2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemtrace.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
37 * The slab_lock protects operations on the object of a particular
38 * slab and its metadata in the page struct. If the slab lock
39 * has been taken then no allocations nor frees can be performed
40 * on the objects in the slab nor can the slab be added or removed
41 * from the partial or full lists since this would mean modifying
42 * the page_struct of the slab.
44 * The list_lock protects the partial and full list on each node and
45 * the partial slab counter. If taken then no new slabs may be added or
46 * removed from the lists nor make the number of partial slabs be modified.
47 * (Note that the total number of slabs is an atomic value that may be
48 * modified without taking the list lock).
50 * The list_lock is a centralized lock and thus we avoid taking it as
51 * much as possible. As long as SLUB does not have to handle partial
52 * slabs, operations can continue without any centralized lock. F.e.
53 * allocating a long series of objects that fill up slabs does not require
56 * The lock order is sometimes inverted when we are trying to get a slab
57 * off a list. We take the list_lock and then look for a page on the list
58 * to use. While we do that objects in the slabs may be freed. We can
59 * only operate on the slab if we have also taken the slab_lock. So we use
60 * a slab_trylock() on the slab. If trylock was successful then no frees
61 * can occur anymore and we can use the slab for allocations etc. If the
62 * slab_trylock() does not succeed then frees are in progress in the slab and
63 * we must stay away from it for a while since we may cause a bouncing
64 * cacheline if we try to acquire the lock. So go onto the next slab.
65 * If all pages are busy then we may allocate a new slab instead of reusing
66 * a partial slab. A new slab has noone operating on it and thus there is
67 * no danger of cacheline contention.
69 * Interrupts are disabled during allocation and deallocation in order to
70 * make the slab allocator safe to use in the context of an irq. In addition
71 * interrupts are disabled to ensure that the processor does not change
72 * while handling per_cpu slabs, due to kernel preemption.
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
77 * Slabs with free elements are kept on a partial list and during regular
78 * operations no list for full slabs is used. If an object in a full slab is
79 * freed then the slab will show up again on the partial lists.
80 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
83 * Slabs are freed when they become empty. Teardown and setup is
84 * minimal so we rely on the page allocators per cpu caches for
85 * fast frees and allocs.
87 * Overloading of page flags that are otherwise used for LRU management.
89 * PageActive The slab is frozen and exempt from list processing.
90 * This means that the slab is dedicated to a purpose
91 * such as satisfying allocations for a specific
92 * processor. Objects may be freed in the slab while
93 * it is frozen but slab_free will then skip the usual
94 * list operations. It is up to the processor holding
95 * the slab to integrate the slab into the slab lists
96 * when the slab is no longer needed.
98 * One use of this flag is to mark slabs that are
99 * used for allocations. Then such a slab becomes a cpu
100 * slab. The cpu slab may be equipped with an additional
101 * freelist that allows lockless access to
102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
107 * the fast path and disables lockless freelists.
110 #ifdef CONFIG_SLUB_DEBUG
117 * Issues still to be resolved:
119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
121 * - Variable sizing of the per node arrays
124 /* Enable to test recovery from slab corruption on boot */
125 #undef SLUB_RESILIENCY_TEST
128 * Mininum number of partial slabs. These will be left on the partial
129 * lists even if they are empty. kmem_cache_shrink may reclaim them.
131 #define MIN_PARTIAL 5
134 * Maximum number of desirable partial slabs.
135 * The existence of more partial slabs makes kmem_cache_shrink
136 * sort the partial list by the number of objects in the.
138 #define MAX_PARTIAL 10
140 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
141 SLAB_POISON | SLAB_STORE_USER)
144 * Debugging flags that require metadata to be stored in the slab. These get
145 * disabled when slub_debug=O is used and a cache's min order increases with
148 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
151 * Set of flags that will prevent slab merging
153 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
154 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
156 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
157 SLAB_CACHE_DMA | SLAB_NOTRACK)
159 #ifndef ARCH_KMALLOC_MINALIGN
160 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
163 #ifndef ARCH_SLAB_MINALIGN
164 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
168 #define OO_MASK ((1 << OO_SHIFT) - 1)
169 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
171 /* Internal SLUB flags */
172 #define __OBJECT_POISON 0x80000000 /* Poison object */
173 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
175 static int kmem_size = sizeof(struct kmem_cache);
178 static struct notifier_block slab_notifier;
182 DOWN, /* No slab functionality available */
183 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
184 UP, /* Everything works but does not show up in sysfs */
188 /* A list of all slab caches on the system */
189 static DECLARE_RWSEM(slub_lock);
190 static LIST_HEAD(slab_caches);
193 * Tracking user of a slab.
196 unsigned long addr; /* Called from address */
197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
202 enum track_item { TRACK_ALLOC, TRACK_FREE };
204 #ifdef CONFIG_SLUB_DEBUG
205 static int sysfs_slab_add(struct kmem_cache *);
206 static int sysfs_slab_alias(struct kmem_cache *, const char *);
207 static void sysfs_slab_remove(struct kmem_cache *);
210 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
213 static inline void sysfs_slab_remove(struct kmem_cache *s)
220 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
222 #ifdef CONFIG_SLUB_STATS
227 /********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
231 int slab_is_available(void)
233 return slab_state >= UP;
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
239 return s->node[node];
241 return &s->local_node;
245 /* Verify that a pointer has an address that is valid within a slab page */
246 static inline int check_valid_pointer(struct kmem_cache *s,
247 struct page *page, const void *object)
254 base = page_address(page);
255 if (object < base || object >= base + page->objects * s->size ||
256 (object - base) % s->size) {
263 static inline void *get_freepointer(struct kmem_cache *s, void *object)
265 return *(void **)(object + s->offset);
268 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
270 *(void **)(object + s->offset) = fp;
273 /* Loop over all objects in a slab */
274 #define for_each_object(__p, __s, __addr, __objects) \
275 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
279 #define for_each_free_object(__p, __s, __free) \
280 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
282 /* Determine object index from a given position */
283 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
285 return (p - addr) / s->size;
288 static inline struct kmem_cache_order_objects oo_make(int order,
291 struct kmem_cache_order_objects x = {
292 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
298 static inline int oo_order(struct kmem_cache_order_objects x)
300 return x.x >> OO_SHIFT;
303 static inline int oo_objects(struct kmem_cache_order_objects x)
305 return x.x & OO_MASK;
308 #ifdef CONFIG_SLUB_DEBUG
312 #ifdef CONFIG_SLUB_DEBUG_ON
313 static int slub_debug = DEBUG_DEFAULT_FLAGS;
315 static int slub_debug;
318 static char *slub_debug_slabs;
319 static int disable_higher_order_debug;
324 static void print_section(char *text, u8 *addr, unsigned int length)
332 for (i = 0; i < length; i++) {
334 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
337 printk(KERN_CONT " %02x", addr[i]);
339 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
341 printk(KERN_CONT " %s\n", ascii);
348 printk(KERN_CONT " ");
352 printk(KERN_CONT " %s\n", ascii);
356 static struct track *get_track(struct kmem_cache *s, void *object,
357 enum track_item alloc)
362 p = object + s->offset + sizeof(void *);
364 p = object + s->inuse;
369 static void set_track(struct kmem_cache *s, void *object,
370 enum track_item alloc, unsigned long addr)
372 struct track *p = get_track(s, object, alloc);
376 p->cpu = smp_processor_id();
377 p->pid = current->pid;
380 memset(p, 0, sizeof(struct track));
383 static void init_tracking(struct kmem_cache *s, void *object)
385 if (!(s->flags & SLAB_STORE_USER))
388 set_track(s, object, TRACK_FREE, 0UL);
389 set_track(s, object, TRACK_ALLOC, 0UL);
392 static void print_track(const char *s, struct track *t)
397 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
398 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
401 static void print_tracking(struct kmem_cache *s, void *object)
403 if (!(s->flags & SLAB_STORE_USER))
406 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
407 print_track("Freed", get_track(s, object, TRACK_FREE));
410 static void print_page_info(struct page *page)
412 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
413 page, page->objects, page->inuse, page->freelist, page->flags);
417 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
423 vsnprintf(buf, sizeof(buf), fmt, args);
425 printk(KERN_ERR "========================================"
426 "=====================================\n");
427 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
428 printk(KERN_ERR "----------------------------------------"
429 "-------------------------------------\n\n");
432 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
438 vsnprintf(buf, sizeof(buf), fmt, args);
440 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
443 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
445 unsigned int off; /* Offset of last byte */
446 u8 *addr = page_address(page);
448 print_tracking(s, p);
450 print_page_info(page);
452 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
453 p, p - addr, get_freepointer(s, p));
456 print_section("Bytes b4", p - 16, 16);
458 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
460 if (s->flags & SLAB_RED_ZONE)
461 print_section("Redzone", p + s->objsize,
462 s->inuse - s->objsize);
465 off = s->offset + sizeof(void *);
469 if (s->flags & SLAB_STORE_USER)
470 off += 2 * sizeof(struct track);
473 /* Beginning of the filler is the free pointer */
474 print_section("Padding", p + off, s->size - off);
479 static void object_err(struct kmem_cache *s, struct page *page,
480 u8 *object, char *reason)
482 slab_bug(s, "%s", reason);
483 print_trailer(s, page, object);
486 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
492 vsnprintf(buf, sizeof(buf), fmt, args);
494 slab_bug(s, "%s", buf);
495 print_page_info(page);
499 static void init_object(struct kmem_cache *s, void *object, int active)
503 if (s->flags & __OBJECT_POISON) {
504 memset(p, POISON_FREE, s->objsize - 1);
505 p[s->objsize - 1] = POISON_END;
508 if (s->flags & SLAB_RED_ZONE)
509 memset(p + s->objsize,
510 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
511 s->inuse - s->objsize);
514 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
517 if (*start != (u8)value)
525 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
526 void *from, void *to)
528 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
529 memset(from, data, to - from);
532 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
533 u8 *object, char *what,
534 u8 *start, unsigned int value, unsigned int bytes)
539 fault = check_bytes(start, value, bytes);
544 while (end > fault && end[-1] == value)
547 slab_bug(s, "%s overwritten", what);
548 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
549 fault, end - 1, fault[0], value);
550 print_trailer(s, page, object);
552 restore_bytes(s, what, value, fault, end);
560 * Bytes of the object to be managed.
561 * If the freepointer may overlay the object then the free
562 * pointer is the first word of the object.
564 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
567 * object + s->objsize
568 * Padding to reach word boundary. This is also used for Redzoning.
569 * Padding is extended by another word if Redzoning is enabled and
572 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
573 * 0xcc (RED_ACTIVE) for objects in use.
576 * Meta data starts here.
578 * A. Free pointer (if we cannot overwrite object on free)
579 * B. Tracking data for SLAB_STORE_USER
580 * C. Padding to reach required alignment boundary or at mininum
581 * one word if debugging is on to be able to detect writes
582 * before the word boundary.
584 * Padding is done using 0x5a (POISON_INUSE)
587 * Nothing is used beyond s->size.
589 * If slabcaches are merged then the objsize and inuse boundaries are mostly
590 * ignored. And therefore no slab options that rely on these boundaries
591 * may be used with merged slabcaches.
594 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
596 unsigned long off = s->inuse; /* The end of info */
599 /* Freepointer is placed after the object. */
600 off += sizeof(void *);
602 if (s->flags & SLAB_STORE_USER)
603 /* We also have user information there */
604 off += 2 * sizeof(struct track);
609 return check_bytes_and_report(s, page, p, "Object padding",
610 p + off, POISON_INUSE, s->size - off);
613 /* Check the pad bytes at the end of a slab page */
614 static int slab_pad_check(struct kmem_cache *s, struct page *page)
622 if (!(s->flags & SLAB_POISON))
625 start = page_address(page);
626 length = (PAGE_SIZE << compound_order(page));
627 end = start + length;
628 remainder = length % s->size;
632 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
635 while (end > fault && end[-1] == POISON_INUSE)
638 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
639 print_section("Padding", end - remainder, remainder);
641 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
645 static int check_object(struct kmem_cache *s, struct page *page,
646 void *object, int active)
649 u8 *endobject = object + s->objsize;
651 if (s->flags & SLAB_RED_ZONE) {
653 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
655 if (!check_bytes_and_report(s, page, object, "Redzone",
656 endobject, red, s->inuse - s->objsize))
659 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
660 check_bytes_and_report(s, page, p, "Alignment padding",
661 endobject, POISON_INUSE, s->inuse - s->objsize);
665 if (s->flags & SLAB_POISON) {
666 if (!active && (s->flags & __OBJECT_POISON) &&
667 (!check_bytes_and_report(s, page, p, "Poison", p,
668 POISON_FREE, s->objsize - 1) ||
669 !check_bytes_and_report(s, page, p, "Poison",
670 p + s->objsize - 1, POISON_END, 1)))
673 * check_pad_bytes cleans up on its own.
675 check_pad_bytes(s, page, p);
678 if (!s->offset && active)
680 * Object and freepointer overlap. Cannot check
681 * freepointer while object is allocated.
685 /* Check free pointer validity */
686 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
687 object_err(s, page, p, "Freepointer corrupt");
689 * No choice but to zap it and thus lose the remainder
690 * of the free objects in this slab. May cause
691 * another error because the object count is now wrong.
693 set_freepointer(s, p, NULL);
699 static int check_slab(struct kmem_cache *s, struct page *page)
703 VM_BUG_ON(!irqs_disabled());
705 if (!PageSlab(page)) {
706 slab_err(s, page, "Not a valid slab page");
710 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
711 if (page->objects > maxobj) {
712 slab_err(s, page, "objects %u > max %u",
713 s->name, page->objects, maxobj);
716 if (page->inuse > page->objects) {
717 slab_err(s, page, "inuse %u > max %u",
718 s->name, page->inuse, page->objects);
721 /* Slab_pad_check fixes things up after itself */
722 slab_pad_check(s, page);
727 * Determine if a certain object on a page is on the freelist. Must hold the
728 * slab lock to guarantee that the chains are in a consistent state.
730 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
733 void *fp = page->freelist;
735 unsigned long max_objects;
737 while (fp && nr <= page->objects) {
740 if (!check_valid_pointer(s, page, fp)) {
742 object_err(s, page, object,
743 "Freechain corrupt");
744 set_freepointer(s, object, NULL);
747 slab_err(s, page, "Freepointer corrupt");
748 page->freelist = NULL;
749 page->inuse = page->objects;
750 slab_fix(s, "Freelist cleared");
756 fp = get_freepointer(s, object);
760 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
761 if (max_objects > MAX_OBJS_PER_PAGE)
762 max_objects = MAX_OBJS_PER_PAGE;
764 if (page->objects != max_objects) {
765 slab_err(s, page, "Wrong number of objects. Found %d but "
766 "should be %d", page->objects, max_objects);
767 page->objects = max_objects;
768 slab_fix(s, "Number of objects adjusted.");
770 if (page->inuse != page->objects - nr) {
771 slab_err(s, page, "Wrong object count. Counter is %d but "
772 "counted were %d", page->inuse, page->objects - nr);
773 page->inuse = page->objects - nr;
774 slab_fix(s, "Object count adjusted.");
776 return search == NULL;
779 static void trace(struct kmem_cache *s, struct page *page, void *object,
782 if (s->flags & SLAB_TRACE) {
783 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
785 alloc ? "alloc" : "free",
790 print_section("Object", (void *)object, s->objsize);
797 * Tracking of fully allocated slabs for debugging purposes.
799 static void add_full(struct kmem_cache_node *n, struct page *page)
801 spin_lock(&n->list_lock);
802 list_add(&page->lru, &n->full);
803 spin_unlock(&n->list_lock);
806 static void remove_full(struct kmem_cache *s, struct page *page)
808 struct kmem_cache_node *n;
810 if (!(s->flags & SLAB_STORE_USER))
813 n = get_node(s, page_to_nid(page));
815 spin_lock(&n->list_lock);
816 list_del(&page->lru);
817 spin_unlock(&n->list_lock);
820 /* Tracking of the number of slabs for debugging purposes */
821 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
823 struct kmem_cache_node *n = get_node(s, node);
825 return atomic_long_read(&n->nr_slabs);
828 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
830 return atomic_long_read(&n->nr_slabs);
833 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
835 struct kmem_cache_node *n = get_node(s, node);
838 * May be called early in order to allocate a slab for the
839 * kmem_cache_node structure. Solve the chicken-egg
840 * dilemma by deferring the increment of the count during
841 * bootstrap (see early_kmem_cache_node_alloc).
843 if (!NUMA_BUILD || n) {
844 atomic_long_inc(&n->nr_slabs);
845 atomic_long_add(objects, &n->total_objects);
848 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
850 struct kmem_cache_node *n = get_node(s, node);
852 atomic_long_dec(&n->nr_slabs);
853 atomic_long_sub(objects, &n->total_objects);
856 /* Object debug checks for alloc/free paths */
857 static void setup_object_debug(struct kmem_cache *s, struct page *page,
860 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
863 init_object(s, object, 0);
864 init_tracking(s, object);
867 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
868 void *object, unsigned long addr)
870 if (!check_slab(s, page))
873 if (!on_freelist(s, page, object)) {
874 object_err(s, page, object, "Object already allocated");
878 if (!check_valid_pointer(s, page, object)) {
879 object_err(s, page, object, "Freelist Pointer check fails");
883 if (!check_object(s, page, object, 0))
886 /* Success perform special debug activities for allocs */
887 if (s->flags & SLAB_STORE_USER)
888 set_track(s, object, TRACK_ALLOC, addr);
889 trace(s, page, object, 1);
890 init_object(s, object, 1);
894 if (PageSlab(page)) {
896 * If this is a slab page then lets do the best we can
897 * to avoid issues in the future. Marking all objects
898 * as used avoids touching the remaining objects.
900 slab_fix(s, "Marking all objects used");
901 page->inuse = page->objects;
902 page->freelist = NULL;
907 static int free_debug_processing(struct kmem_cache *s, struct page *page,
908 void *object, unsigned long addr)
910 if (!check_slab(s, page))
913 if (!check_valid_pointer(s, page, object)) {
914 slab_err(s, page, "Invalid object pointer 0x%p", object);
918 if (on_freelist(s, page, object)) {
919 object_err(s, page, object, "Object already free");
923 if (!check_object(s, page, object, 1))
926 if (unlikely(s != page->slab)) {
927 if (!PageSlab(page)) {
928 slab_err(s, page, "Attempt to free object(0x%p) "
929 "outside of slab", object);
930 } else if (!page->slab) {
932 "SLUB <none>: no slab for object 0x%p.\n",
936 object_err(s, page, object,
937 "page slab pointer corrupt.");
941 /* Special debug activities for freeing objects */
942 if (!PageSlubFrozen(page) && !page->freelist)
943 remove_full(s, page);
944 if (s->flags & SLAB_STORE_USER)
945 set_track(s, object, TRACK_FREE, addr);
946 trace(s, page, object, 0);
947 init_object(s, object, 0);
951 slab_fix(s, "Object at 0x%p not freed", object);
955 static int __init setup_slub_debug(char *str)
957 slub_debug = DEBUG_DEFAULT_FLAGS;
958 if (*str++ != '=' || !*str)
960 * No options specified. Switch on full debugging.
966 * No options but restriction on slabs. This means full
967 * debugging for slabs matching a pattern.
971 if (tolower(*str) == 'o') {
973 * Avoid enabling debugging on caches if its minimum order
974 * would increase as a result.
976 disable_higher_order_debug = 1;
983 * Switch off all debugging measures.
988 * Determine which debug features should be switched on
990 for (; *str && *str != ','; str++) {
991 switch (tolower(*str)) {
993 slub_debug |= SLAB_DEBUG_FREE;
996 slub_debug |= SLAB_RED_ZONE;
999 slub_debug |= SLAB_POISON;
1002 slub_debug |= SLAB_STORE_USER;
1005 slub_debug |= SLAB_TRACE;
1008 printk(KERN_ERR "slub_debug option '%c' "
1009 "unknown. skipped\n", *str);
1015 slub_debug_slabs = str + 1;
1020 __setup("slub_debug", setup_slub_debug);
1022 static unsigned long kmem_cache_flags(unsigned long objsize,
1023 unsigned long flags, const char *name,
1024 void (*ctor)(void *))
1027 * Enable debugging if selected on the kernel commandline.
1029 if (slub_debug && (!slub_debug_slabs ||
1030 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1031 flags |= slub_debug;
1036 static inline void setup_object_debug(struct kmem_cache *s,
1037 struct page *page, void *object) {}
1039 static inline int alloc_debug_processing(struct kmem_cache *s,
1040 struct page *page, void *object, unsigned long addr) { return 0; }
1042 static inline int free_debug_processing(struct kmem_cache *s,
1043 struct page *page, void *object, unsigned long addr) { return 0; }
1045 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1047 static inline int check_object(struct kmem_cache *s, struct page *page,
1048 void *object, int active) { return 1; }
1049 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1050 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1051 unsigned long flags, const char *name,
1052 void (*ctor)(void *))
1056 #define slub_debug 0
1058 #define disable_higher_order_debug 0
1060 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1062 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1064 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1066 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1071 * Slab allocation and freeing
1073 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1074 struct kmem_cache_order_objects oo)
1076 int order = oo_order(oo);
1078 flags |= __GFP_NOTRACK;
1081 return alloc_pages(flags, order);
1083 return alloc_pages_node(node, flags, order);
1086 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1089 struct kmem_cache_order_objects oo = s->oo;
1092 flags |= s->allocflags;
1095 * Let the initial higher-order allocation fail under memory pressure
1096 * so we fall-back to the minimum order allocation.
1098 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1100 page = alloc_slab_page(alloc_gfp, node, oo);
1101 if (unlikely(!page)) {
1104 * Allocation may have failed due to fragmentation.
1105 * Try a lower order alloc if possible
1107 page = alloc_slab_page(flags, node, oo);
1111 stat(this_cpu_ptr(s->cpu_slab), ORDER_FALLBACK);
1114 if (kmemcheck_enabled
1115 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1116 int pages = 1 << oo_order(oo);
1118 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1121 * Objects from caches that have a constructor don't get
1122 * cleared when they're allocated, so we need to do it here.
1125 kmemcheck_mark_uninitialized_pages(page, pages);
1127 kmemcheck_mark_unallocated_pages(page, pages);
1130 page->objects = oo_objects(oo);
1131 mod_zone_page_state(page_zone(page),
1132 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1133 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1139 static void setup_object(struct kmem_cache *s, struct page *page,
1142 setup_object_debug(s, page, object);
1143 if (unlikely(s->ctor))
1147 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1154 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1156 page = allocate_slab(s,
1157 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1161 inc_slabs_node(s, page_to_nid(page), page->objects);
1163 page->flags |= 1 << PG_slab;
1164 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1165 SLAB_STORE_USER | SLAB_TRACE))
1166 __SetPageSlubDebug(page);
1168 start = page_address(page);
1170 if (unlikely(s->flags & SLAB_POISON))
1171 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1174 for_each_object(p, s, start, page->objects) {
1175 setup_object(s, page, last);
1176 set_freepointer(s, last, p);
1179 setup_object(s, page, last);
1180 set_freepointer(s, last, NULL);
1182 page->freelist = start;
1188 static void __free_slab(struct kmem_cache *s, struct page *page)
1190 int order = compound_order(page);
1191 int pages = 1 << order;
1193 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1196 slab_pad_check(s, page);
1197 for_each_object(p, s, page_address(page),
1199 check_object(s, page, p, 0);
1200 __ClearPageSlubDebug(page);
1203 kmemcheck_free_shadow(page, compound_order(page));
1205 mod_zone_page_state(page_zone(page),
1206 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1207 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1210 __ClearPageSlab(page);
1211 reset_page_mapcount(page);
1212 if (current->reclaim_state)
1213 current->reclaim_state->reclaimed_slab += pages;
1214 __free_pages(page, order);
1217 static void rcu_free_slab(struct rcu_head *h)
1221 page = container_of((struct list_head *)h, struct page, lru);
1222 __free_slab(page->slab, page);
1225 static void free_slab(struct kmem_cache *s, struct page *page)
1227 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1229 * RCU free overloads the RCU head over the LRU
1231 struct rcu_head *head = (void *)&page->lru;
1233 call_rcu(head, rcu_free_slab);
1235 __free_slab(s, page);
1238 static void discard_slab(struct kmem_cache *s, struct page *page)
1240 dec_slabs_node(s, page_to_nid(page), page->objects);
1245 * Per slab locking using the pagelock
1247 static __always_inline void slab_lock(struct page *page)
1249 bit_spin_lock(PG_locked, &page->flags);
1252 static __always_inline void slab_unlock(struct page *page)
1254 __bit_spin_unlock(PG_locked, &page->flags);
1257 static __always_inline int slab_trylock(struct page *page)
1261 rc = bit_spin_trylock(PG_locked, &page->flags);
1266 * Management of partially allocated slabs
1268 static void add_partial(struct kmem_cache_node *n,
1269 struct page *page, int tail)
1271 spin_lock(&n->list_lock);
1274 list_add_tail(&page->lru, &n->partial);
1276 list_add(&page->lru, &n->partial);
1277 spin_unlock(&n->list_lock);
1280 static void remove_partial(struct kmem_cache *s, struct page *page)
1282 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1284 spin_lock(&n->list_lock);
1285 list_del(&page->lru);
1287 spin_unlock(&n->list_lock);
1291 * Lock slab and remove from the partial list.
1293 * Must hold list_lock.
1295 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1298 if (slab_trylock(page)) {
1299 list_del(&page->lru);
1301 __SetPageSlubFrozen(page);
1308 * Try to allocate a partial slab from a specific node.
1310 static struct page *get_partial_node(struct kmem_cache_node *n)
1315 * Racy check. If we mistakenly see no partial slabs then we
1316 * just allocate an empty slab. If we mistakenly try to get a
1317 * partial slab and there is none available then get_partials()
1320 if (!n || !n->nr_partial)
1323 spin_lock(&n->list_lock);
1324 list_for_each_entry(page, &n->partial, lru)
1325 if (lock_and_freeze_slab(n, page))
1329 spin_unlock(&n->list_lock);
1334 * Get a page from somewhere. Search in increasing NUMA distances.
1336 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1339 struct zonelist *zonelist;
1342 enum zone_type high_zoneidx = gfp_zone(flags);
1346 * The defrag ratio allows a configuration of the tradeoffs between
1347 * inter node defragmentation and node local allocations. A lower
1348 * defrag_ratio increases the tendency to do local allocations
1349 * instead of attempting to obtain partial slabs from other nodes.
1351 * If the defrag_ratio is set to 0 then kmalloc() always
1352 * returns node local objects. If the ratio is higher then kmalloc()
1353 * may return off node objects because partial slabs are obtained
1354 * from other nodes and filled up.
1356 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1357 * defrag_ratio = 1000) then every (well almost) allocation will
1358 * first attempt to defrag slab caches on other nodes. This means
1359 * scanning over all nodes to look for partial slabs which may be
1360 * expensive if we do it every time we are trying to find a slab
1361 * with available objects.
1363 if (!s->remote_node_defrag_ratio ||
1364 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1367 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1368 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1369 struct kmem_cache_node *n;
1371 n = get_node(s, zone_to_nid(zone));
1373 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1374 n->nr_partial > s->min_partial) {
1375 page = get_partial_node(n);
1385 * Get a partial page, lock it and return it.
1387 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1390 int searchnode = (node == -1) ? numa_node_id() : node;
1392 page = get_partial_node(get_node(s, searchnode));
1393 if (page || (flags & __GFP_THISNODE))
1396 return get_any_partial(s, flags);
1400 * Move a page back to the lists.
1402 * Must be called with the slab lock held.
1404 * On exit the slab lock will have been dropped.
1406 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1408 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1409 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1411 __ClearPageSlubFrozen(page);
1414 if (page->freelist) {
1415 add_partial(n, page, tail);
1416 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1418 stat(c, DEACTIVATE_FULL);
1419 if (SLABDEBUG && PageSlubDebug(page) &&
1420 (s->flags & SLAB_STORE_USER))
1425 stat(c, DEACTIVATE_EMPTY);
1426 if (n->nr_partial < s->min_partial) {
1428 * Adding an empty slab to the partial slabs in order
1429 * to avoid page allocator overhead. This slab needs
1430 * to come after the other slabs with objects in
1431 * so that the others get filled first. That way the
1432 * size of the partial list stays small.
1434 * kmem_cache_shrink can reclaim any empty slabs from
1437 add_partial(n, page, 1);
1441 stat(__this_cpu_ptr(s->cpu_slab), FREE_SLAB);
1442 discard_slab(s, page);
1448 * Remove the cpu slab
1450 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1452 struct page *page = c->page;
1456 stat(c, DEACTIVATE_REMOTE_FREES);
1458 * Merge cpu freelist into slab freelist. Typically we get here
1459 * because both freelists are empty. So this is unlikely
1462 while (unlikely(c->freelist)) {
1465 tail = 0; /* Hot objects. Put the slab first */
1467 /* Retrieve object from cpu_freelist */
1468 object = c->freelist;
1469 c->freelist = get_freepointer(s, c->freelist);
1471 /* And put onto the regular freelist */
1472 set_freepointer(s, object, page->freelist);
1473 page->freelist = object;
1477 unfreeze_slab(s, page, tail);
1480 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1482 stat(c, CPUSLAB_FLUSH);
1484 deactivate_slab(s, c);
1490 * Called from IPI handler with interrupts disabled.
1492 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1494 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1496 if (likely(c && c->page))
1500 static void flush_cpu_slab(void *d)
1502 struct kmem_cache *s = d;
1504 __flush_cpu_slab(s, smp_processor_id());
1507 static void flush_all(struct kmem_cache *s)
1509 on_each_cpu(flush_cpu_slab, s, 1);
1513 * Check if the objects in a per cpu structure fit numa
1514 * locality expectations.
1516 static inline int node_match(struct kmem_cache_cpu *c, int node)
1519 if (node != -1 && c->node != node)
1525 static int count_free(struct page *page)
1527 return page->objects - page->inuse;
1530 static unsigned long count_partial(struct kmem_cache_node *n,
1531 int (*get_count)(struct page *))
1533 unsigned long flags;
1534 unsigned long x = 0;
1537 spin_lock_irqsave(&n->list_lock, flags);
1538 list_for_each_entry(page, &n->partial, lru)
1539 x += get_count(page);
1540 spin_unlock_irqrestore(&n->list_lock, flags);
1544 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1546 #ifdef CONFIG_SLUB_DEBUG
1547 return atomic_long_read(&n->total_objects);
1553 static noinline void
1554 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1559 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1561 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1562 "default order: %d, min order: %d\n", s->name, s->objsize,
1563 s->size, oo_order(s->oo), oo_order(s->min));
1565 if (oo_order(s->min) > get_order(s->objsize))
1566 printk(KERN_WARNING " %s debugging increased min order, use "
1567 "slub_debug=O to disable.\n", s->name);
1569 for_each_online_node(node) {
1570 struct kmem_cache_node *n = get_node(s, node);
1571 unsigned long nr_slabs;
1572 unsigned long nr_objs;
1573 unsigned long nr_free;
1578 nr_free = count_partial(n, count_free);
1579 nr_slabs = node_nr_slabs(n);
1580 nr_objs = node_nr_objs(n);
1583 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1584 node, nr_slabs, nr_objs, nr_free);
1589 * Slow path. The lockless freelist is empty or we need to perform
1592 * Interrupts are disabled.
1594 * Processing is still very fast if new objects have been freed to the
1595 * regular freelist. In that case we simply take over the regular freelist
1596 * as the lockless freelist and zap the regular freelist.
1598 * If that is not working then we fall back to the partial lists. We take the
1599 * first element of the freelist as the object to allocate now and move the
1600 * rest of the freelist to the lockless freelist.
1602 * And if we were unable to get a new slab from the partial slab lists then
1603 * we need to allocate a new slab. This is the slowest path since it involves
1604 * a call to the page allocator and the setup of a new slab.
1606 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1607 unsigned long addr, struct kmem_cache_cpu *c)
1612 /* We handle __GFP_ZERO in the caller */
1613 gfpflags &= ~__GFP_ZERO;
1619 if (unlikely(!node_match(c, node)))
1622 stat(c, ALLOC_REFILL);
1625 object = c->page->freelist;
1626 if (unlikely(!object))
1628 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1631 c->freelist = get_freepointer(s, object);
1632 c->page->inuse = c->page->objects;
1633 c->page->freelist = NULL;
1634 c->node = page_to_nid(c->page);
1636 slab_unlock(c->page);
1637 stat(c, ALLOC_SLOWPATH);
1641 deactivate_slab(s, c);
1644 new = get_partial(s, gfpflags, node);
1647 stat(c, ALLOC_FROM_PARTIAL);
1651 if (gfpflags & __GFP_WAIT)
1654 new = new_slab(s, gfpflags, node);
1656 if (gfpflags & __GFP_WAIT)
1657 local_irq_disable();
1660 c = __this_cpu_ptr(s->cpu_slab);
1661 stat(c, ALLOC_SLAB);
1665 __SetPageSlubFrozen(new);
1669 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1670 slab_out_of_memory(s, gfpflags, node);
1673 if (!alloc_debug_processing(s, c->page, object, addr))
1677 c->page->freelist = get_freepointer(s, object);
1683 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1684 * have the fastpath folded into their functions. So no function call
1685 * overhead for requests that can be satisfied on the fastpath.
1687 * The fastpath works by first checking if the lockless freelist can be used.
1688 * If not then __slab_alloc is called for slow processing.
1690 * Otherwise we can simply pick the next object from the lockless free list.
1692 static __always_inline void *slab_alloc(struct kmem_cache *s,
1693 gfp_t gfpflags, int node, unsigned long addr)
1696 struct kmem_cache_cpu *c;
1697 unsigned long flags;
1699 gfpflags &= gfp_allowed_mask;
1701 lockdep_trace_alloc(gfpflags);
1702 might_sleep_if(gfpflags & __GFP_WAIT);
1704 if (should_failslab(s->objsize, gfpflags))
1707 local_irq_save(flags);
1708 c = __this_cpu_ptr(s->cpu_slab);
1709 object = c->freelist;
1710 if (unlikely(!object || !node_match(c, node)))
1712 object = __slab_alloc(s, gfpflags, node, addr, c);
1715 c->freelist = get_freepointer(s, object);
1716 stat(c, ALLOC_FASTPATH);
1718 local_irq_restore(flags);
1720 if (unlikely(gfpflags & __GFP_ZERO) && object)
1721 memset(object, 0, s->objsize);
1723 kmemcheck_slab_alloc(s, gfpflags, object, s->objsize);
1724 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, gfpflags);
1729 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1731 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1733 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1737 EXPORT_SYMBOL(kmem_cache_alloc);
1739 #ifdef CONFIG_TRACING
1740 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1742 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1744 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1748 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1750 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1752 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1753 s->objsize, s->size, gfpflags, node);
1757 EXPORT_SYMBOL(kmem_cache_alloc_node);
1760 #ifdef CONFIG_TRACING
1761 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1765 return slab_alloc(s, gfpflags, node, _RET_IP_);
1767 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1771 * Slow patch handling. This may still be called frequently since objects
1772 * have a longer lifetime than the cpu slabs in most processing loads.
1774 * So we still attempt to reduce cache line usage. Just take the slab
1775 * lock and free the item. If there is no additional partial page
1776 * handling required then we can return immediately.
1778 static void __slab_free(struct kmem_cache *s, struct page *page,
1779 void *x, unsigned long addr)
1782 void **object = (void *)x;
1783 struct kmem_cache_cpu *c;
1785 c = __this_cpu_ptr(s->cpu_slab);
1786 stat(c, FREE_SLOWPATH);
1789 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1793 prior = page->freelist;
1794 set_freepointer(s, object, prior);
1795 page->freelist = object;
1798 if (unlikely(PageSlubFrozen(page))) {
1799 stat(c, FREE_FROZEN);
1803 if (unlikely(!page->inuse))
1807 * Objects left in the slab. If it was not on the partial list before
1810 if (unlikely(!prior)) {
1811 add_partial(get_node(s, page_to_nid(page)), page, 1);
1812 stat(c, FREE_ADD_PARTIAL);
1822 * Slab still on the partial list.
1824 remove_partial(s, page);
1825 stat(c, FREE_REMOVE_PARTIAL);
1829 discard_slab(s, page);
1833 if (!free_debug_processing(s, page, x, addr))
1839 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1840 * can perform fastpath freeing without additional function calls.
1842 * The fastpath is only possible if we are freeing to the current cpu slab
1843 * of this processor. This typically the case if we have just allocated
1846 * If fastpath is not possible then fall back to __slab_free where we deal
1847 * with all sorts of special processing.
1849 static __always_inline void slab_free(struct kmem_cache *s,
1850 struct page *page, void *x, unsigned long addr)
1852 void **object = (void *)x;
1853 struct kmem_cache_cpu *c;
1854 unsigned long flags;
1856 kmemleak_free_recursive(x, s->flags);
1857 local_irq_save(flags);
1858 c = __this_cpu_ptr(s->cpu_slab);
1859 kmemcheck_slab_free(s, object, s->objsize);
1860 debug_check_no_locks_freed(object, s->objsize);
1861 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1862 debug_check_no_obj_freed(object, s->objsize);
1863 if (likely(page == c->page && c->node >= 0)) {
1864 set_freepointer(s, object, c->freelist);
1865 c->freelist = object;
1866 stat(c, FREE_FASTPATH);
1868 __slab_free(s, page, x, addr);
1870 local_irq_restore(flags);
1873 void kmem_cache_free(struct kmem_cache *s, void *x)
1877 page = virt_to_head_page(x);
1879 slab_free(s, page, x, _RET_IP_);
1881 trace_kmem_cache_free(_RET_IP_, x);
1883 EXPORT_SYMBOL(kmem_cache_free);
1885 /* Figure out on which slab page the object resides */
1886 static struct page *get_object_page(const void *x)
1888 struct page *page = virt_to_head_page(x);
1890 if (!PageSlab(page))
1897 * Object placement in a slab is made very easy because we always start at
1898 * offset 0. If we tune the size of the object to the alignment then we can
1899 * get the required alignment by putting one properly sized object after
1902 * Notice that the allocation order determines the sizes of the per cpu
1903 * caches. Each processor has always one slab available for allocations.
1904 * Increasing the allocation order reduces the number of times that slabs
1905 * must be moved on and off the partial lists and is therefore a factor in
1910 * Mininum / Maximum order of slab pages. This influences locking overhead
1911 * and slab fragmentation. A higher order reduces the number of partial slabs
1912 * and increases the number of allocations possible without having to
1913 * take the list_lock.
1915 static int slub_min_order;
1916 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1917 static int slub_min_objects;
1920 * Merge control. If this is set then no merging of slab caches will occur.
1921 * (Could be removed. This was introduced to pacify the merge skeptics.)
1923 static int slub_nomerge;
1926 * Calculate the order of allocation given an slab object size.
1928 * The order of allocation has significant impact on performance and other
1929 * system components. Generally order 0 allocations should be preferred since
1930 * order 0 does not cause fragmentation in the page allocator. Larger objects
1931 * be problematic to put into order 0 slabs because there may be too much
1932 * unused space left. We go to a higher order if more than 1/16th of the slab
1935 * In order to reach satisfactory performance we must ensure that a minimum
1936 * number of objects is in one slab. Otherwise we may generate too much
1937 * activity on the partial lists which requires taking the list_lock. This is
1938 * less a concern for large slabs though which are rarely used.
1940 * slub_max_order specifies the order where we begin to stop considering the
1941 * number of objects in a slab as critical. If we reach slub_max_order then
1942 * we try to keep the page order as low as possible. So we accept more waste
1943 * of space in favor of a small page order.
1945 * Higher order allocations also allow the placement of more objects in a
1946 * slab and thereby reduce object handling overhead. If the user has
1947 * requested a higher mininum order then we start with that one instead of
1948 * the smallest order which will fit the object.
1950 static inline int slab_order(int size, int min_objects,
1951 int max_order, int fract_leftover)
1955 int min_order = slub_min_order;
1957 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1958 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1960 for (order = max(min_order,
1961 fls(min_objects * size - 1) - PAGE_SHIFT);
1962 order <= max_order; order++) {
1964 unsigned long slab_size = PAGE_SIZE << order;
1966 if (slab_size < min_objects * size)
1969 rem = slab_size % size;
1971 if (rem <= slab_size / fract_leftover)
1979 static inline int calculate_order(int size)
1987 * Attempt to find best configuration for a slab. This
1988 * works by first attempting to generate a layout with
1989 * the best configuration and backing off gradually.
1991 * First we reduce the acceptable waste in a slab. Then
1992 * we reduce the minimum objects required in a slab.
1994 min_objects = slub_min_objects;
1996 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1997 max_objects = (PAGE_SIZE << slub_max_order)/size;
1998 min_objects = min(min_objects, max_objects);
2000 while (min_objects > 1) {
2002 while (fraction >= 4) {
2003 order = slab_order(size, min_objects,
2004 slub_max_order, fraction);
2005 if (order <= slub_max_order)
2013 * We were unable to place multiple objects in a slab. Now
2014 * lets see if we can place a single object there.
2016 order = slab_order(size, 1, slub_max_order, 1);
2017 if (order <= slub_max_order)
2021 * Doh this slab cannot be placed using slub_max_order.
2023 order = slab_order(size, 1, MAX_ORDER, 1);
2024 if (order < MAX_ORDER)
2030 * Figure out what the alignment of the objects will be.
2032 static unsigned long calculate_alignment(unsigned long flags,
2033 unsigned long align, unsigned long size)
2036 * If the user wants hardware cache aligned objects then follow that
2037 * suggestion if the object is sufficiently large.
2039 * The hardware cache alignment cannot override the specified
2040 * alignment though. If that is greater then use it.
2042 if (flags & SLAB_HWCACHE_ALIGN) {
2043 unsigned long ralign = cache_line_size();
2044 while (size <= ralign / 2)
2046 align = max(align, ralign);
2049 if (align < ARCH_SLAB_MINALIGN)
2050 align = ARCH_SLAB_MINALIGN;
2052 return ALIGN(align, sizeof(void *));
2056 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2059 spin_lock_init(&n->list_lock);
2060 INIT_LIST_HEAD(&n->partial);
2061 #ifdef CONFIG_SLUB_DEBUG
2062 atomic_long_set(&n->nr_slabs, 0);
2063 atomic_long_set(&n->total_objects, 0);
2064 INIT_LIST_HEAD(&n->full);
2068 static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[SLUB_PAGE_SHIFT]);
2070 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2072 if (s < kmalloc_caches + KMALLOC_CACHES && s >= kmalloc_caches)
2074 * Boot time creation of the kmalloc array. Use static per cpu data
2075 * since the per cpu allocator is not available yet.
2077 s->cpu_slab = per_cpu_var(kmalloc_percpu) + (s - kmalloc_caches);
2079 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2089 * No kmalloc_node yet so do it by hand. We know that this is the first
2090 * slab on the node for this slabcache. There are no concurrent accesses
2093 * Note that this function only works on the kmalloc_node_cache
2094 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2095 * memory on a fresh node that has no slab structures yet.
2097 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2100 struct kmem_cache_node *n;
2101 unsigned long flags;
2103 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2105 page = new_slab(kmalloc_caches, gfpflags, node);
2108 if (page_to_nid(page) != node) {
2109 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2111 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2112 "in order to be able to continue\n");
2117 page->freelist = get_freepointer(kmalloc_caches, n);
2119 kmalloc_caches->node[node] = n;
2120 #ifdef CONFIG_SLUB_DEBUG
2121 init_object(kmalloc_caches, n, 1);
2122 init_tracking(kmalloc_caches, n);
2124 init_kmem_cache_node(n, kmalloc_caches);
2125 inc_slabs_node(kmalloc_caches, node, page->objects);
2128 * lockdep requires consistent irq usage for each lock
2129 * so even though there cannot be a race this early in
2130 * the boot sequence, we still disable irqs.
2132 local_irq_save(flags);
2133 add_partial(n, page, 0);
2134 local_irq_restore(flags);
2137 static void free_kmem_cache_nodes(struct kmem_cache *s)
2141 for_each_node_state(node, N_NORMAL_MEMORY) {
2142 struct kmem_cache_node *n = s->node[node];
2143 if (n && n != &s->local_node)
2144 kmem_cache_free(kmalloc_caches, n);
2145 s->node[node] = NULL;
2149 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2154 if (slab_state >= UP)
2155 local_node = page_to_nid(virt_to_page(s));
2159 for_each_node_state(node, N_NORMAL_MEMORY) {
2160 struct kmem_cache_node *n;
2162 if (local_node == node)
2165 if (slab_state == DOWN) {
2166 early_kmem_cache_node_alloc(gfpflags, node);
2169 n = kmem_cache_alloc_node(kmalloc_caches,
2173 free_kmem_cache_nodes(s);
2179 init_kmem_cache_node(n, s);
2184 static void free_kmem_cache_nodes(struct kmem_cache *s)
2188 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2190 init_kmem_cache_node(&s->local_node, s);
2195 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2197 if (min < MIN_PARTIAL)
2199 else if (min > MAX_PARTIAL)
2201 s->min_partial = min;
2205 * calculate_sizes() determines the order and the distribution of data within
2208 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2210 unsigned long flags = s->flags;
2211 unsigned long size = s->objsize;
2212 unsigned long align = s->align;
2216 * Round up object size to the next word boundary. We can only
2217 * place the free pointer at word boundaries and this determines
2218 * the possible location of the free pointer.
2220 size = ALIGN(size, sizeof(void *));
2222 #ifdef CONFIG_SLUB_DEBUG
2224 * Determine if we can poison the object itself. If the user of
2225 * the slab may touch the object after free or before allocation
2226 * then we should never poison the object itself.
2228 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2230 s->flags |= __OBJECT_POISON;
2232 s->flags &= ~__OBJECT_POISON;
2236 * If we are Redzoning then check if there is some space between the
2237 * end of the object and the free pointer. If not then add an
2238 * additional word to have some bytes to store Redzone information.
2240 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2241 size += sizeof(void *);
2245 * With that we have determined the number of bytes in actual use
2246 * by the object. This is the potential offset to the free pointer.
2250 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2253 * Relocate free pointer after the object if it is not
2254 * permitted to overwrite the first word of the object on
2257 * This is the case if we do RCU, have a constructor or
2258 * destructor or are poisoning the objects.
2261 size += sizeof(void *);
2264 #ifdef CONFIG_SLUB_DEBUG
2265 if (flags & SLAB_STORE_USER)
2267 * Need to store information about allocs and frees after
2270 size += 2 * sizeof(struct track);
2272 if (flags & SLAB_RED_ZONE)
2274 * Add some empty padding so that we can catch
2275 * overwrites from earlier objects rather than let
2276 * tracking information or the free pointer be
2277 * corrupted if a user writes before the start
2280 size += sizeof(void *);
2284 * Determine the alignment based on various parameters that the
2285 * user specified and the dynamic determination of cache line size
2288 align = calculate_alignment(flags, align, s->objsize);
2292 * SLUB stores one object immediately after another beginning from
2293 * offset 0. In order to align the objects we have to simply size
2294 * each object to conform to the alignment.
2296 size = ALIGN(size, align);
2298 if (forced_order >= 0)
2299 order = forced_order;
2301 order = calculate_order(size);
2308 s->allocflags |= __GFP_COMP;
2310 if (s->flags & SLAB_CACHE_DMA)
2311 s->allocflags |= SLUB_DMA;
2313 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2314 s->allocflags |= __GFP_RECLAIMABLE;
2317 * Determine the number of objects per slab
2319 s->oo = oo_make(order, size);
2320 s->min = oo_make(get_order(size), size);
2321 if (oo_objects(s->oo) > oo_objects(s->max))
2324 return !!oo_objects(s->oo);
2328 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2329 const char *name, size_t size,
2330 size_t align, unsigned long flags,
2331 void (*ctor)(void *))
2333 memset(s, 0, kmem_size);
2338 s->flags = kmem_cache_flags(size, flags, name, ctor);
2340 if (!calculate_sizes(s, -1))
2342 if (disable_higher_order_debug) {
2344 * Disable debugging flags that store metadata if the min slab
2347 if (get_order(s->size) > get_order(s->objsize)) {
2348 s->flags &= ~DEBUG_METADATA_FLAGS;
2350 if (!calculate_sizes(s, -1))
2356 * The larger the object size is, the more pages we want on the partial
2357 * list to avoid pounding the page allocator excessively.
2359 set_min_partial(s, ilog2(s->size));
2362 s->remote_node_defrag_ratio = 1000;
2364 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2367 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2370 free_kmem_cache_nodes(s);
2372 if (flags & SLAB_PANIC)
2373 panic("Cannot create slab %s size=%lu realsize=%u "
2374 "order=%u offset=%u flags=%lx\n",
2375 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2381 * Check if a given pointer is valid
2383 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2387 page = get_object_page(object);
2389 if (!page || s != page->slab)
2390 /* No slab or wrong slab */
2393 if (!check_valid_pointer(s, page, object))
2397 * We could also check if the object is on the slabs freelist.
2398 * But this would be too expensive and it seems that the main
2399 * purpose of kmem_ptr_valid() is to check if the object belongs
2400 * to a certain slab.
2404 EXPORT_SYMBOL(kmem_ptr_validate);
2407 * Determine the size of a slab object
2409 unsigned int kmem_cache_size(struct kmem_cache *s)
2413 EXPORT_SYMBOL(kmem_cache_size);
2415 const char *kmem_cache_name(struct kmem_cache *s)
2419 EXPORT_SYMBOL(kmem_cache_name);
2421 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2424 #ifdef CONFIG_SLUB_DEBUG
2425 void *addr = page_address(page);
2427 DECLARE_BITMAP(map, page->objects);
2429 bitmap_zero(map, page->objects);
2430 slab_err(s, page, "%s", text);
2432 for_each_free_object(p, s, page->freelist)
2433 set_bit(slab_index(p, s, addr), map);
2435 for_each_object(p, s, addr, page->objects) {
2437 if (!test_bit(slab_index(p, s, addr), map)) {
2438 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2440 print_tracking(s, p);
2448 * Attempt to free all partial slabs on a node.
2450 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2452 unsigned long flags;
2453 struct page *page, *h;
2455 spin_lock_irqsave(&n->list_lock, flags);
2456 list_for_each_entry_safe(page, h, &n->partial, lru) {
2458 list_del(&page->lru);
2459 discard_slab(s, page);
2462 list_slab_objects(s, page,
2463 "Objects remaining on kmem_cache_close()");
2466 spin_unlock_irqrestore(&n->list_lock, flags);
2470 * Release all resources used by a slab cache.
2472 static inline int kmem_cache_close(struct kmem_cache *s)
2477 free_percpu(s->cpu_slab);
2478 /* Attempt to free all objects */
2479 for_each_node_state(node, N_NORMAL_MEMORY) {
2480 struct kmem_cache_node *n = get_node(s, node);
2483 if (n->nr_partial || slabs_node(s, node))
2486 free_kmem_cache_nodes(s);
2491 * Close a cache and release the kmem_cache structure
2492 * (must be used for caches created using kmem_cache_create)
2494 void kmem_cache_destroy(struct kmem_cache *s)
2496 down_write(&slub_lock);
2500 up_write(&slub_lock);
2501 if (kmem_cache_close(s)) {
2502 printk(KERN_ERR "SLUB %s: %s called for cache that "
2503 "still has objects.\n", s->name, __func__);
2506 if (s->flags & SLAB_DESTROY_BY_RCU)
2508 sysfs_slab_remove(s);
2510 up_write(&slub_lock);
2512 EXPORT_SYMBOL(kmem_cache_destroy);
2514 /********************************************************************
2516 *******************************************************************/
2518 struct kmem_cache kmalloc_caches[KMALLOC_CACHES] __cacheline_aligned;
2519 EXPORT_SYMBOL(kmalloc_caches);
2521 static int __init setup_slub_min_order(char *str)
2523 get_option(&str, &slub_min_order);
2528 __setup("slub_min_order=", setup_slub_min_order);
2530 static int __init setup_slub_max_order(char *str)
2532 get_option(&str, &slub_max_order);
2533 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2538 __setup("slub_max_order=", setup_slub_max_order);
2540 static int __init setup_slub_min_objects(char *str)
2542 get_option(&str, &slub_min_objects);
2547 __setup("slub_min_objects=", setup_slub_min_objects);
2549 static int __init setup_slub_nomerge(char *str)
2555 __setup("slub_nomerge", setup_slub_nomerge);
2557 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2558 const char *name, int size, gfp_t gfp_flags)
2560 unsigned int flags = 0;
2562 if (gfp_flags & SLUB_DMA)
2563 flags = SLAB_CACHE_DMA;
2566 * This function is called with IRQs disabled during early-boot on
2567 * single CPU so there's no need to take slub_lock here.
2569 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2573 list_add(&s->list, &slab_caches);
2575 if (sysfs_slab_add(s))
2580 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2583 #ifdef CONFIG_ZONE_DMA
2584 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2586 static void sysfs_add_func(struct work_struct *w)
2588 struct kmem_cache *s;
2590 down_write(&slub_lock);
2591 list_for_each_entry(s, &slab_caches, list) {
2592 if (s->flags & __SYSFS_ADD_DEFERRED) {
2593 s->flags &= ~__SYSFS_ADD_DEFERRED;
2597 up_write(&slub_lock);
2600 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2602 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2604 struct kmem_cache *s;
2607 unsigned long slabflags;
2610 s = kmalloc_caches_dma[index];
2614 /* Dynamically create dma cache */
2615 if (flags & __GFP_WAIT)
2616 down_write(&slub_lock);
2618 if (!down_write_trylock(&slub_lock))
2622 if (kmalloc_caches_dma[index])
2625 realsize = kmalloc_caches[index].objsize;
2626 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2627 (unsigned int)realsize);
2630 for (i = 0; i < KMALLOC_CACHES; i++)
2631 if (!kmalloc_caches[i].size)
2634 BUG_ON(i >= KMALLOC_CACHES);
2635 s = kmalloc_caches + i;
2638 * Must defer sysfs creation to a workqueue because we don't know
2639 * what context we are called from. Before sysfs comes up, we don't
2640 * need to do anything because our sysfs initcall will start by
2641 * adding all existing slabs to sysfs.
2643 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
2644 if (slab_state >= SYSFS)
2645 slabflags |= __SYSFS_ADD_DEFERRED;
2647 if (!s || !text || !kmem_cache_open(s, flags, text,
2648 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
2654 list_add(&s->list, &slab_caches);
2655 kmalloc_caches_dma[index] = s;
2657 if (slab_state >= SYSFS)
2658 schedule_work(&sysfs_add_work);
2661 up_write(&slub_lock);
2663 return kmalloc_caches_dma[index];
2668 * Conversion table for small slabs sizes / 8 to the index in the
2669 * kmalloc array. This is necessary for slabs < 192 since we have non power
2670 * of two cache sizes there. The size of larger slabs can be determined using
2673 static s8 size_index[24] = {
2700 static inline int size_index_elem(size_t bytes)
2702 return (bytes - 1) / 8;
2705 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2711 return ZERO_SIZE_PTR;
2713 index = size_index[size_index_elem(size)];
2715 index = fls(size - 1);
2717 #ifdef CONFIG_ZONE_DMA
2718 if (unlikely((flags & SLUB_DMA)))
2719 return dma_kmalloc_cache(index, flags);
2722 return &kmalloc_caches[index];
2725 void *__kmalloc(size_t size, gfp_t flags)
2727 struct kmem_cache *s;
2730 if (unlikely(size > SLUB_MAX_SIZE))
2731 return kmalloc_large(size, flags);
2733 s = get_slab(size, flags);
2735 if (unlikely(ZERO_OR_NULL_PTR(s)))
2738 ret = slab_alloc(s, flags, -1, _RET_IP_);
2740 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2744 EXPORT_SYMBOL(__kmalloc);
2746 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2751 flags |= __GFP_COMP | __GFP_NOTRACK;
2752 page = alloc_pages_node(node, flags, get_order(size));
2754 ptr = page_address(page);
2756 kmemleak_alloc(ptr, size, 1, flags);
2761 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2763 struct kmem_cache *s;
2766 if (unlikely(size > SLUB_MAX_SIZE)) {
2767 ret = kmalloc_large_node(size, flags, node);
2769 trace_kmalloc_node(_RET_IP_, ret,
2770 size, PAGE_SIZE << get_order(size),
2776 s = get_slab(size, flags);
2778 if (unlikely(ZERO_OR_NULL_PTR(s)))
2781 ret = slab_alloc(s, flags, node, _RET_IP_);
2783 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2787 EXPORT_SYMBOL(__kmalloc_node);
2790 size_t ksize(const void *object)
2793 struct kmem_cache *s;
2795 if (unlikely(object == ZERO_SIZE_PTR))
2798 page = virt_to_head_page(object);
2800 if (unlikely(!PageSlab(page))) {
2801 WARN_ON(!PageCompound(page));
2802 return PAGE_SIZE << compound_order(page);
2806 #ifdef CONFIG_SLUB_DEBUG
2808 * Debugging requires use of the padding between object
2809 * and whatever may come after it.
2811 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2816 * If we have the need to store the freelist pointer
2817 * back there or track user information then we can
2818 * only use the space before that information.
2820 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2823 * Else we can use all the padding etc for the allocation
2827 EXPORT_SYMBOL(ksize);
2829 void kfree(const void *x)
2832 void *object = (void *)x;
2834 trace_kfree(_RET_IP_, x);
2836 if (unlikely(ZERO_OR_NULL_PTR(x)))
2839 page = virt_to_head_page(x);
2840 if (unlikely(!PageSlab(page))) {
2841 BUG_ON(!PageCompound(page));
2846 slab_free(page->slab, page, object, _RET_IP_);
2848 EXPORT_SYMBOL(kfree);
2851 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2852 * the remaining slabs by the number of items in use. The slabs with the
2853 * most items in use come first. New allocations will then fill those up
2854 * and thus they can be removed from the partial lists.
2856 * The slabs with the least items are placed last. This results in them
2857 * being allocated from last increasing the chance that the last objects
2858 * are freed in them.
2860 int kmem_cache_shrink(struct kmem_cache *s)
2864 struct kmem_cache_node *n;
2867 int objects = oo_objects(s->max);
2868 struct list_head *slabs_by_inuse =
2869 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2870 unsigned long flags;
2872 if (!slabs_by_inuse)
2876 for_each_node_state(node, N_NORMAL_MEMORY) {
2877 n = get_node(s, node);
2882 for (i = 0; i < objects; i++)
2883 INIT_LIST_HEAD(slabs_by_inuse + i);
2885 spin_lock_irqsave(&n->list_lock, flags);
2888 * Build lists indexed by the items in use in each slab.
2890 * Note that concurrent frees may occur while we hold the
2891 * list_lock. page->inuse here is the upper limit.
2893 list_for_each_entry_safe(page, t, &n->partial, lru) {
2894 if (!page->inuse && slab_trylock(page)) {
2896 * Must hold slab lock here because slab_free
2897 * may have freed the last object and be
2898 * waiting to release the slab.
2900 list_del(&page->lru);
2903 discard_slab(s, page);
2905 list_move(&page->lru,
2906 slabs_by_inuse + page->inuse);
2911 * Rebuild the partial list with the slabs filled up most
2912 * first and the least used slabs at the end.
2914 for (i = objects - 1; i >= 0; i--)
2915 list_splice(slabs_by_inuse + i, n->partial.prev);
2917 spin_unlock_irqrestore(&n->list_lock, flags);
2920 kfree(slabs_by_inuse);
2923 EXPORT_SYMBOL(kmem_cache_shrink);
2925 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2926 static int slab_mem_going_offline_callback(void *arg)
2928 struct kmem_cache *s;
2930 down_read(&slub_lock);
2931 list_for_each_entry(s, &slab_caches, list)
2932 kmem_cache_shrink(s);
2933 up_read(&slub_lock);
2938 static void slab_mem_offline_callback(void *arg)
2940 struct kmem_cache_node *n;
2941 struct kmem_cache *s;
2942 struct memory_notify *marg = arg;
2945 offline_node = marg->status_change_nid;
2948 * If the node still has available memory. we need kmem_cache_node
2951 if (offline_node < 0)
2954 down_read(&slub_lock);
2955 list_for_each_entry(s, &slab_caches, list) {
2956 n = get_node(s, offline_node);
2959 * if n->nr_slabs > 0, slabs still exist on the node
2960 * that is going down. We were unable to free them,
2961 * and offline_pages() function shoudn't call this
2962 * callback. So, we must fail.
2964 BUG_ON(slabs_node(s, offline_node));
2966 s->node[offline_node] = NULL;
2967 kmem_cache_free(kmalloc_caches, n);
2970 up_read(&slub_lock);
2973 static int slab_mem_going_online_callback(void *arg)
2975 struct kmem_cache_node *n;
2976 struct kmem_cache *s;
2977 struct memory_notify *marg = arg;
2978 int nid = marg->status_change_nid;
2982 * If the node's memory is already available, then kmem_cache_node is
2983 * already created. Nothing to do.
2989 * We are bringing a node online. No memory is available yet. We must
2990 * allocate a kmem_cache_node structure in order to bring the node
2993 down_read(&slub_lock);
2994 list_for_each_entry(s, &slab_caches, list) {
2996 * XXX: kmem_cache_alloc_node will fallback to other nodes
2997 * since memory is not yet available from the node that
3000 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3005 init_kmem_cache_node(n, s);
3009 up_read(&slub_lock);
3013 static int slab_memory_callback(struct notifier_block *self,
3014 unsigned long action, void *arg)
3019 case MEM_GOING_ONLINE:
3020 ret = slab_mem_going_online_callback(arg);
3022 case MEM_GOING_OFFLINE:
3023 ret = slab_mem_going_offline_callback(arg);
3026 case MEM_CANCEL_ONLINE:
3027 slab_mem_offline_callback(arg);
3030 case MEM_CANCEL_OFFLINE:
3034 ret = notifier_from_errno(ret);
3040 #endif /* CONFIG_MEMORY_HOTPLUG */
3042 /********************************************************************
3043 * Basic setup of slabs
3044 *******************************************************************/
3046 void __init kmem_cache_init(void)
3053 * Must first have the slab cache available for the allocations of the
3054 * struct kmem_cache_node's. There is special bootstrap code in
3055 * kmem_cache_open for slab_state == DOWN.
3057 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3058 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3059 kmalloc_caches[0].refcount = -1;
3062 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3065 /* Able to allocate the per node structures */
3066 slab_state = PARTIAL;
3068 /* Caches that are not of the two-to-the-power-of size */
3069 if (KMALLOC_MIN_SIZE <= 32) {
3070 create_kmalloc_cache(&kmalloc_caches[1],
3071 "kmalloc-96", 96, GFP_NOWAIT);
3074 if (KMALLOC_MIN_SIZE <= 64) {
3075 create_kmalloc_cache(&kmalloc_caches[2],
3076 "kmalloc-192", 192, GFP_NOWAIT);
3080 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3081 create_kmalloc_cache(&kmalloc_caches[i],
3082 "kmalloc", 1 << i, GFP_NOWAIT);
3088 * Patch up the size_index table if we have strange large alignment
3089 * requirements for the kmalloc array. This is only the case for
3090 * MIPS it seems. The standard arches will not generate any code here.
3092 * Largest permitted alignment is 256 bytes due to the way we
3093 * handle the index determination for the smaller caches.
3095 * Make sure that nothing crazy happens if someone starts tinkering
3096 * around with ARCH_KMALLOC_MINALIGN
3098 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3099 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3101 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3102 int elem = size_index_elem(i);
3103 if (elem >= ARRAY_SIZE(size_index))
3105 size_index[elem] = KMALLOC_SHIFT_LOW;
3108 if (KMALLOC_MIN_SIZE == 64) {
3110 * The 96 byte size cache is not used if the alignment
3113 for (i = 64 + 8; i <= 96; i += 8)
3114 size_index[size_index_elem(i)] = 7;
3115 } else if (KMALLOC_MIN_SIZE == 128) {
3117 * The 192 byte sized cache is not used if the alignment
3118 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3121 for (i = 128 + 8; i <= 192; i += 8)
3122 size_index[size_index_elem(i)] = 8;
3127 /* Provide the correct kmalloc names now that the caches are up */
3128 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3129 kmalloc_caches[i]. name =
3130 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3133 register_cpu_notifier(&slab_notifier);
3136 kmem_size = offsetof(struct kmem_cache, node) +
3137 nr_node_ids * sizeof(struct kmem_cache_node *);
3139 kmem_size = sizeof(struct kmem_cache);
3143 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3144 " CPUs=%d, Nodes=%d\n",
3145 caches, cache_line_size(),
3146 slub_min_order, slub_max_order, slub_min_objects,
3147 nr_cpu_ids, nr_node_ids);
3150 void __init kmem_cache_init_late(void)
3155 * Find a mergeable slab cache
3157 static int slab_unmergeable(struct kmem_cache *s)
3159 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3166 * We may have set a slab to be unmergeable during bootstrap.
3168 if (s->refcount < 0)
3174 static struct kmem_cache *find_mergeable(size_t size,
3175 size_t align, unsigned long flags, const char *name,
3176 void (*ctor)(void *))
3178 struct kmem_cache *s;
3180 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3186 size = ALIGN(size, sizeof(void *));
3187 align = calculate_alignment(flags, align, size);
3188 size = ALIGN(size, align);
3189 flags = kmem_cache_flags(size, flags, name, NULL);
3191 list_for_each_entry(s, &slab_caches, list) {
3192 if (slab_unmergeable(s))
3198 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3201 * Check if alignment is compatible.
3202 * Courtesy of Adrian Drzewiecki
3204 if ((s->size & ~(align - 1)) != s->size)
3207 if (s->size - size >= sizeof(void *))
3215 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3216 size_t align, unsigned long flags, void (*ctor)(void *))
3218 struct kmem_cache *s;
3223 down_write(&slub_lock);
3224 s = find_mergeable(size, align, flags, name, ctor);
3228 * Adjust the object sizes so that we clear
3229 * the complete object on kzalloc.
3231 s->objsize = max(s->objsize, (int)size);
3232 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3233 up_write(&slub_lock);
3235 if (sysfs_slab_alias(s, name)) {
3236 down_write(&slub_lock);
3238 up_write(&slub_lock);
3244 s = kmalloc(kmem_size, GFP_KERNEL);
3246 if (kmem_cache_open(s, GFP_KERNEL, name,
3247 size, align, flags, ctor)) {
3248 list_add(&s->list, &slab_caches);
3249 up_write(&slub_lock);
3250 if (sysfs_slab_add(s)) {
3251 down_write(&slub_lock);
3253 up_write(&slub_lock);
3261 up_write(&slub_lock);
3264 if (flags & SLAB_PANIC)
3265 panic("Cannot create slabcache %s\n", name);
3270 EXPORT_SYMBOL(kmem_cache_create);
3274 * Use the cpu notifier to insure that the cpu slabs are flushed when
3277 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3278 unsigned long action, void *hcpu)
3280 long cpu = (long)hcpu;
3281 struct kmem_cache *s;
3282 unsigned long flags;
3285 case CPU_UP_CANCELED:
3286 case CPU_UP_CANCELED_FROZEN:
3288 case CPU_DEAD_FROZEN:
3289 down_read(&slub_lock);
3290 list_for_each_entry(s, &slab_caches, list) {
3291 local_irq_save(flags);
3292 __flush_cpu_slab(s, cpu);
3293 local_irq_restore(flags);
3295 up_read(&slub_lock);
3303 static struct notifier_block __cpuinitdata slab_notifier = {
3304 .notifier_call = slab_cpuup_callback
3309 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3311 struct kmem_cache *s;
3314 if (unlikely(size > SLUB_MAX_SIZE))
3315 return kmalloc_large(size, gfpflags);
3317 s = get_slab(size, gfpflags);
3319 if (unlikely(ZERO_OR_NULL_PTR(s)))
3322 ret = slab_alloc(s, gfpflags, -1, caller);
3324 /* Honor the call site pointer we recieved. */
3325 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3330 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3331 int node, unsigned long caller)
3333 struct kmem_cache *s;
3336 if (unlikely(size > SLUB_MAX_SIZE))
3337 return kmalloc_large_node(size, gfpflags, node);
3339 s = get_slab(size, gfpflags);
3341 if (unlikely(ZERO_OR_NULL_PTR(s)))
3344 ret = slab_alloc(s, gfpflags, node, caller);
3346 /* Honor the call site pointer we recieved. */
3347 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3352 #ifdef CONFIG_SLUB_DEBUG
3353 static int count_inuse(struct page *page)
3358 static int count_total(struct page *page)
3360 return page->objects;
3363 static int validate_slab(struct kmem_cache *s, struct page *page,
3367 void *addr = page_address(page);
3369 if (!check_slab(s, page) ||
3370 !on_freelist(s, page, NULL))
3373 /* Now we know that a valid freelist exists */
3374 bitmap_zero(map, page->objects);
3376 for_each_free_object(p, s, page->freelist) {
3377 set_bit(slab_index(p, s, addr), map);
3378 if (!check_object(s, page, p, 0))
3382 for_each_object(p, s, addr, page->objects)
3383 if (!test_bit(slab_index(p, s, addr), map))
3384 if (!check_object(s, page, p, 1))
3389 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3392 if (slab_trylock(page)) {
3393 validate_slab(s, page, map);
3396 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3399 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3400 if (!PageSlubDebug(page))
3401 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3402 "on slab 0x%p\n", s->name, page);
3404 if (PageSlubDebug(page))
3405 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3406 "slab 0x%p\n", s->name, page);
3410 static int validate_slab_node(struct kmem_cache *s,
3411 struct kmem_cache_node *n, unsigned long *map)
3413 unsigned long count = 0;
3415 unsigned long flags;
3417 spin_lock_irqsave(&n->list_lock, flags);
3419 list_for_each_entry(page, &n->partial, lru) {
3420 validate_slab_slab(s, page, map);
3423 if (count != n->nr_partial)
3424 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3425 "counter=%ld\n", s->name, count, n->nr_partial);
3427 if (!(s->flags & SLAB_STORE_USER))
3430 list_for_each_entry(page, &n->full, lru) {
3431 validate_slab_slab(s, page, map);
3434 if (count != atomic_long_read(&n->nr_slabs))
3435 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3436 "counter=%ld\n", s->name, count,
3437 atomic_long_read(&n->nr_slabs));
3440 spin_unlock_irqrestore(&n->list_lock, flags);
3444 static long validate_slab_cache(struct kmem_cache *s)
3447 unsigned long count = 0;
3448 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3449 sizeof(unsigned long), GFP_KERNEL);
3455 for_each_node_state(node, N_NORMAL_MEMORY) {
3456 struct kmem_cache_node *n = get_node(s, node);
3458 count += validate_slab_node(s, n, map);
3464 #ifdef SLUB_RESILIENCY_TEST
3465 static void resiliency_test(void)
3469 printk(KERN_ERR "SLUB resiliency testing\n");
3470 printk(KERN_ERR "-----------------------\n");
3471 printk(KERN_ERR "A. Corruption after allocation\n");
3473 p = kzalloc(16, GFP_KERNEL);
3475 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3476 " 0x12->0x%p\n\n", p + 16);
3478 validate_slab_cache(kmalloc_caches + 4);
3480 /* Hmmm... The next two are dangerous */
3481 p = kzalloc(32, GFP_KERNEL);
3482 p[32 + sizeof(void *)] = 0x34;
3483 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3484 " 0x34 -> -0x%p\n", p);
3486 "If allocated object is overwritten then not detectable\n\n");
3488 validate_slab_cache(kmalloc_caches + 5);
3489 p = kzalloc(64, GFP_KERNEL);
3490 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3492 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3495 "If allocated object is overwritten then not detectable\n\n");
3496 validate_slab_cache(kmalloc_caches + 6);
3498 printk(KERN_ERR "\nB. Corruption after free\n");
3499 p = kzalloc(128, GFP_KERNEL);
3502 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3503 validate_slab_cache(kmalloc_caches + 7);
3505 p = kzalloc(256, GFP_KERNEL);
3508 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3510 validate_slab_cache(kmalloc_caches + 8);
3512 p = kzalloc(512, GFP_KERNEL);
3515 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3516 validate_slab_cache(kmalloc_caches + 9);
3519 static void resiliency_test(void) {};
3523 * Generate lists of code addresses where slabcache objects are allocated
3528 unsigned long count;
3535 DECLARE_BITMAP(cpus, NR_CPUS);
3541 unsigned long count;
3542 struct location *loc;
3545 static void free_loc_track(struct loc_track *t)
3548 free_pages((unsigned long)t->loc,
3549 get_order(sizeof(struct location) * t->max));
3552 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3557 order = get_order(sizeof(struct location) * max);
3559 l = (void *)__get_free_pages(flags, order);
3564 memcpy(l, t->loc, sizeof(struct location) * t->count);
3572 static int add_location(struct loc_track *t, struct kmem_cache *s,
3573 const struct track *track)
3575 long start, end, pos;
3577 unsigned long caddr;
3578 unsigned long age = jiffies - track->when;
3584 pos = start + (end - start + 1) / 2;
3587 * There is nothing at "end". If we end up there
3588 * we need to add something to before end.
3593 caddr = t->loc[pos].addr;
3594 if (track->addr == caddr) {
3600 if (age < l->min_time)
3602 if (age > l->max_time)
3605 if (track->pid < l->min_pid)
3606 l->min_pid = track->pid;
3607 if (track->pid > l->max_pid)
3608 l->max_pid = track->pid;
3610 cpumask_set_cpu(track->cpu,
3611 to_cpumask(l->cpus));
3613 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3617 if (track->addr < caddr)
3624 * Not found. Insert new tracking element.
3626 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3632 (t->count - pos) * sizeof(struct location));
3635 l->addr = track->addr;
3639 l->min_pid = track->pid;
3640 l->max_pid = track->pid;
3641 cpumask_clear(to_cpumask(l->cpus));
3642 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3643 nodes_clear(l->nodes);
3644 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3648 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3649 struct page *page, enum track_item alloc)
3651 void *addr = page_address(page);
3652 DECLARE_BITMAP(map, page->objects);
3655 bitmap_zero(map, page->objects);
3656 for_each_free_object(p, s, page->freelist)
3657 set_bit(slab_index(p, s, addr), map);
3659 for_each_object(p, s, addr, page->objects)
3660 if (!test_bit(slab_index(p, s, addr), map))
3661 add_location(t, s, get_track(s, p, alloc));
3664 static int list_locations(struct kmem_cache *s, char *buf,
3665 enum track_item alloc)
3669 struct loc_track t = { 0, 0, NULL };
3672 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3674 return sprintf(buf, "Out of memory\n");
3676 /* Push back cpu slabs */
3679 for_each_node_state(node, N_NORMAL_MEMORY) {
3680 struct kmem_cache_node *n = get_node(s, node);
3681 unsigned long flags;
3684 if (!atomic_long_read(&n->nr_slabs))
3687 spin_lock_irqsave(&n->list_lock, flags);
3688 list_for_each_entry(page, &n->partial, lru)
3689 process_slab(&t, s, page, alloc);
3690 list_for_each_entry(page, &n->full, lru)
3691 process_slab(&t, s, page, alloc);
3692 spin_unlock_irqrestore(&n->list_lock, flags);
3695 for (i = 0; i < t.count; i++) {
3696 struct location *l = &t.loc[i];
3698 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3700 len += sprintf(buf + len, "%7ld ", l->count);
3703 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3705 len += sprintf(buf + len, "<not-available>");
3707 if (l->sum_time != l->min_time) {
3708 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3710 (long)div_u64(l->sum_time, l->count),
3713 len += sprintf(buf + len, " age=%ld",
3716 if (l->min_pid != l->max_pid)
3717 len += sprintf(buf + len, " pid=%ld-%ld",
3718 l->min_pid, l->max_pid);
3720 len += sprintf(buf + len, " pid=%ld",
3723 if (num_online_cpus() > 1 &&
3724 !cpumask_empty(to_cpumask(l->cpus)) &&
3725 len < PAGE_SIZE - 60) {
3726 len += sprintf(buf + len, " cpus=");
3727 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3728 to_cpumask(l->cpus));
3731 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3732 len < PAGE_SIZE - 60) {
3733 len += sprintf(buf + len, " nodes=");
3734 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3738 len += sprintf(buf + len, "\n");
3743 len += sprintf(buf, "No data\n");
3747 enum slab_stat_type {
3748 SL_ALL, /* All slabs */
3749 SL_PARTIAL, /* Only partially allocated slabs */
3750 SL_CPU, /* Only slabs used for cpu caches */
3751 SL_OBJECTS, /* Determine allocated objects not slabs */
3752 SL_TOTAL /* Determine object capacity not slabs */
3755 #define SO_ALL (1 << SL_ALL)
3756 #define SO_PARTIAL (1 << SL_PARTIAL)
3757 #define SO_CPU (1 << SL_CPU)
3758 #define SO_OBJECTS (1 << SL_OBJECTS)
3759 #define SO_TOTAL (1 << SL_TOTAL)
3761 static ssize_t show_slab_objects(struct kmem_cache *s,
3762 char *buf, unsigned long flags)
3764 unsigned long total = 0;
3767 unsigned long *nodes;
3768 unsigned long *per_cpu;
3770 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3773 per_cpu = nodes + nr_node_ids;
3775 if (flags & SO_CPU) {
3778 for_each_possible_cpu(cpu) {
3779 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3781 if (!c || c->node < 0)
3785 if (flags & SO_TOTAL)
3786 x = c->page->objects;
3787 else if (flags & SO_OBJECTS)
3793 nodes[c->node] += x;
3799 if (flags & SO_ALL) {
3800 for_each_node_state(node, N_NORMAL_MEMORY) {
3801 struct kmem_cache_node *n = get_node(s, node);
3803 if (flags & SO_TOTAL)
3804 x = atomic_long_read(&n->total_objects);
3805 else if (flags & SO_OBJECTS)
3806 x = atomic_long_read(&n->total_objects) -
3807 count_partial(n, count_free);
3810 x = atomic_long_read(&n->nr_slabs);
3815 } else if (flags & SO_PARTIAL) {
3816 for_each_node_state(node, N_NORMAL_MEMORY) {
3817 struct kmem_cache_node *n = get_node(s, node);
3819 if (flags & SO_TOTAL)
3820 x = count_partial(n, count_total);
3821 else if (flags & SO_OBJECTS)
3822 x = count_partial(n, count_inuse);
3829 x = sprintf(buf, "%lu", total);
3831 for_each_node_state(node, N_NORMAL_MEMORY)
3833 x += sprintf(buf + x, " N%d=%lu",
3837 return x + sprintf(buf + x, "\n");
3840 static int any_slab_objects(struct kmem_cache *s)
3844 for_each_online_node(node) {
3845 struct kmem_cache_node *n = get_node(s, node);
3850 if (atomic_long_read(&n->total_objects))
3856 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3857 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3859 struct slab_attribute {
3860 struct attribute attr;
3861 ssize_t (*show)(struct kmem_cache *s, char *buf);
3862 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3865 #define SLAB_ATTR_RO(_name) \
3866 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3868 #define SLAB_ATTR(_name) \
3869 static struct slab_attribute _name##_attr = \
3870 __ATTR(_name, 0644, _name##_show, _name##_store)
3872 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3874 return sprintf(buf, "%d\n", s->size);
3876 SLAB_ATTR_RO(slab_size);
3878 static ssize_t align_show(struct kmem_cache *s, char *buf)
3880 return sprintf(buf, "%d\n", s->align);
3882 SLAB_ATTR_RO(align);
3884 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3886 return sprintf(buf, "%d\n", s->objsize);
3888 SLAB_ATTR_RO(object_size);
3890 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3892 return sprintf(buf, "%d\n", oo_objects(s->oo));
3894 SLAB_ATTR_RO(objs_per_slab);
3896 static ssize_t order_store(struct kmem_cache *s,
3897 const char *buf, size_t length)
3899 unsigned long order;
3902 err = strict_strtoul(buf, 10, &order);
3906 if (order > slub_max_order || order < slub_min_order)
3909 calculate_sizes(s, order);
3913 static ssize_t order_show(struct kmem_cache *s, char *buf)
3915 return sprintf(buf, "%d\n", oo_order(s->oo));
3919 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3921 return sprintf(buf, "%lu\n", s->min_partial);
3924 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3930 err = strict_strtoul(buf, 10, &min);
3934 set_min_partial(s, min);
3937 SLAB_ATTR(min_partial);
3939 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3942 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3944 return n + sprintf(buf + n, "\n");
3950 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3952 return sprintf(buf, "%d\n", s->refcount - 1);
3954 SLAB_ATTR_RO(aliases);
3956 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3958 return show_slab_objects(s, buf, SO_ALL);
3960 SLAB_ATTR_RO(slabs);
3962 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3964 return show_slab_objects(s, buf, SO_PARTIAL);
3966 SLAB_ATTR_RO(partial);
3968 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3970 return show_slab_objects(s, buf, SO_CPU);
3972 SLAB_ATTR_RO(cpu_slabs);
3974 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3976 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3978 SLAB_ATTR_RO(objects);
3980 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3982 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3984 SLAB_ATTR_RO(objects_partial);
3986 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3988 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3990 SLAB_ATTR_RO(total_objects);
3992 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3994 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3997 static ssize_t sanity_checks_store(struct kmem_cache *s,
3998 const char *buf, size_t length)
4000 s->flags &= ~SLAB_DEBUG_FREE;
4002 s->flags |= SLAB_DEBUG_FREE;
4005 SLAB_ATTR(sanity_checks);
4007 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4009 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4012 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4015 s->flags &= ~SLAB_TRACE;
4017 s->flags |= SLAB_TRACE;
4022 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4024 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4027 static ssize_t reclaim_account_store(struct kmem_cache *s,
4028 const char *buf, size_t length)
4030 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4032 s->flags |= SLAB_RECLAIM_ACCOUNT;
4035 SLAB_ATTR(reclaim_account);
4037 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4039 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4041 SLAB_ATTR_RO(hwcache_align);
4043 #ifdef CONFIG_ZONE_DMA
4044 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4046 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4048 SLAB_ATTR_RO(cache_dma);
4051 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4053 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4055 SLAB_ATTR_RO(destroy_by_rcu);
4057 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4059 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4062 static ssize_t red_zone_store(struct kmem_cache *s,
4063 const char *buf, size_t length)
4065 if (any_slab_objects(s))
4068 s->flags &= ~SLAB_RED_ZONE;
4070 s->flags |= SLAB_RED_ZONE;
4071 calculate_sizes(s, -1);
4074 SLAB_ATTR(red_zone);
4076 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4078 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4081 static ssize_t poison_store(struct kmem_cache *s,
4082 const char *buf, size_t length)
4084 if (any_slab_objects(s))
4087 s->flags &= ~SLAB_POISON;
4089 s->flags |= SLAB_POISON;
4090 calculate_sizes(s, -1);
4095 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4097 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4100 static ssize_t store_user_store(struct kmem_cache *s,
4101 const char *buf, size_t length)
4103 if (any_slab_objects(s))
4106 s->flags &= ~SLAB_STORE_USER;
4108 s->flags |= SLAB_STORE_USER;
4109 calculate_sizes(s, -1);
4112 SLAB_ATTR(store_user);
4114 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4119 static ssize_t validate_store(struct kmem_cache *s,
4120 const char *buf, size_t length)
4124 if (buf[0] == '1') {
4125 ret = validate_slab_cache(s);
4131 SLAB_ATTR(validate);
4133 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4138 static ssize_t shrink_store(struct kmem_cache *s,
4139 const char *buf, size_t length)
4141 if (buf[0] == '1') {
4142 int rc = kmem_cache_shrink(s);
4152 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4154 if (!(s->flags & SLAB_STORE_USER))
4156 return list_locations(s, buf, TRACK_ALLOC);
4158 SLAB_ATTR_RO(alloc_calls);
4160 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4162 if (!(s->flags & SLAB_STORE_USER))
4164 return list_locations(s, buf, TRACK_FREE);
4166 SLAB_ATTR_RO(free_calls);
4169 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4171 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4174 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4175 const char *buf, size_t length)
4177 unsigned long ratio;
4180 err = strict_strtoul(buf, 10, &ratio);
4185 s->remote_node_defrag_ratio = ratio * 10;
4189 SLAB_ATTR(remote_node_defrag_ratio);
4192 #ifdef CONFIG_SLUB_STATS
4193 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4195 unsigned long sum = 0;
4198 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4203 for_each_online_cpu(cpu) {
4204 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4210 len = sprintf(buf, "%lu", sum);
4213 for_each_online_cpu(cpu) {
4214 if (data[cpu] && len < PAGE_SIZE - 20)
4215 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4219 return len + sprintf(buf + len, "\n");
4222 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4226 for_each_online_cpu(cpu)
4227 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4230 #define STAT_ATTR(si, text) \
4231 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4233 return show_stat(s, buf, si); \
4235 static ssize_t text##_store(struct kmem_cache *s, \
4236 const char *buf, size_t length) \
4238 if (buf[0] != '0') \
4240 clear_stat(s, si); \
4245 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4246 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4247 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4248 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4249 STAT_ATTR(FREE_FROZEN, free_frozen);
4250 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4251 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4252 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4253 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4254 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4255 STAT_ATTR(FREE_SLAB, free_slab);
4256 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4257 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4258 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4259 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4260 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4261 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4262 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4265 static struct attribute *slab_attrs[] = {
4266 &slab_size_attr.attr,
4267 &object_size_attr.attr,
4268 &objs_per_slab_attr.attr,
4270 &min_partial_attr.attr,
4272 &objects_partial_attr.attr,
4273 &total_objects_attr.attr,
4276 &cpu_slabs_attr.attr,
4280 &sanity_checks_attr.attr,
4282 &hwcache_align_attr.attr,
4283 &reclaim_account_attr.attr,
4284 &destroy_by_rcu_attr.attr,
4285 &red_zone_attr.attr,
4287 &store_user_attr.attr,
4288 &validate_attr.attr,
4290 &alloc_calls_attr.attr,
4291 &free_calls_attr.attr,
4292 #ifdef CONFIG_ZONE_DMA
4293 &cache_dma_attr.attr,
4296 &remote_node_defrag_ratio_attr.attr,
4298 #ifdef CONFIG_SLUB_STATS
4299 &alloc_fastpath_attr.attr,
4300 &alloc_slowpath_attr.attr,
4301 &free_fastpath_attr.attr,
4302 &free_slowpath_attr.attr,
4303 &free_frozen_attr.attr,
4304 &free_add_partial_attr.attr,
4305 &free_remove_partial_attr.attr,
4306 &alloc_from_partial_attr.attr,
4307 &alloc_slab_attr.attr,
4308 &alloc_refill_attr.attr,
4309 &free_slab_attr.attr,
4310 &cpuslab_flush_attr.attr,
4311 &deactivate_full_attr.attr,
4312 &deactivate_empty_attr.attr,
4313 &deactivate_to_head_attr.attr,
4314 &deactivate_to_tail_attr.attr,
4315 &deactivate_remote_frees_attr.attr,
4316 &order_fallback_attr.attr,
4321 static struct attribute_group slab_attr_group = {
4322 .attrs = slab_attrs,
4325 static ssize_t slab_attr_show(struct kobject *kobj,
4326 struct attribute *attr,
4329 struct slab_attribute *attribute;
4330 struct kmem_cache *s;
4333 attribute = to_slab_attr(attr);
4336 if (!attribute->show)
4339 err = attribute->show(s, buf);
4344 static ssize_t slab_attr_store(struct kobject *kobj,
4345 struct attribute *attr,
4346 const char *buf, size_t len)
4348 struct slab_attribute *attribute;
4349 struct kmem_cache *s;
4352 attribute = to_slab_attr(attr);
4355 if (!attribute->store)
4358 err = attribute->store(s, buf, len);
4363 static void kmem_cache_release(struct kobject *kobj)
4365 struct kmem_cache *s = to_slab(kobj);
4370 static struct sysfs_ops slab_sysfs_ops = {
4371 .show = slab_attr_show,
4372 .store = slab_attr_store,
4375 static struct kobj_type slab_ktype = {
4376 .sysfs_ops = &slab_sysfs_ops,
4377 .release = kmem_cache_release
4380 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4382 struct kobj_type *ktype = get_ktype(kobj);
4384 if (ktype == &slab_ktype)
4389 static struct kset_uevent_ops slab_uevent_ops = {
4390 .filter = uevent_filter,
4393 static struct kset *slab_kset;
4395 #define ID_STR_LENGTH 64
4397 /* Create a unique string id for a slab cache:
4399 * Format :[flags-]size
4401 static char *create_unique_id(struct kmem_cache *s)
4403 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4410 * First flags affecting slabcache operations. We will only
4411 * get here for aliasable slabs so we do not need to support
4412 * too many flags. The flags here must cover all flags that
4413 * are matched during merging to guarantee that the id is
4416 if (s->flags & SLAB_CACHE_DMA)
4418 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4420 if (s->flags & SLAB_DEBUG_FREE)
4422 if (!(s->flags & SLAB_NOTRACK))
4426 p += sprintf(p, "%07d", s->size);
4427 BUG_ON(p > name + ID_STR_LENGTH - 1);
4431 static int sysfs_slab_add(struct kmem_cache *s)
4437 if (slab_state < SYSFS)
4438 /* Defer until later */
4441 unmergeable = slab_unmergeable(s);
4444 * Slabcache can never be merged so we can use the name proper.
4445 * This is typically the case for debug situations. In that
4446 * case we can catch duplicate names easily.
4448 sysfs_remove_link(&slab_kset->kobj, s->name);
4452 * Create a unique name for the slab as a target
4455 name = create_unique_id(s);
4458 s->kobj.kset = slab_kset;
4459 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4461 kobject_put(&s->kobj);
4465 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4467 kobject_del(&s->kobj);
4468 kobject_put(&s->kobj);
4471 kobject_uevent(&s->kobj, KOBJ_ADD);
4473 /* Setup first alias */
4474 sysfs_slab_alias(s, s->name);
4480 static void sysfs_slab_remove(struct kmem_cache *s)
4482 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4483 kobject_del(&s->kobj);
4484 kobject_put(&s->kobj);
4488 * Need to buffer aliases during bootup until sysfs becomes
4489 * available lest we lose that information.
4491 struct saved_alias {
4492 struct kmem_cache *s;
4494 struct saved_alias *next;
4497 static struct saved_alias *alias_list;
4499 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4501 struct saved_alias *al;
4503 if (slab_state == SYSFS) {
4505 * If we have a leftover link then remove it.
4507 sysfs_remove_link(&slab_kset->kobj, name);
4508 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4511 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4517 al->next = alias_list;
4522 static int __init slab_sysfs_init(void)
4524 struct kmem_cache *s;
4527 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4529 printk(KERN_ERR "Cannot register slab subsystem.\n");
4535 list_for_each_entry(s, &slab_caches, list) {
4536 err = sysfs_slab_add(s);
4538 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4539 " to sysfs\n", s->name);
4542 while (alias_list) {
4543 struct saved_alias *al = alias_list;
4545 alias_list = alias_list->next;
4546 err = sysfs_slab_alias(al->s, al->name);
4548 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4549 " %s to sysfs\n", s->name);
4557 __initcall(slab_sysfs_init);
4561 * The /proc/slabinfo ABI
4563 #ifdef CONFIG_SLABINFO
4564 static void print_slabinfo_header(struct seq_file *m)
4566 seq_puts(m, "slabinfo - version: 2.1\n");
4567 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4568 "<objperslab> <pagesperslab>");
4569 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4570 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4574 static void *s_start(struct seq_file *m, loff_t *pos)
4578 down_read(&slub_lock);
4580 print_slabinfo_header(m);
4582 return seq_list_start(&slab_caches, *pos);
4585 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4587 return seq_list_next(p, &slab_caches, pos);
4590 static void s_stop(struct seq_file *m, void *p)
4592 up_read(&slub_lock);
4595 static int s_show(struct seq_file *m, void *p)
4597 unsigned long nr_partials = 0;
4598 unsigned long nr_slabs = 0;
4599 unsigned long nr_inuse = 0;
4600 unsigned long nr_objs = 0;
4601 unsigned long nr_free = 0;
4602 struct kmem_cache *s;
4605 s = list_entry(p, struct kmem_cache, list);
4607 for_each_online_node(node) {
4608 struct kmem_cache_node *n = get_node(s, node);
4613 nr_partials += n->nr_partial;
4614 nr_slabs += atomic_long_read(&n->nr_slabs);
4615 nr_objs += atomic_long_read(&n->total_objects);
4616 nr_free += count_partial(n, count_free);
4619 nr_inuse = nr_objs - nr_free;
4621 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4622 nr_objs, s->size, oo_objects(s->oo),
4623 (1 << oo_order(s->oo)));
4624 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4625 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4631 static const struct seq_operations slabinfo_op = {
4638 static int slabinfo_open(struct inode *inode, struct file *file)
4640 return seq_open(file, &slabinfo_op);
4643 static const struct file_operations proc_slabinfo_operations = {
4644 .open = slabinfo_open,
4646 .llseek = seq_lseek,
4647 .release = seq_release,
4650 static int __init slab_proc_init(void)
4652 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4655 module_init(slab_proc_init);
4656 #endif /* CONFIG_SLABINFO */