2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <trace/kmemtrace.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/mempolicy.h>
24 #include <linux/ctype.h>
25 #include <linux/debugobjects.h>
26 #include <linux/kallsyms.h>
27 #include <linux/memory.h>
28 #include <linux/math64.h>
29 #include <linux/fault-inject.h>
36 * The slab_lock protects operations on the object of a particular
37 * slab and its metadata in the page struct. If the slab lock
38 * has been taken then no allocations nor frees can be performed
39 * on the objects in the slab nor can the slab be added or removed
40 * from the partial or full lists since this would mean modifying
41 * the page_struct of the slab.
43 * The list_lock protects the partial and full list on each node and
44 * the partial slab counter. If taken then no new slabs may be added or
45 * removed from the lists nor make the number of partial slabs be modified.
46 * (Note that the total number of slabs is an atomic value that may be
47 * modified without taking the list lock).
49 * The list_lock is a centralized lock and thus we avoid taking it as
50 * much as possible. As long as SLUB does not have to handle partial
51 * slabs, operations can continue without any centralized lock. F.e.
52 * allocating a long series of objects that fill up slabs does not require
55 * The lock order is sometimes inverted when we are trying to get a slab
56 * off a list. We take the list_lock and then look for a page on the list
57 * to use. While we do that objects in the slabs may be freed. We can
58 * only operate on the slab if we have also taken the slab_lock. So we use
59 * a slab_trylock() on the slab. If trylock was successful then no frees
60 * can occur anymore and we can use the slab for allocations etc. If the
61 * slab_trylock() does not succeed then frees are in progress in the slab and
62 * we must stay away from it for a while since we may cause a bouncing
63 * cacheline if we try to acquire the lock. So go onto the next slab.
64 * If all pages are busy then we may allocate a new slab instead of reusing
65 * a partial slab. A new slab has noone operating on it and thus there is
66 * no danger of cacheline contention.
68 * Interrupts are disabled during allocation and deallocation in order to
69 * make the slab allocator safe to use in the context of an irq. In addition
70 * interrupts are disabled to ensure that the processor does not change
71 * while handling per_cpu slabs, due to kernel preemption.
73 * SLUB assigns one slab for allocation to each processor.
74 * Allocations only occur from these slabs called cpu slabs.
76 * Slabs with free elements are kept on a partial list and during regular
77 * operations no list for full slabs is used. If an object in a full slab is
78 * freed then the slab will show up again on the partial lists.
79 * We track full slabs for debugging purposes though because otherwise we
80 * cannot scan all objects.
82 * Slabs are freed when they become empty. Teardown and setup is
83 * minimal so we rely on the page allocators per cpu caches for
84 * fast frees and allocs.
86 * Overloading of page flags that are otherwise used for LRU management.
88 * PageActive The slab is frozen and exempt from list processing.
89 * This means that the slab is dedicated to a purpose
90 * such as satisfying allocations for a specific
91 * processor. Objects may be freed in the slab while
92 * it is frozen but slab_free will then skip the usual
93 * list operations. It is up to the processor holding
94 * the slab to integrate the slab into the slab lists
95 * when the slab is no longer needed.
97 * One use of this flag is to mark slabs that are
98 * used for allocations. Then such a slab becomes a cpu
99 * slab. The cpu slab may be equipped with an additional
100 * freelist that allows lockless access to
101 * free objects in addition to the regular freelist
102 * that requires the slab lock.
104 * PageError Slab requires special handling due to debug
105 * options set. This moves slab handling out of
106 * the fast path and disables lockless freelists.
109 #ifdef CONFIG_SLUB_DEBUG
116 * Issues still to be resolved:
118 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
120 * - Variable sizing of the per node arrays
123 /* Enable to test recovery from slab corruption on boot */
124 #undef SLUB_RESILIENCY_TEST
127 * Mininum number of partial slabs. These will be left on the partial
128 * lists even if they are empty. kmem_cache_shrink may reclaim them.
130 #define MIN_PARTIAL 5
133 * Maximum number of desirable partial slabs.
134 * The existence of more partial slabs makes kmem_cache_shrink
135 * sort the partial list by the number of objects in the.
137 #define MAX_PARTIAL 10
139 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
140 SLAB_POISON | SLAB_STORE_USER)
143 * Set of flags that will prevent slab merging
145 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
146 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
148 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
151 #ifndef ARCH_KMALLOC_MINALIGN
152 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
155 #ifndef ARCH_SLAB_MINALIGN
156 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
160 #define OO_MASK ((1 << OO_SHIFT) - 1)
161 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
163 /* Internal SLUB flags */
164 #define __OBJECT_POISON 0x80000000 /* Poison object */
165 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
167 static int kmem_size = sizeof(struct kmem_cache);
170 static struct notifier_block slab_notifier;
174 DOWN, /* No slab functionality available */
175 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
176 UP, /* Everything works but does not show up in sysfs */
180 /* A list of all slab caches on the system */
181 static DECLARE_RWSEM(slub_lock);
182 static LIST_HEAD(slab_caches);
185 * Tracking user of a slab.
188 unsigned long addr; /* Called from address */
189 int cpu; /* Was running on cpu */
190 int pid; /* Pid context */
191 unsigned long when; /* When did the operation occur */
194 enum track_item { TRACK_ALLOC, TRACK_FREE };
196 #ifdef CONFIG_SLUB_DEBUG
197 static int sysfs_slab_add(struct kmem_cache *);
198 static int sysfs_slab_alias(struct kmem_cache *, const char *);
199 static void sysfs_slab_remove(struct kmem_cache *);
202 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
203 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
205 static inline void sysfs_slab_remove(struct kmem_cache *s)
212 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
219 /********************************************************************
220 * Core slab cache functions
221 *******************************************************************/
223 int slab_is_available(void)
225 return slab_state >= UP;
228 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
231 return s->node[node];
233 return &s->local_node;
237 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
240 return s->cpu_slab[cpu];
246 /* Verify that a pointer has an address that is valid within a slab page */
247 static inline int check_valid_pointer(struct kmem_cache *s,
248 struct page *page, const void *object)
255 base = page_address(page);
256 if (object < base || object >= base + page->objects * s->size ||
257 (object - base) % s->size) {
265 * Slow version of get and set free pointer.
267 * This version requires touching the cache lines of kmem_cache which
268 * we avoid to do in the fast alloc free paths. There we obtain the offset
269 * from the page struct.
271 static inline void *get_freepointer(struct kmem_cache *s, void *object)
273 return *(void **)(object + s->offset);
276 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
278 *(void **)(object + s->offset) = fp;
281 /* Loop over all objects in a slab */
282 #define for_each_object(__p, __s, __addr, __objects) \
283 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
287 #define for_each_free_object(__p, __s, __free) \
288 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 return (p - addr) / s->size;
296 static inline struct kmem_cache_order_objects oo_make(int order,
299 struct kmem_cache_order_objects x = {
300 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
306 static inline int oo_order(struct kmem_cache_order_objects x)
308 return x.x >> OO_SHIFT;
311 static inline int oo_objects(struct kmem_cache_order_objects x)
313 return x.x & OO_MASK;
316 #ifdef CONFIG_SLUB_DEBUG
320 #ifdef CONFIG_SLUB_DEBUG_ON
321 static int slub_debug = DEBUG_DEFAULT_FLAGS;
323 static int slub_debug;
326 static char *slub_debug_slabs;
331 static void print_section(char *text, u8 *addr, unsigned int length)
339 for (i = 0; i < length; i++) {
341 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
344 printk(KERN_CONT " %02x", addr[i]);
346 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
348 printk(KERN_CONT " %s\n", ascii);
355 printk(KERN_CONT " ");
359 printk(KERN_CONT " %s\n", ascii);
363 static struct track *get_track(struct kmem_cache *s, void *object,
364 enum track_item alloc)
369 p = object + s->offset + sizeof(void *);
371 p = object + s->inuse;
376 static void set_track(struct kmem_cache *s, void *object,
377 enum track_item alloc, unsigned long addr)
379 struct track *p = get_track(s, object, alloc);
383 p->cpu = smp_processor_id();
384 p->pid = current->pid;
387 memset(p, 0, sizeof(struct track));
390 static void init_tracking(struct kmem_cache *s, void *object)
392 if (!(s->flags & SLAB_STORE_USER))
395 set_track(s, object, TRACK_FREE, 0UL);
396 set_track(s, object, TRACK_ALLOC, 0UL);
399 static void print_track(const char *s, struct track *t)
404 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
405 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
408 static void print_tracking(struct kmem_cache *s, void *object)
410 if (!(s->flags & SLAB_STORE_USER))
413 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
414 print_track("Freed", get_track(s, object, TRACK_FREE));
417 static void print_page_info(struct page *page)
419 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
420 page, page->objects, page->inuse, page->freelist, page->flags);
424 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
430 vsnprintf(buf, sizeof(buf), fmt, args);
432 printk(KERN_ERR "========================================"
433 "=====================================\n");
434 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
435 printk(KERN_ERR "----------------------------------------"
436 "-------------------------------------\n\n");
439 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
445 vsnprintf(buf, sizeof(buf), fmt, args);
447 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
450 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
452 unsigned int off; /* Offset of last byte */
453 u8 *addr = page_address(page);
455 print_tracking(s, p);
457 print_page_info(page);
459 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
460 p, p - addr, get_freepointer(s, p));
463 print_section("Bytes b4", p - 16, 16);
465 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
467 if (s->flags & SLAB_RED_ZONE)
468 print_section("Redzone", p + s->objsize,
469 s->inuse - s->objsize);
472 off = s->offset + sizeof(void *);
476 if (s->flags & SLAB_STORE_USER)
477 off += 2 * sizeof(struct track);
480 /* Beginning of the filler is the free pointer */
481 print_section("Padding", p + off, s->size - off);
486 static void object_err(struct kmem_cache *s, struct page *page,
487 u8 *object, char *reason)
489 slab_bug(s, "%s", reason);
490 print_trailer(s, page, object);
493 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
499 vsnprintf(buf, sizeof(buf), fmt, args);
501 slab_bug(s, "%s", buf);
502 print_page_info(page);
506 static void init_object(struct kmem_cache *s, void *object, int active)
510 if (s->flags & __OBJECT_POISON) {
511 memset(p, POISON_FREE, s->objsize - 1);
512 p[s->objsize - 1] = POISON_END;
515 if (s->flags & SLAB_RED_ZONE)
516 memset(p + s->objsize,
517 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
518 s->inuse - s->objsize);
521 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
524 if (*start != (u8)value)
532 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
533 void *from, void *to)
535 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
536 memset(from, data, to - from);
539 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
540 u8 *object, char *what,
541 u8 *start, unsigned int value, unsigned int bytes)
546 fault = check_bytes(start, value, bytes);
551 while (end > fault && end[-1] == value)
554 slab_bug(s, "%s overwritten", what);
555 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
556 fault, end - 1, fault[0], value);
557 print_trailer(s, page, object);
559 restore_bytes(s, what, value, fault, end);
567 * Bytes of the object to be managed.
568 * If the freepointer may overlay the object then the free
569 * pointer is the first word of the object.
571 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
574 * object + s->objsize
575 * Padding to reach word boundary. This is also used for Redzoning.
576 * Padding is extended by another word if Redzoning is enabled and
579 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
580 * 0xcc (RED_ACTIVE) for objects in use.
583 * Meta data starts here.
585 * A. Free pointer (if we cannot overwrite object on free)
586 * B. Tracking data for SLAB_STORE_USER
587 * C. Padding to reach required alignment boundary or at mininum
588 * one word if debugging is on to be able to detect writes
589 * before the word boundary.
591 * Padding is done using 0x5a (POISON_INUSE)
594 * Nothing is used beyond s->size.
596 * If slabcaches are merged then the objsize and inuse boundaries are mostly
597 * ignored. And therefore no slab options that rely on these boundaries
598 * may be used with merged slabcaches.
601 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
603 unsigned long off = s->inuse; /* The end of info */
606 /* Freepointer is placed after the object. */
607 off += sizeof(void *);
609 if (s->flags & SLAB_STORE_USER)
610 /* We also have user information there */
611 off += 2 * sizeof(struct track);
616 return check_bytes_and_report(s, page, p, "Object padding",
617 p + off, POISON_INUSE, s->size - off);
620 /* Check the pad bytes at the end of a slab page */
621 static int slab_pad_check(struct kmem_cache *s, struct page *page)
629 if (!(s->flags & SLAB_POISON))
632 start = page_address(page);
633 length = (PAGE_SIZE << compound_order(page));
634 end = start + length;
635 remainder = length % s->size;
639 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
642 while (end > fault && end[-1] == POISON_INUSE)
645 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
646 print_section("Padding", end - remainder, remainder);
648 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
652 static int check_object(struct kmem_cache *s, struct page *page,
653 void *object, int active)
656 u8 *endobject = object + s->objsize;
658 if (s->flags & SLAB_RED_ZONE) {
660 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
662 if (!check_bytes_and_report(s, page, object, "Redzone",
663 endobject, red, s->inuse - s->objsize))
666 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
667 check_bytes_and_report(s, page, p, "Alignment padding",
668 endobject, POISON_INUSE, s->inuse - s->objsize);
672 if (s->flags & SLAB_POISON) {
673 if (!active && (s->flags & __OBJECT_POISON) &&
674 (!check_bytes_and_report(s, page, p, "Poison", p,
675 POISON_FREE, s->objsize - 1) ||
676 !check_bytes_and_report(s, page, p, "Poison",
677 p + s->objsize - 1, POISON_END, 1)))
680 * check_pad_bytes cleans up on its own.
682 check_pad_bytes(s, page, p);
685 if (!s->offset && active)
687 * Object and freepointer overlap. Cannot check
688 * freepointer while object is allocated.
692 /* Check free pointer validity */
693 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
694 object_err(s, page, p, "Freepointer corrupt");
696 * No choice but to zap it and thus lose the remainder
697 * of the free objects in this slab. May cause
698 * another error because the object count is now wrong.
700 set_freepointer(s, p, NULL);
706 static int check_slab(struct kmem_cache *s, struct page *page)
710 VM_BUG_ON(!irqs_disabled());
712 if (!PageSlab(page)) {
713 slab_err(s, page, "Not a valid slab page");
717 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
718 if (page->objects > maxobj) {
719 slab_err(s, page, "objects %u > max %u",
720 s->name, page->objects, maxobj);
723 if (page->inuse > page->objects) {
724 slab_err(s, page, "inuse %u > max %u",
725 s->name, page->inuse, page->objects);
728 /* Slab_pad_check fixes things up after itself */
729 slab_pad_check(s, page);
734 * Determine if a certain object on a page is on the freelist. Must hold the
735 * slab lock to guarantee that the chains are in a consistent state.
737 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
740 void *fp = page->freelist;
742 unsigned long max_objects;
744 while (fp && nr <= page->objects) {
747 if (!check_valid_pointer(s, page, fp)) {
749 object_err(s, page, object,
750 "Freechain corrupt");
751 set_freepointer(s, object, NULL);
754 slab_err(s, page, "Freepointer corrupt");
755 page->freelist = NULL;
756 page->inuse = page->objects;
757 slab_fix(s, "Freelist cleared");
763 fp = get_freepointer(s, object);
767 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
768 if (max_objects > MAX_OBJS_PER_PAGE)
769 max_objects = MAX_OBJS_PER_PAGE;
771 if (page->objects != max_objects) {
772 slab_err(s, page, "Wrong number of objects. Found %d but "
773 "should be %d", page->objects, max_objects);
774 page->objects = max_objects;
775 slab_fix(s, "Number of objects adjusted.");
777 if (page->inuse != page->objects - nr) {
778 slab_err(s, page, "Wrong object count. Counter is %d but "
779 "counted were %d", page->inuse, page->objects - nr);
780 page->inuse = page->objects - nr;
781 slab_fix(s, "Object count adjusted.");
783 return search == NULL;
786 static void trace(struct kmem_cache *s, struct page *page, void *object,
789 if (s->flags & SLAB_TRACE) {
790 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
792 alloc ? "alloc" : "free",
797 print_section("Object", (void *)object, s->objsize);
804 * Tracking of fully allocated slabs for debugging purposes.
806 static void add_full(struct kmem_cache_node *n, struct page *page)
808 spin_lock(&n->list_lock);
809 list_add(&page->lru, &n->full);
810 spin_unlock(&n->list_lock);
813 static void remove_full(struct kmem_cache *s, struct page *page)
815 struct kmem_cache_node *n;
817 if (!(s->flags & SLAB_STORE_USER))
820 n = get_node(s, page_to_nid(page));
822 spin_lock(&n->list_lock);
823 list_del(&page->lru);
824 spin_unlock(&n->list_lock);
827 /* Tracking of the number of slabs for debugging purposes */
828 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
830 struct kmem_cache_node *n = get_node(s, node);
832 return atomic_long_read(&n->nr_slabs);
835 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
837 struct kmem_cache_node *n = get_node(s, node);
840 * May be called early in order to allocate a slab for the
841 * kmem_cache_node structure. Solve the chicken-egg
842 * dilemma by deferring the increment of the count during
843 * bootstrap (see early_kmem_cache_node_alloc).
845 if (!NUMA_BUILD || n) {
846 atomic_long_inc(&n->nr_slabs);
847 atomic_long_add(objects, &n->total_objects);
850 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
852 struct kmem_cache_node *n = get_node(s, node);
854 atomic_long_dec(&n->nr_slabs);
855 atomic_long_sub(objects, &n->total_objects);
858 /* Object debug checks for alloc/free paths */
859 static void setup_object_debug(struct kmem_cache *s, struct page *page,
862 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
865 init_object(s, object, 0);
866 init_tracking(s, object);
869 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
870 void *object, unsigned long addr)
872 if (!check_slab(s, page))
875 if (!on_freelist(s, page, object)) {
876 object_err(s, page, object, "Object already allocated");
880 if (!check_valid_pointer(s, page, object)) {
881 object_err(s, page, object, "Freelist Pointer check fails");
885 if (!check_object(s, page, object, 0))
888 /* Success perform special debug activities for allocs */
889 if (s->flags & SLAB_STORE_USER)
890 set_track(s, object, TRACK_ALLOC, addr);
891 trace(s, page, object, 1);
892 init_object(s, object, 1);
896 if (PageSlab(page)) {
898 * If this is a slab page then lets do the best we can
899 * to avoid issues in the future. Marking all objects
900 * as used avoids touching the remaining objects.
902 slab_fix(s, "Marking all objects used");
903 page->inuse = page->objects;
904 page->freelist = NULL;
909 static int free_debug_processing(struct kmem_cache *s, struct page *page,
910 void *object, unsigned long addr)
912 if (!check_slab(s, page))
915 if (!check_valid_pointer(s, page, object)) {
916 slab_err(s, page, "Invalid object pointer 0x%p", object);
920 if (on_freelist(s, page, object)) {
921 object_err(s, page, object, "Object already free");
925 if (!check_object(s, page, object, 1))
928 if (unlikely(s != page->slab)) {
929 if (!PageSlab(page)) {
930 slab_err(s, page, "Attempt to free object(0x%p) "
931 "outside of slab", object);
932 } else if (!page->slab) {
934 "SLUB <none>: no slab for object 0x%p.\n",
938 object_err(s, page, object,
939 "page slab pointer corrupt.");
943 /* Special debug activities for freeing objects */
944 if (!PageSlubFrozen(page) && !page->freelist)
945 remove_full(s, page);
946 if (s->flags & SLAB_STORE_USER)
947 set_track(s, object, TRACK_FREE, addr);
948 trace(s, page, object, 0);
949 init_object(s, object, 0);
953 slab_fix(s, "Object at 0x%p not freed", object);
957 static int __init setup_slub_debug(char *str)
959 slub_debug = DEBUG_DEFAULT_FLAGS;
960 if (*str++ != '=' || !*str)
962 * No options specified. Switch on full debugging.
968 * No options but restriction on slabs. This means full
969 * debugging for slabs matching a pattern.
976 * Switch off all debugging measures.
981 * Determine which debug features should be switched on
983 for (; *str && *str != ','; str++) {
984 switch (tolower(*str)) {
986 slub_debug |= SLAB_DEBUG_FREE;
989 slub_debug |= SLAB_RED_ZONE;
992 slub_debug |= SLAB_POISON;
995 slub_debug |= SLAB_STORE_USER;
998 slub_debug |= SLAB_TRACE;
1001 printk(KERN_ERR "slub_debug option '%c' "
1002 "unknown. skipped\n", *str);
1008 slub_debug_slabs = str + 1;
1013 __setup("slub_debug", setup_slub_debug);
1015 static unsigned long kmem_cache_flags(unsigned long objsize,
1016 unsigned long flags, const char *name,
1017 void (*ctor)(void *))
1020 * Enable debugging if selected on the kernel commandline.
1022 if (slub_debug && (!slub_debug_slabs ||
1023 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1024 flags |= slub_debug;
1029 static inline void setup_object_debug(struct kmem_cache *s,
1030 struct page *page, void *object) {}
1032 static inline int alloc_debug_processing(struct kmem_cache *s,
1033 struct page *page, void *object, unsigned long addr) { return 0; }
1035 static inline int free_debug_processing(struct kmem_cache *s,
1036 struct page *page, void *object, unsigned long addr) { return 0; }
1038 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1040 static inline int check_object(struct kmem_cache *s, struct page *page,
1041 void *object, int active) { return 1; }
1042 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1043 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1044 unsigned long flags, const char *name,
1045 void (*ctor)(void *))
1049 #define slub_debug 0
1051 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1053 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1055 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1060 * Slab allocation and freeing
1062 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1063 struct kmem_cache_order_objects oo)
1065 int order = oo_order(oo);
1068 return alloc_pages(flags, order);
1070 return alloc_pages_node(node, flags, order);
1073 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1076 struct kmem_cache_order_objects oo = s->oo;
1078 flags |= s->allocflags;
1080 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1082 if (unlikely(!page)) {
1085 * Allocation may have failed due to fragmentation.
1086 * Try a lower order alloc if possible
1088 page = alloc_slab_page(flags, node, oo);
1092 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1094 page->objects = oo_objects(oo);
1095 mod_zone_page_state(page_zone(page),
1096 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1097 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1103 static void setup_object(struct kmem_cache *s, struct page *page,
1106 setup_object_debug(s, page, object);
1107 if (unlikely(s->ctor))
1111 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1118 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1120 page = allocate_slab(s,
1121 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1125 inc_slabs_node(s, page_to_nid(page), page->objects);
1127 page->flags |= 1 << PG_slab;
1128 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1129 SLAB_STORE_USER | SLAB_TRACE))
1130 __SetPageSlubDebug(page);
1132 start = page_address(page);
1134 if (unlikely(s->flags & SLAB_POISON))
1135 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1138 for_each_object(p, s, start, page->objects) {
1139 setup_object(s, page, last);
1140 set_freepointer(s, last, p);
1143 setup_object(s, page, last);
1144 set_freepointer(s, last, NULL);
1146 page->freelist = start;
1152 static void __free_slab(struct kmem_cache *s, struct page *page)
1154 int order = compound_order(page);
1155 int pages = 1 << order;
1157 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1160 slab_pad_check(s, page);
1161 for_each_object(p, s, page_address(page),
1163 check_object(s, page, p, 0);
1164 __ClearPageSlubDebug(page);
1167 mod_zone_page_state(page_zone(page),
1168 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1169 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1172 __ClearPageSlab(page);
1173 reset_page_mapcount(page);
1174 if (current->reclaim_state)
1175 current->reclaim_state->reclaimed_slab += pages;
1176 __free_pages(page, order);
1179 static void rcu_free_slab(struct rcu_head *h)
1183 page = container_of((struct list_head *)h, struct page, lru);
1184 __free_slab(page->slab, page);
1187 static void free_slab(struct kmem_cache *s, struct page *page)
1189 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1191 * RCU free overloads the RCU head over the LRU
1193 struct rcu_head *head = (void *)&page->lru;
1195 call_rcu(head, rcu_free_slab);
1197 __free_slab(s, page);
1200 static void discard_slab(struct kmem_cache *s, struct page *page)
1202 dec_slabs_node(s, page_to_nid(page), page->objects);
1207 * Per slab locking using the pagelock
1209 static __always_inline void slab_lock(struct page *page)
1211 bit_spin_lock(PG_locked, &page->flags);
1214 static __always_inline void slab_unlock(struct page *page)
1216 __bit_spin_unlock(PG_locked, &page->flags);
1219 static __always_inline int slab_trylock(struct page *page)
1223 rc = bit_spin_trylock(PG_locked, &page->flags);
1228 * Management of partially allocated slabs
1230 static void add_partial(struct kmem_cache_node *n,
1231 struct page *page, int tail)
1233 spin_lock(&n->list_lock);
1236 list_add_tail(&page->lru, &n->partial);
1238 list_add(&page->lru, &n->partial);
1239 spin_unlock(&n->list_lock);
1242 static void remove_partial(struct kmem_cache *s, struct page *page)
1244 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1246 spin_lock(&n->list_lock);
1247 list_del(&page->lru);
1249 spin_unlock(&n->list_lock);
1253 * Lock slab and remove from the partial list.
1255 * Must hold list_lock.
1257 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1260 if (slab_trylock(page)) {
1261 list_del(&page->lru);
1263 __SetPageSlubFrozen(page);
1270 * Try to allocate a partial slab from a specific node.
1272 static struct page *get_partial_node(struct kmem_cache_node *n)
1277 * Racy check. If we mistakenly see no partial slabs then we
1278 * just allocate an empty slab. If we mistakenly try to get a
1279 * partial slab and there is none available then get_partials()
1282 if (!n || !n->nr_partial)
1285 spin_lock(&n->list_lock);
1286 list_for_each_entry(page, &n->partial, lru)
1287 if (lock_and_freeze_slab(n, page))
1291 spin_unlock(&n->list_lock);
1296 * Get a page from somewhere. Search in increasing NUMA distances.
1298 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1301 struct zonelist *zonelist;
1304 enum zone_type high_zoneidx = gfp_zone(flags);
1308 * The defrag ratio allows a configuration of the tradeoffs between
1309 * inter node defragmentation and node local allocations. A lower
1310 * defrag_ratio increases the tendency to do local allocations
1311 * instead of attempting to obtain partial slabs from other nodes.
1313 * If the defrag_ratio is set to 0 then kmalloc() always
1314 * returns node local objects. If the ratio is higher then kmalloc()
1315 * may return off node objects because partial slabs are obtained
1316 * from other nodes and filled up.
1318 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1319 * defrag_ratio = 1000) then every (well almost) allocation will
1320 * first attempt to defrag slab caches on other nodes. This means
1321 * scanning over all nodes to look for partial slabs which may be
1322 * expensive if we do it every time we are trying to find a slab
1323 * with available objects.
1325 if (!s->remote_node_defrag_ratio ||
1326 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1329 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1330 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1331 struct kmem_cache_node *n;
1333 n = get_node(s, zone_to_nid(zone));
1335 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1336 n->nr_partial > s->min_partial) {
1337 page = get_partial_node(n);
1347 * Get a partial page, lock it and return it.
1349 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1352 int searchnode = (node == -1) ? numa_node_id() : node;
1354 page = get_partial_node(get_node(s, searchnode));
1355 if (page || (flags & __GFP_THISNODE))
1358 return get_any_partial(s, flags);
1362 * Move a page back to the lists.
1364 * Must be called with the slab lock held.
1366 * On exit the slab lock will have been dropped.
1368 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1370 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1371 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1373 __ClearPageSlubFrozen(page);
1376 if (page->freelist) {
1377 add_partial(n, page, tail);
1378 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1380 stat(c, DEACTIVATE_FULL);
1381 if (SLABDEBUG && PageSlubDebug(page) &&
1382 (s->flags & SLAB_STORE_USER))
1387 stat(c, DEACTIVATE_EMPTY);
1388 if (n->nr_partial < s->min_partial) {
1390 * Adding an empty slab to the partial slabs in order
1391 * to avoid page allocator overhead. This slab needs
1392 * to come after the other slabs with objects in
1393 * so that the others get filled first. That way the
1394 * size of the partial list stays small.
1396 * kmem_cache_shrink can reclaim any empty slabs from
1399 add_partial(n, page, 1);
1403 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1404 discard_slab(s, page);
1410 * Remove the cpu slab
1412 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1414 struct page *page = c->page;
1418 stat(c, DEACTIVATE_REMOTE_FREES);
1420 * Merge cpu freelist into slab freelist. Typically we get here
1421 * because both freelists are empty. So this is unlikely
1424 while (unlikely(c->freelist)) {
1427 tail = 0; /* Hot objects. Put the slab first */
1429 /* Retrieve object from cpu_freelist */
1430 object = c->freelist;
1431 c->freelist = c->freelist[c->offset];
1433 /* And put onto the regular freelist */
1434 object[c->offset] = page->freelist;
1435 page->freelist = object;
1439 unfreeze_slab(s, page, tail);
1442 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1444 stat(c, CPUSLAB_FLUSH);
1446 deactivate_slab(s, c);
1452 * Called from IPI handler with interrupts disabled.
1454 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1456 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1458 if (likely(c && c->page))
1462 static void flush_cpu_slab(void *d)
1464 struct kmem_cache *s = d;
1466 __flush_cpu_slab(s, smp_processor_id());
1469 static void flush_all(struct kmem_cache *s)
1471 on_each_cpu(flush_cpu_slab, s, 1);
1475 * Check if the objects in a per cpu structure fit numa
1476 * locality expectations.
1478 static inline int node_match(struct kmem_cache_cpu *c, int node)
1481 if (node != -1 && c->node != node)
1488 * Slow path. The lockless freelist is empty or we need to perform
1491 * Interrupts are disabled.
1493 * Processing is still very fast if new objects have been freed to the
1494 * regular freelist. In that case we simply take over the regular freelist
1495 * as the lockless freelist and zap the regular freelist.
1497 * If that is not working then we fall back to the partial lists. We take the
1498 * first element of the freelist as the object to allocate now and move the
1499 * rest of the freelist to the lockless freelist.
1501 * And if we were unable to get a new slab from the partial slab lists then
1502 * we need to allocate a new slab. This is the slowest path since it involves
1503 * a call to the page allocator and the setup of a new slab.
1505 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1506 unsigned long addr, struct kmem_cache_cpu *c)
1511 /* We handle __GFP_ZERO in the caller */
1512 gfpflags &= ~__GFP_ZERO;
1518 if (unlikely(!node_match(c, node)))
1521 stat(c, ALLOC_REFILL);
1524 object = c->page->freelist;
1525 if (unlikely(!object))
1527 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1530 c->freelist = object[c->offset];
1531 c->page->inuse = c->page->objects;
1532 c->page->freelist = NULL;
1533 c->node = page_to_nid(c->page);
1535 slab_unlock(c->page);
1536 stat(c, ALLOC_SLOWPATH);
1540 deactivate_slab(s, c);
1543 new = get_partial(s, gfpflags, node);
1546 stat(c, ALLOC_FROM_PARTIAL);
1550 if (gfpflags & __GFP_WAIT)
1553 new = new_slab(s, gfpflags, node);
1555 if (gfpflags & __GFP_WAIT)
1556 local_irq_disable();
1559 c = get_cpu_slab(s, smp_processor_id());
1560 stat(c, ALLOC_SLAB);
1564 __SetPageSlubFrozen(new);
1570 if (!alloc_debug_processing(s, c->page, object, addr))
1574 c->page->freelist = object[c->offset];
1580 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1581 * have the fastpath folded into their functions. So no function call
1582 * overhead for requests that can be satisfied on the fastpath.
1584 * The fastpath works by first checking if the lockless freelist can be used.
1585 * If not then __slab_alloc is called for slow processing.
1587 * Otherwise we can simply pick the next object from the lockless free list.
1589 static __always_inline void *slab_alloc(struct kmem_cache *s,
1590 gfp_t gfpflags, int node, unsigned long addr)
1593 struct kmem_cache_cpu *c;
1594 unsigned long flags;
1595 unsigned int objsize;
1597 lockdep_trace_alloc(gfpflags);
1598 might_sleep_if(gfpflags & __GFP_WAIT);
1600 if (should_failslab(s->objsize, gfpflags))
1603 local_irq_save(flags);
1604 c = get_cpu_slab(s, smp_processor_id());
1605 objsize = c->objsize;
1606 if (unlikely(!c->freelist || !node_match(c, node)))
1608 object = __slab_alloc(s, gfpflags, node, addr, c);
1611 object = c->freelist;
1612 c->freelist = object[c->offset];
1613 stat(c, ALLOC_FASTPATH);
1615 local_irq_restore(flags);
1617 if (unlikely((gfpflags & __GFP_ZERO) && object))
1618 memset(object, 0, objsize);
1623 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1625 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1627 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1631 EXPORT_SYMBOL(kmem_cache_alloc);
1633 #ifdef CONFIG_KMEMTRACE
1634 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1636 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1638 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1642 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1644 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1646 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1647 s->objsize, s->size, gfpflags, node);
1651 EXPORT_SYMBOL(kmem_cache_alloc_node);
1654 #ifdef CONFIG_KMEMTRACE
1655 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1659 return slab_alloc(s, gfpflags, node, _RET_IP_);
1661 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1665 * Slow patch handling. This may still be called frequently since objects
1666 * have a longer lifetime than the cpu slabs in most processing loads.
1668 * So we still attempt to reduce cache line usage. Just take the slab
1669 * lock and free the item. If there is no additional partial page
1670 * handling required then we can return immediately.
1672 static void __slab_free(struct kmem_cache *s, struct page *page,
1673 void *x, unsigned long addr, unsigned int offset)
1676 void **object = (void *)x;
1677 struct kmem_cache_cpu *c;
1679 c = get_cpu_slab(s, raw_smp_processor_id());
1680 stat(c, FREE_SLOWPATH);
1683 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1687 prior = object[offset] = page->freelist;
1688 page->freelist = object;
1691 if (unlikely(PageSlubFrozen(page))) {
1692 stat(c, FREE_FROZEN);
1696 if (unlikely(!page->inuse))
1700 * Objects left in the slab. If it was not on the partial list before
1703 if (unlikely(!prior)) {
1704 add_partial(get_node(s, page_to_nid(page)), page, 1);
1705 stat(c, FREE_ADD_PARTIAL);
1715 * Slab still on the partial list.
1717 remove_partial(s, page);
1718 stat(c, FREE_REMOVE_PARTIAL);
1722 discard_slab(s, page);
1726 if (!free_debug_processing(s, page, x, addr))
1732 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1733 * can perform fastpath freeing without additional function calls.
1735 * The fastpath is only possible if we are freeing to the current cpu slab
1736 * of this processor. This typically the case if we have just allocated
1739 * If fastpath is not possible then fall back to __slab_free where we deal
1740 * with all sorts of special processing.
1742 static __always_inline void slab_free(struct kmem_cache *s,
1743 struct page *page, void *x, unsigned long addr)
1745 void **object = (void *)x;
1746 struct kmem_cache_cpu *c;
1747 unsigned long flags;
1749 local_irq_save(flags);
1750 c = get_cpu_slab(s, smp_processor_id());
1751 debug_check_no_locks_freed(object, c->objsize);
1752 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1753 debug_check_no_obj_freed(object, c->objsize);
1754 if (likely(page == c->page && c->node >= 0)) {
1755 object[c->offset] = c->freelist;
1756 c->freelist = object;
1757 stat(c, FREE_FASTPATH);
1759 __slab_free(s, page, x, addr, c->offset);
1761 local_irq_restore(flags);
1764 void kmem_cache_free(struct kmem_cache *s, void *x)
1768 page = virt_to_head_page(x);
1770 slab_free(s, page, x, _RET_IP_);
1772 trace_kmem_cache_free(_RET_IP_, x);
1774 EXPORT_SYMBOL(kmem_cache_free);
1776 /* Figure out on which slab page the object resides */
1777 static struct page *get_object_page(const void *x)
1779 struct page *page = virt_to_head_page(x);
1781 if (!PageSlab(page))
1788 * Object placement in a slab is made very easy because we always start at
1789 * offset 0. If we tune the size of the object to the alignment then we can
1790 * get the required alignment by putting one properly sized object after
1793 * Notice that the allocation order determines the sizes of the per cpu
1794 * caches. Each processor has always one slab available for allocations.
1795 * Increasing the allocation order reduces the number of times that slabs
1796 * must be moved on and off the partial lists and is therefore a factor in
1801 * Mininum / Maximum order of slab pages. This influences locking overhead
1802 * and slab fragmentation. A higher order reduces the number of partial slabs
1803 * and increases the number of allocations possible without having to
1804 * take the list_lock.
1806 static int slub_min_order;
1807 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1808 static int slub_min_objects;
1811 * Merge control. If this is set then no merging of slab caches will occur.
1812 * (Could be removed. This was introduced to pacify the merge skeptics.)
1814 static int slub_nomerge;
1817 * Calculate the order of allocation given an slab object size.
1819 * The order of allocation has significant impact on performance and other
1820 * system components. Generally order 0 allocations should be preferred since
1821 * order 0 does not cause fragmentation in the page allocator. Larger objects
1822 * be problematic to put into order 0 slabs because there may be too much
1823 * unused space left. We go to a higher order if more than 1/16th of the slab
1826 * In order to reach satisfactory performance we must ensure that a minimum
1827 * number of objects is in one slab. Otherwise we may generate too much
1828 * activity on the partial lists which requires taking the list_lock. This is
1829 * less a concern for large slabs though which are rarely used.
1831 * slub_max_order specifies the order where we begin to stop considering the
1832 * number of objects in a slab as critical. If we reach slub_max_order then
1833 * we try to keep the page order as low as possible. So we accept more waste
1834 * of space in favor of a small page order.
1836 * Higher order allocations also allow the placement of more objects in a
1837 * slab and thereby reduce object handling overhead. If the user has
1838 * requested a higher mininum order then we start with that one instead of
1839 * the smallest order which will fit the object.
1841 static inline int slab_order(int size, int min_objects,
1842 int max_order, int fract_leftover)
1846 int min_order = slub_min_order;
1848 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1849 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1851 for (order = max(min_order,
1852 fls(min_objects * size - 1) - PAGE_SHIFT);
1853 order <= max_order; order++) {
1855 unsigned long slab_size = PAGE_SIZE << order;
1857 if (slab_size < min_objects * size)
1860 rem = slab_size % size;
1862 if (rem <= slab_size / fract_leftover)
1870 static inline int calculate_order(int size)
1878 * Attempt to find best configuration for a slab. This
1879 * works by first attempting to generate a layout with
1880 * the best configuration and backing off gradually.
1882 * First we reduce the acceptable waste in a slab. Then
1883 * we reduce the minimum objects required in a slab.
1885 min_objects = slub_min_objects;
1887 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1888 max_objects = (PAGE_SIZE << slub_max_order)/size;
1889 min_objects = min(min_objects, max_objects);
1891 while (min_objects > 1) {
1893 while (fraction >= 4) {
1894 order = slab_order(size, min_objects,
1895 slub_max_order, fraction);
1896 if (order <= slub_max_order)
1904 * We were unable to place multiple objects in a slab. Now
1905 * lets see if we can place a single object there.
1907 order = slab_order(size, 1, slub_max_order, 1);
1908 if (order <= slub_max_order)
1912 * Doh this slab cannot be placed using slub_max_order.
1914 order = slab_order(size, 1, MAX_ORDER, 1);
1915 if (order < MAX_ORDER)
1921 * Figure out what the alignment of the objects will be.
1923 static unsigned long calculate_alignment(unsigned long flags,
1924 unsigned long align, unsigned long size)
1927 * If the user wants hardware cache aligned objects then follow that
1928 * suggestion if the object is sufficiently large.
1930 * The hardware cache alignment cannot override the specified
1931 * alignment though. If that is greater then use it.
1933 if (flags & SLAB_HWCACHE_ALIGN) {
1934 unsigned long ralign = cache_line_size();
1935 while (size <= ralign / 2)
1937 align = max(align, ralign);
1940 if (align < ARCH_SLAB_MINALIGN)
1941 align = ARCH_SLAB_MINALIGN;
1943 return ALIGN(align, sizeof(void *));
1946 static void init_kmem_cache_cpu(struct kmem_cache *s,
1947 struct kmem_cache_cpu *c)
1952 c->offset = s->offset / sizeof(void *);
1953 c->objsize = s->objsize;
1954 #ifdef CONFIG_SLUB_STATS
1955 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1960 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1963 spin_lock_init(&n->list_lock);
1964 INIT_LIST_HEAD(&n->partial);
1965 #ifdef CONFIG_SLUB_DEBUG
1966 atomic_long_set(&n->nr_slabs, 0);
1967 atomic_long_set(&n->total_objects, 0);
1968 INIT_LIST_HEAD(&n->full);
1974 * Per cpu array for per cpu structures.
1976 * The per cpu array places all kmem_cache_cpu structures from one processor
1977 * close together meaning that it becomes possible that multiple per cpu
1978 * structures are contained in one cacheline. This may be particularly
1979 * beneficial for the kmalloc caches.
1981 * A desktop system typically has around 60-80 slabs. With 100 here we are
1982 * likely able to get per cpu structures for all caches from the array defined
1983 * here. We must be able to cover all kmalloc caches during bootstrap.
1985 * If the per cpu array is exhausted then fall back to kmalloc
1986 * of individual cachelines. No sharing is possible then.
1988 #define NR_KMEM_CACHE_CPU 100
1990 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1991 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1993 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1994 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
1996 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1997 int cpu, gfp_t flags)
1999 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2002 per_cpu(kmem_cache_cpu_free, cpu) =
2003 (void *)c->freelist;
2005 /* Table overflow: So allocate ourselves */
2007 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2008 flags, cpu_to_node(cpu));
2013 init_kmem_cache_cpu(s, c);
2017 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2019 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2020 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2024 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2025 per_cpu(kmem_cache_cpu_free, cpu) = c;
2028 static void free_kmem_cache_cpus(struct kmem_cache *s)
2032 for_each_online_cpu(cpu) {
2033 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2036 s->cpu_slab[cpu] = NULL;
2037 free_kmem_cache_cpu(c, cpu);
2042 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2046 for_each_online_cpu(cpu) {
2047 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2052 c = alloc_kmem_cache_cpu(s, cpu, flags);
2054 free_kmem_cache_cpus(s);
2057 s->cpu_slab[cpu] = c;
2063 * Initialize the per cpu array.
2065 static void init_alloc_cpu_cpu(int cpu)
2069 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2072 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2073 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2075 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2078 static void __init init_alloc_cpu(void)
2082 for_each_online_cpu(cpu)
2083 init_alloc_cpu_cpu(cpu);
2087 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2088 static inline void init_alloc_cpu(void) {}
2090 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2092 init_kmem_cache_cpu(s, &s->cpu_slab);
2099 * No kmalloc_node yet so do it by hand. We know that this is the first
2100 * slab on the node for this slabcache. There are no concurrent accesses
2103 * Note that this function only works on the kmalloc_node_cache
2104 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2105 * memory on a fresh node that has no slab structures yet.
2107 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2110 struct kmem_cache_node *n;
2111 unsigned long flags;
2113 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2115 page = new_slab(kmalloc_caches, gfpflags, node);
2118 if (page_to_nid(page) != node) {
2119 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2121 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2122 "in order to be able to continue\n");
2127 page->freelist = get_freepointer(kmalloc_caches, n);
2129 kmalloc_caches->node[node] = n;
2130 #ifdef CONFIG_SLUB_DEBUG
2131 init_object(kmalloc_caches, n, 1);
2132 init_tracking(kmalloc_caches, n);
2134 init_kmem_cache_node(n, kmalloc_caches);
2135 inc_slabs_node(kmalloc_caches, node, page->objects);
2138 * lockdep requires consistent irq usage for each lock
2139 * so even though there cannot be a race this early in
2140 * the boot sequence, we still disable irqs.
2142 local_irq_save(flags);
2143 add_partial(n, page, 0);
2144 local_irq_restore(flags);
2147 static void free_kmem_cache_nodes(struct kmem_cache *s)
2151 for_each_node_state(node, N_NORMAL_MEMORY) {
2152 struct kmem_cache_node *n = s->node[node];
2153 if (n && n != &s->local_node)
2154 kmem_cache_free(kmalloc_caches, n);
2155 s->node[node] = NULL;
2159 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2164 if (slab_state >= UP)
2165 local_node = page_to_nid(virt_to_page(s));
2169 for_each_node_state(node, N_NORMAL_MEMORY) {
2170 struct kmem_cache_node *n;
2172 if (local_node == node)
2175 if (slab_state == DOWN) {
2176 early_kmem_cache_node_alloc(gfpflags, node);
2179 n = kmem_cache_alloc_node(kmalloc_caches,
2183 free_kmem_cache_nodes(s);
2189 init_kmem_cache_node(n, s);
2194 static void free_kmem_cache_nodes(struct kmem_cache *s)
2198 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2200 init_kmem_cache_node(&s->local_node, s);
2205 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2207 if (min < MIN_PARTIAL)
2209 else if (min > MAX_PARTIAL)
2211 s->min_partial = min;
2215 * calculate_sizes() determines the order and the distribution of data within
2218 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2220 unsigned long flags = s->flags;
2221 unsigned long size = s->objsize;
2222 unsigned long align = s->align;
2226 * Round up object size to the next word boundary. We can only
2227 * place the free pointer at word boundaries and this determines
2228 * the possible location of the free pointer.
2230 size = ALIGN(size, sizeof(void *));
2232 #ifdef CONFIG_SLUB_DEBUG
2234 * Determine if we can poison the object itself. If the user of
2235 * the slab may touch the object after free or before allocation
2236 * then we should never poison the object itself.
2238 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2240 s->flags |= __OBJECT_POISON;
2242 s->flags &= ~__OBJECT_POISON;
2246 * If we are Redzoning then check if there is some space between the
2247 * end of the object and the free pointer. If not then add an
2248 * additional word to have some bytes to store Redzone information.
2250 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2251 size += sizeof(void *);
2255 * With that we have determined the number of bytes in actual use
2256 * by the object. This is the potential offset to the free pointer.
2260 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2263 * Relocate free pointer after the object if it is not
2264 * permitted to overwrite the first word of the object on
2267 * This is the case if we do RCU, have a constructor or
2268 * destructor or are poisoning the objects.
2271 size += sizeof(void *);
2274 #ifdef CONFIG_SLUB_DEBUG
2275 if (flags & SLAB_STORE_USER)
2277 * Need to store information about allocs and frees after
2280 size += 2 * sizeof(struct track);
2282 if (flags & SLAB_RED_ZONE)
2284 * Add some empty padding so that we can catch
2285 * overwrites from earlier objects rather than let
2286 * tracking information or the free pointer be
2287 * corrupted if a user writes before the start
2290 size += sizeof(void *);
2294 * Determine the alignment based on various parameters that the
2295 * user specified and the dynamic determination of cache line size
2298 align = calculate_alignment(flags, align, s->objsize);
2301 * SLUB stores one object immediately after another beginning from
2302 * offset 0. In order to align the objects we have to simply size
2303 * each object to conform to the alignment.
2305 size = ALIGN(size, align);
2307 if (forced_order >= 0)
2308 order = forced_order;
2310 order = calculate_order(size);
2317 s->allocflags |= __GFP_COMP;
2319 if (s->flags & SLAB_CACHE_DMA)
2320 s->allocflags |= SLUB_DMA;
2322 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2323 s->allocflags |= __GFP_RECLAIMABLE;
2326 * Determine the number of objects per slab
2328 s->oo = oo_make(order, size);
2329 s->min = oo_make(get_order(size), size);
2330 if (oo_objects(s->oo) > oo_objects(s->max))
2333 return !!oo_objects(s->oo);
2337 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2338 const char *name, size_t size,
2339 size_t align, unsigned long flags,
2340 void (*ctor)(void *))
2342 memset(s, 0, kmem_size);
2347 s->flags = kmem_cache_flags(size, flags, name, ctor);
2349 if (!calculate_sizes(s, -1))
2353 * The larger the object size is, the more pages we want on the partial
2354 * list to avoid pounding the page allocator excessively.
2356 set_min_partial(s, ilog2(s->size));
2359 s->remote_node_defrag_ratio = 1000;
2361 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2364 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2366 free_kmem_cache_nodes(s);
2368 if (flags & SLAB_PANIC)
2369 panic("Cannot create slab %s size=%lu realsize=%u "
2370 "order=%u offset=%u flags=%lx\n",
2371 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2377 * Check if a given pointer is valid
2379 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2383 page = get_object_page(object);
2385 if (!page || s != page->slab)
2386 /* No slab or wrong slab */
2389 if (!check_valid_pointer(s, page, object))
2393 * We could also check if the object is on the slabs freelist.
2394 * But this would be too expensive and it seems that the main
2395 * purpose of kmem_ptr_valid() is to check if the object belongs
2396 * to a certain slab.
2400 EXPORT_SYMBOL(kmem_ptr_validate);
2403 * Determine the size of a slab object
2405 unsigned int kmem_cache_size(struct kmem_cache *s)
2409 EXPORT_SYMBOL(kmem_cache_size);
2411 const char *kmem_cache_name(struct kmem_cache *s)
2415 EXPORT_SYMBOL(kmem_cache_name);
2417 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2420 #ifdef CONFIG_SLUB_DEBUG
2421 void *addr = page_address(page);
2423 DECLARE_BITMAP(map, page->objects);
2425 bitmap_zero(map, page->objects);
2426 slab_err(s, page, "%s", text);
2428 for_each_free_object(p, s, page->freelist)
2429 set_bit(slab_index(p, s, addr), map);
2431 for_each_object(p, s, addr, page->objects) {
2433 if (!test_bit(slab_index(p, s, addr), map)) {
2434 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2436 print_tracking(s, p);
2444 * Attempt to free all partial slabs on a node.
2446 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2448 unsigned long flags;
2449 struct page *page, *h;
2451 spin_lock_irqsave(&n->list_lock, flags);
2452 list_for_each_entry_safe(page, h, &n->partial, lru) {
2454 list_del(&page->lru);
2455 discard_slab(s, page);
2458 list_slab_objects(s, page,
2459 "Objects remaining on kmem_cache_close()");
2462 spin_unlock_irqrestore(&n->list_lock, flags);
2466 * Release all resources used by a slab cache.
2468 static inline int kmem_cache_close(struct kmem_cache *s)
2474 /* Attempt to free all objects */
2475 free_kmem_cache_cpus(s);
2476 for_each_node_state(node, N_NORMAL_MEMORY) {
2477 struct kmem_cache_node *n = get_node(s, node);
2480 if (n->nr_partial || slabs_node(s, node))
2483 free_kmem_cache_nodes(s);
2488 * Close a cache and release the kmem_cache structure
2489 * (must be used for caches created using kmem_cache_create)
2491 void kmem_cache_destroy(struct kmem_cache *s)
2493 down_write(&slub_lock);
2497 up_write(&slub_lock);
2498 if (kmem_cache_close(s)) {
2499 printk(KERN_ERR "SLUB %s: %s called for cache that "
2500 "still has objects.\n", s->name, __func__);
2503 sysfs_slab_remove(s);
2505 up_write(&slub_lock);
2507 EXPORT_SYMBOL(kmem_cache_destroy);
2509 /********************************************************************
2511 *******************************************************************/
2513 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2514 EXPORT_SYMBOL(kmalloc_caches);
2516 static int __init setup_slub_min_order(char *str)
2518 get_option(&str, &slub_min_order);
2523 __setup("slub_min_order=", setup_slub_min_order);
2525 static int __init setup_slub_max_order(char *str)
2527 get_option(&str, &slub_max_order);
2528 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2533 __setup("slub_max_order=", setup_slub_max_order);
2535 static int __init setup_slub_min_objects(char *str)
2537 get_option(&str, &slub_min_objects);
2542 __setup("slub_min_objects=", setup_slub_min_objects);
2544 static int __init setup_slub_nomerge(char *str)
2550 __setup("slub_nomerge", setup_slub_nomerge);
2552 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2553 const char *name, int size, gfp_t gfp_flags)
2555 unsigned int flags = 0;
2557 if (gfp_flags & SLUB_DMA)
2558 flags = SLAB_CACHE_DMA;
2560 down_write(&slub_lock);
2561 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2565 list_add(&s->list, &slab_caches);
2566 up_write(&slub_lock);
2567 if (sysfs_slab_add(s))
2572 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2575 #ifdef CONFIG_ZONE_DMA
2576 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2578 static void sysfs_add_func(struct work_struct *w)
2580 struct kmem_cache *s;
2582 down_write(&slub_lock);
2583 list_for_each_entry(s, &slab_caches, list) {
2584 if (s->flags & __SYSFS_ADD_DEFERRED) {
2585 s->flags &= ~__SYSFS_ADD_DEFERRED;
2589 up_write(&slub_lock);
2592 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2594 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2596 struct kmem_cache *s;
2600 s = kmalloc_caches_dma[index];
2604 /* Dynamically create dma cache */
2605 if (flags & __GFP_WAIT)
2606 down_write(&slub_lock);
2608 if (!down_write_trylock(&slub_lock))
2612 if (kmalloc_caches_dma[index])
2615 realsize = kmalloc_caches[index].objsize;
2616 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2617 (unsigned int)realsize);
2618 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2620 if (!s || !text || !kmem_cache_open(s, flags, text,
2621 realsize, ARCH_KMALLOC_MINALIGN,
2622 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2628 list_add(&s->list, &slab_caches);
2629 kmalloc_caches_dma[index] = s;
2631 schedule_work(&sysfs_add_work);
2634 up_write(&slub_lock);
2636 return kmalloc_caches_dma[index];
2641 * Conversion table for small slabs sizes / 8 to the index in the
2642 * kmalloc array. This is necessary for slabs < 192 since we have non power
2643 * of two cache sizes there. The size of larger slabs can be determined using
2646 static s8 size_index[24] = {
2673 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2679 return ZERO_SIZE_PTR;
2681 index = size_index[(size - 1) / 8];
2683 index = fls(size - 1);
2685 #ifdef CONFIG_ZONE_DMA
2686 if (unlikely((flags & SLUB_DMA)))
2687 return dma_kmalloc_cache(index, flags);
2690 return &kmalloc_caches[index];
2693 void *__kmalloc(size_t size, gfp_t flags)
2695 struct kmem_cache *s;
2698 if (unlikely(size > SLUB_MAX_SIZE))
2699 return kmalloc_large(size, flags);
2701 s = get_slab(size, flags);
2703 if (unlikely(ZERO_OR_NULL_PTR(s)))
2706 ret = slab_alloc(s, flags, -1, _RET_IP_);
2708 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2712 EXPORT_SYMBOL(__kmalloc);
2714 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2716 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2720 return page_address(page);
2726 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2728 struct kmem_cache *s;
2731 if (unlikely(size > SLUB_MAX_SIZE)) {
2732 ret = kmalloc_large_node(size, flags, node);
2734 trace_kmalloc_node(_RET_IP_, ret,
2735 size, PAGE_SIZE << get_order(size),
2741 s = get_slab(size, flags);
2743 if (unlikely(ZERO_OR_NULL_PTR(s)))
2746 ret = slab_alloc(s, flags, node, _RET_IP_);
2748 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2752 EXPORT_SYMBOL(__kmalloc_node);
2755 size_t ksize(const void *object)
2758 struct kmem_cache *s;
2760 if (unlikely(object == ZERO_SIZE_PTR))
2763 page = virt_to_head_page(object);
2765 if (unlikely(!PageSlab(page))) {
2766 WARN_ON(!PageCompound(page));
2767 return PAGE_SIZE << compound_order(page);
2771 #ifdef CONFIG_SLUB_DEBUG
2773 * Debugging requires use of the padding between object
2774 * and whatever may come after it.
2776 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2781 * If we have the need to store the freelist pointer
2782 * back there or track user information then we can
2783 * only use the space before that information.
2785 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2788 * Else we can use all the padding etc for the allocation
2792 EXPORT_SYMBOL(ksize);
2794 void kfree(const void *x)
2797 void *object = (void *)x;
2799 trace_kfree(_RET_IP_, x);
2801 if (unlikely(ZERO_OR_NULL_PTR(x)))
2804 page = virt_to_head_page(x);
2805 if (unlikely(!PageSlab(page))) {
2806 BUG_ON(!PageCompound(page));
2810 slab_free(page->slab, page, object, _RET_IP_);
2812 EXPORT_SYMBOL(kfree);
2815 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2816 * the remaining slabs by the number of items in use. The slabs with the
2817 * most items in use come first. New allocations will then fill those up
2818 * and thus they can be removed from the partial lists.
2820 * The slabs with the least items are placed last. This results in them
2821 * being allocated from last increasing the chance that the last objects
2822 * are freed in them.
2824 int kmem_cache_shrink(struct kmem_cache *s)
2828 struct kmem_cache_node *n;
2831 int objects = oo_objects(s->max);
2832 struct list_head *slabs_by_inuse =
2833 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2834 unsigned long flags;
2836 if (!slabs_by_inuse)
2840 for_each_node_state(node, N_NORMAL_MEMORY) {
2841 n = get_node(s, node);
2846 for (i = 0; i < objects; i++)
2847 INIT_LIST_HEAD(slabs_by_inuse + i);
2849 spin_lock_irqsave(&n->list_lock, flags);
2852 * Build lists indexed by the items in use in each slab.
2854 * Note that concurrent frees may occur while we hold the
2855 * list_lock. page->inuse here is the upper limit.
2857 list_for_each_entry_safe(page, t, &n->partial, lru) {
2858 if (!page->inuse && slab_trylock(page)) {
2860 * Must hold slab lock here because slab_free
2861 * may have freed the last object and be
2862 * waiting to release the slab.
2864 list_del(&page->lru);
2867 discard_slab(s, page);
2869 list_move(&page->lru,
2870 slabs_by_inuse + page->inuse);
2875 * Rebuild the partial list with the slabs filled up most
2876 * first and the least used slabs at the end.
2878 for (i = objects - 1; i >= 0; i--)
2879 list_splice(slabs_by_inuse + i, n->partial.prev);
2881 spin_unlock_irqrestore(&n->list_lock, flags);
2884 kfree(slabs_by_inuse);
2887 EXPORT_SYMBOL(kmem_cache_shrink);
2889 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2890 static int slab_mem_going_offline_callback(void *arg)
2892 struct kmem_cache *s;
2894 down_read(&slub_lock);
2895 list_for_each_entry(s, &slab_caches, list)
2896 kmem_cache_shrink(s);
2897 up_read(&slub_lock);
2902 static void slab_mem_offline_callback(void *arg)
2904 struct kmem_cache_node *n;
2905 struct kmem_cache *s;
2906 struct memory_notify *marg = arg;
2909 offline_node = marg->status_change_nid;
2912 * If the node still has available memory. we need kmem_cache_node
2915 if (offline_node < 0)
2918 down_read(&slub_lock);
2919 list_for_each_entry(s, &slab_caches, list) {
2920 n = get_node(s, offline_node);
2923 * if n->nr_slabs > 0, slabs still exist on the node
2924 * that is going down. We were unable to free them,
2925 * and offline_pages() function shoudn't call this
2926 * callback. So, we must fail.
2928 BUG_ON(slabs_node(s, offline_node));
2930 s->node[offline_node] = NULL;
2931 kmem_cache_free(kmalloc_caches, n);
2934 up_read(&slub_lock);
2937 static int slab_mem_going_online_callback(void *arg)
2939 struct kmem_cache_node *n;
2940 struct kmem_cache *s;
2941 struct memory_notify *marg = arg;
2942 int nid = marg->status_change_nid;
2946 * If the node's memory is already available, then kmem_cache_node is
2947 * already created. Nothing to do.
2953 * We are bringing a node online. No memory is available yet. We must
2954 * allocate a kmem_cache_node structure in order to bring the node
2957 down_read(&slub_lock);
2958 list_for_each_entry(s, &slab_caches, list) {
2960 * XXX: kmem_cache_alloc_node will fallback to other nodes
2961 * since memory is not yet available from the node that
2964 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2969 init_kmem_cache_node(n, s);
2973 up_read(&slub_lock);
2977 static int slab_memory_callback(struct notifier_block *self,
2978 unsigned long action, void *arg)
2983 case MEM_GOING_ONLINE:
2984 ret = slab_mem_going_online_callback(arg);
2986 case MEM_GOING_OFFLINE:
2987 ret = slab_mem_going_offline_callback(arg);
2990 case MEM_CANCEL_ONLINE:
2991 slab_mem_offline_callback(arg);
2994 case MEM_CANCEL_OFFLINE:
2998 ret = notifier_from_errno(ret);
3004 #endif /* CONFIG_MEMORY_HOTPLUG */
3006 /********************************************************************
3007 * Basic setup of slabs
3008 *******************************************************************/
3010 void __init kmem_cache_init(void)
3019 * Must first have the slab cache available for the allocations of the
3020 * struct kmem_cache_node's. There is special bootstrap code in
3021 * kmem_cache_open for slab_state == DOWN.
3023 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3024 sizeof(struct kmem_cache_node), GFP_KERNEL);
3025 kmalloc_caches[0].refcount = -1;
3028 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3031 /* Able to allocate the per node structures */
3032 slab_state = PARTIAL;
3034 /* Caches that are not of the two-to-the-power-of size */
3035 if (KMALLOC_MIN_SIZE <= 64) {
3036 create_kmalloc_cache(&kmalloc_caches[1],
3037 "kmalloc-96", 96, GFP_KERNEL);
3039 create_kmalloc_cache(&kmalloc_caches[2],
3040 "kmalloc-192", 192, GFP_KERNEL);
3044 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3045 create_kmalloc_cache(&kmalloc_caches[i],
3046 "kmalloc", 1 << i, GFP_KERNEL);
3052 * Patch up the size_index table if we have strange large alignment
3053 * requirements for the kmalloc array. This is only the case for
3054 * MIPS it seems. The standard arches will not generate any code here.
3056 * Largest permitted alignment is 256 bytes due to the way we
3057 * handle the index determination for the smaller caches.
3059 * Make sure that nothing crazy happens if someone starts tinkering
3060 * around with ARCH_KMALLOC_MINALIGN
3062 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3063 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3065 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3066 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3068 if (KMALLOC_MIN_SIZE == 128) {
3070 * The 192 byte sized cache is not used if the alignment
3071 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3074 for (i = 128 + 8; i <= 192; i += 8)
3075 size_index[(i - 1) / 8] = 8;
3080 /* Provide the correct kmalloc names now that the caches are up */
3081 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3082 kmalloc_caches[i]. name =
3083 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3086 register_cpu_notifier(&slab_notifier);
3087 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3088 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3090 kmem_size = sizeof(struct kmem_cache);
3094 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3095 " CPUs=%d, Nodes=%d\n",
3096 caches, cache_line_size(),
3097 slub_min_order, slub_max_order, slub_min_objects,
3098 nr_cpu_ids, nr_node_ids);
3102 * Find a mergeable slab cache
3104 static int slab_unmergeable(struct kmem_cache *s)
3106 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3113 * We may have set a slab to be unmergeable during bootstrap.
3115 if (s->refcount < 0)
3121 static struct kmem_cache *find_mergeable(size_t size,
3122 size_t align, unsigned long flags, const char *name,
3123 void (*ctor)(void *))
3125 struct kmem_cache *s;
3127 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3133 size = ALIGN(size, sizeof(void *));
3134 align = calculate_alignment(flags, align, size);
3135 size = ALIGN(size, align);
3136 flags = kmem_cache_flags(size, flags, name, NULL);
3138 list_for_each_entry(s, &slab_caches, list) {
3139 if (slab_unmergeable(s))
3145 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3148 * Check if alignment is compatible.
3149 * Courtesy of Adrian Drzewiecki
3151 if ((s->size & ~(align - 1)) != s->size)
3154 if (s->size - size >= sizeof(void *))
3162 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3163 size_t align, unsigned long flags, void (*ctor)(void *))
3165 struct kmem_cache *s;
3167 down_write(&slub_lock);
3168 s = find_mergeable(size, align, flags, name, ctor);
3174 * Adjust the object sizes so that we clear
3175 * the complete object on kzalloc.
3177 s->objsize = max(s->objsize, (int)size);
3180 * And then we need to update the object size in the
3181 * per cpu structures
3183 for_each_online_cpu(cpu)
3184 get_cpu_slab(s, cpu)->objsize = s->objsize;
3186 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3187 up_write(&slub_lock);
3189 if (sysfs_slab_alias(s, name)) {
3190 down_write(&slub_lock);
3192 up_write(&slub_lock);
3198 s = kmalloc(kmem_size, GFP_KERNEL);
3200 if (kmem_cache_open(s, GFP_KERNEL, name,
3201 size, align, flags, ctor)) {
3202 list_add(&s->list, &slab_caches);
3203 up_write(&slub_lock);
3204 if (sysfs_slab_add(s)) {
3205 down_write(&slub_lock);
3207 up_write(&slub_lock);
3215 up_write(&slub_lock);
3218 if (flags & SLAB_PANIC)
3219 panic("Cannot create slabcache %s\n", name);
3224 EXPORT_SYMBOL(kmem_cache_create);
3228 * Use the cpu notifier to insure that the cpu slabs are flushed when
3231 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3232 unsigned long action, void *hcpu)
3234 long cpu = (long)hcpu;
3235 struct kmem_cache *s;
3236 unsigned long flags;
3239 case CPU_UP_PREPARE:
3240 case CPU_UP_PREPARE_FROZEN:
3241 init_alloc_cpu_cpu(cpu);
3242 down_read(&slub_lock);
3243 list_for_each_entry(s, &slab_caches, list)
3244 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3246 up_read(&slub_lock);
3249 case CPU_UP_CANCELED:
3250 case CPU_UP_CANCELED_FROZEN:
3252 case CPU_DEAD_FROZEN:
3253 down_read(&slub_lock);
3254 list_for_each_entry(s, &slab_caches, list) {
3255 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3257 local_irq_save(flags);
3258 __flush_cpu_slab(s, cpu);
3259 local_irq_restore(flags);
3260 free_kmem_cache_cpu(c, cpu);
3261 s->cpu_slab[cpu] = NULL;
3263 up_read(&slub_lock);
3271 static struct notifier_block __cpuinitdata slab_notifier = {
3272 .notifier_call = slab_cpuup_callback
3277 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3279 struct kmem_cache *s;
3282 if (unlikely(size > SLUB_MAX_SIZE))
3283 return kmalloc_large(size, gfpflags);
3285 s = get_slab(size, gfpflags);
3287 if (unlikely(ZERO_OR_NULL_PTR(s)))
3290 ret = slab_alloc(s, gfpflags, -1, caller);
3292 /* Honor the call site pointer we recieved. */
3293 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3298 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3299 int node, unsigned long caller)
3301 struct kmem_cache *s;
3304 if (unlikely(size > SLUB_MAX_SIZE))
3305 return kmalloc_large_node(size, gfpflags, node);
3307 s = get_slab(size, gfpflags);
3309 if (unlikely(ZERO_OR_NULL_PTR(s)))
3312 ret = slab_alloc(s, gfpflags, node, caller);
3314 /* Honor the call site pointer we recieved. */
3315 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3320 #ifdef CONFIG_SLUB_DEBUG
3321 static unsigned long count_partial(struct kmem_cache_node *n,
3322 int (*get_count)(struct page *))
3324 unsigned long flags;
3325 unsigned long x = 0;
3328 spin_lock_irqsave(&n->list_lock, flags);
3329 list_for_each_entry(page, &n->partial, lru)
3330 x += get_count(page);
3331 spin_unlock_irqrestore(&n->list_lock, flags);
3335 static int count_inuse(struct page *page)
3340 static int count_total(struct page *page)
3342 return page->objects;
3345 static int count_free(struct page *page)
3347 return page->objects - page->inuse;
3350 static int validate_slab(struct kmem_cache *s, struct page *page,
3354 void *addr = page_address(page);
3356 if (!check_slab(s, page) ||
3357 !on_freelist(s, page, NULL))
3360 /* Now we know that a valid freelist exists */
3361 bitmap_zero(map, page->objects);
3363 for_each_free_object(p, s, page->freelist) {
3364 set_bit(slab_index(p, s, addr), map);
3365 if (!check_object(s, page, p, 0))
3369 for_each_object(p, s, addr, page->objects)
3370 if (!test_bit(slab_index(p, s, addr), map))
3371 if (!check_object(s, page, p, 1))
3376 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3379 if (slab_trylock(page)) {
3380 validate_slab(s, page, map);
3383 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3386 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3387 if (!PageSlubDebug(page))
3388 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3389 "on slab 0x%p\n", s->name, page);
3391 if (PageSlubDebug(page))
3392 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3393 "slab 0x%p\n", s->name, page);
3397 static int validate_slab_node(struct kmem_cache *s,
3398 struct kmem_cache_node *n, unsigned long *map)
3400 unsigned long count = 0;
3402 unsigned long flags;
3404 spin_lock_irqsave(&n->list_lock, flags);
3406 list_for_each_entry(page, &n->partial, lru) {
3407 validate_slab_slab(s, page, map);
3410 if (count != n->nr_partial)
3411 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3412 "counter=%ld\n", s->name, count, n->nr_partial);
3414 if (!(s->flags & SLAB_STORE_USER))
3417 list_for_each_entry(page, &n->full, lru) {
3418 validate_slab_slab(s, page, map);
3421 if (count != atomic_long_read(&n->nr_slabs))
3422 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3423 "counter=%ld\n", s->name, count,
3424 atomic_long_read(&n->nr_slabs));
3427 spin_unlock_irqrestore(&n->list_lock, flags);
3431 static long validate_slab_cache(struct kmem_cache *s)
3434 unsigned long count = 0;
3435 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3436 sizeof(unsigned long), GFP_KERNEL);
3442 for_each_node_state(node, N_NORMAL_MEMORY) {
3443 struct kmem_cache_node *n = get_node(s, node);
3445 count += validate_slab_node(s, n, map);
3451 #ifdef SLUB_RESILIENCY_TEST
3452 static void resiliency_test(void)
3456 printk(KERN_ERR "SLUB resiliency testing\n");
3457 printk(KERN_ERR "-----------------------\n");
3458 printk(KERN_ERR "A. Corruption after allocation\n");
3460 p = kzalloc(16, GFP_KERNEL);
3462 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3463 " 0x12->0x%p\n\n", p + 16);
3465 validate_slab_cache(kmalloc_caches + 4);
3467 /* Hmmm... The next two are dangerous */
3468 p = kzalloc(32, GFP_KERNEL);
3469 p[32 + sizeof(void *)] = 0x34;
3470 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3471 " 0x34 -> -0x%p\n", p);
3473 "If allocated object is overwritten then not detectable\n\n");
3475 validate_slab_cache(kmalloc_caches + 5);
3476 p = kzalloc(64, GFP_KERNEL);
3477 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3479 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3482 "If allocated object is overwritten then not detectable\n\n");
3483 validate_slab_cache(kmalloc_caches + 6);
3485 printk(KERN_ERR "\nB. Corruption after free\n");
3486 p = kzalloc(128, GFP_KERNEL);
3489 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3490 validate_slab_cache(kmalloc_caches + 7);
3492 p = kzalloc(256, GFP_KERNEL);
3495 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3497 validate_slab_cache(kmalloc_caches + 8);
3499 p = kzalloc(512, GFP_KERNEL);
3502 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3503 validate_slab_cache(kmalloc_caches + 9);
3506 static void resiliency_test(void) {};
3510 * Generate lists of code addresses where slabcache objects are allocated
3515 unsigned long count;
3522 DECLARE_BITMAP(cpus, NR_CPUS);
3528 unsigned long count;
3529 struct location *loc;
3532 static void free_loc_track(struct loc_track *t)
3535 free_pages((unsigned long)t->loc,
3536 get_order(sizeof(struct location) * t->max));
3539 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3544 order = get_order(sizeof(struct location) * max);
3546 l = (void *)__get_free_pages(flags, order);
3551 memcpy(l, t->loc, sizeof(struct location) * t->count);
3559 static int add_location(struct loc_track *t, struct kmem_cache *s,
3560 const struct track *track)
3562 long start, end, pos;
3564 unsigned long caddr;
3565 unsigned long age = jiffies - track->when;
3571 pos = start + (end - start + 1) / 2;
3574 * There is nothing at "end". If we end up there
3575 * we need to add something to before end.
3580 caddr = t->loc[pos].addr;
3581 if (track->addr == caddr) {
3587 if (age < l->min_time)
3589 if (age > l->max_time)
3592 if (track->pid < l->min_pid)
3593 l->min_pid = track->pid;
3594 if (track->pid > l->max_pid)
3595 l->max_pid = track->pid;
3597 cpumask_set_cpu(track->cpu,
3598 to_cpumask(l->cpus));
3600 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3604 if (track->addr < caddr)
3611 * Not found. Insert new tracking element.
3613 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3619 (t->count - pos) * sizeof(struct location));
3622 l->addr = track->addr;
3626 l->min_pid = track->pid;
3627 l->max_pid = track->pid;
3628 cpumask_clear(to_cpumask(l->cpus));
3629 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3630 nodes_clear(l->nodes);
3631 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3635 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3636 struct page *page, enum track_item alloc)
3638 void *addr = page_address(page);
3639 DECLARE_BITMAP(map, page->objects);
3642 bitmap_zero(map, page->objects);
3643 for_each_free_object(p, s, page->freelist)
3644 set_bit(slab_index(p, s, addr), map);
3646 for_each_object(p, s, addr, page->objects)
3647 if (!test_bit(slab_index(p, s, addr), map))
3648 add_location(t, s, get_track(s, p, alloc));
3651 static int list_locations(struct kmem_cache *s, char *buf,
3652 enum track_item alloc)
3656 struct loc_track t = { 0, 0, NULL };
3659 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3661 return sprintf(buf, "Out of memory\n");
3663 /* Push back cpu slabs */
3666 for_each_node_state(node, N_NORMAL_MEMORY) {
3667 struct kmem_cache_node *n = get_node(s, node);
3668 unsigned long flags;
3671 if (!atomic_long_read(&n->nr_slabs))
3674 spin_lock_irqsave(&n->list_lock, flags);
3675 list_for_each_entry(page, &n->partial, lru)
3676 process_slab(&t, s, page, alloc);
3677 list_for_each_entry(page, &n->full, lru)
3678 process_slab(&t, s, page, alloc);
3679 spin_unlock_irqrestore(&n->list_lock, flags);
3682 for (i = 0; i < t.count; i++) {
3683 struct location *l = &t.loc[i];
3685 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3687 len += sprintf(buf + len, "%7ld ", l->count);
3690 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3692 len += sprintf(buf + len, "<not-available>");
3694 if (l->sum_time != l->min_time) {
3695 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3697 (long)div_u64(l->sum_time, l->count),
3700 len += sprintf(buf + len, " age=%ld",
3703 if (l->min_pid != l->max_pid)
3704 len += sprintf(buf + len, " pid=%ld-%ld",
3705 l->min_pid, l->max_pid);
3707 len += sprintf(buf + len, " pid=%ld",
3710 if (num_online_cpus() > 1 &&
3711 !cpumask_empty(to_cpumask(l->cpus)) &&
3712 len < PAGE_SIZE - 60) {
3713 len += sprintf(buf + len, " cpus=");
3714 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3715 to_cpumask(l->cpus));
3718 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3719 len < PAGE_SIZE - 60) {
3720 len += sprintf(buf + len, " nodes=");
3721 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3725 len += sprintf(buf + len, "\n");
3730 len += sprintf(buf, "No data\n");
3734 enum slab_stat_type {
3735 SL_ALL, /* All slabs */
3736 SL_PARTIAL, /* Only partially allocated slabs */
3737 SL_CPU, /* Only slabs used for cpu caches */
3738 SL_OBJECTS, /* Determine allocated objects not slabs */
3739 SL_TOTAL /* Determine object capacity not slabs */
3742 #define SO_ALL (1 << SL_ALL)
3743 #define SO_PARTIAL (1 << SL_PARTIAL)
3744 #define SO_CPU (1 << SL_CPU)
3745 #define SO_OBJECTS (1 << SL_OBJECTS)
3746 #define SO_TOTAL (1 << SL_TOTAL)
3748 static ssize_t show_slab_objects(struct kmem_cache *s,
3749 char *buf, unsigned long flags)
3751 unsigned long total = 0;
3754 unsigned long *nodes;
3755 unsigned long *per_cpu;
3757 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3760 per_cpu = nodes + nr_node_ids;
3762 if (flags & SO_CPU) {
3765 for_each_possible_cpu(cpu) {
3766 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3768 if (!c || c->node < 0)
3772 if (flags & SO_TOTAL)
3773 x = c->page->objects;
3774 else if (flags & SO_OBJECTS)
3780 nodes[c->node] += x;
3786 if (flags & SO_ALL) {
3787 for_each_node_state(node, N_NORMAL_MEMORY) {
3788 struct kmem_cache_node *n = get_node(s, node);
3790 if (flags & SO_TOTAL)
3791 x = atomic_long_read(&n->total_objects);
3792 else if (flags & SO_OBJECTS)
3793 x = atomic_long_read(&n->total_objects) -
3794 count_partial(n, count_free);
3797 x = atomic_long_read(&n->nr_slabs);
3802 } else if (flags & SO_PARTIAL) {
3803 for_each_node_state(node, N_NORMAL_MEMORY) {
3804 struct kmem_cache_node *n = get_node(s, node);
3806 if (flags & SO_TOTAL)
3807 x = count_partial(n, count_total);
3808 else if (flags & SO_OBJECTS)
3809 x = count_partial(n, count_inuse);
3816 x = sprintf(buf, "%lu", total);
3818 for_each_node_state(node, N_NORMAL_MEMORY)
3820 x += sprintf(buf + x, " N%d=%lu",
3824 return x + sprintf(buf + x, "\n");
3827 static int any_slab_objects(struct kmem_cache *s)
3831 for_each_online_node(node) {
3832 struct kmem_cache_node *n = get_node(s, node);
3837 if (atomic_long_read(&n->total_objects))
3843 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3844 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3846 struct slab_attribute {
3847 struct attribute attr;
3848 ssize_t (*show)(struct kmem_cache *s, char *buf);
3849 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3852 #define SLAB_ATTR_RO(_name) \
3853 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3855 #define SLAB_ATTR(_name) \
3856 static struct slab_attribute _name##_attr = \
3857 __ATTR(_name, 0644, _name##_show, _name##_store)
3859 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3861 return sprintf(buf, "%d\n", s->size);
3863 SLAB_ATTR_RO(slab_size);
3865 static ssize_t align_show(struct kmem_cache *s, char *buf)
3867 return sprintf(buf, "%d\n", s->align);
3869 SLAB_ATTR_RO(align);
3871 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3873 return sprintf(buf, "%d\n", s->objsize);
3875 SLAB_ATTR_RO(object_size);
3877 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3879 return sprintf(buf, "%d\n", oo_objects(s->oo));
3881 SLAB_ATTR_RO(objs_per_slab);
3883 static ssize_t order_store(struct kmem_cache *s,
3884 const char *buf, size_t length)
3886 unsigned long order;
3889 err = strict_strtoul(buf, 10, &order);
3893 if (order > slub_max_order || order < slub_min_order)
3896 calculate_sizes(s, order);
3900 static ssize_t order_show(struct kmem_cache *s, char *buf)
3902 return sprintf(buf, "%d\n", oo_order(s->oo));
3906 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3908 return sprintf(buf, "%lu\n", s->min_partial);
3911 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3917 err = strict_strtoul(buf, 10, &min);
3921 set_min_partial(s, min);
3924 SLAB_ATTR(min_partial);
3926 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3929 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3931 return n + sprintf(buf + n, "\n");
3937 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3939 return sprintf(buf, "%d\n", s->refcount - 1);
3941 SLAB_ATTR_RO(aliases);
3943 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3945 return show_slab_objects(s, buf, SO_ALL);
3947 SLAB_ATTR_RO(slabs);
3949 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3951 return show_slab_objects(s, buf, SO_PARTIAL);
3953 SLAB_ATTR_RO(partial);
3955 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3957 return show_slab_objects(s, buf, SO_CPU);
3959 SLAB_ATTR_RO(cpu_slabs);
3961 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3963 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3965 SLAB_ATTR_RO(objects);
3967 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3969 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3971 SLAB_ATTR_RO(objects_partial);
3973 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3975 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3977 SLAB_ATTR_RO(total_objects);
3979 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3981 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3984 static ssize_t sanity_checks_store(struct kmem_cache *s,
3985 const char *buf, size_t length)
3987 s->flags &= ~SLAB_DEBUG_FREE;
3989 s->flags |= SLAB_DEBUG_FREE;
3992 SLAB_ATTR(sanity_checks);
3994 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3996 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3999 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4002 s->flags &= ~SLAB_TRACE;
4004 s->flags |= SLAB_TRACE;
4009 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4011 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4014 static ssize_t reclaim_account_store(struct kmem_cache *s,
4015 const char *buf, size_t length)
4017 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4019 s->flags |= SLAB_RECLAIM_ACCOUNT;
4022 SLAB_ATTR(reclaim_account);
4024 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4026 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4028 SLAB_ATTR_RO(hwcache_align);
4030 #ifdef CONFIG_ZONE_DMA
4031 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4033 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4035 SLAB_ATTR_RO(cache_dma);
4038 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4040 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4042 SLAB_ATTR_RO(destroy_by_rcu);
4044 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4046 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4049 static ssize_t red_zone_store(struct kmem_cache *s,
4050 const char *buf, size_t length)
4052 if (any_slab_objects(s))
4055 s->flags &= ~SLAB_RED_ZONE;
4057 s->flags |= SLAB_RED_ZONE;
4058 calculate_sizes(s, -1);
4061 SLAB_ATTR(red_zone);
4063 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4065 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4068 static ssize_t poison_store(struct kmem_cache *s,
4069 const char *buf, size_t length)
4071 if (any_slab_objects(s))
4074 s->flags &= ~SLAB_POISON;
4076 s->flags |= SLAB_POISON;
4077 calculate_sizes(s, -1);
4082 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4084 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4087 static ssize_t store_user_store(struct kmem_cache *s,
4088 const char *buf, size_t length)
4090 if (any_slab_objects(s))
4093 s->flags &= ~SLAB_STORE_USER;
4095 s->flags |= SLAB_STORE_USER;
4096 calculate_sizes(s, -1);
4099 SLAB_ATTR(store_user);
4101 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4106 static ssize_t validate_store(struct kmem_cache *s,
4107 const char *buf, size_t length)
4111 if (buf[0] == '1') {
4112 ret = validate_slab_cache(s);
4118 SLAB_ATTR(validate);
4120 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4125 static ssize_t shrink_store(struct kmem_cache *s,
4126 const char *buf, size_t length)
4128 if (buf[0] == '1') {
4129 int rc = kmem_cache_shrink(s);
4139 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4141 if (!(s->flags & SLAB_STORE_USER))
4143 return list_locations(s, buf, TRACK_ALLOC);
4145 SLAB_ATTR_RO(alloc_calls);
4147 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4149 if (!(s->flags & SLAB_STORE_USER))
4151 return list_locations(s, buf, TRACK_FREE);
4153 SLAB_ATTR_RO(free_calls);
4156 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4158 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4161 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4162 const char *buf, size_t length)
4164 unsigned long ratio;
4167 err = strict_strtoul(buf, 10, &ratio);
4172 s->remote_node_defrag_ratio = ratio * 10;
4176 SLAB_ATTR(remote_node_defrag_ratio);
4179 #ifdef CONFIG_SLUB_STATS
4180 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4182 unsigned long sum = 0;
4185 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4190 for_each_online_cpu(cpu) {
4191 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4197 len = sprintf(buf, "%lu", sum);
4200 for_each_online_cpu(cpu) {
4201 if (data[cpu] && len < PAGE_SIZE - 20)
4202 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4206 return len + sprintf(buf + len, "\n");
4209 #define STAT_ATTR(si, text) \
4210 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4212 return show_stat(s, buf, si); \
4214 SLAB_ATTR_RO(text); \
4216 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4217 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4218 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4219 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4220 STAT_ATTR(FREE_FROZEN, free_frozen);
4221 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4222 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4223 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4224 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4225 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4226 STAT_ATTR(FREE_SLAB, free_slab);
4227 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4228 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4229 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4230 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4231 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4232 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4233 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4236 static struct attribute *slab_attrs[] = {
4237 &slab_size_attr.attr,
4238 &object_size_attr.attr,
4239 &objs_per_slab_attr.attr,
4241 &min_partial_attr.attr,
4243 &objects_partial_attr.attr,
4244 &total_objects_attr.attr,
4247 &cpu_slabs_attr.attr,
4251 &sanity_checks_attr.attr,
4253 &hwcache_align_attr.attr,
4254 &reclaim_account_attr.attr,
4255 &destroy_by_rcu_attr.attr,
4256 &red_zone_attr.attr,
4258 &store_user_attr.attr,
4259 &validate_attr.attr,
4261 &alloc_calls_attr.attr,
4262 &free_calls_attr.attr,
4263 #ifdef CONFIG_ZONE_DMA
4264 &cache_dma_attr.attr,
4267 &remote_node_defrag_ratio_attr.attr,
4269 #ifdef CONFIG_SLUB_STATS
4270 &alloc_fastpath_attr.attr,
4271 &alloc_slowpath_attr.attr,
4272 &free_fastpath_attr.attr,
4273 &free_slowpath_attr.attr,
4274 &free_frozen_attr.attr,
4275 &free_add_partial_attr.attr,
4276 &free_remove_partial_attr.attr,
4277 &alloc_from_partial_attr.attr,
4278 &alloc_slab_attr.attr,
4279 &alloc_refill_attr.attr,
4280 &free_slab_attr.attr,
4281 &cpuslab_flush_attr.attr,
4282 &deactivate_full_attr.attr,
4283 &deactivate_empty_attr.attr,
4284 &deactivate_to_head_attr.attr,
4285 &deactivate_to_tail_attr.attr,
4286 &deactivate_remote_frees_attr.attr,
4287 &order_fallback_attr.attr,
4292 static struct attribute_group slab_attr_group = {
4293 .attrs = slab_attrs,
4296 static ssize_t slab_attr_show(struct kobject *kobj,
4297 struct attribute *attr,
4300 struct slab_attribute *attribute;
4301 struct kmem_cache *s;
4304 attribute = to_slab_attr(attr);
4307 if (!attribute->show)
4310 err = attribute->show(s, buf);
4315 static ssize_t slab_attr_store(struct kobject *kobj,
4316 struct attribute *attr,
4317 const char *buf, size_t len)
4319 struct slab_attribute *attribute;
4320 struct kmem_cache *s;
4323 attribute = to_slab_attr(attr);
4326 if (!attribute->store)
4329 err = attribute->store(s, buf, len);
4334 static void kmem_cache_release(struct kobject *kobj)
4336 struct kmem_cache *s = to_slab(kobj);
4341 static struct sysfs_ops slab_sysfs_ops = {
4342 .show = slab_attr_show,
4343 .store = slab_attr_store,
4346 static struct kobj_type slab_ktype = {
4347 .sysfs_ops = &slab_sysfs_ops,
4348 .release = kmem_cache_release
4351 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4353 struct kobj_type *ktype = get_ktype(kobj);
4355 if (ktype == &slab_ktype)
4360 static struct kset_uevent_ops slab_uevent_ops = {
4361 .filter = uevent_filter,
4364 static struct kset *slab_kset;
4366 #define ID_STR_LENGTH 64
4368 /* Create a unique string id for a slab cache:
4370 * Format :[flags-]size
4372 static char *create_unique_id(struct kmem_cache *s)
4374 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4381 * First flags affecting slabcache operations. We will only
4382 * get here for aliasable slabs so we do not need to support
4383 * too many flags. The flags here must cover all flags that
4384 * are matched during merging to guarantee that the id is
4387 if (s->flags & SLAB_CACHE_DMA)
4389 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4391 if (s->flags & SLAB_DEBUG_FREE)
4395 p += sprintf(p, "%07d", s->size);
4396 BUG_ON(p > name + ID_STR_LENGTH - 1);
4400 static int sysfs_slab_add(struct kmem_cache *s)
4406 if (slab_state < SYSFS)
4407 /* Defer until later */
4410 unmergeable = slab_unmergeable(s);
4413 * Slabcache can never be merged so we can use the name proper.
4414 * This is typically the case for debug situations. In that
4415 * case we can catch duplicate names easily.
4417 sysfs_remove_link(&slab_kset->kobj, s->name);
4421 * Create a unique name for the slab as a target
4424 name = create_unique_id(s);
4427 s->kobj.kset = slab_kset;
4428 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4430 kobject_put(&s->kobj);
4434 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4437 kobject_uevent(&s->kobj, KOBJ_ADD);
4439 /* Setup first alias */
4440 sysfs_slab_alias(s, s->name);
4446 static void sysfs_slab_remove(struct kmem_cache *s)
4448 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4449 kobject_del(&s->kobj);
4450 kobject_put(&s->kobj);
4454 * Need to buffer aliases during bootup until sysfs becomes
4455 * available lest we lose that information.
4457 struct saved_alias {
4458 struct kmem_cache *s;
4460 struct saved_alias *next;
4463 static struct saved_alias *alias_list;
4465 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4467 struct saved_alias *al;
4469 if (slab_state == SYSFS) {
4471 * If we have a leftover link then remove it.
4473 sysfs_remove_link(&slab_kset->kobj, name);
4474 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4477 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4483 al->next = alias_list;
4488 static int __init slab_sysfs_init(void)
4490 struct kmem_cache *s;
4493 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4495 printk(KERN_ERR "Cannot register slab subsystem.\n");
4501 list_for_each_entry(s, &slab_caches, list) {
4502 err = sysfs_slab_add(s);
4504 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4505 " to sysfs\n", s->name);
4508 while (alias_list) {
4509 struct saved_alias *al = alias_list;
4511 alias_list = alias_list->next;
4512 err = sysfs_slab_alias(al->s, al->name);
4514 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4515 " %s to sysfs\n", s->name);
4523 __initcall(slab_sysfs_init);
4527 * The /proc/slabinfo ABI
4529 #ifdef CONFIG_SLABINFO
4530 static void print_slabinfo_header(struct seq_file *m)
4532 seq_puts(m, "slabinfo - version: 2.1\n");
4533 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4534 "<objperslab> <pagesperslab>");
4535 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4536 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4540 static void *s_start(struct seq_file *m, loff_t *pos)
4544 down_read(&slub_lock);
4546 print_slabinfo_header(m);
4548 return seq_list_start(&slab_caches, *pos);
4551 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4553 return seq_list_next(p, &slab_caches, pos);
4556 static void s_stop(struct seq_file *m, void *p)
4558 up_read(&slub_lock);
4561 static int s_show(struct seq_file *m, void *p)
4563 unsigned long nr_partials = 0;
4564 unsigned long nr_slabs = 0;
4565 unsigned long nr_inuse = 0;
4566 unsigned long nr_objs = 0;
4567 unsigned long nr_free = 0;
4568 struct kmem_cache *s;
4571 s = list_entry(p, struct kmem_cache, list);
4573 for_each_online_node(node) {
4574 struct kmem_cache_node *n = get_node(s, node);
4579 nr_partials += n->nr_partial;
4580 nr_slabs += atomic_long_read(&n->nr_slabs);
4581 nr_objs += atomic_long_read(&n->total_objects);
4582 nr_free += count_partial(n, count_free);
4585 nr_inuse = nr_objs - nr_free;
4587 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4588 nr_objs, s->size, oo_objects(s->oo),
4589 (1 << oo_order(s->oo)));
4590 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4591 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4597 static const struct seq_operations slabinfo_op = {
4604 static int slabinfo_open(struct inode *inode, struct file *file)
4606 return seq_open(file, &slabinfo_op);
4609 static const struct file_operations proc_slabinfo_operations = {
4610 .open = slabinfo_open,
4612 .llseek = seq_lseek,
4613 .release = seq_release,
4616 static int __init slab_proc_init(void)
4618 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4621 module_init(slab_proc_init);
4622 #endif /* CONFIG_SLABINFO */