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 <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
23 #include <linux/memory.h>
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
72 * freed then the slab will show up again on the partial lists.
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
80 * Overloading of page flags that are otherwise used for LRU management.
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
94 * freelist that allows lockless access to
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
100 * the fast path and disables lockless freelists.
103 #define FROZEN (1 << PG_active)
105 #ifdef CONFIG_SLUB_DEBUG
106 #define SLABDEBUG (1 << PG_error)
111 static inline int SlabFrozen(struct page *page)
113 return page->flags & FROZEN;
116 static inline void SetSlabFrozen(struct page *page)
118 page->flags |= FROZEN;
121 static inline void ClearSlabFrozen(struct page *page)
123 page->flags &= ~FROZEN;
126 static inline int SlabDebug(struct page *page)
128 return page->flags & SLABDEBUG;
131 static inline void SetSlabDebug(struct page *page)
133 page->flags |= SLABDEBUG;
136 static inline void ClearSlabDebug(struct page *page)
138 page->flags &= ~SLABDEBUG;
142 * Issues still to be resolved:
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
146 * - Variable sizing of the per node arrays
149 /* Enable to test recovery from slab corruption on boot */
150 #undef SLUB_RESILIENCY_TEST
155 * Small page size. Make sure that we do not fragment memory
157 #define DEFAULT_MAX_ORDER 1
158 #define DEFAULT_MIN_OBJECTS 4
163 * Large page machines are customarily able to handle larger
166 #define DEFAULT_MAX_ORDER 2
167 #define DEFAULT_MIN_OBJECTS 8
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
175 #define MIN_PARTIAL 5
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
182 #define MAX_PARTIAL 10
184 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
188 * Set of flags that will prevent slab merging
190 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
193 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
196 #ifndef ARCH_KMALLOC_MINALIGN
197 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
200 #ifndef ARCH_SLAB_MINALIGN
201 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 /* Internal SLUB flags */
205 #define __OBJECT_POISON 0x80000000 /* Poison object */
206 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
207 #define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208 #define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
210 /* Not all arches define cache_line_size */
211 #ifndef cache_line_size
212 #define cache_line_size() L1_CACHE_BYTES
215 static int kmem_size = sizeof(struct kmem_cache);
218 static struct notifier_block slab_notifier;
222 DOWN, /* No slab functionality available */
223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
224 UP, /* Everything works but does not show up in sysfs */
228 /* A list of all slab caches on the system */
229 static DECLARE_RWSEM(slub_lock);
230 static LIST_HEAD(slab_caches);
233 * Tracking user of a slab.
236 void *addr; /* Called from address */
237 int cpu; /* Was running on cpu */
238 int pid; /* Pid context */
239 unsigned long when; /* When did the operation occur */
242 enum track_item { TRACK_ALLOC, TRACK_FREE };
244 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
245 static int sysfs_slab_add(struct kmem_cache *);
246 static int sysfs_slab_alias(struct kmem_cache *, const char *);
247 static void sysfs_slab_remove(struct kmem_cache *);
250 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
253 static inline void sysfs_slab_remove(struct kmem_cache *s)
260 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
262 #ifdef CONFIG_SLUB_STATS
267 /********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
271 int slab_is_available(void)
273 return slab_state >= UP;
276 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
279 return s->node[node];
281 return &s->local_node;
285 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
288 return s->cpu_slab[cpu];
294 /* Verify that a pointer has an address that is valid within a slab page */
295 static inline int check_valid_pointer(struct kmem_cache *s,
296 struct page *page, const void *object)
303 base = page_address(page);
304 if (object < base || object >= base + page->objects * s->size ||
305 (object - base) % s->size) {
313 * Slow version of get and set free pointer.
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
319 static inline void *get_freepointer(struct kmem_cache *s, void *object)
321 return *(void **)(object + s->offset);
324 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
326 *(void **)(object + s->offset) = fp;
329 /* Loop over all objects in a slab */
330 #define for_each_object(__p, __s, __addr, __objects) \
331 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
335 #define for_each_free_object(__p, __s, __free) \
336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
338 /* Determine object index from a given position */
339 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
341 return (p - addr) / s->size;
344 static inline struct kmem_cache_order_objects oo_make(int order,
347 struct kmem_cache_order_objects x = {
348 (order << 16) + (PAGE_SIZE << order) / size
354 static inline int oo_order(struct kmem_cache_order_objects x)
359 static inline int oo_objects(struct kmem_cache_order_objects x)
361 return x.x & ((1 << 16) - 1);
364 #ifdef CONFIG_SLUB_DEBUG
368 #ifdef CONFIG_SLUB_DEBUG_ON
369 static int slub_debug = DEBUG_DEFAULT_FLAGS;
371 static int slub_debug;
374 static char *slub_debug_slabs;
379 static void print_section(char *text, u8 *addr, unsigned int length)
387 for (i = 0; i < length; i++) {
389 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
392 printk(KERN_CONT " %02x", addr[i]);
394 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
396 printk(KERN_CONT " %s\n", ascii);
403 printk(KERN_CONT " ");
407 printk(KERN_CONT " %s\n", ascii);
411 static struct track *get_track(struct kmem_cache *s, void *object,
412 enum track_item alloc)
417 p = object + s->offset + sizeof(void *);
419 p = object + s->inuse;
424 static void set_track(struct kmem_cache *s, void *object,
425 enum track_item alloc, void *addr)
430 p = object + s->offset + sizeof(void *);
432 p = object + s->inuse;
437 p->cpu = smp_processor_id();
438 p->pid = current ? current->pid : -1;
441 memset(p, 0, sizeof(struct track));
444 static void init_tracking(struct kmem_cache *s, void *object)
446 if (!(s->flags & SLAB_STORE_USER))
449 set_track(s, object, TRACK_FREE, NULL);
450 set_track(s, object, TRACK_ALLOC, NULL);
453 static void print_track(const char *s, struct track *t)
458 printk(KERN_ERR "INFO: %s in ", s);
459 __print_symbol("%s", (unsigned long)t->addr);
460 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
463 static void print_tracking(struct kmem_cache *s, void *object)
465 if (!(s->flags & SLAB_STORE_USER))
468 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
469 print_track("Freed", get_track(s, object, TRACK_FREE));
472 static void print_page_info(struct page *page)
474 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
475 page, page->objects, page->inuse, page->freelist, page->flags);
479 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
485 vsnprintf(buf, sizeof(buf), fmt, args);
487 printk(KERN_ERR "========================================"
488 "=====================================\n");
489 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
490 printk(KERN_ERR "----------------------------------------"
491 "-------------------------------------\n\n");
494 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
500 vsnprintf(buf, sizeof(buf), fmt, args);
502 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
505 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
507 unsigned int off; /* Offset of last byte */
508 u8 *addr = page_address(page);
510 print_tracking(s, p);
512 print_page_info(page);
514 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
515 p, p - addr, get_freepointer(s, p));
518 print_section("Bytes b4", p - 16, 16);
520 print_section("Object", p, min(s->objsize, 128));
522 if (s->flags & SLAB_RED_ZONE)
523 print_section("Redzone", p + s->objsize,
524 s->inuse - s->objsize);
527 off = s->offset + sizeof(void *);
531 if (s->flags & SLAB_STORE_USER)
532 off += 2 * sizeof(struct track);
535 /* Beginning of the filler is the free pointer */
536 print_section("Padding", p + off, s->size - off);
541 static void object_err(struct kmem_cache *s, struct page *page,
542 u8 *object, char *reason)
544 slab_bug(s, "%s", reason);
545 print_trailer(s, page, object);
548 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
554 vsnprintf(buf, sizeof(buf), fmt, args);
556 slab_bug(s, "%s", buf);
557 print_page_info(page);
561 static void init_object(struct kmem_cache *s, void *object, int active)
565 if (s->flags & __OBJECT_POISON) {
566 memset(p, POISON_FREE, s->objsize - 1);
567 p[s->objsize - 1] = POISON_END;
570 if (s->flags & SLAB_RED_ZONE)
571 memset(p + s->objsize,
572 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
573 s->inuse - s->objsize);
576 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
579 if (*start != (u8)value)
587 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
588 void *from, void *to)
590 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
591 memset(from, data, to - from);
594 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
595 u8 *object, char *what,
596 u8 *start, unsigned int value, unsigned int bytes)
601 fault = check_bytes(start, value, bytes);
606 while (end > fault && end[-1] == value)
609 slab_bug(s, "%s overwritten", what);
610 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
611 fault, end - 1, fault[0], value);
612 print_trailer(s, page, object);
614 restore_bytes(s, what, value, fault, end);
622 * Bytes of the object to be managed.
623 * If the freepointer may overlay the object then the free
624 * pointer is the first word of the object.
626 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
629 * object + s->objsize
630 * Padding to reach word boundary. This is also used for Redzoning.
631 * Padding is extended by another word if Redzoning is enabled and
634 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
635 * 0xcc (RED_ACTIVE) for objects in use.
638 * Meta data starts here.
640 * A. Free pointer (if we cannot overwrite object on free)
641 * B. Tracking data for SLAB_STORE_USER
642 * C. Padding to reach required alignment boundary or at mininum
643 * one word if debugging is on to be able to detect writes
644 * before the word boundary.
646 * Padding is done using 0x5a (POISON_INUSE)
649 * Nothing is used beyond s->size.
651 * If slabcaches are merged then the objsize and inuse boundaries are mostly
652 * ignored. And therefore no slab options that rely on these boundaries
653 * may be used with merged slabcaches.
656 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
658 unsigned long off = s->inuse; /* The end of info */
661 /* Freepointer is placed after the object. */
662 off += sizeof(void *);
664 if (s->flags & SLAB_STORE_USER)
665 /* We also have user information there */
666 off += 2 * sizeof(struct track);
671 return check_bytes_and_report(s, page, p, "Object padding",
672 p + off, POISON_INUSE, s->size - off);
675 /* Check the pad bytes at the end of a slab page */
676 static int slab_pad_check(struct kmem_cache *s, struct page *page)
684 if (!(s->flags & SLAB_POISON))
687 start = page_address(page);
688 length = (PAGE_SIZE << compound_order(page));
689 end = start + length;
690 remainder = length % s->size;
694 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
697 while (end > fault && end[-1] == POISON_INUSE)
700 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
701 print_section("Padding", end - remainder, remainder);
703 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
707 static int check_object(struct kmem_cache *s, struct page *page,
708 void *object, int active)
711 u8 *endobject = object + s->objsize;
713 if (s->flags & SLAB_RED_ZONE) {
715 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
717 if (!check_bytes_and_report(s, page, object, "Redzone",
718 endobject, red, s->inuse - s->objsize))
721 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
722 check_bytes_and_report(s, page, p, "Alignment padding",
723 endobject, POISON_INUSE, s->inuse - s->objsize);
727 if (s->flags & SLAB_POISON) {
728 if (!active && (s->flags & __OBJECT_POISON) &&
729 (!check_bytes_and_report(s, page, p, "Poison", p,
730 POISON_FREE, s->objsize - 1) ||
731 !check_bytes_and_report(s, page, p, "Poison",
732 p + s->objsize - 1, POISON_END, 1)))
735 * check_pad_bytes cleans up on its own.
737 check_pad_bytes(s, page, p);
740 if (!s->offset && active)
742 * Object and freepointer overlap. Cannot check
743 * freepointer while object is allocated.
747 /* Check free pointer validity */
748 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
749 object_err(s, page, p, "Freepointer corrupt");
751 * No choice but to zap it and thus loose the remainder
752 * of the free objects in this slab. May cause
753 * another error because the object count is now wrong.
755 set_freepointer(s, p, NULL);
761 static int check_slab(struct kmem_cache *s, struct page *page)
765 VM_BUG_ON(!irqs_disabled());
767 if (!PageSlab(page)) {
768 slab_err(s, page, "Not a valid slab page");
772 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
773 if (page->objects > maxobj) {
774 slab_err(s, page, "objects %u > max %u",
775 s->name, page->objects, maxobj);
778 if (page->inuse > page->objects) {
779 slab_err(s, page, "inuse %u > max %u",
780 s->name, page->inuse, page->objects);
783 /* Slab_pad_check fixes things up after itself */
784 slab_pad_check(s, page);
789 * Determine if a certain object on a page is on the freelist. Must hold the
790 * slab lock to guarantee that the chains are in a consistent state.
792 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
795 void *fp = page->freelist;
797 unsigned long max_objects;
799 while (fp && nr <= page->objects) {
802 if (!check_valid_pointer(s, page, fp)) {
804 object_err(s, page, object,
805 "Freechain corrupt");
806 set_freepointer(s, object, NULL);
809 slab_err(s, page, "Freepointer corrupt");
810 page->freelist = NULL;
811 page->inuse = page->objects;
812 slab_fix(s, "Freelist cleared");
818 fp = get_freepointer(s, object);
822 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
823 if (max_objects > 65535)
826 if (page->objects != max_objects) {
827 slab_err(s, page, "Wrong number of objects. Found %d but "
828 "should be %d", page->objects, max_objects);
829 page->objects = max_objects;
830 slab_fix(s, "Number of objects adjusted.");
832 if (page->inuse != page->objects - nr) {
833 slab_err(s, page, "Wrong object count. Counter is %d but "
834 "counted were %d", page->inuse, page->objects - nr);
835 page->inuse = page->objects - nr;
836 slab_fix(s, "Object count adjusted.");
838 return search == NULL;
841 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
843 if (s->flags & SLAB_TRACE) {
844 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
846 alloc ? "alloc" : "free",
851 print_section("Object", (void *)object, s->objsize);
858 * Tracking of fully allocated slabs for debugging purposes.
860 static void add_full(struct kmem_cache_node *n, struct page *page)
862 spin_lock(&n->list_lock);
863 list_add(&page->lru, &n->full);
864 spin_unlock(&n->list_lock);
867 static void remove_full(struct kmem_cache *s, struct page *page)
869 struct kmem_cache_node *n;
871 if (!(s->flags & SLAB_STORE_USER))
874 n = get_node(s, page_to_nid(page));
876 spin_lock(&n->list_lock);
877 list_del(&page->lru);
878 spin_unlock(&n->list_lock);
881 /* Tracking of the number of slabs for debugging purposes */
882 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
884 struct kmem_cache_node *n = get_node(s, node);
886 return atomic_long_read(&n->nr_slabs);
889 static inline void inc_slabs_node(struct kmem_cache *s, int node)
891 struct kmem_cache_node *n = get_node(s, node);
894 * May be called early in order to allocate a slab for the
895 * kmem_cache_node structure. Solve the chicken-egg
896 * dilemma by deferring the increment of the count during
897 * bootstrap (see early_kmem_cache_node_alloc).
899 if (!NUMA_BUILD || n)
900 atomic_long_inc(&n->nr_slabs);
902 static inline void dec_slabs_node(struct kmem_cache *s, int node)
904 struct kmem_cache_node *n = get_node(s, node);
906 atomic_long_dec(&n->nr_slabs);
909 /* Object debug checks for alloc/free paths */
910 static void setup_object_debug(struct kmem_cache *s, struct page *page,
913 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
916 init_object(s, object, 0);
917 init_tracking(s, object);
920 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
921 void *object, void *addr)
923 if (!check_slab(s, page))
926 if (!on_freelist(s, page, object)) {
927 object_err(s, page, object, "Object already allocated");
931 if (!check_valid_pointer(s, page, object)) {
932 object_err(s, page, object, "Freelist Pointer check fails");
936 if (!check_object(s, page, object, 0))
939 /* Success perform special debug activities for allocs */
940 if (s->flags & SLAB_STORE_USER)
941 set_track(s, object, TRACK_ALLOC, addr);
942 trace(s, page, object, 1);
943 init_object(s, object, 1);
947 if (PageSlab(page)) {
949 * If this is a slab page then lets do the best we can
950 * to avoid issues in the future. Marking all objects
951 * as used avoids touching the remaining objects.
953 slab_fix(s, "Marking all objects used");
954 page->inuse = page->objects;
955 page->freelist = NULL;
960 static int free_debug_processing(struct kmem_cache *s, struct page *page,
961 void *object, void *addr)
963 if (!check_slab(s, page))
966 if (!check_valid_pointer(s, page, object)) {
967 slab_err(s, page, "Invalid object pointer 0x%p", object);
971 if (on_freelist(s, page, object)) {
972 object_err(s, page, object, "Object already free");
976 if (!check_object(s, page, object, 1))
979 if (unlikely(s != page->slab)) {
980 if (!PageSlab(page)) {
981 slab_err(s, page, "Attempt to free object(0x%p) "
982 "outside of slab", object);
983 } else if (!page->slab) {
985 "SLUB <none>: no slab for object 0x%p.\n",
989 object_err(s, page, object,
990 "page slab pointer corrupt.");
994 /* Special debug activities for freeing objects */
995 if (!SlabFrozen(page) && !page->freelist)
996 remove_full(s, page);
997 if (s->flags & SLAB_STORE_USER)
998 set_track(s, object, TRACK_FREE, addr);
999 trace(s, page, object, 0);
1000 init_object(s, object, 0);
1004 slab_fix(s, "Object at 0x%p not freed", object);
1008 static int __init setup_slub_debug(char *str)
1010 slub_debug = DEBUG_DEFAULT_FLAGS;
1011 if (*str++ != '=' || !*str)
1013 * No options specified. Switch on full debugging.
1019 * No options but restriction on slabs. This means full
1020 * debugging for slabs matching a pattern.
1027 * Switch off all debugging measures.
1032 * Determine which debug features should be switched on
1034 for (; *str && *str != ','; str++) {
1035 switch (tolower(*str)) {
1037 slub_debug |= SLAB_DEBUG_FREE;
1040 slub_debug |= SLAB_RED_ZONE;
1043 slub_debug |= SLAB_POISON;
1046 slub_debug |= SLAB_STORE_USER;
1049 slub_debug |= SLAB_TRACE;
1052 printk(KERN_ERR "slub_debug option '%c' "
1053 "unknown. skipped\n", *str);
1059 slub_debug_slabs = str + 1;
1064 __setup("slub_debug", setup_slub_debug);
1066 static unsigned long kmem_cache_flags(unsigned long objsize,
1067 unsigned long flags, const char *name,
1068 void (*ctor)(struct kmem_cache *, void *))
1071 * Enable debugging if selected on the kernel commandline.
1073 if (slub_debug && (!slub_debug_slabs ||
1074 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1075 flags |= slub_debug;
1080 static inline void setup_object_debug(struct kmem_cache *s,
1081 struct page *page, void *object) {}
1083 static inline int alloc_debug_processing(struct kmem_cache *s,
1084 struct page *page, void *object, void *addr) { return 0; }
1086 static inline int free_debug_processing(struct kmem_cache *s,
1087 struct page *page, void *object, void *addr) { return 0; }
1089 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1091 static inline int check_object(struct kmem_cache *s, struct page *page,
1092 void *object, int active) { return 1; }
1093 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1094 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1095 unsigned long flags, const char *name,
1096 void (*ctor)(struct kmem_cache *, void *))
1100 #define slub_debug 0
1102 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1104 static inline void inc_slabs_node(struct kmem_cache *s, int node) {}
1105 static inline void dec_slabs_node(struct kmem_cache *s, int node) {}
1108 * Slab allocation and freeing
1110 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1113 struct kmem_cache_order_objects oo = s->oo;
1114 int order = oo_order(oo);
1115 int pages = 1 << order;
1117 flags |= s->allocflags;
1120 page = alloc_pages(flags, order);
1122 page = alloc_pages_node(node, flags, order);
1127 page->objects = oo_objects(oo);
1128 mod_zone_page_state(page_zone(page),
1129 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1130 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1136 static void setup_object(struct kmem_cache *s, struct page *page,
1139 setup_object_debug(s, page, object);
1140 if (unlikely(s->ctor))
1144 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1151 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1153 page = allocate_slab(s,
1154 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1158 inc_slabs_node(s, page_to_nid(page));
1160 page->flags |= 1 << PG_slab;
1161 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1162 SLAB_STORE_USER | SLAB_TRACE))
1165 start = page_address(page);
1167 if (unlikely(s->flags & SLAB_POISON))
1168 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1171 for_each_object(p, s, start, page->objects) {
1172 setup_object(s, page, last);
1173 set_freepointer(s, last, p);
1176 setup_object(s, page, last);
1177 set_freepointer(s, last, NULL);
1179 page->freelist = start;
1185 static void __free_slab(struct kmem_cache *s, struct page *page)
1187 int order = compound_order(page);
1188 int pages = 1 << order;
1190 if (unlikely(SlabDebug(page))) {
1193 slab_pad_check(s, page);
1194 for_each_object(p, s, page_address(page),
1196 check_object(s, page, p, 0);
1197 ClearSlabDebug(page);
1200 mod_zone_page_state(page_zone(page),
1201 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1202 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1205 __ClearPageSlab(page);
1206 reset_page_mapcount(page);
1207 __free_pages(page, order);
1210 static void rcu_free_slab(struct rcu_head *h)
1214 page = container_of((struct list_head *)h, struct page, lru);
1215 __free_slab(page->slab, page);
1218 static void free_slab(struct kmem_cache *s, struct page *page)
1220 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1222 * RCU free overloads the RCU head over the LRU
1224 struct rcu_head *head = (void *)&page->lru;
1226 call_rcu(head, rcu_free_slab);
1228 __free_slab(s, page);
1231 static void discard_slab(struct kmem_cache *s, struct page *page)
1233 dec_slabs_node(s, page_to_nid(page));
1238 * Per slab locking using the pagelock
1240 static __always_inline void slab_lock(struct page *page)
1242 bit_spin_lock(PG_locked, &page->flags);
1245 static __always_inline void slab_unlock(struct page *page)
1247 __bit_spin_unlock(PG_locked, &page->flags);
1250 static __always_inline int slab_trylock(struct page *page)
1254 rc = bit_spin_trylock(PG_locked, &page->flags);
1259 * Management of partially allocated slabs
1261 static void add_partial(struct kmem_cache_node *n,
1262 struct page *page, int tail)
1264 spin_lock(&n->list_lock);
1267 list_add_tail(&page->lru, &n->partial);
1269 list_add(&page->lru, &n->partial);
1270 spin_unlock(&n->list_lock);
1273 static void remove_partial(struct kmem_cache *s,
1276 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1278 spin_lock(&n->list_lock);
1279 list_del(&page->lru);
1281 spin_unlock(&n->list_lock);
1285 * Lock slab and remove from the partial list.
1287 * Must hold list_lock.
1289 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1291 if (slab_trylock(page)) {
1292 list_del(&page->lru);
1294 SetSlabFrozen(page);
1301 * Try to allocate a partial slab from a specific node.
1303 static struct page *get_partial_node(struct kmem_cache_node *n)
1308 * Racy check. If we mistakenly see no partial slabs then we
1309 * just allocate an empty slab. If we mistakenly try to get a
1310 * partial slab and there is none available then get_partials()
1313 if (!n || !n->nr_partial)
1316 spin_lock(&n->list_lock);
1317 list_for_each_entry(page, &n->partial, lru)
1318 if (lock_and_freeze_slab(n, page))
1322 spin_unlock(&n->list_lock);
1327 * Get a page from somewhere. Search in increasing NUMA distances.
1329 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1332 struct zonelist *zonelist;
1337 * The defrag ratio allows a configuration of the tradeoffs between
1338 * inter node defragmentation and node local allocations. A lower
1339 * defrag_ratio increases the tendency to do local allocations
1340 * instead of attempting to obtain partial slabs from other nodes.
1342 * If the defrag_ratio is set to 0 then kmalloc() always
1343 * returns node local objects. If the ratio is higher then kmalloc()
1344 * may return off node objects because partial slabs are obtained
1345 * from other nodes and filled up.
1347 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1348 * defrag_ratio = 1000) then every (well almost) allocation will
1349 * first attempt to defrag slab caches on other nodes. This means
1350 * scanning over all nodes to look for partial slabs which may be
1351 * expensive if we do it every time we are trying to find a slab
1352 * with available objects.
1354 if (!s->remote_node_defrag_ratio ||
1355 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1358 zonelist = &NODE_DATA(
1359 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
1360 for (z = zonelist->zones; *z; z++) {
1361 struct kmem_cache_node *n;
1363 n = get_node(s, zone_to_nid(*z));
1365 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1366 n->nr_partial > MIN_PARTIAL) {
1367 page = get_partial_node(n);
1377 * Get a partial page, lock it and return it.
1379 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1382 int searchnode = (node == -1) ? numa_node_id() : node;
1384 page = get_partial_node(get_node(s, searchnode));
1385 if (page || (flags & __GFP_THISNODE))
1388 return get_any_partial(s, flags);
1392 * Move a page back to the lists.
1394 * Must be called with the slab lock held.
1396 * On exit the slab lock will have been dropped.
1398 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1400 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1401 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1403 ClearSlabFrozen(page);
1406 if (page->freelist) {
1407 add_partial(n, page, tail);
1408 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1410 stat(c, DEACTIVATE_FULL);
1411 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1416 stat(c, DEACTIVATE_EMPTY);
1417 if (n->nr_partial < MIN_PARTIAL) {
1419 * Adding an empty slab to the partial slabs in order
1420 * to avoid page allocator overhead. This slab needs
1421 * to come after the other slabs with objects in
1422 * so that the others get filled first. That way the
1423 * size of the partial list stays small.
1425 * kmem_cache_shrink can reclaim any empty slabs from the
1428 add_partial(n, page, 1);
1432 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1433 discard_slab(s, page);
1439 * Remove the cpu slab
1441 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1443 struct page *page = c->page;
1447 stat(c, DEACTIVATE_REMOTE_FREES);
1449 * Merge cpu freelist into slab freelist. Typically we get here
1450 * because both freelists are empty. So this is unlikely
1453 while (unlikely(c->freelist)) {
1456 tail = 0; /* Hot objects. Put the slab first */
1458 /* Retrieve object from cpu_freelist */
1459 object = c->freelist;
1460 c->freelist = c->freelist[c->offset];
1462 /* And put onto the regular freelist */
1463 object[c->offset] = page->freelist;
1464 page->freelist = object;
1468 unfreeze_slab(s, page, tail);
1471 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1473 stat(c, CPUSLAB_FLUSH);
1475 deactivate_slab(s, c);
1481 * Called from IPI handler with interrupts disabled.
1483 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1485 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1487 if (likely(c && c->page))
1491 static void flush_cpu_slab(void *d)
1493 struct kmem_cache *s = d;
1495 __flush_cpu_slab(s, smp_processor_id());
1498 static void flush_all(struct kmem_cache *s)
1501 on_each_cpu(flush_cpu_slab, s, 1, 1);
1503 unsigned long flags;
1505 local_irq_save(flags);
1507 local_irq_restore(flags);
1512 * Check if the objects in a per cpu structure fit numa
1513 * locality expectations.
1515 static inline int node_match(struct kmem_cache_cpu *c, int node)
1518 if (node != -1 && c->node != node)
1525 * Slow path. The lockless freelist is empty or we need to perform
1528 * Interrupts are disabled.
1530 * Processing is still very fast if new objects have been freed to the
1531 * regular freelist. In that case we simply take over the regular freelist
1532 * as the lockless freelist and zap the regular freelist.
1534 * If that is not working then we fall back to the partial lists. We take the
1535 * first element of the freelist as the object to allocate now and move the
1536 * rest of the freelist to the lockless freelist.
1538 * And if we were unable to get a new slab from the partial slab lists then
1539 * we need to allocate a new slab. This is the slowest path since it involves
1540 * a call to the page allocator and the setup of a new slab.
1542 static void *__slab_alloc(struct kmem_cache *s,
1543 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1548 /* We handle __GFP_ZERO in the caller */
1549 gfpflags &= ~__GFP_ZERO;
1555 if (unlikely(!node_match(c, node)))
1558 stat(c, ALLOC_REFILL);
1561 object = c->page->freelist;
1562 if (unlikely(!object))
1564 if (unlikely(SlabDebug(c->page)))
1567 c->freelist = object[c->offset];
1568 c->page->inuse = c->page->objects;
1569 c->page->freelist = NULL;
1570 c->node = page_to_nid(c->page);
1572 slab_unlock(c->page);
1573 stat(c, ALLOC_SLOWPATH);
1577 deactivate_slab(s, c);
1580 new = get_partial(s, gfpflags, node);
1583 stat(c, ALLOC_FROM_PARTIAL);
1587 if (gfpflags & __GFP_WAIT)
1590 new = new_slab(s, gfpflags, node);
1592 if (gfpflags & __GFP_WAIT)
1593 local_irq_disable();
1596 c = get_cpu_slab(s, smp_processor_id());
1597 stat(c, ALLOC_SLAB);
1607 * No memory available.
1609 * If the slab uses higher order allocs but the object is
1610 * smaller than a page size then we can fallback in emergencies
1611 * to the page allocator via kmalloc_large. The page allocator may
1612 * have failed to obtain a higher order page and we can try to
1613 * allocate a single page if the object fits into a single page.
1614 * That is only possible if certain conditions are met that are being
1615 * checked when a slab is created.
1617 if (!(gfpflags & __GFP_NORETRY) &&
1618 (s->flags & __PAGE_ALLOC_FALLBACK)) {
1619 if (gfpflags & __GFP_WAIT)
1621 object = kmalloc_large(s->objsize, gfpflags);
1622 if (gfpflags & __GFP_WAIT)
1623 local_irq_disable();
1628 if (!alloc_debug_processing(s, c->page, object, addr))
1632 c->page->freelist = object[c->offset];
1638 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1639 * have the fastpath folded into their functions. So no function call
1640 * overhead for requests that can be satisfied on the fastpath.
1642 * The fastpath works by first checking if the lockless freelist can be used.
1643 * If not then __slab_alloc is called for slow processing.
1645 * Otherwise we can simply pick the next object from the lockless free list.
1647 static __always_inline void *slab_alloc(struct kmem_cache *s,
1648 gfp_t gfpflags, int node, void *addr)
1651 struct kmem_cache_cpu *c;
1652 unsigned long flags;
1654 local_irq_save(flags);
1655 c = get_cpu_slab(s, smp_processor_id());
1656 if (unlikely(!c->freelist || !node_match(c, node)))
1658 object = __slab_alloc(s, gfpflags, node, addr, c);
1661 object = c->freelist;
1662 c->freelist = object[c->offset];
1663 stat(c, ALLOC_FASTPATH);
1665 local_irq_restore(flags);
1667 if (unlikely((gfpflags & __GFP_ZERO) && object))
1668 memset(object, 0, c->objsize);
1673 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1675 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1677 EXPORT_SYMBOL(kmem_cache_alloc);
1680 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1682 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1684 EXPORT_SYMBOL(kmem_cache_alloc_node);
1688 * Slow patch handling. This may still be called frequently since objects
1689 * have a longer lifetime than the cpu slabs in most processing loads.
1691 * So we still attempt to reduce cache line usage. Just take the slab
1692 * lock and free the item. If there is no additional partial page
1693 * handling required then we can return immediately.
1695 static void __slab_free(struct kmem_cache *s, struct page *page,
1696 void *x, void *addr, unsigned int offset)
1699 void **object = (void *)x;
1700 struct kmem_cache_cpu *c;
1702 c = get_cpu_slab(s, raw_smp_processor_id());
1703 stat(c, FREE_SLOWPATH);
1706 if (unlikely(SlabDebug(page)))
1710 prior = object[offset] = page->freelist;
1711 page->freelist = object;
1714 if (unlikely(SlabFrozen(page))) {
1715 stat(c, FREE_FROZEN);
1719 if (unlikely(!page->inuse))
1723 * Objects left in the slab. If it was not on the partial list before
1726 if (unlikely(!prior)) {
1727 add_partial(get_node(s, page_to_nid(page)), page, 1);
1728 stat(c, FREE_ADD_PARTIAL);
1738 * Slab still on the partial list.
1740 remove_partial(s, page);
1741 stat(c, FREE_REMOVE_PARTIAL);
1745 discard_slab(s, page);
1749 if (!free_debug_processing(s, page, x, addr))
1755 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1756 * can perform fastpath freeing without additional function calls.
1758 * The fastpath is only possible if we are freeing to the current cpu slab
1759 * of this processor. This typically the case if we have just allocated
1762 * If fastpath is not possible then fall back to __slab_free where we deal
1763 * with all sorts of special processing.
1765 static __always_inline void slab_free(struct kmem_cache *s,
1766 struct page *page, void *x, void *addr)
1768 void **object = (void *)x;
1769 struct kmem_cache_cpu *c;
1770 unsigned long flags;
1772 local_irq_save(flags);
1773 c = get_cpu_slab(s, smp_processor_id());
1774 debug_check_no_locks_freed(object, c->objsize);
1775 if (likely(page == c->page && c->node >= 0)) {
1776 object[c->offset] = c->freelist;
1777 c->freelist = object;
1778 stat(c, FREE_FASTPATH);
1780 __slab_free(s, page, x, addr, c->offset);
1782 local_irq_restore(flags);
1785 void kmem_cache_free(struct kmem_cache *s, void *x)
1789 page = virt_to_head_page(x);
1791 slab_free(s, page, x, __builtin_return_address(0));
1793 EXPORT_SYMBOL(kmem_cache_free);
1795 /* Figure out on which slab object the object resides */
1796 static struct page *get_object_page(const void *x)
1798 struct page *page = virt_to_head_page(x);
1800 if (!PageSlab(page))
1807 * Object placement in a slab is made very easy because we always start at
1808 * offset 0. If we tune the size of the object to the alignment then we can
1809 * get the required alignment by putting one properly sized object after
1812 * Notice that the allocation order determines the sizes of the per cpu
1813 * caches. Each processor has always one slab available for allocations.
1814 * Increasing the allocation order reduces the number of times that slabs
1815 * must be moved on and off the partial lists and is therefore a factor in
1820 * Mininum / Maximum order of slab pages. This influences locking overhead
1821 * and slab fragmentation. A higher order reduces the number of partial slabs
1822 * and increases the number of allocations possible without having to
1823 * take the list_lock.
1825 static int slub_min_order;
1826 static int slub_max_order = DEFAULT_MAX_ORDER;
1827 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1830 * Merge control. If this is set then no merging of slab caches will occur.
1831 * (Could be removed. This was introduced to pacify the merge skeptics.)
1833 static int slub_nomerge;
1836 * Calculate the order of allocation given an slab object size.
1838 * The order of allocation has significant impact on performance and other
1839 * system components. Generally order 0 allocations should be preferred since
1840 * order 0 does not cause fragmentation in the page allocator. Larger objects
1841 * be problematic to put into order 0 slabs because there may be too much
1842 * unused space left. We go to a higher order if more than 1/8th of the slab
1845 * In order to reach satisfactory performance we must ensure that a minimum
1846 * number of objects is in one slab. Otherwise we may generate too much
1847 * activity on the partial lists which requires taking the list_lock. This is
1848 * less a concern for large slabs though which are rarely used.
1850 * slub_max_order specifies the order where we begin to stop considering the
1851 * number of objects in a slab as critical. If we reach slub_max_order then
1852 * we try to keep the page order as low as possible. So we accept more waste
1853 * of space in favor of a small page order.
1855 * Higher order allocations also allow the placement of more objects in a
1856 * slab and thereby reduce object handling overhead. If the user has
1857 * requested a higher mininum order then we start with that one instead of
1858 * the smallest order which will fit the object.
1860 static inline int slab_order(int size, int min_objects,
1861 int max_order, int fract_leftover)
1865 int min_order = slub_min_order;
1867 if ((PAGE_SIZE << min_order) / size > 65535)
1868 return get_order(size * 65535) - 1;
1870 for (order = max(min_order,
1871 fls(min_objects * size - 1) - PAGE_SHIFT);
1872 order <= max_order; order++) {
1874 unsigned long slab_size = PAGE_SIZE << order;
1876 if (slab_size < min_objects * size)
1879 rem = slab_size % size;
1881 if (rem <= slab_size / fract_leftover)
1889 static inline int calculate_order(int size)
1896 * Attempt to find best configuration for a slab. This
1897 * works by first attempting to generate a layout with
1898 * the best configuration and backing off gradually.
1900 * First we reduce the acceptable waste in a slab. Then
1901 * we reduce the minimum objects required in a slab.
1903 min_objects = slub_min_objects;
1904 while (min_objects > 1) {
1906 while (fraction >= 4) {
1907 order = slab_order(size, min_objects,
1908 slub_max_order, fraction);
1909 if (order <= slub_max_order)
1917 * We were unable to place multiple objects in a slab. Now
1918 * lets see if we can place a single object there.
1920 order = slab_order(size, 1, slub_max_order, 1);
1921 if (order <= slub_max_order)
1925 * Doh this slab cannot be placed using slub_max_order.
1927 order = slab_order(size, 1, MAX_ORDER, 1);
1928 if (order <= MAX_ORDER)
1934 * Figure out what the alignment of the objects will be.
1936 static unsigned long calculate_alignment(unsigned long flags,
1937 unsigned long align, unsigned long size)
1940 * If the user wants hardware cache aligned objects then follow that
1941 * suggestion if the object is sufficiently large.
1943 * The hardware cache alignment cannot override the specified
1944 * alignment though. If that is greater then use it.
1946 if (flags & SLAB_HWCACHE_ALIGN) {
1947 unsigned long ralign = cache_line_size();
1948 while (size <= ralign / 2)
1950 align = max(align, ralign);
1953 if (align < ARCH_SLAB_MINALIGN)
1954 align = ARCH_SLAB_MINALIGN;
1956 return ALIGN(align, sizeof(void *));
1959 static void init_kmem_cache_cpu(struct kmem_cache *s,
1960 struct kmem_cache_cpu *c)
1965 c->offset = s->offset / sizeof(void *);
1966 c->objsize = s->objsize;
1967 #ifdef CONFIG_SLUB_STATS
1968 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1972 static void init_kmem_cache_node(struct kmem_cache_node *n)
1975 spin_lock_init(&n->list_lock);
1976 INIT_LIST_HEAD(&n->partial);
1977 #ifdef CONFIG_SLUB_DEBUG
1978 atomic_long_set(&n->nr_slabs, 0);
1979 INIT_LIST_HEAD(&n->full);
1985 * Per cpu array for per cpu structures.
1987 * The per cpu array places all kmem_cache_cpu structures from one processor
1988 * close together meaning that it becomes possible that multiple per cpu
1989 * structures are contained in one cacheline. This may be particularly
1990 * beneficial for the kmalloc caches.
1992 * A desktop system typically has around 60-80 slabs. With 100 here we are
1993 * likely able to get per cpu structures for all caches from the array defined
1994 * here. We must be able to cover all kmalloc caches during bootstrap.
1996 * If the per cpu array is exhausted then fall back to kmalloc
1997 * of individual cachelines. No sharing is possible then.
1999 #define NR_KMEM_CACHE_CPU 100
2001 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2002 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2004 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2005 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
2007 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2008 int cpu, gfp_t flags)
2010 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2013 per_cpu(kmem_cache_cpu_free, cpu) =
2014 (void *)c->freelist;
2016 /* Table overflow: So allocate ourselves */
2018 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2019 flags, cpu_to_node(cpu));
2024 init_kmem_cache_cpu(s, c);
2028 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2030 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2031 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2035 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2036 per_cpu(kmem_cache_cpu_free, cpu) = c;
2039 static void free_kmem_cache_cpus(struct kmem_cache *s)
2043 for_each_online_cpu(cpu) {
2044 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2047 s->cpu_slab[cpu] = NULL;
2048 free_kmem_cache_cpu(c, cpu);
2053 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2057 for_each_online_cpu(cpu) {
2058 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2063 c = alloc_kmem_cache_cpu(s, cpu, flags);
2065 free_kmem_cache_cpus(s);
2068 s->cpu_slab[cpu] = c;
2074 * Initialize the per cpu array.
2076 static void init_alloc_cpu_cpu(int cpu)
2080 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2083 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2084 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2086 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2089 static void __init init_alloc_cpu(void)
2093 for_each_online_cpu(cpu)
2094 init_alloc_cpu_cpu(cpu);
2098 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2099 static inline void init_alloc_cpu(void) {}
2101 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2103 init_kmem_cache_cpu(s, &s->cpu_slab);
2110 * No kmalloc_node yet so do it by hand. We know that this is the first
2111 * slab on the node for this slabcache. There are no concurrent accesses
2114 * Note that this function only works on the kmalloc_node_cache
2115 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2116 * memory on a fresh node that has no slab structures yet.
2118 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2122 struct kmem_cache_node *n;
2123 unsigned long flags;
2125 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2127 page = new_slab(kmalloc_caches, gfpflags, node);
2130 if (page_to_nid(page) != node) {
2131 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2133 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2134 "in order to be able to continue\n");
2139 page->freelist = get_freepointer(kmalloc_caches, n);
2141 kmalloc_caches->node[node] = n;
2142 #ifdef CONFIG_SLUB_DEBUG
2143 init_object(kmalloc_caches, n, 1);
2144 init_tracking(kmalloc_caches, n);
2146 init_kmem_cache_node(n);
2147 inc_slabs_node(kmalloc_caches, node);
2150 * lockdep requires consistent irq usage for each lock
2151 * so even though there cannot be a race this early in
2152 * the boot sequence, we still disable irqs.
2154 local_irq_save(flags);
2155 add_partial(n, page, 0);
2156 local_irq_restore(flags);
2160 static void free_kmem_cache_nodes(struct kmem_cache *s)
2164 for_each_node_state(node, N_NORMAL_MEMORY) {
2165 struct kmem_cache_node *n = s->node[node];
2166 if (n && n != &s->local_node)
2167 kmem_cache_free(kmalloc_caches, n);
2168 s->node[node] = NULL;
2172 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2177 if (slab_state >= UP)
2178 local_node = page_to_nid(virt_to_page(s));
2182 for_each_node_state(node, N_NORMAL_MEMORY) {
2183 struct kmem_cache_node *n;
2185 if (local_node == node)
2188 if (slab_state == DOWN) {
2189 n = early_kmem_cache_node_alloc(gfpflags,
2193 n = kmem_cache_alloc_node(kmalloc_caches,
2197 free_kmem_cache_nodes(s);
2203 init_kmem_cache_node(n);
2208 static void free_kmem_cache_nodes(struct kmem_cache *s)
2212 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2214 init_kmem_cache_node(&s->local_node);
2220 * calculate_sizes() determines the order and the distribution of data within
2223 static int calculate_sizes(struct kmem_cache *s)
2225 unsigned long flags = s->flags;
2226 unsigned long size = s->objsize;
2227 unsigned long align = s->align;
2231 * Round up object size to the next word boundary. We can only
2232 * place the free pointer at word boundaries and this determines
2233 * the possible location of the free pointer.
2235 size = ALIGN(size, sizeof(void *));
2237 #ifdef CONFIG_SLUB_DEBUG
2239 * Determine if we can poison the object itself. If the user of
2240 * the slab may touch the object after free or before allocation
2241 * then we should never poison the object itself.
2243 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2245 s->flags |= __OBJECT_POISON;
2247 s->flags &= ~__OBJECT_POISON;
2251 * If we are Redzoning then check if there is some space between the
2252 * end of the object and the free pointer. If not then add an
2253 * additional word to have some bytes to store Redzone information.
2255 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2256 size += sizeof(void *);
2260 * With that we have determined the number of bytes in actual use
2261 * by the object. This is the potential offset to the free pointer.
2265 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2268 * Relocate free pointer after the object if it is not
2269 * permitted to overwrite the first word of the object on
2272 * This is the case if we do RCU, have a constructor or
2273 * destructor or are poisoning the objects.
2276 size += sizeof(void *);
2279 #ifdef CONFIG_SLUB_DEBUG
2280 if (flags & SLAB_STORE_USER)
2282 * Need to store information about allocs and frees after
2285 size += 2 * sizeof(struct track);
2287 if (flags & SLAB_RED_ZONE)
2289 * Add some empty padding so that we can catch
2290 * overwrites from earlier objects rather than let
2291 * tracking information or the free pointer be
2292 * corrupted if an user writes before the start
2295 size += sizeof(void *);
2299 * Determine the alignment based on various parameters that the
2300 * user specified and the dynamic determination of cache line size
2303 align = calculate_alignment(flags, align, s->objsize);
2306 * SLUB stores one object immediately after another beginning from
2307 * offset 0. In order to align the objects we have to simply size
2308 * each object to conform to the alignment.
2310 size = ALIGN(size, align);
2313 if ((flags & __KMALLOC_CACHE) &&
2314 PAGE_SIZE / size < slub_min_objects) {
2316 * Kmalloc cache that would not have enough objects in
2317 * an order 0 page. Kmalloc slabs can fallback to
2318 * page allocator order 0 allocs so take a reasonably large
2319 * order that will allows us a good number of objects.
2321 order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2322 s->flags |= __PAGE_ALLOC_FALLBACK;
2323 s->allocflags |= __GFP_NOWARN;
2325 order = calculate_order(size);
2332 s->allocflags |= __GFP_COMP;
2334 if (s->flags & SLAB_CACHE_DMA)
2335 s->allocflags |= SLUB_DMA;
2337 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2338 s->allocflags |= __GFP_RECLAIMABLE;
2341 * Determine the number of objects per slab
2343 s->oo = oo_make(order, size);
2345 return !!oo_objects(s->oo);
2349 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2350 const char *name, size_t size,
2351 size_t align, unsigned long flags,
2352 void (*ctor)(struct kmem_cache *, void *))
2354 memset(s, 0, kmem_size);
2359 s->flags = kmem_cache_flags(size, flags, name, ctor);
2361 if (!calculate_sizes(s))
2366 s->remote_node_defrag_ratio = 100;
2368 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2371 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2373 free_kmem_cache_nodes(s);
2375 if (flags & SLAB_PANIC)
2376 panic("Cannot create slab %s size=%lu realsize=%u "
2377 "order=%u offset=%u flags=%lx\n",
2378 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2384 * Check if a given pointer is valid
2386 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2390 page = get_object_page(object);
2392 if (!page || s != page->slab)
2393 /* No slab or wrong slab */
2396 if (!check_valid_pointer(s, page, object))
2400 * We could also check if the object is on the slabs freelist.
2401 * But this would be too expensive and it seems that the main
2402 * purpose of kmem_ptr_valid() is to check if the object belongs
2403 * to a certain slab.
2407 EXPORT_SYMBOL(kmem_ptr_validate);
2410 * Determine the size of a slab object
2412 unsigned int kmem_cache_size(struct kmem_cache *s)
2416 EXPORT_SYMBOL(kmem_cache_size);
2418 const char *kmem_cache_name(struct kmem_cache *s)
2422 EXPORT_SYMBOL(kmem_cache_name);
2424 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2427 #ifdef CONFIG_SLUB_DEBUG
2428 void *addr = page_address(page);
2430 DECLARE_BITMAP(map, page->objects);
2432 bitmap_zero(map, page->objects);
2433 slab_err(s, page, "%s", text);
2435 for_each_free_object(p, s, page->freelist)
2436 set_bit(slab_index(p, s, addr), map);
2438 for_each_object(p, s, addr, page->objects) {
2440 if (!test_bit(slab_index(p, s, addr), map)) {
2441 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2443 print_tracking(s, p);
2451 * Attempt to free all partial slabs on a node.
2453 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2455 unsigned long flags;
2456 struct page *page, *h;
2458 spin_lock_irqsave(&n->list_lock, flags);
2459 list_for_each_entry_safe(page, h, &n->partial, lru) {
2461 list_del(&page->lru);
2462 discard_slab(s, page);
2465 list_slab_objects(s, page,
2466 "Objects remaining on kmem_cache_close()");
2469 spin_unlock_irqrestore(&n->list_lock, flags);
2473 * Release all resources used by a slab cache.
2475 static inline int kmem_cache_close(struct kmem_cache *s)
2481 /* Attempt to free all objects */
2482 free_kmem_cache_cpus(s);
2483 for_each_node_state(node, N_NORMAL_MEMORY) {
2484 struct kmem_cache_node *n = get_node(s, node);
2487 if (n->nr_partial || slabs_node(s, node))
2490 free_kmem_cache_nodes(s);
2495 * Close a cache and release the kmem_cache structure
2496 * (must be used for caches created using kmem_cache_create)
2498 void kmem_cache_destroy(struct kmem_cache *s)
2500 down_write(&slub_lock);
2504 up_write(&slub_lock);
2505 if (kmem_cache_close(s)) {
2506 printk(KERN_ERR "SLUB %s: %s called for cache that "
2507 "still has objects.\n", s->name, __func__);
2510 sysfs_slab_remove(s);
2512 up_write(&slub_lock);
2514 EXPORT_SYMBOL(kmem_cache_destroy);
2516 /********************************************************************
2518 *******************************************************************/
2520 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2521 EXPORT_SYMBOL(kmalloc_caches);
2523 static int __init setup_slub_min_order(char *str)
2525 get_option(&str, &slub_min_order);
2530 __setup("slub_min_order=", setup_slub_min_order);
2532 static int __init setup_slub_max_order(char *str)
2534 get_option(&str, &slub_max_order);
2539 __setup("slub_max_order=", setup_slub_max_order);
2541 static int __init setup_slub_min_objects(char *str)
2543 get_option(&str, &slub_min_objects);
2548 __setup("slub_min_objects=", setup_slub_min_objects);
2550 static int __init setup_slub_nomerge(char *str)
2556 __setup("slub_nomerge", setup_slub_nomerge);
2558 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2559 const char *name, int size, gfp_t gfp_flags)
2561 unsigned int flags = 0;
2563 if (gfp_flags & SLUB_DMA)
2564 flags = SLAB_CACHE_DMA;
2566 down_write(&slub_lock);
2567 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2568 flags | __KMALLOC_CACHE, NULL))
2571 list_add(&s->list, &slab_caches);
2572 up_write(&slub_lock);
2573 if (sysfs_slab_add(s))
2578 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2581 #ifdef CONFIG_ZONE_DMA
2582 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2584 static void sysfs_add_func(struct work_struct *w)
2586 struct kmem_cache *s;
2588 down_write(&slub_lock);
2589 list_for_each_entry(s, &slab_caches, list) {
2590 if (s->flags & __SYSFS_ADD_DEFERRED) {
2591 s->flags &= ~__SYSFS_ADD_DEFERRED;
2595 up_write(&slub_lock);
2598 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2600 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2602 struct kmem_cache *s;
2606 s = kmalloc_caches_dma[index];
2610 /* Dynamically create dma cache */
2611 if (flags & __GFP_WAIT)
2612 down_write(&slub_lock);
2614 if (!down_write_trylock(&slub_lock))
2618 if (kmalloc_caches_dma[index])
2621 realsize = kmalloc_caches[index].objsize;
2622 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2623 (unsigned int)realsize);
2624 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2626 if (!s || !text || !kmem_cache_open(s, flags, text,
2627 realsize, ARCH_KMALLOC_MINALIGN,
2628 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2634 list_add(&s->list, &slab_caches);
2635 kmalloc_caches_dma[index] = s;
2637 schedule_work(&sysfs_add_work);
2640 up_write(&slub_lock);
2642 return kmalloc_caches_dma[index];
2647 * Conversion table for small slabs sizes / 8 to the index in the
2648 * kmalloc array. This is necessary for slabs < 192 since we have non power
2649 * of two cache sizes there. The size of larger slabs can be determined using
2652 static s8 size_index[24] = {
2679 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2685 return ZERO_SIZE_PTR;
2687 index = size_index[(size - 1) / 8];
2689 index = fls(size - 1);
2691 #ifdef CONFIG_ZONE_DMA
2692 if (unlikely((flags & SLUB_DMA)))
2693 return dma_kmalloc_cache(index, flags);
2696 return &kmalloc_caches[index];
2699 void *__kmalloc(size_t size, gfp_t flags)
2701 struct kmem_cache *s;
2703 if (unlikely(size > PAGE_SIZE))
2704 return kmalloc_large(size, flags);
2706 s = get_slab(size, flags);
2708 if (unlikely(ZERO_OR_NULL_PTR(s)))
2711 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2713 EXPORT_SYMBOL(__kmalloc);
2715 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2717 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2721 return page_address(page);
2727 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2729 struct kmem_cache *s;
2731 if (unlikely(size > PAGE_SIZE))
2732 return kmalloc_large_node(size, flags, node);
2734 s = get_slab(size, flags);
2736 if (unlikely(ZERO_OR_NULL_PTR(s)))
2739 return slab_alloc(s, flags, node, __builtin_return_address(0));
2741 EXPORT_SYMBOL(__kmalloc_node);
2744 size_t ksize(const void *object)
2747 struct kmem_cache *s;
2749 if (unlikely(object == ZERO_SIZE_PTR))
2752 page = virt_to_head_page(object);
2754 if (unlikely(!PageSlab(page)))
2755 return PAGE_SIZE << compound_order(page);
2759 #ifdef CONFIG_SLUB_DEBUG
2761 * Debugging requires use of the padding between object
2762 * and whatever may come after it.
2764 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2769 * If we have the need to store the freelist pointer
2770 * back there or track user information then we can
2771 * only use the space before that information.
2773 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2776 * Else we can use all the padding etc for the allocation
2780 EXPORT_SYMBOL(ksize);
2782 void kfree(const void *x)
2785 void *object = (void *)x;
2787 if (unlikely(ZERO_OR_NULL_PTR(x)))
2790 page = virt_to_head_page(x);
2791 if (unlikely(!PageSlab(page))) {
2795 slab_free(page->slab, page, object, __builtin_return_address(0));
2797 EXPORT_SYMBOL(kfree);
2800 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2801 * the remaining slabs by the number of items in use. The slabs with the
2802 * most items in use come first. New allocations will then fill those up
2803 * and thus they can be removed from the partial lists.
2805 * The slabs with the least items are placed last. This results in them
2806 * being allocated from last increasing the chance that the last objects
2807 * are freed in them.
2809 int kmem_cache_shrink(struct kmem_cache *s)
2813 struct kmem_cache_node *n;
2816 int objects = oo_objects(s->oo);
2817 struct list_head *slabs_by_inuse =
2818 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2819 unsigned long flags;
2821 if (!slabs_by_inuse)
2825 for_each_node_state(node, N_NORMAL_MEMORY) {
2826 n = get_node(s, node);
2831 for (i = 0; i < objects; i++)
2832 INIT_LIST_HEAD(slabs_by_inuse + i);
2834 spin_lock_irqsave(&n->list_lock, flags);
2837 * Build lists indexed by the items in use in each slab.
2839 * Note that concurrent frees may occur while we hold the
2840 * list_lock. page->inuse here is the upper limit.
2842 list_for_each_entry_safe(page, t, &n->partial, lru) {
2843 if (!page->inuse && slab_trylock(page)) {
2845 * Must hold slab lock here because slab_free
2846 * may have freed the last object and be
2847 * waiting to release the slab.
2849 list_del(&page->lru);
2852 discard_slab(s, page);
2854 list_move(&page->lru,
2855 slabs_by_inuse + page->inuse);
2860 * Rebuild the partial list with the slabs filled up most
2861 * first and the least used slabs at the end.
2863 for (i = objects - 1; i >= 0; i--)
2864 list_splice(slabs_by_inuse + i, n->partial.prev);
2866 spin_unlock_irqrestore(&n->list_lock, flags);
2869 kfree(slabs_by_inuse);
2872 EXPORT_SYMBOL(kmem_cache_shrink);
2874 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2875 static int slab_mem_going_offline_callback(void *arg)
2877 struct kmem_cache *s;
2879 down_read(&slub_lock);
2880 list_for_each_entry(s, &slab_caches, list)
2881 kmem_cache_shrink(s);
2882 up_read(&slub_lock);
2887 static void slab_mem_offline_callback(void *arg)
2889 struct kmem_cache_node *n;
2890 struct kmem_cache *s;
2891 struct memory_notify *marg = arg;
2894 offline_node = marg->status_change_nid;
2897 * If the node still has available memory. we need kmem_cache_node
2900 if (offline_node < 0)
2903 down_read(&slub_lock);
2904 list_for_each_entry(s, &slab_caches, list) {
2905 n = get_node(s, offline_node);
2908 * if n->nr_slabs > 0, slabs still exist on the node
2909 * that is going down. We were unable to free them,
2910 * and offline_pages() function shoudn't call this
2911 * callback. So, we must fail.
2913 BUG_ON(slabs_node(s, offline_node));
2915 s->node[offline_node] = NULL;
2916 kmem_cache_free(kmalloc_caches, n);
2919 up_read(&slub_lock);
2922 static int slab_mem_going_online_callback(void *arg)
2924 struct kmem_cache_node *n;
2925 struct kmem_cache *s;
2926 struct memory_notify *marg = arg;
2927 int nid = marg->status_change_nid;
2931 * If the node's memory is already available, then kmem_cache_node is
2932 * already created. Nothing to do.
2938 * We are bringing a node online. No memory is availabe yet. We must
2939 * allocate a kmem_cache_node structure in order to bring the node
2942 down_read(&slub_lock);
2943 list_for_each_entry(s, &slab_caches, list) {
2945 * XXX: kmem_cache_alloc_node will fallback to other nodes
2946 * since memory is not yet available from the node that
2949 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2954 init_kmem_cache_node(n);
2958 up_read(&slub_lock);
2962 static int slab_memory_callback(struct notifier_block *self,
2963 unsigned long action, void *arg)
2968 case MEM_GOING_ONLINE:
2969 ret = slab_mem_going_online_callback(arg);
2971 case MEM_GOING_OFFLINE:
2972 ret = slab_mem_going_offline_callback(arg);
2975 case MEM_CANCEL_ONLINE:
2976 slab_mem_offline_callback(arg);
2979 case MEM_CANCEL_OFFLINE:
2983 ret = notifier_from_errno(ret);
2987 #endif /* CONFIG_MEMORY_HOTPLUG */
2989 /********************************************************************
2990 * Basic setup of slabs
2991 *******************************************************************/
2993 void __init kmem_cache_init(void)
3002 * Must first have the slab cache available for the allocations of the
3003 * struct kmem_cache_node's. There is special bootstrap code in
3004 * kmem_cache_open for slab_state == DOWN.
3006 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3007 sizeof(struct kmem_cache_node), GFP_KERNEL);
3008 kmalloc_caches[0].refcount = -1;
3011 hotplug_memory_notifier(slab_memory_callback, 1);
3014 /* Able to allocate the per node structures */
3015 slab_state = PARTIAL;
3017 /* Caches that are not of the two-to-the-power-of size */
3018 if (KMALLOC_MIN_SIZE <= 64) {
3019 create_kmalloc_cache(&kmalloc_caches[1],
3020 "kmalloc-96", 96, GFP_KERNEL);
3023 if (KMALLOC_MIN_SIZE <= 128) {
3024 create_kmalloc_cache(&kmalloc_caches[2],
3025 "kmalloc-192", 192, GFP_KERNEL);
3029 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
3030 create_kmalloc_cache(&kmalloc_caches[i],
3031 "kmalloc", 1 << i, GFP_KERNEL);
3037 * Patch up the size_index table if we have strange large alignment
3038 * requirements for the kmalloc array. This is only the case for
3039 * MIPS it seems. The standard arches will not generate any code here.
3041 * Largest permitted alignment is 256 bytes due to the way we
3042 * handle the index determination for the smaller caches.
3044 * Make sure that nothing crazy happens if someone starts tinkering
3045 * around with ARCH_KMALLOC_MINALIGN
3047 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3048 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3050 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3051 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3055 /* Provide the correct kmalloc names now that the caches are up */
3056 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3057 kmalloc_caches[i]. name =
3058 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3061 register_cpu_notifier(&slab_notifier);
3062 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3063 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3065 kmem_size = sizeof(struct kmem_cache);
3069 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3070 " CPUs=%d, Nodes=%d\n",
3071 caches, cache_line_size(),
3072 slub_min_order, slub_max_order, slub_min_objects,
3073 nr_cpu_ids, nr_node_ids);
3077 * Find a mergeable slab cache
3079 static int slab_unmergeable(struct kmem_cache *s)
3081 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3084 if ((s->flags & __PAGE_ALLOC_FALLBACK))
3091 * We may have set a slab to be unmergeable during bootstrap.
3093 if (s->refcount < 0)
3099 static struct kmem_cache *find_mergeable(size_t size,
3100 size_t align, unsigned long flags, const char *name,
3101 void (*ctor)(struct kmem_cache *, void *))
3103 struct kmem_cache *s;
3105 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3111 size = ALIGN(size, sizeof(void *));
3112 align = calculate_alignment(flags, align, size);
3113 size = ALIGN(size, align);
3114 flags = kmem_cache_flags(size, flags, name, NULL);
3116 list_for_each_entry(s, &slab_caches, list) {
3117 if (slab_unmergeable(s))
3123 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3126 * Check if alignment is compatible.
3127 * Courtesy of Adrian Drzewiecki
3129 if ((s->size & ~(align - 1)) != s->size)
3132 if (s->size - size >= sizeof(void *))
3140 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3141 size_t align, unsigned long flags,
3142 void (*ctor)(struct kmem_cache *, void *))
3144 struct kmem_cache *s;
3146 down_write(&slub_lock);
3147 s = find_mergeable(size, align, flags, name, ctor);
3153 * Adjust the object sizes so that we clear
3154 * the complete object on kzalloc.
3156 s->objsize = max(s->objsize, (int)size);
3159 * And then we need to update the object size in the
3160 * per cpu structures
3162 for_each_online_cpu(cpu)
3163 get_cpu_slab(s, cpu)->objsize = s->objsize;
3165 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3166 up_write(&slub_lock);
3168 if (sysfs_slab_alias(s, name))
3173 s = kmalloc(kmem_size, GFP_KERNEL);
3175 if (kmem_cache_open(s, GFP_KERNEL, name,
3176 size, align, flags, ctor)) {
3177 list_add(&s->list, &slab_caches);
3178 up_write(&slub_lock);
3179 if (sysfs_slab_add(s))
3185 up_write(&slub_lock);
3188 if (flags & SLAB_PANIC)
3189 panic("Cannot create slabcache %s\n", name);
3194 EXPORT_SYMBOL(kmem_cache_create);
3198 * Use the cpu notifier to insure that the cpu slabs are flushed when
3201 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3202 unsigned long action, void *hcpu)
3204 long cpu = (long)hcpu;
3205 struct kmem_cache *s;
3206 unsigned long flags;
3209 case CPU_UP_PREPARE:
3210 case CPU_UP_PREPARE_FROZEN:
3211 init_alloc_cpu_cpu(cpu);
3212 down_read(&slub_lock);
3213 list_for_each_entry(s, &slab_caches, list)
3214 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3216 up_read(&slub_lock);
3219 case CPU_UP_CANCELED:
3220 case CPU_UP_CANCELED_FROZEN:
3222 case CPU_DEAD_FROZEN:
3223 down_read(&slub_lock);
3224 list_for_each_entry(s, &slab_caches, list) {
3225 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3227 local_irq_save(flags);
3228 __flush_cpu_slab(s, cpu);
3229 local_irq_restore(flags);
3230 free_kmem_cache_cpu(c, cpu);
3231 s->cpu_slab[cpu] = NULL;
3233 up_read(&slub_lock);
3241 static struct notifier_block __cpuinitdata slab_notifier = {
3242 .notifier_call = slab_cpuup_callback
3247 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3249 struct kmem_cache *s;
3251 if (unlikely(size > PAGE_SIZE))
3252 return kmalloc_large(size, gfpflags);
3254 s = get_slab(size, gfpflags);
3256 if (unlikely(ZERO_OR_NULL_PTR(s)))
3259 return slab_alloc(s, gfpflags, -1, caller);
3262 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3263 int node, void *caller)
3265 struct kmem_cache *s;
3267 if (unlikely(size > PAGE_SIZE))
3268 return kmalloc_large_node(size, gfpflags, node);
3270 s = get_slab(size, gfpflags);
3272 if (unlikely(ZERO_OR_NULL_PTR(s)))
3275 return slab_alloc(s, gfpflags, node, caller);
3278 #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3279 static unsigned long count_partial(struct kmem_cache_node *n)
3281 unsigned long flags;
3282 unsigned long x = 0;
3285 spin_lock_irqsave(&n->list_lock, flags);
3286 list_for_each_entry(page, &n->partial, lru)
3288 spin_unlock_irqrestore(&n->list_lock, flags);
3293 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3294 static int validate_slab(struct kmem_cache *s, struct page *page,
3298 void *addr = page_address(page);
3300 if (!check_slab(s, page) ||
3301 !on_freelist(s, page, NULL))
3304 /* Now we know that a valid freelist exists */
3305 bitmap_zero(map, page->objects);
3307 for_each_free_object(p, s, page->freelist) {
3308 set_bit(slab_index(p, s, addr), map);
3309 if (!check_object(s, page, p, 0))
3313 for_each_object(p, s, addr, page->objects)
3314 if (!test_bit(slab_index(p, s, addr), map))
3315 if (!check_object(s, page, p, 1))
3320 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3323 if (slab_trylock(page)) {
3324 validate_slab(s, page, map);
3327 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3330 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3331 if (!SlabDebug(page))
3332 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3333 "on slab 0x%p\n", s->name, page);
3335 if (SlabDebug(page))
3336 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3337 "slab 0x%p\n", s->name, page);
3341 static int validate_slab_node(struct kmem_cache *s,
3342 struct kmem_cache_node *n, unsigned long *map)
3344 unsigned long count = 0;
3346 unsigned long flags;
3348 spin_lock_irqsave(&n->list_lock, flags);
3350 list_for_each_entry(page, &n->partial, lru) {
3351 validate_slab_slab(s, page, map);
3354 if (count != n->nr_partial)
3355 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3356 "counter=%ld\n", s->name, count, n->nr_partial);
3358 if (!(s->flags & SLAB_STORE_USER))
3361 list_for_each_entry(page, &n->full, lru) {
3362 validate_slab_slab(s, page, map);
3365 if (count != atomic_long_read(&n->nr_slabs))
3366 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3367 "counter=%ld\n", s->name, count,
3368 atomic_long_read(&n->nr_slabs));
3371 spin_unlock_irqrestore(&n->list_lock, flags);
3375 static long validate_slab_cache(struct kmem_cache *s)
3378 unsigned long count = 0;
3379 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->oo)) *
3380 sizeof(unsigned long), GFP_KERNEL);
3386 for_each_node_state(node, N_NORMAL_MEMORY) {
3387 struct kmem_cache_node *n = get_node(s, node);
3389 count += validate_slab_node(s, n, map);
3395 #ifdef SLUB_RESILIENCY_TEST
3396 static void resiliency_test(void)
3400 printk(KERN_ERR "SLUB resiliency testing\n");
3401 printk(KERN_ERR "-----------------------\n");
3402 printk(KERN_ERR "A. Corruption after allocation\n");
3404 p = kzalloc(16, GFP_KERNEL);
3406 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3407 " 0x12->0x%p\n\n", p + 16);
3409 validate_slab_cache(kmalloc_caches + 4);
3411 /* Hmmm... The next two are dangerous */
3412 p = kzalloc(32, GFP_KERNEL);
3413 p[32 + sizeof(void *)] = 0x34;
3414 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3415 " 0x34 -> -0x%p\n", p);
3417 "If allocated object is overwritten then not detectable\n\n");
3419 validate_slab_cache(kmalloc_caches + 5);
3420 p = kzalloc(64, GFP_KERNEL);
3421 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3423 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3426 "If allocated object is overwritten then not detectable\n\n");
3427 validate_slab_cache(kmalloc_caches + 6);
3429 printk(KERN_ERR "\nB. Corruption after free\n");
3430 p = kzalloc(128, GFP_KERNEL);
3433 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3434 validate_slab_cache(kmalloc_caches + 7);
3436 p = kzalloc(256, GFP_KERNEL);
3439 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3441 validate_slab_cache(kmalloc_caches + 8);
3443 p = kzalloc(512, GFP_KERNEL);
3446 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3447 validate_slab_cache(kmalloc_caches + 9);
3450 static void resiliency_test(void) {};
3454 * Generate lists of code addresses where slabcache objects are allocated
3459 unsigned long count;
3472 unsigned long count;
3473 struct location *loc;
3476 static void free_loc_track(struct loc_track *t)
3479 free_pages((unsigned long)t->loc,
3480 get_order(sizeof(struct location) * t->max));
3483 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3488 order = get_order(sizeof(struct location) * max);
3490 l = (void *)__get_free_pages(flags, order);
3495 memcpy(l, t->loc, sizeof(struct location) * t->count);
3503 static int add_location(struct loc_track *t, struct kmem_cache *s,
3504 const struct track *track)
3506 long start, end, pos;
3509 unsigned long age = jiffies - track->when;
3515 pos = start + (end - start + 1) / 2;
3518 * There is nothing at "end". If we end up there
3519 * we need to add something to before end.
3524 caddr = t->loc[pos].addr;
3525 if (track->addr == caddr) {
3531 if (age < l->min_time)
3533 if (age > l->max_time)
3536 if (track->pid < l->min_pid)
3537 l->min_pid = track->pid;
3538 if (track->pid > l->max_pid)
3539 l->max_pid = track->pid;
3541 cpu_set(track->cpu, l->cpus);
3543 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3547 if (track->addr < caddr)
3554 * Not found. Insert new tracking element.
3556 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3562 (t->count - pos) * sizeof(struct location));
3565 l->addr = track->addr;
3569 l->min_pid = track->pid;
3570 l->max_pid = track->pid;
3571 cpus_clear(l->cpus);
3572 cpu_set(track->cpu, l->cpus);
3573 nodes_clear(l->nodes);
3574 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3578 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3579 struct page *page, enum track_item alloc)
3581 void *addr = page_address(page);
3582 DECLARE_BITMAP(map, page->objects);
3585 bitmap_zero(map, page->objects);
3586 for_each_free_object(p, s, page->freelist)
3587 set_bit(slab_index(p, s, addr), map);
3589 for_each_object(p, s, addr, page->objects)
3590 if (!test_bit(slab_index(p, s, addr), map))
3591 add_location(t, s, get_track(s, p, alloc));
3594 static int list_locations(struct kmem_cache *s, char *buf,
3595 enum track_item alloc)
3599 struct loc_track t = { 0, 0, NULL };
3602 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3604 return sprintf(buf, "Out of memory\n");
3606 /* Push back cpu slabs */
3609 for_each_node_state(node, N_NORMAL_MEMORY) {
3610 struct kmem_cache_node *n = get_node(s, node);
3611 unsigned long flags;
3614 if (!atomic_long_read(&n->nr_slabs))
3617 spin_lock_irqsave(&n->list_lock, flags);
3618 list_for_each_entry(page, &n->partial, lru)
3619 process_slab(&t, s, page, alloc);
3620 list_for_each_entry(page, &n->full, lru)
3621 process_slab(&t, s, page, alloc);
3622 spin_unlock_irqrestore(&n->list_lock, flags);
3625 for (i = 0; i < t.count; i++) {
3626 struct location *l = &t.loc[i];
3628 if (len > PAGE_SIZE - 100)
3630 len += sprintf(buf + len, "%7ld ", l->count);
3633 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3635 len += sprintf(buf + len, "<not-available>");
3637 if (l->sum_time != l->min_time) {
3638 unsigned long remainder;
3640 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3642 div_long_long_rem(l->sum_time, l->count, &remainder),
3645 len += sprintf(buf + len, " age=%ld",
3648 if (l->min_pid != l->max_pid)
3649 len += sprintf(buf + len, " pid=%ld-%ld",
3650 l->min_pid, l->max_pid);
3652 len += sprintf(buf + len, " pid=%ld",
3655 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3656 len < PAGE_SIZE - 60) {
3657 len += sprintf(buf + len, " cpus=");
3658 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3662 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3663 len < PAGE_SIZE - 60) {
3664 len += sprintf(buf + len, " nodes=");
3665 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3669 len += sprintf(buf + len, "\n");
3674 len += sprintf(buf, "No data\n");
3678 enum slab_stat_type {
3685 #define SO_FULL (1 << SL_FULL)
3686 #define SO_PARTIAL (1 << SL_PARTIAL)
3687 #define SO_CPU (1 << SL_CPU)
3688 #define SO_OBJECTS (1 << SL_OBJECTS)
3690 static ssize_t show_slab_objects(struct kmem_cache *s,
3691 char *buf, unsigned long flags)
3693 unsigned long total = 0;
3697 unsigned long *nodes;
3698 unsigned long *per_cpu;
3700 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3703 per_cpu = nodes + nr_node_ids;
3705 for_each_possible_cpu(cpu) {
3707 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3717 if (flags & SO_CPU) {
3718 if (flags & SO_OBJECTS)
3729 for_each_node_state(node, N_NORMAL_MEMORY) {
3730 struct kmem_cache_node *n = get_node(s, node);
3732 if (flags & SO_PARTIAL) {
3733 if (flags & SO_OBJECTS)
3734 x = count_partial(n);
3741 if (flags & SO_FULL) {
3742 int full_slabs = atomic_long_read(&n->nr_slabs)
3746 if (flags & SO_OBJECTS)
3747 x = full_slabs * oo_objects(s->oo);
3755 x = sprintf(buf, "%lu", total);
3757 for_each_node_state(node, N_NORMAL_MEMORY)
3759 x += sprintf(buf + x, " N%d=%lu",
3763 return x + sprintf(buf + x, "\n");
3766 static int any_slab_objects(struct kmem_cache *s)
3771 for_each_possible_cpu(cpu) {
3772 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3778 for_each_online_node(node) {
3779 struct kmem_cache_node *n = get_node(s, node);
3784 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3790 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3791 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3793 struct slab_attribute {
3794 struct attribute attr;
3795 ssize_t (*show)(struct kmem_cache *s, char *buf);
3796 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3799 #define SLAB_ATTR_RO(_name) \
3800 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3802 #define SLAB_ATTR(_name) \
3803 static struct slab_attribute _name##_attr = \
3804 __ATTR(_name, 0644, _name##_show, _name##_store)
3806 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3808 return sprintf(buf, "%d\n", s->size);
3810 SLAB_ATTR_RO(slab_size);
3812 static ssize_t align_show(struct kmem_cache *s, char *buf)
3814 return sprintf(buf, "%d\n", s->align);
3816 SLAB_ATTR_RO(align);
3818 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3820 return sprintf(buf, "%d\n", s->objsize);
3822 SLAB_ATTR_RO(object_size);
3824 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3826 return sprintf(buf, "%d\n", oo_objects(s->oo));
3828 SLAB_ATTR_RO(objs_per_slab);
3830 static ssize_t order_show(struct kmem_cache *s, char *buf)
3832 return sprintf(buf, "%d\n", oo_order(s->oo));
3834 SLAB_ATTR_RO(order);
3836 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3839 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3841 return n + sprintf(buf + n, "\n");
3847 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3849 return sprintf(buf, "%d\n", s->refcount - 1);
3851 SLAB_ATTR_RO(aliases);
3853 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3855 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3857 SLAB_ATTR_RO(slabs);
3859 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3861 return show_slab_objects(s, buf, SO_PARTIAL);
3863 SLAB_ATTR_RO(partial);
3865 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3867 return show_slab_objects(s, buf, SO_CPU);
3869 SLAB_ATTR_RO(cpu_slabs);
3871 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3873 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3875 SLAB_ATTR_RO(objects);
3877 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3879 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3882 static ssize_t sanity_checks_store(struct kmem_cache *s,
3883 const char *buf, size_t length)
3885 s->flags &= ~SLAB_DEBUG_FREE;
3887 s->flags |= SLAB_DEBUG_FREE;
3890 SLAB_ATTR(sanity_checks);
3892 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3894 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3897 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3900 s->flags &= ~SLAB_TRACE;
3902 s->flags |= SLAB_TRACE;
3907 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3909 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3912 static ssize_t reclaim_account_store(struct kmem_cache *s,
3913 const char *buf, size_t length)
3915 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3917 s->flags |= SLAB_RECLAIM_ACCOUNT;
3920 SLAB_ATTR(reclaim_account);
3922 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3924 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3926 SLAB_ATTR_RO(hwcache_align);
3928 #ifdef CONFIG_ZONE_DMA
3929 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3931 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3933 SLAB_ATTR_RO(cache_dma);
3936 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3938 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3940 SLAB_ATTR_RO(destroy_by_rcu);
3942 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3944 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3947 static ssize_t red_zone_store(struct kmem_cache *s,
3948 const char *buf, size_t length)
3950 if (any_slab_objects(s))
3953 s->flags &= ~SLAB_RED_ZONE;
3955 s->flags |= SLAB_RED_ZONE;
3959 SLAB_ATTR(red_zone);
3961 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3963 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3966 static ssize_t poison_store(struct kmem_cache *s,
3967 const char *buf, size_t length)
3969 if (any_slab_objects(s))
3972 s->flags &= ~SLAB_POISON;
3974 s->flags |= SLAB_POISON;
3980 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3982 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3985 static ssize_t store_user_store(struct kmem_cache *s,
3986 const char *buf, size_t length)
3988 if (any_slab_objects(s))
3991 s->flags &= ~SLAB_STORE_USER;
3993 s->flags |= SLAB_STORE_USER;
3997 SLAB_ATTR(store_user);
3999 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4004 static ssize_t validate_store(struct kmem_cache *s,
4005 const char *buf, size_t length)
4009 if (buf[0] == '1') {
4010 ret = validate_slab_cache(s);
4016 SLAB_ATTR(validate);
4018 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4023 static ssize_t shrink_store(struct kmem_cache *s,
4024 const char *buf, size_t length)
4026 if (buf[0] == '1') {
4027 int rc = kmem_cache_shrink(s);
4037 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4039 if (!(s->flags & SLAB_STORE_USER))
4041 return list_locations(s, buf, TRACK_ALLOC);
4043 SLAB_ATTR_RO(alloc_calls);
4045 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4047 if (!(s->flags & SLAB_STORE_USER))
4049 return list_locations(s, buf, TRACK_FREE);
4051 SLAB_ATTR_RO(free_calls);
4054 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4056 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4059 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4060 const char *buf, size_t length)
4062 int n = simple_strtoul(buf, NULL, 10);
4065 s->remote_node_defrag_ratio = n * 10;
4068 SLAB_ATTR(remote_node_defrag_ratio);
4071 #ifdef CONFIG_SLUB_STATS
4072 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4074 unsigned long sum = 0;
4077 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4082 for_each_online_cpu(cpu) {
4083 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4089 len = sprintf(buf, "%lu", sum);
4092 for_each_online_cpu(cpu) {
4093 if (data[cpu] && len < PAGE_SIZE - 20)
4094 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4098 return len + sprintf(buf + len, "\n");
4101 #define STAT_ATTR(si, text) \
4102 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4104 return show_stat(s, buf, si); \
4106 SLAB_ATTR_RO(text); \
4108 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4109 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4110 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4111 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4112 STAT_ATTR(FREE_FROZEN, free_frozen);
4113 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4114 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4115 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4116 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4117 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4118 STAT_ATTR(FREE_SLAB, free_slab);
4119 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4120 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4121 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4122 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4123 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4124 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4128 static struct attribute *slab_attrs[] = {
4129 &slab_size_attr.attr,
4130 &object_size_attr.attr,
4131 &objs_per_slab_attr.attr,
4136 &cpu_slabs_attr.attr,
4140 &sanity_checks_attr.attr,
4142 &hwcache_align_attr.attr,
4143 &reclaim_account_attr.attr,
4144 &destroy_by_rcu_attr.attr,
4145 &red_zone_attr.attr,
4147 &store_user_attr.attr,
4148 &validate_attr.attr,
4150 &alloc_calls_attr.attr,
4151 &free_calls_attr.attr,
4152 #ifdef CONFIG_ZONE_DMA
4153 &cache_dma_attr.attr,
4156 &remote_node_defrag_ratio_attr.attr,
4158 #ifdef CONFIG_SLUB_STATS
4159 &alloc_fastpath_attr.attr,
4160 &alloc_slowpath_attr.attr,
4161 &free_fastpath_attr.attr,
4162 &free_slowpath_attr.attr,
4163 &free_frozen_attr.attr,
4164 &free_add_partial_attr.attr,
4165 &free_remove_partial_attr.attr,
4166 &alloc_from_partial_attr.attr,
4167 &alloc_slab_attr.attr,
4168 &alloc_refill_attr.attr,
4169 &free_slab_attr.attr,
4170 &cpuslab_flush_attr.attr,
4171 &deactivate_full_attr.attr,
4172 &deactivate_empty_attr.attr,
4173 &deactivate_to_head_attr.attr,
4174 &deactivate_to_tail_attr.attr,
4175 &deactivate_remote_frees_attr.attr,
4180 static struct attribute_group slab_attr_group = {
4181 .attrs = slab_attrs,
4184 static ssize_t slab_attr_show(struct kobject *kobj,
4185 struct attribute *attr,
4188 struct slab_attribute *attribute;
4189 struct kmem_cache *s;
4192 attribute = to_slab_attr(attr);
4195 if (!attribute->show)
4198 err = attribute->show(s, buf);
4203 static ssize_t slab_attr_store(struct kobject *kobj,
4204 struct attribute *attr,
4205 const char *buf, size_t len)
4207 struct slab_attribute *attribute;
4208 struct kmem_cache *s;
4211 attribute = to_slab_attr(attr);
4214 if (!attribute->store)
4217 err = attribute->store(s, buf, len);
4222 static void kmem_cache_release(struct kobject *kobj)
4224 struct kmem_cache *s = to_slab(kobj);
4229 static struct sysfs_ops slab_sysfs_ops = {
4230 .show = slab_attr_show,
4231 .store = slab_attr_store,
4234 static struct kobj_type slab_ktype = {
4235 .sysfs_ops = &slab_sysfs_ops,
4236 .release = kmem_cache_release
4239 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4241 struct kobj_type *ktype = get_ktype(kobj);
4243 if (ktype == &slab_ktype)
4248 static struct kset_uevent_ops slab_uevent_ops = {
4249 .filter = uevent_filter,
4252 static struct kset *slab_kset;
4254 #define ID_STR_LENGTH 64
4256 /* Create a unique string id for a slab cache:
4258 * Format :[flags-]size
4260 static char *create_unique_id(struct kmem_cache *s)
4262 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4269 * First flags affecting slabcache operations. We will only
4270 * get here for aliasable slabs so we do not need to support
4271 * too many flags. The flags here must cover all flags that
4272 * are matched during merging to guarantee that the id is
4275 if (s->flags & SLAB_CACHE_DMA)
4277 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4279 if (s->flags & SLAB_DEBUG_FREE)
4283 p += sprintf(p, "%07d", s->size);
4284 BUG_ON(p > name + ID_STR_LENGTH - 1);
4288 static int sysfs_slab_add(struct kmem_cache *s)
4294 if (slab_state < SYSFS)
4295 /* Defer until later */
4298 unmergeable = slab_unmergeable(s);
4301 * Slabcache can never be merged so we can use the name proper.
4302 * This is typically the case for debug situations. In that
4303 * case we can catch duplicate names easily.
4305 sysfs_remove_link(&slab_kset->kobj, s->name);
4309 * Create a unique name for the slab as a target
4312 name = create_unique_id(s);
4315 s->kobj.kset = slab_kset;
4316 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4318 kobject_put(&s->kobj);
4322 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4325 kobject_uevent(&s->kobj, KOBJ_ADD);
4327 /* Setup first alias */
4328 sysfs_slab_alias(s, s->name);
4334 static void sysfs_slab_remove(struct kmem_cache *s)
4336 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4337 kobject_del(&s->kobj);
4338 kobject_put(&s->kobj);
4342 * Need to buffer aliases during bootup until sysfs becomes
4343 * available lest we loose that information.
4345 struct saved_alias {
4346 struct kmem_cache *s;
4348 struct saved_alias *next;
4351 static struct saved_alias *alias_list;
4353 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4355 struct saved_alias *al;
4357 if (slab_state == SYSFS) {
4359 * If we have a leftover link then remove it.
4361 sysfs_remove_link(&slab_kset->kobj, name);
4362 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4365 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4371 al->next = alias_list;
4376 static int __init slab_sysfs_init(void)
4378 struct kmem_cache *s;
4381 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4383 printk(KERN_ERR "Cannot register slab subsystem.\n");
4389 list_for_each_entry(s, &slab_caches, list) {
4390 err = sysfs_slab_add(s);
4392 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4393 " to sysfs\n", s->name);
4396 while (alias_list) {
4397 struct saved_alias *al = alias_list;
4399 alias_list = alias_list->next;
4400 err = sysfs_slab_alias(al->s, al->name);
4402 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4403 " %s to sysfs\n", s->name);
4411 __initcall(slab_sysfs_init);
4415 * The /proc/slabinfo ABI
4417 #ifdef CONFIG_SLABINFO
4419 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4420 size_t count, loff_t *ppos)
4426 static void print_slabinfo_header(struct seq_file *m)
4428 seq_puts(m, "slabinfo - version: 2.1\n");
4429 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4430 "<objperslab> <pagesperslab>");
4431 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4432 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4436 static void *s_start(struct seq_file *m, loff_t *pos)
4440 down_read(&slub_lock);
4442 print_slabinfo_header(m);
4444 return seq_list_start(&slab_caches, *pos);
4447 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4449 return seq_list_next(p, &slab_caches, pos);
4452 static void s_stop(struct seq_file *m, void *p)
4454 up_read(&slub_lock);
4457 static int s_show(struct seq_file *m, void *p)
4459 unsigned long nr_partials = 0;
4460 unsigned long nr_slabs = 0;
4461 unsigned long nr_inuse = 0;
4462 unsigned long nr_objs;
4463 struct kmem_cache *s;
4466 s = list_entry(p, struct kmem_cache, list);
4468 for_each_online_node(node) {
4469 struct kmem_cache_node *n = get_node(s, node);
4474 nr_partials += n->nr_partial;
4475 nr_slabs += atomic_long_read(&n->nr_slabs);
4476 nr_inuse += count_partial(n);
4479 nr_objs = nr_slabs * oo_objects(s->oo);
4480 nr_inuse += (nr_slabs - nr_partials) * oo_objects(s->oo);
4482 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4483 nr_objs, s->size, oo_objects(s->oo),
4484 (1 << oo_order(s->oo)));
4485 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4486 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4492 const struct seq_operations slabinfo_op = {
4499 #endif /* CONFIG_SLABINFO */