitmap_zero(map, page->objects);
3625 for_each_free_object(p, s, page->freelist) {
3626 set_bit(slab_index(p, s, addr), map);
3627 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3631 for_each_object(p, s, addr, page->objects)
3632 if (!test_bit(slab_index(p, s, addr), map))
3633 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3638 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3641 if (slab_trylock(page)) {
3642 validate_slab(s, page, map);
3645 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3649 static int validate_slab_node(struct kmem_cache *s,
3650 struct kmem_cache_node *n, unsigned long *map)
3652 unsigned long count = 0;
3654 unsigned long flags;
3656 spin_lock_irqsave(&n->list_lock, flags);
3658 list_for_each_entry(page, &n->partial, lru) {
3659 validate_slab_slab(s, page, map);
3662 if (count != n->nr_partial)
3663 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3664 "counter=%ld\n", s->name, count, n->nr_partial);
3666 if (!(s->flags & SLAB_STORE_USER))
3669 list_for_each_entry(page, &n->full, lru) {
3670 validate_slab_slab(s, page, map);
3673 if (count != atomic_long_read(&n->nr_slabs))
3674 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3675 "counter=%ld\n", s->name, count,
3676 atomic_long_read(&n->nr_slabs));
3679 spin_unlock_irqrestore(&n->list_lock, flags);
3683 static long validate_slab_cache(struct kmem_cache *s)
3686 unsigned long count = 0;
3687 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3688 sizeof(unsigned long), GFP_KERNEL);
3694 for_each_node_state(node, N_NORMAL_MEMORY) {
3695 struct kmem_cache_node *n = get_node(s, node);
3697 count += validate_slab_node(s, n, map);
3703 * Generate lists of code addresses where slabcache objects are allocated
3708 unsigned long count;
3715 DECLARE_BITMAP(cpus, NR_CPUS);
3721 unsigned long count;
3722 struct location *loc;
3725 static void free_loc_track(struct loc_track *t)
3728 free_pages((unsigned long)t->loc,
3729 get_order(sizeof(struct location) * t->max));
3732 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3737 order = get_order(sizeof(struct location) * max);
3739 l = (void *)__get_free_pages(flags, order);
3744 memcpy(l, t->loc, sizeof(struct location) * t->count);
3752 static int add_location(struct loc_track *t, struct kmem_cache *s,
3753 const struct track *track)
3755 long start, end, pos;
3757 unsigned long caddr;
3758 unsigned long age = jiffies - track->when;
3764 pos = start + (end - start + 1) / 2;
3767 * There is nothing at "end". If we end up there
3768 * we need to add something to before end.
3773 caddr = t->loc[pos].addr;
3774 if (track->addr == caddr) {
3780 if (age < l->min_time)
3782 if (age > l->max_time)
3785 if (track->pid < l->min_pid)
3786 l->min_pid = track->pid;
3787 if (track->pid > l->max_pid)
3788 l->max_pid = track->pid;
3790 cpumask_set_cpu(track->cpu,
3791 to_cpumask(l->cpus));
3793 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3797 if (track->addr < caddr)
3804 * Not found. Insert new tracking element.
3806 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3812 (t->count - pos) * sizeof(struct location));
3815 l->addr = track->addr;
3819 l->min_pid = track->pid;
3820 l->max_pid = track->pid;
3821 cpumask_clear(to_cpumask(l->cpus));
3822 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3823 nodes_clear(l->nodes);
3824 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3828 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3829 struct page *page, enum track_item alloc,
3832 void *addr = page_address(page);
3835 bitmap_zero(map, page->objects);
3836 for_each_free_object(p, s, page->freelist)
3837 set_bit(slab_index(p, s, addr), map);
3839 for_each_object(p, s, addr, page->objects)
3840 if (!test_bit(slab_index(p, s, addr), map))
3841 add_location(t, s, get_track(s, p, alloc));
3844 static int list_locations(struct kmem_cache *s, char *buf,
3845 enum track_item alloc)
3849 struct loc_track t = { 0, 0, NULL };
3851 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3852 sizeof(unsigned long), GFP_KERNEL);
3854 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3857 return sprintf(buf, "Out of memory\n");
3859 /* Push back cpu slabs */
3862 for_each_node_state(node, N_NORMAL_MEMORY) {
3863 struct kmem_cache_node *n = get_node(s, node);
3864 unsigned long flags;
3867 if (!atomic_long_read(&n->nr_slabs))
3870 spin_lock_irqsave(&n->list_lock, flags);
3871 list_for_each_entry(page, &n->partial, lru)
3872 process_slab(&t, s, page, alloc, map);
3873 list_for_each_entry(page, &n->full, lru)
3874 process_slab(&t, s, page, alloc, map);
3875 spin_unlock_irqrestore(&n->list_lock, flags);
3878 for (i = 0; i < t.count; i++) {
3879 struct location *l = &t.loc[i];
3881 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3883 len += sprintf(buf + len, "%7ld ", l->count);
3886 len += sprintf(buf + len, "%pS", (void *)l->addr);
3888 len += sprintf(buf + len, "<not-available>");
3890 if (l->sum_time != l->min_time) {
3891 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3893 (long)div_u64(l->sum_time, l->count),
3896 len += sprintf(buf + len, " age=%ld",
3899 if (l->min_pid != l->max_pid)
3900 len += sprintf(buf + len, " pid=%ld-%ld",
3901 l->min_pid, l->max_pid);
3903 len += sprintf(buf + len, " pid=%ld",
3906 if (num_online_cpus() > 1 &&
3907 !cpumask_empty(to_cpumask(l->cpus)) &&
3908 len < PAGE_SIZE - 60) {
3909 len += sprintf(buf + len, " cpus=");
3910 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3911 to_cpumask(l->cpus));
3914 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3915 len < PAGE_SIZE - 60) {
3916 len += sprintf(buf + len, " nodes=");
3917 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3921 len += sprintf(buf + len, "\n");
3927 len += sprintf(buf, "No data\n");
3932 #ifdef SLUB_RESILIENCY_TEST
3933 static void resiliency_test(void)
3937 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3939 printk(KERN_ERR "SLUB resiliency testing\n");
3940 printk(KERN_ERR "-----------------------\n");
3941 printk(KERN_ERR "A. Corruption after allocation\n");
3943 p = kzalloc(16, GFP_KERNEL);
3945 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3946 " 0x12->0x%p\n\n", p + 16);
3948 validate_slab_cache(kmalloc_caches[4]);
3950 /* Hmmm... The next two are dangerous */
3951 p = kzalloc(32, GFP_KERNEL);
3952 p[32 + sizeof(void *)] = 0x34;
3953 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3954 " 0x34 -> -0x%p\n", p);
3956 "If allocated object is overwritten then not detectable\n\n");
3958 validate_slab_cache(kmalloc_caches[5]);
3959 p = kzalloc(64, GFP_KERNEL);
3960 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3962 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3965 "If allocated object is overwritten then not detectable\n\n");
3966 validate_slab_cache(kmalloc_caches[6]);
3968 printk(KERN_ERR "\nB. Corruption after free\n");
3969 p = kzalloc(128, GFP_KERNEL);
3972 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3973 validate_slab_cache(kmalloc_caches[7]);
3975 p = kzalloc(256, GFP_KERNEL);
3978 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3980 validate_slab_cache(kmalloc_caches[8]);
3982 p = kzalloc(512, GFP_KERNEL);
3985 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3986 validate_slab_cache(kmalloc_caches[9]);
3990 static void resiliency_test(void) {};
3995 enum slab_stat_type {
3996 SL_ALL, /* All slabs */
3997 SL_PARTIAL, /* Only partially allocated slabs */
3998 SL_CPU, /* Only slabs used for cpu caches */
3999 SL_OBJECTS, /* Determine allocated objects not slabs */
4000 SL_TOTAL /* Determine object capacity not slabs */
4003 #define SO_ALL (1 << SL_ALL)
4004 #define SO_PARTIAL (1 << SL_PARTIAL)
4005 #define SO_CPU (1 << SL_CPU)
4006 #define SO_OBJECTS (1 << SL_OBJECTS)
4007 #define SO_TOTAL (1 << SL_TOTAL)
4009 static ssize_t show_slab_objects(struct kmem_cache *s,
4010 char *buf, unsigned long flags)
4012 unsigned long total = 0;
4015 unsigned long *nodes;
4016 unsigned long *per_cpu;
4018 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4021 per_cpu = nodes + nr_node_ids;
4023 if (flags & SO_CPU) {
4026 for_each_possible_cpu(cpu) {
4027 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4029 if (!c || c->node < 0)
4033 if (flags & SO_TOTAL)
4034 x = c->page->objects;
4035 else if (flags & SO_OBJECTS)
4041 nodes[c->node] += x;
4047 lock_memory_hotplug();
4048 #ifdef CONFIG_SLUB_DEBUG
4049 if (flags & SO_ALL) {
4050 for_each_node_state(node, N_NORMAL_MEMORY) {
4051 struct kmem_cache_node *n = get_node(s, node);
4053 if (flags & SO_TOTAL)
4054 x = atomic_long_read(&n->total_objects);
4055 else if (flags & SO_OBJECTS)
4056 x = atomic_long_read(&n->total_objects) -
4057 count_partial(n, count_free);
4060 x = atomic_long_read(&n->nr_slabs);
4067 if (flags & SO_PARTIAL) {
4068 for_each_node_state(node, N_NORMAL_MEMORY) {
4069 struct kmem_cache_node *n = get_node(s, node);
4071 if (flags & SO_TOTAL)
4072 x = count_partial(n, count_total);
4073 else if (flags & SO_OBJECTS)
4074 x = count_partial(n, count_inuse);
4081 x = sprintf(buf, "%lu", total);
4083 for_each_node_state(node, N_NORMAL_MEMORY)
4085 x += sprintf(buf + x, " N%d=%lu",
4088 unlock_memory_hotplug();
4090 return x + sprintf(buf + x, "\n");
4093 #ifdef CONFIG_SLUB_DEBUG
4094 static int any_slab_objects(struct kmem_cache *s)
4098 for_each_online_node(node) {
4099 struct kmem_cache_node *n = get_node(s, node);
4104 if (atomic_long_read(&n->total_objects))
4111 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4112 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4114 struct slab_attribute {
4115 struct attribute attr;
4116 ssize_t (*show)(struct kmem_cache *s, char *buf);
4117 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4120 #define SLAB_ATTR_RO(_name) \
4121 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4123 #define SLAB_ATTR(_name) \
4124 static struct slab_attribute _name##_attr = \
4125 __ATTR(_name, 0644, _name##_show, _name##_store)
4127 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4129 return sprintf(buf, "%d\n", s->size);
4131 SLAB_ATTR_RO(slab_size);
4133 static ssize_t align_show(struct kmem_cache *s, char *buf)
4135 return sprintf(buf, "%d\n", s->align);
4137 SLAB_ATTR_RO(align);
4139 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4141 return sprintf(buf, "%d\n", s->objsize);
4143 SLAB_ATTR_RO(object_size);
4145 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4147 return sprintf(buf, "%d\n", oo_objects(s->oo));
4149 SLAB_ATTR_RO(objs_per_slab);
4151 static ssize_t order_store(struct kmem_cache *s,
4152 const char *buf, size_t length)
4154 unsigned long order;
4157 err = strict_strtoul(buf, 10, &order);
4161 if (order > slub_max_order || order < slub_min_order)
4164 calculate_sizes(s, order);
4168 static ssize_t order_show(struct kmem_cache *s, char *buf)
4170 return sprintf(buf, "%d\n", oo_order(s->oo));
4174 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4176 return sprintf(buf, "%lu\n", s->min_partial);
4179 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4185 err = strict_strtoul(buf, 10, &min);
4189 set_min_partial(s, min);
4192 SLAB_ATTR(min_partial);
4194 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4198 return sprintf(buf, "%pS\n", s->ctor);
4202 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4204 return sprintf(buf, "%d\n", s->refcount - 1);
4206 SLAB_ATTR_RO(aliases);
4208 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4210 return show_slab_objects(s, buf, SO_PARTIAL);
4212 SLAB_ATTR_RO(partial);
4214 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4216 return show_slab_objects(s, buf, SO_CPU);
4218 SLAB_ATTR_RO(cpu_slabs);
4220 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4222 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4224 SLAB_ATTR_RO(objects);
4226 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4228 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4230 SLAB_ATTR_RO(objects_partial);
4232 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4234 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4237 static ssize_t reclaim_account_store(struct kmem_cache *s,
4238 const char *buf, size_t length)
4240 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4242 s->flags |= SLAB_RECLAIM_ACCOUNT;
4245 SLAB_ATTR(reclaim_account);
4247 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4249 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4251 SLAB_ATTR_RO(hwcache_align);
4253 #ifdef CONFIG_ZONE_DMA
4254 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4256 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4258 SLAB_ATTR_RO(cache_dma);
4261 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4263 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4265 SLAB_ATTR_RO(destroy_by_rcu);
4267 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4269 return sprintf(buf, "%d\n", s->reserved);
4271 SLAB_ATTR_RO(reserved);
4273 #ifdef CONFIG_SLUB_DEBUG
4274 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4276 return show_slab_objects(s, buf, SO_ALL);
4278 SLAB_ATTR_RO(slabs);
4280 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4282 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4284 SLAB_ATTR_RO(total_objects);
4286 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4288 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4291 static ssize_t sanity_checks_store(struct kmem_cache *s,
4292 const char *buf, size_t length)
4294 s->flags &= ~SLAB_DEBUG_FREE;
4296 s->flags |= SLAB_DEBUG_FREE;
4299 SLAB_ATTR(sanity_checks);
4301 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4303 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4306 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4309 s->flags &= ~SLAB_TRACE;
4311 s->flags |= SLAB_TRACE;
4316 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4318 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4321 static ssize_t red_zone_store(struct kmem_cache *s,
4322 const char *buf, size_t length)
4324 if (any_slab_objects(s))
4327 s->flags &= ~SLAB_RED_ZONE;
4329 s->flags |= SLAB_RED_ZONE;
4330 calculate_sizes(s, -1);
4333 SLAB_ATTR(red_zone);
4335 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4337 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4340 static ssize_t poison_store(struct kmem_cache *s,
4341 const char *buf, size_t length)
4343 if (any_slab_objects(s))
4346 s->flags &= ~SLAB_POISON;
4348 s->flags |= SLAB_POISON;
4349 calculate_sizes(s, -1);
4354 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4356 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4359 static ssize_t store_user_store(struct kmem_cache *s,
4360 const char *buf, size_t length)
4362 if (any_slab_objects(s))
4365 s->flags &= ~SLAB_STORE_USER;
4367 s->flags |= SLAB_STORE_USER;
4368 calculate_sizes(s, -1);
4371 SLAB_ATTR(store_user);
4373 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4378 static ssize_t validate_store(struct kmem_cache *s,
4379 const char *buf, size_t length)
4383 if (buf[0] == '1') {
4384 ret = validate_slab_cache(s);
4390 SLAB_ATTR(validate);
4392 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4394 if (!(s->flags & SLAB_STORE_USER))
4396 return list_locations(s, buf, TRACK_ALLOC);
4398 SLAB_ATTR_RO(alloc_calls);
4400 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4402 if (!(s->flags & SLAB_STORE_USER))
4404 return list_locations(s, buf, TRACK_FREE);
4406 SLAB_ATTR_RO(free_calls);
4407 #endif /* CONFIG_SLUB_DEBUG */
4409 #ifdef CONFIG_FAILSLAB
4410 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4412 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4415 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4418 s->flags &= ~SLAB_FAILSLAB;
4420 s->flags |= SLAB_FAILSLAB;
4423 SLAB_ATTR(failslab);
4426 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4431 static ssize_t shrink_store(struct kmem_cache *s,
4432 const char *buf, size_t length)
4434 if (buf[0] == '1') {
4435 int rc = kmem_cache_shrink(s);
4446 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4448 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4451 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4452 const char *buf, size_t length)
4454 unsigned long ratio;
4457 err = strict_strtoul(buf, 10, &ratio);
4462 s->remote_node_defrag_ratio = ratio * 10;
4466 SLAB_ATTR(remote_node_defrag_ratio);
4469 #ifdef CONFIG_SLUB_STATS
4470 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4472 unsigned long sum = 0;
4475 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4480 for_each_online_cpu(cpu) {
4481 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4487 len = sprintf(buf, "%lu", sum);
4490 for_each_online_cpu(cpu) {
4491 if (data[cpu] && len < PAGE_SIZE - 20)
4492 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4496 return len + sprintf(buf + len, "\n");
4499 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4503 for_each_online_cpu(cpu)
4504 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4507 #define STAT_ATTR(si, text) \
4508 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4510 return show_stat(s, buf, si); \
4512 static ssize_t text##_store(struct kmem_cache *s, \
4513 const char *buf, size_t length) \
4515 if (buf[0] != '0') \
4517 clear_stat(s, si); \
4522 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4523 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4524 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4525 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4526 STAT_ATTR(FREE_FROZEN, free_frozen);
4527 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4528 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4529 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4530 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4531 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4532 STAT_ATTR(FREE_SLAB, free_slab);
4533 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4534 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4535 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4536 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4537 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4538 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4539 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4542 static struct attribute *slab_attrs[] = {
4543 &slab_size_attr.attr,
4544 &object_size_attr.attr,
4545 &objs_per_slab_attr.attr,
4547 &min_partial_attr.attr,
4549 &objects_partial_attr.attr,
4551 &cpu_slabs_attr.attr,
4555 &hwcache_align_attr.attr,
4556 &reclaim_account_attr.attr,
4557 &destroy_by_rcu_attr.attr,
4559 &reserved_attr.attr,
4560 #ifdef CONFIG_SLUB_DEBUG
4561 &total_objects_attr.attr,
4563 &sanity_checks_attr.attr,
4565 &red_zone_attr.attr,
4567 &store_user_attr.attr,
4568 &validate_attr.attr,
4569 &alloc_calls_attr.attr,
4570 &free_calls_attr.attr,
4572 #ifdef CONFIG_ZONE_DMA
4573 &cache_dma_attr.attr,
4576 &remote_node_defrag_ratio_attr.attr,
4578 #ifdef CONFIG_SLUB_STATS
4579 &alloc_fastpath_attr.attr,
4580 &alloc_slowpath_attr.attr,
4581 &free_fastpath_attr.attr,
4582 &free_slowpath_attr.attr,
4583 &free_frozen_attr.attr,
4584 &free_add_partial_attr.attr,
4585 &free_remove_partial_attr.attr,
4586 &alloc_from_partial_attr.attr,
4587 &alloc_slab_attr.attr,
4588 &alloc_refill_attr.attr,
4589 &free_slab_attr.attr,
4590 &cpuslab_flush_attr.attr,
4591 &deactivate_full_attr.attr,
4592 &deactivate_empty_attr.attr,
4593 &deactivate_to_head_attr.attr,
4594 &deactivate_to_tail_attr.attr,
4595 &deactivate_remote_frees_attr.attr,
4596 &order_fallback_attr.attr,
4598 #ifdef CONFIG_FAILSLAB
4599 &failslab_attr.attr,
4605 static struct attribute_group slab_attr_group = {
4606 .attrs = slab_attrs,
4609 static ssize_t slab_attr_show(struct kobject *kobj,
4610 struct attribute *attr,
4613 struct slab_attribute *attribute;
4614 struct kmem_cache *s;
4617 attribute = to_slab_attr(attr);
4620 if (!attribute->show)
4623 err = attribute->show(s, buf);
4628 static ssize_t slab_attr_store(struct kobject *kobj,
4629 struct attribute *attr,
4630 const char *buf, size_t len)
4632 struct slab_attribute *attribute;
4633 struct kmem_cache *s;
4636 attribute = to_slab_attr(attr);
4639 if (!attribute->store)
4642 err = attribute->store(s, buf, len);
4647 static void kmem_cache_release(struct kobject *kobj)
4649 struct kmem_cache *s = to_slab(kobj);
4655 static const struct sysfs_ops slab_sysfs_ops = {
4656 .show = slab_attr_show,
4657 .store = slab_attr_store,
4660 static struct kobj_type slab_ktype = {
4661 .sysfs_ops = &slab_sysfs_ops,
4662 .release = kmem_cache_release
4665 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4667 struct kobj_type *ktype = get_ktype(kobj);
4669 if (ktype == &slab_ktype)
4674 static const struct kset_uevent_ops slab_uevent_ops = {
4675 .filter = uevent_filter,
4678 static struct kset *slab_kset;
4680 #define ID_STR_LENGTH 64
4682 /* Create a unique string id for a slab cache:
4684 * Format :[flags-]size
4686 static char *create_unique_id(struct kmem_cache *s)
4688 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4695 * First flags affecting slabcache operations. We will only
4696 * get here for aliasable slabs so we do not need to support
4697 * too many flags. The flags here must cover all flags that
4698 * are matched during merging to guarantee that the id is
4701 if (s->flags & SLAB_CACHE_DMA)
4703 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4705 if (s->flags & SLAB_DEBUG_FREE)
4707 if (!(s->flags & SLAB_NOTRACK))
4711 p += sprintf(p, "%07d", s->size);
4712 BUG_ON(p > name + ID_STR_LENGTH - 1);
4716 static int sysfs_slab_add(struct kmem_cache *s)
4722 if (slab_state < SYSFS)
4723 /* Defer until later */
4726 unmergeable = slab_unmergeable(s);
4729 * Slabcache can never be merged so we can use the name proper.
4730 * This is typically the case for debug situations. In that
4731 * case we can catch duplicate names easily.
4733 sysfs_remove_link(&slab_kset->kobj, s->name);
4737 * Create a unique name for the slab as a target
4740 name = create_unique_id(s);
4743 s->kobj.kset = slab_kset;
4744 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4746 kobject_put(&s->kobj);
4750 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4752 kobject_del(&s->kobj);
4753 kobject_put(&s->kobj);
4756 kobject_uevent(&s->kobj, KOBJ_ADD);
4758 /* Setup first alias */
4759 sysfs_slab_alias(s, s->name);
4765 static void sysfs_slab_remove(struct kmem_cache *s)
4767 if (slab_state < SYSFS)
4769 * Sysfs has not been setup yet so no need to remove the
4774 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4775 kobject_del(&s->kobj);
4776 kobject_put(&s->kobj);
4780 * Need to buffer aliases during bootup until sysfs becomes
4781 * available lest we lose that information.
4783 struct saved_alias {
4784 struct kmem_cache *s;
4786 struct saved_alias *next;
4789 static struct saved_alias *alias_list;
4791 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4793 struct saved_alias *al;
4795 if (slab_state == SYSFS) {
4797 * If we have a leftover link then remove it.
4799 sysfs_remove_link(&slab_kset->kobj, name);
4800 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4803 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4809 al->next = alias_list;
4814 static int __init slab_sysfs_init(void)
4816 struct kmem_cache *s;
4819 down_write(&slub_lock);
4821 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4823 up_write(&slub_lock);
4824 printk(KERN_ERR "Cannot register slab subsystem.\n");
4830 list_for_each_entry(s, &slab_caches, list) {
4831 err = sysfs_slab_add(s);
4833 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4834 " to sysfs\n", s->name);
4837 while (alias_list) {
4838 struct saved_alias *al = alias_list;
4840 alias_list = alias_list->next;
4841 err = sysfs_slab_alias(al->s, al->name);
4843 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4844 " %s to sysfs\n", s->name);
4848 up_write(&slub_lock);
4853 __initcall(slab_sysfs_init);
4854 #endif /* CONFIG_SYSFS */
4857 * The /proc/slabinfo ABI
4859 #ifdef CONFIG_SLABINFO
4860 static void print_slabinfo_header(struct seq_file *m)
4862 seq_puts(m, "slabinfo - version: 2.1\n");
4863 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4864 "<objperslab> <pagesperslab>");
4865 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4866 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4870 static void *s_start(struct seq_file *m, loff_t *pos)
4874 down_read(&slub_lock);
4876 print_slabinfo_header(m);
4878 return seq_list_start(&slab_caches, *pos);
4881 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4883 return seq_list_next(p, &slab_caches, pos);
4886 static void s_stop(struct seq_file *m, void *p)
4888 up_read(&slub_lock);
4891 static int s_show(struct seq_file *m, void *p)
4893 unsigned long nr_partials = 0;
4894 unsigned long nr_slabs = 0;
4895 unsigned long nr_inuse = 0;
4896 unsigned long nr_objs = 0;
4897 unsigned long nr_free = 0;
4898 struct kmem_cache *s;
4901 s = list_entry(p, struct kmem_cache, list);
4903 for_each_online_node(node) {
4904 struct kmem_cache_node *n = get_node(s, node);
4909 nr_partials += n->nr_partial;
4910 nr_slabs += atomic_long_read(&n->nr_slabs);
4911 nr_objs += atomic_long_read(&n->total_objects);
4912 nr_free += count_partial(n, count_free);
4915 nr_inuse = nr_objs - nr_free;
4917 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4918 nr_objs, s->size, oo_objects(s->oo),
4919 (1 << oo_order(s->oo)));
4920 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4921 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4927 static const struct seq_operations slabinfo_op = {
4934 static int slabinfo_open(struct inode *inode, struct file *file)
4936 return seq_open(file, &slabinfo_op);
4939 static const struct file_operations proc_slabinfo_operations = {
4940 .open = slabinfo_open,
4942 .llseek = seq_lseek,
4943 .release = seq_release,
4946 static int __init slab_proc_init(void)
4948 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4951 module_init(slab_proc_init);
4952 #endif /* CONFIG_SLABINFO */