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/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/proc_fs.h>
18 #include <linux/seq_file.h>
19 #include <trace/kmemtrace.h>
20 #include <linux/cpu.h>
21 #include <linux/cpuset.h>
22 #include <linux/mempolicy.h>
23 #include <linux/ctype.h>
24 #include <linux/debugobjects.h>
25 #include <linux/kallsyms.h>
26 #include <linux/memory.h>
27 #include <linux/math64.h>
28 #include <linux/fault-inject.h>
35 * The slab_lock protects operations on the object of a particular
36 * slab and its metadata in the page struct. If the slab lock
37 * has been taken then no allocations nor frees can be performed
38 * on the objects in the slab nor can the slab be added or removed
39 * from the partial or full lists since this would mean modifying
40 * the page_struct of the slab.
42 * The list_lock protects the partial and full list on each node and
43 * the partial slab counter. If taken then no new slabs may be added or
44 * removed from the lists nor make the number of partial slabs be modified.
45 * (Note that the total number of slabs is an atomic value that may be
46 * modified without taking the list lock).
48 * The list_lock is a centralized lock and thus we avoid taking it as
49 * much as possible. As long as SLUB does not have to handle partial
50 * slabs, operations can continue without any centralized lock. F.e.
51 * allocating a long series of objects that fill up slabs does not require
54 * The lock order is sometimes inverted when we are trying to get a slab
55 * off a list. We take the list_lock and then look for a page on the list
56 * to use. While we do that objects in the slabs may be freed. We can
57 * only operate on the slab if we have also taken the slab_lock. So we use
58 * a slab_trylock() on the slab. If trylock was successful then no frees
59 * can occur anymore and we can use the slab for allocations etc. If the
60 * slab_trylock() does not succeed then frees are in progress in the slab and
61 * we must stay away from it for a while since we may cause a bouncing
62 * cacheline if we try to acquire the lock. So go onto the next slab.
63 * If all pages are busy then we may allocate a new slab instead of reusing
64 * a partial slab. A new slab has noone operating on it and thus there is
65 * no danger of cacheline contention.
67 * Interrupts are disabled during allocation and deallocation in order to
68 * make the slab allocator safe to use in the context of an irq. In addition
69 * interrupts are disabled to ensure that the processor does not change
70 * while handling per_cpu slabs, due to kernel preemption.
72 * SLUB assigns one slab for allocation to each processor.
73 * Allocations only occur from these slabs called cpu slabs.
75 * Slabs with free elements are kept on a partial list and during regular
76 * operations no list for full slabs is used. If an object in a full slab is
77 * freed then the slab will show up again on the partial lists.
78 * We track full slabs for debugging purposes though because otherwise we
79 * cannot scan all objects.
81 * Slabs are freed when they become empty. Teardown and setup is
82 * minimal so we rely on the page allocators per cpu caches for
83 * fast frees and allocs.
85 * Overloading of page flags that are otherwise used for LRU management.
87 * PageActive The slab is frozen and exempt from list processing.
88 * This means that the slab is dedicated to a purpose
89 * such as satisfying allocations for a specific
90 * processor. Objects may be freed in the slab while
91 * it is frozen but slab_free will then skip the usual
92 * list operations. It is up to the processor holding
93 * the slab to integrate the slab into the slab lists
94 * when the slab is no longer needed.
96 * One use of this flag is to mark slabs that are
97 * used for allocations. Then such a slab becomes a cpu
98 * slab. The cpu slab may be equipped with an additional
99 * freelist that allows lockless access to
100 * free objects in addition to the regular freelist
101 * that requires the slab lock.
103 * PageError Slab requires special handling due to debug
104 * options set. This moves slab handling out of
105 * the fast path and disables lockless freelists.
108 #ifdef CONFIG_SLUB_DEBUG
115 * Issues still to be resolved:
117 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
119 * - Variable sizing of the per node arrays
122 /* Enable to test recovery from slab corruption on boot */
123 #undef SLUB_RESILIENCY_TEST
126 * Mininum number of partial slabs. These will be left on the partial
127 * lists even if they are empty. kmem_cache_shrink may reclaim them.
129 #define MIN_PARTIAL 5
132 * Maximum number of desirable partial slabs.
133 * The existence of more partial slabs makes kmem_cache_shrink
134 * sort the partial list by the number of objects in the.
136 #define MAX_PARTIAL 10
138 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
139 SLAB_POISON | SLAB_STORE_USER)
142 * Set of flags that will prevent slab merging
144 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
145 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
147 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
150 #ifndef ARCH_KMALLOC_MINALIGN
151 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
154 #ifndef ARCH_SLAB_MINALIGN
155 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
159 #define OO_MASK ((1 << OO_SHIFT) - 1)
160 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
162 /* Internal SLUB flags */
163 #define __OBJECT_POISON 0x80000000 /* Poison object */
164 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
166 static int kmem_size = sizeof(struct kmem_cache);
169 static struct notifier_block slab_notifier;
173 DOWN, /* No slab functionality available */
174 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
175 UP, /* Everything works but does not show up in sysfs */
179 /* A list of all slab caches on the system */
180 static DECLARE_RWSEM(slub_lock);
181 static LIST_HEAD(slab_caches);
184 * Tracking user of a slab.
187 unsigned long addr; /* Called from address */
188 int cpu; /* Was running on cpu */
189 int pid; /* Pid context */
190 unsigned long when; /* When did the operation occur */
193 enum track_item { TRACK_ALLOC, TRACK_FREE };
195 #ifdef CONFIG_SLUB_DEBUG
196 static int sysfs_slab_add(struct kmem_cache *);
197 static int sysfs_slab_alias(struct kmem_cache *, const char *);
198 static void sysfs_slab_remove(struct kmem_cache *);
201 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
202 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
204 static inline void sysfs_slab_remove(struct kmem_cache *s)
211 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
213 #ifdef CONFIG_SLUB_STATS
218 /********************************************************************
219 * Core slab cache functions
220 *******************************************************************/
222 int slab_is_available(void)
224 return slab_state >= UP;
227 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
230 return s->node[node];
232 return &s->local_node;
236 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
239 return s->cpu_slab[cpu];
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) {
264 * Slow version of get and set free pointer.
266 * This version requires touching the cache lines of kmem_cache which
267 * we avoid to do in the fast alloc free paths. There we obtain the offset
268 * from the page struct.
270 static inline void *get_freepointer(struct kmem_cache *s, void *object)
272 return *(void **)(object + s->offset);
275 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
277 *(void **)(object + s->offset) = fp;
280 /* Loop over all objects in a slab */
281 #define for_each_object(__p, __s, __addr, __objects) \
282 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
286 #define for_each_free_object(__p, __s, __free) \
287 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 static inline struct kmem_cache_order_objects oo_make(int order,
298 struct kmem_cache_order_objects x = {
299 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
305 static inline int oo_order(struct kmem_cache_order_objects x)
307 return x.x >> OO_SHIFT;
310 static inline int oo_objects(struct kmem_cache_order_objects x)
312 return x.x & OO_MASK;
315 #ifdef CONFIG_SLUB_DEBUG
319 #ifdef CONFIG_SLUB_DEBUG_ON
320 static int slub_debug = DEBUG_DEFAULT_FLAGS;
322 static int slub_debug;
325 static char *slub_debug_slabs;
330 static void print_section(char *text, u8 *addr, unsigned int length)
338 for (i = 0; i < length; i++) {
340 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
343 printk(KERN_CONT " %02x", addr[i]);
345 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
347 printk(KERN_CONT " %s\n", ascii);
354 printk(KERN_CONT " ");
358 printk(KERN_CONT " %s\n", ascii);
362 static struct track *get_track(struct kmem_cache *s, void *object,
363 enum track_item alloc)
368 p = object + s->offset + sizeof(void *);
370 p = object + s->inuse;
375 static void set_track(struct kmem_cache *s, void *object,
376 enum track_item alloc, unsigned long addr)
381 p = object + s->offset + sizeof(void *);
383 p = object + s->inuse;
388 p->cpu = smp_processor_id();
389 p->pid = current->pid;
392 memset(p, 0, sizeof(struct track));
395 static void init_tracking(struct kmem_cache *s, void *object)
397 if (!(s->flags & SLAB_STORE_USER))
400 set_track(s, object, TRACK_FREE, 0UL);
401 set_track(s, object, TRACK_ALLOC, 0UL);
404 static void print_track(const char *s, struct track *t)
409 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
410 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
413 static void print_tracking(struct kmem_cache *s, void *object)
415 if (!(s->flags & SLAB_STORE_USER))
418 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
419 print_track("Freed", get_track(s, object, TRACK_FREE));
422 static void print_page_info(struct page *page)
424 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
425 page, page->objects, page->inuse, page->freelist, page->flags);
429 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
435 vsnprintf(buf, sizeof(buf), fmt, args);
437 printk(KERN_ERR "========================================"
438 "=====================================\n");
439 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
440 printk(KERN_ERR "----------------------------------------"
441 "-------------------------------------\n\n");
444 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
450 vsnprintf(buf, sizeof(buf), fmt, args);
452 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
455 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
457 unsigned int off; /* Offset of last byte */
458 u8 *addr = page_address(page);
460 print_tracking(s, p);
462 print_page_info(page);
464 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
465 p, p - addr, get_freepointer(s, p));
468 print_section("Bytes b4", p - 16, 16);
470 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
472 if (s->flags & SLAB_RED_ZONE)
473 print_section("Redzone", p + s->objsize,
474 s->inuse - s->objsize);
477 off = s->offset + sizeof(void *);
481 if (s->flags & SLAB_STORE_USER)
482 off += 2 * sizeof(struct track);
485 /* Beginning of the filler is the free pointer */
486 print_section("Padding", p + off, s->size - off);
491 static void object_err(struct kmem_cache *s, struct page *page,
492 u8 *object, char *reason)
494 slab_bug(s, "%s", reason);
495 print_trailer(s, page, object);
498 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
504 vsnprintf(buf, sizeof(buf), fmt, args);
506 slab_bug(s, "%s", buf);
507 print_page_info(page);
511 static void init_object(struct kmem_cache *s, void *object, int active)
515 if (s->flags & __OBJECT_POISON) {
516 memset(p, POISON_FREE, s->objsize - 1);
517 p[s->objsize - 1] = POISON_END;
520 if (s->flags & SLAB_RED_ZONE)
521 memset(p + s->objsize,
522 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
523 s->inuse - s->objsize);
526 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
529 if (*start != (u8)value)
537 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
538 void *from, void *to)
540 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
541 memset(from, data, to - from);
544 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
545 u8 *object, char *what,
546 u8 *start, unsigned int value, unsigned int bytes)
551 fault = check_bytes(start, value, bytes);
556 while (end > fault && end[-1] == value)
559 slab_bug(s, "%s overwritten", what);
560 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
561 fault, end - 1, fault[0], value);
562 print_trailer(s, page, object);
564 restore_bytes(s, what, value, fault, end);
572 * Bytes of the object to be managed.
573 * If the freepointer may overlay the object then the free
574 * pointer is the first word of the object.
576 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
579 * object + s->objsize
580 * Padding to reach word boundary. This is also used for Redzoning.
581 * Padding is extended by another word if Redzoning is enabled and
584 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
585 * 0xcc (RED_ACTIVE) for objects in use.
588 * Meta data starts here.
590 * A. Free pointer (if we cannot overwrite object on free)
591 * B. Tracking data for SLAB_STORE_USER
592 * C. Padding to reach required alignment boundary or at mininum
593 * one word if debugging is on to be able to detect writes
594 * before the word boundary.
596 * Padding is done using 0x5a (POISON_INUSE)
599 * Nothing is used beyond s->size.
601 * If slabcaches are merged then the objsize and inuse boundaries are mostly
602 * ignored. And therefore no slab options that rely on these boundaries
603 * may be used with merged slabcaches.
606 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
608 unsigned long off = s->inuse; /* The end of info */
611 /* Freepointer is placed after the object. */
612 off += sizeof(void *);
614 if (s->flags & SLAB_STORE_USER)
615 /* We also have user information there */
616 off += 2 * sizeof(struct track);
621 return check_bytes_and_report(s, page, p, "Object padding",
622 p + off, POISON_INUSE, s->size - off);
625 /* Check the pad bytes at the end of a slab page */
626 static int slab_pad_check(struct kmem_cache *s, struct page *page)
634 if (!(s->flags & SLAB_POISON))
637 start = page_address(page);
638 length = (PAGE_SIZE << compound_order(page));
639 end = start + length;
640 remainder = length % s->size;
644 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
647 while (end > fault && end[-1] == POISON_INUSE)
650 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
651 print_section("Padding", end - remainder, remainder);
653 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
657 static int check_object(struct kmem_cache *s, struct page *page,
658 void *object, int active)
661 u8 *endobject = object + s->objsize;
663 if (s->flags & SLAB_RED_ZONE) {
665 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
667 if (!check_bytes_and_report(s, page, object, "Redzone",
668 endobject, red, s->inuse - s->objsize))
671 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
672 check_bytes_and_report(s, page, p, "Alignment padding",
673 endobject, POISON_INUSE, s->inuse - s->objsize);
677 if (s->flags & SLAB_POISON) {
678 if (!active && (s->flags & __OBJECT_POISON) &&
679 (!check_bytes_and_report(s, page, p, "Poison", p,
680 POISON_FREE, s->objsize - 1) ||
681 !check_bytes_and_report(s, page, p, "Poison",
682 p + s->objsize - 1, POISON_END, 1)))
685 * check_pad_bytes cleans up on its own.
687 check_pad_bytes(s, page, p);
690 if (!s->offset && active)
692 * Object and freepointer overlap. Cannot check
693 * freepointer while object is allocated.
697 /* Check free pointer validity */
698 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
699 object_err(s, page, p, "Freepointer corrupt");
701 * No choice but to zap it and thus lose the remainder
702 * of the free objects in this slab. May cause
703 * another error because the object count is now wrong.
705 set_freepointer(s, p, NULL);
711 static int check_slab(struct kmem_cache *s, struct page *page)
715 VM_BUG_ON(!irqs_disabled());
717 if (!PageSlab(page)) {
718 slab_err(s, page, "Not a valid slab page");
722 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
723 if (page->objects > maxobj) {
724 slab_err(s, page, "objects %u > max %u",
725 s->name, page->objects, maxobj);
728 if (page->inuse > page->objects) {
729 slab_err(s, page, "inuse %u > max %u",
730 s->name, page->inuse, page->objects);
733 /* Slab_pad_check fixes things up after itself */
734 slab_pad_check(s, page);
739 * Determine if a certain object on a page is on the freelist. Must hold the
740 * slab lock to guarantee that the chains are in a consistent state.
742 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
745 void *fp = page->freelist;
747 unsigned long max_objects;
749 while (fp && nr <= page->objects) {
752 if (!check_valid_pointer(s, page, fp)) {
754 object_err(s, page, object,
755 "Freechain corrupt");
756 set_freepointer(s, object, NULL);
759 slab_err(s, page, "Freepointer corrupt");
760 page->freelist = NULL;
761 page->inuse = page->objects;
762 slab_fix(s, "Freelist cleared");
768 fp = get_freepointer(s, object);
772 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
773 if (max_objects > MAX_OBJS_PER_PAGE)
774 max_objects = MAX_OBJS_PER_PAGE;
776 if (page->objects != max_objects) {
777 slab_err(s, page, "Wrong number of objects. Found %d but "
778 "should be %d", page->objects, max_objects);
779 page->objects = max_objects;
780 slab_fix(s, "Number of objects adjusted.");
782 if (page->inuse != page->objects - nr) {
783 slab_err(s, page, "Wrong object count. Counter is %d but "
784 "counted were %d", page->inuse, page->objects - nr);
785 page->inuse = page->objects - nr;
786 slab_fix(s, "Object count adjusted.");
788 return search == NULL;
791 static void trace(struct kmem_cache *s, struct page *page, void *object,
794 if (s->flags & SLAB_TRACE) {
795 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
797 alloc ? "alloc" : "free",
802 print_section("Object", (void *)object, s->objsize);
809 * Tracking of fully allocated slabs for debugging purposes.
811 static void add_full(struct kmem_cache_node *n, struct page *page)
813 spin_lock(&n->list_lock);
814 list_add(&page->lru, &n->full);
815 spin_unlock(&n->list_lock);
818 static void remove_full(struct kmem_cache *s, struct page *page)
820 struct kmem_cache_node *n;
822 if (!(s->flags & SLAB_STORE_USER))
825 n = get_node(s, page_to_nid(page));
827 spin_lock(&n->list_lock);
828 list_del(&page->lru);
829 spin_unlock(&n->list_lock);
832 /* Tracking of the number of slabs for debugging purposes */
833 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
835 struct kmem_cache_node *n = get_node(s, node);
837 return atomic_long_read(&n->nr_slabs);
840 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
842 struct kmem_cache_node *n = get_node(s, node);
845 * May be called early in order to allocate a slab for the
846 * kmem_cache_node structure. Solve the chicken-egg
847 * dilemma by deferring the increment of the count during
848 * bootstrap (see early_kmem_cache_node_alloc).
850 if (!NUMA_BUILD || n) {
851 atomic_long_inc(&n->nr_slabs);
852 atomic_long_add(objects, &n->total_objects);
855 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
857 struct kmem_cache_node *n = get_node(s, node);
859 atomic_long_dec(&n->nr_slabs);
860 atomic_long_sub(objects, &n->total_objects);
863 /* Object debug checks for alloc/free paths */
864 static void setup_object_debug(struct kmem_cache *s, struct page *page,
867 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
870 init_object(s, object, 0);
871 init_tracking(s, object);
874 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
875 void *object, unsigned long addr)
877 if (!check_slab(s, page))
880 if (!on_freelist(s, page, object)) {
881 object_err(s, page, object, "Object already allocated");
885 if (!check_valid_pointer(s, page, object)) {
886 object_err(s, page, object, "Freelist Pointer check fails");
890 if (!check_object(s, page, object, 0))
893 /* Success perform special debug activities for allocs */
894 if (s->flags & SLAB_STORE_USER)
895 set_track(s, object, TRACK_ALLOC, addr);
896 trace(s, page, object, 1);
897 init_object(s, object, 1);
901 if (PageSlab(page)) {
903 * If this is a slab page then lets do the best we can
904 * to avoid issues in the future. Marking all objects
905 * as used avoids touching the remaining objects.
907 slab_fix(s, "Marking all objects used");
908 page->inuse = page->objects;
909 page->freelist = NULL;
914 static int free_debug_processing(struct kmem_cache *s, struct page *page,
915 void *object, unsigned long addr)
917 if (!check_slab(s, page))
920 if (!check_valid_pointer(s, page, object)) {
921 slab_err(s, page, "Invalid object pointer 0x%p", object);
925 if (on_freelist(s, page, object)) {
926 object_err(s, page, object, "Object already free");
930 if (!check_object(s, page, object, 1))
933 if (unlikely(s != page->slab)) {
934 if (!PageSlab(page)) {
935 slab_err(s, page, "Attempt to free object(0x%p) "
936 "outside of slab", object);
937 } else if (!page->slab) {
939 "SLUB <none>: no slab for object 0x%p.\n",
943 object_err(s, page, object,
944 "page slab pointer corrupt.");
948 /* Special debug activities for freeing objects */
949 if (!PageSlubFrozen(page) && !page->freelist)
950 remove_full(s, page);
951 if (s->flags & SLAB_STORE_USER)
952 set_track(s, object, TRACK_FREE, addr);
953 trace(s, page, object, 0);
954 init_object(s, object, 0);
958 slab_fix(s, "Object at 0x%p not freed", object);
962 static int __init setup_slub_debug(char *str)
964 slub_debug = DEBUG_DEFAULT_FLAGS;
965 if (*str++ != '=' || !*str)
967 * No options specified. Switch on full debugging.
973 * No options but restriction on slabs. This means full
974 * debugging for slabs matching a pattern.
981 * Switch off all debugging measures.
986 * Determine which debug features should be switched on
988 for (; *str && *str != ','; str++) {
989 switch (tolower(*str)) {
991 slub_debug |= SLAB_DEBUG_FREE;
994 slub_debug |= SLAB_RED_ZONE;
997 slub_debug |= SLAB_POISON;
1000 slub_debug |= SLAB_STORE_USER;
1003 slub_debug |= SLAB_TRACE;
1006 printk(KERN_ERR "slub_debug option '%c' "
1007 "unknown. skipped\n", *str);
1013 slub_debug_slabs = str + 1;
1018 __setup("slub_debug", setup_slub_debug);
1020 static unsigned long kmem_cache_flags(unsigned long objsize,
1021 unsigned long flags, const char *name,
1022 void (*ctor)(void *))
1025 * Enable debugging if selected on the kernel commandline.
1027 if (slub_debug && (!slub_debug_slabs ||
1028 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1029 flags |= slub_debug;
1034 static inline void setup_object_debug(struct kmem_cache *s,
1035 struct page *page, void *object) {}
1037 static inline int alloc_debug_processing(struct kmem_cache *s,
1038 struct page *page, void *object, unsigned long addr) { return 0; }
1040 static inline int free_debug_processing(struct kmem_cache *s,
1041 struct page *page, void *object, unsigned long addr) { return 0; }
1043 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1045 static inline int check_object(struct kmem_cache *s, struct page *page,
1046 void *object, int active) { return 1; }
1047 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1048 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1049 unsigned long flags, const char *name,
1050 void (*ctor)(void *))
1054 #define slub_debug 0
1056 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1058 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1060 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1065 * Slab allocation and freeing
1067 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1068 struct kmem_cache_order_objects oo)
1070 int order = oo_order(oo);
1073 return alloc_pages(flags, order);
1075 return alloc_pages_node(node, flags, order);
1078 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1081 struct kmem_cache_order_objects oo = s->oo;
1083 flags |= s->allocflags;
1085 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1087 if (unlikely(!page)) {
1090 * Allocation may have failed due to fragmentation.
1091 * Try a lower order alloc if possible
1093 page = alloc_slab_page(flags, node, oo);
1097 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1099 page->objects = oo_objects(oo);
1100 mod_zone_page_state(page_zone(page),
1101 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1102 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1108 static void setup_object(struct kmem_cache *s, struct page *page,
1111 setup_object_debug(s, page, object);
1112 if (unlikely(s->ctor))
1116 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1123 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1125 page = allocate_slab(s,
1126 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1130 inc_slabs_node(s, page_to_nid(page), page->objects);
1132 page->flags |= 1 << PG_slab;
1133 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1134 SLAB_STORE_USER | SLAB_TRACE))
1135 __SetPageSlubDebug(page);
1137 start = page_address(page);
1139 if (unlikely(s->flags & SLAB_POISON))
1140 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1143 for_each_object(p, s, start, page->objects) {
1144 setup_object(s, page, last);
1145 set_freepointer(s, last, p);
1148 setup_object(s, page, last);
1149 set_freepointer(s, last, NULL);
1151 page->freelist = start;
1157 static void __free_slab(struct kmem_cache *s, struct page *page)
1159 int order = compound_order(page);
1160 int pages = 1 << order;
1162 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1165 slab_pad_check(s, page);
1166 for_each_object(p, s, page_address(page),
1168 check_object(s, page, p, 0);
1169 __ClearPageSlubDebug(page);
1172 mod_zone_page_state(page_zone(page),
1173 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1174 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1177 __ClearPageSlab(page);
1178 reset_page_mapcount(page);
1179 __free_pages(page, order);
1182 static void rcu_free_slab(struct rcu_head *h)
1186 page = container_of((struct list_head *)h, struct page, lru);
1187 __free_slab(page->slab, page);
1190 static void free_slab(struct kmem_cache *s, struct page *page)
1192 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1194 * RCU free overloads the RCU head over the LRU
1196 struct rcu_head *head = (void *)&page->lru;
1198 call_rcu(head, rcu_free_slab);
1200 __free_slab(s, page);
1203 static void discard_slab(struct kmem_cache *s, struct page *page)
1205 dec_slabs_node(s, page_to_nid(page), page->objects);
1210 * Per slab locking using the pagelock
1212 static __always_inline void slab_lock(struct page *page)
1214 bit_spin_lock(PG_locked, &page->flags);
1217 static __always_inline void slab_unlock(struct page *page)
1219 __bit_spin_unlock(PG_locked, &page->flags);
1222 static __always_inline int slab_trylock(struct page *page)
1226 rc = bit_spin_trylock(PG_locked, &page->flags);
1231 * Management of partially allocated slabs
1233 static void add_partial(struct kmem_cache_node *n,
1234 struct page *page, int tail)
1236 spin_lock(&n->list_lock);
1239 list_add_tail(&page->lru, &n->partial);
1241 list_add(&page->lru, &n->partial);
1242 spin_unlock(&n->list_lock);
1245 static void remove_partial(struct kmem_cache *s, struct page *page)
1247 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1249 spin_lock(&n->list_lock);
1250 list_del(&page->lru);
1252 spin_unlock(&n->list_lock);
1256 * Lock slab and remove from the partial list.
1258 * Must hold list_lock.
1260 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1263 if (slab_trylock(page)) {
1264 list_del(&page->lru);
1266 __SetPageSlubFrozen(page);
1273 * Try to allocate a partial slab from a specific node.
1275 static struct page *get_partial_node(struct kmem_cache_node *n)
1280 * Racy check. If we mistakenly see no partial slabs then we
1281 * just allocate an empty slab. If we mistakenly try to get a
1282 * partial slab and there is none available then get_partials()
1285 if (!n || !n->nr_partial)
1288 spin_lock(&n->list_lock);
1289 list_for_each_entry(page, &n->partial, lru)
1290 if (lock_and_freeze_slab(n, page))
1294 spin_unlock(&n->list_lock);
1299 * Get a page from somewhere. Search in increasing NUMA distances.
1301 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1304 struct zonelist *zonelist;
1307 enum zone_type high_zoneidx = gfp_zone(flags);
1311 * The defrag ratio allows a configuration of the tradeoffs between
1312 * inter node defragmentation and node local allocations. A lower
1313 * defrag_ratio increases the tendency to do local allocations
1314 * instead of attempting to obtain partial slabs from other nodes.
1316 * If the defrag_ratio is set to 0 then kmalloc() always
1317 * returns node local objects. If the ratio is higher then kmalloc()
1318 * may return off node objects because partial slabs are obtained
1319 * from other nodes and filled up.
1321 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1322 * defrag_ratio = 1000) then every (well almost) allocation will
1323 * first attempt to defrag slab caches on other nodes. This means
1324 * scanning over all nodes to look for partial slabs which may be
1325 * expensive if we do it every time we are trying to find a slab
1326 * with available objects.
1328 if (!s->remote_node_defrag_ratio ||
1329 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1332 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1333 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1334 struct kmem_cache_node *n;
1336 n = get_node(s, zone_to_nid(zone));
1338 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1339 n->nr_partial > n->min_partial) {
1340 page = get_partial_node(n);
1350 * Get a partial page, lock it and return it.
1352 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1355 int searchnode = (node == -1) ? numa_node_id() : node;
1357 page = get_partial_node(get_node(s, searchnode));
1358 if (page || (flags & __GFP_THISNODE))
1361 return get_any_partial(s, flags);
1365 * Move a page back to the lists.
1367 * Must be called with the slab lock held.
1369 * On exit the slab lock will have been dropped.
1371 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1373 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1374 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1376 __ClearPageSlubFrozen(page);
1379 if (page->freelist) {
1380 add_partial(n, page, tail);
1381 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1383 stat(c, DEACTIVATE_FULL);
1384 if (SLABDEBUG && PageSlubDebug(page) &&
1385 (s->flags & SLAB_STORE_USER))
1390 stat(c, DEACTIVATE_EMPTY);
1391 if (n->nr_partial < n->min_partial) {
1393 * Adding an empty slab to the partial slabs in order
1394 * to avoid page allocator overhead. This slab needs
1395 * to come after the other slabs with objects in
1396 * so that the others get filled first. That way the
1397 * size of the partial list stays small.
1399 * kmem_cache_shrink can reclaim any empty slabs from
1402 add_partial(n, page, 1);
1406 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1407 discard_slab(s, page);
1413 * Remove the cpu slab
1415 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1417 struct page *page = c->page;
1421 stat(c, DEACTIVATE_REMOTE_FREES);
1423 * Merge cpu freelist into slab freelist. Typically we get here
1424 * because both freelists are empty. So this is unlikely
1427 while (unlikely(c->freelist)) {
1430 tail = 0; /* Hot objects. Put the slab first */
1432 /* Retrieve object from cpu_freelist */
1433 object = c->freelist;
1434 c->freelist = c->freelist[c->offset];
1436 /* And put onto the regular freelist */
1437 object[c->offset] = page->freelist;
1438 page->freelist = object;
1442 unfreeze_slab(s, page, tail);
1445 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1447 stat(c, CPUSLAB_FLUSH);
1449 deactivate_slab(s, c);
1455 * Called from IPI handler with interrupts disabled.
1457 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1459 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1461 if (likely(c && c->page))
1465 static void flush_cpu_slab(void *d)
1467 struct kmem_cache *s = d;
1469 __flush_cpu_slab(s, smp_processor_id());
1472 static void flush_all(struct kmem_cache *s)
1474 on_each_cpu(flush_cpu_slab, s, 1);
1478 * Check if the objects in a per cpu structure fit numa
1479 * locality expectations.
1481 static inline int node_match(struct kmem_cache_cpu *c, int node)
1484 if (node != -1 && c->node != node)
1491 * Slow path. The lockless freelist is empty or we need to perform
1494 * Interrupts are disabled.
1496 * Processing is still very fast if new objects have been freed to the
1497 * regular freelist. In that case we simply take over the regular freelist
1498 * as the lockless freelist and zap the regular freelist.
1500 * If that is not working then we fall back to the partial lists. We take the
1501 * first element of the freelist as the object to allocate now and move the
1502 * rest of the freelist to the lockless freelist.
1504 * And if we were unable to get a new slab from the partial slab lists then
1505 * we need to allocate a new slab. This is the slowest path since it involves
1506 * a call to the page allocator and the setup of a new slab.
1508 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1509 unsigned long addr, struct kmem_cache_cpu *c)
1514 /* We handle __GFP_ZERO in the caller */
1515 gfpflags &= ~__GFP_ZERO;
1521 if (unlikely(!node_match(c, node)))
1524 stat(c, ALLOC_REFILL);
1527 object = c->page->freelist;
1528 if (unlikely(!object))
1530 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1533 c->freelist = object[c->offset];
1534 c->page->inuse = c->page->objects;
1535 c->page->freelist = NULL;
1536 c->node = page_to_nid(c->page);
1538 slab_unlock(c->page);
1539 stat(c, ALLOC_SLOWPATH);
1543 deactivate_slab(s, c);
1546 new = get_partial(s, gfpflags, node);
1549 stat(c, ALLOC_FROM_PARTIAL);
1553 if (gfpflags & __GFP_WAIT)
1556 new = new_slab(s, gfpflags, node);
1558 if (gfpflags & __GFP_WAIT)
1559 local_irq_disable();
1562 c = get_cpu_slab(s, smp_processor_id());
1563 stat(c, ALLOC_SLAB);
1567 __SetPageSlubFrozen(new);
1573 if (!alloc_debug_processing(s, c->page, object, addr))
1577 c->page->freelist = object[c->offset];
1583 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1584 * have the fastpath folded into their functions. So no function call
1585 * overhead for requests that can be satisfied on the fastpath.
1587 * The fastpath works by first checking if the lockless freelist can be used.
1588 * If not then __slab_alloc is called for slow processing.
1590 * Otherwise we can simply pick the next object from the lockless free list.
1592 static __always_inline void *slab_alloc(struct kmem_cache *s,
1593 gfp_t gfpflags, int node, unsigned long addr)
1596 struct kmem_cache_cpu *c;
1597 unsigned long flags;
1598 unsigned int objsize;
1600 might_sleep_if(gfpflags & __GFP_WAIT);
1602 if (should_failslab(s->objsize, gfpflags))
1605 local_irq_save(flags);
1606 c = get_cpu_slab(s, smp_processor_id());
1607 objsize = c->objsize;
1608 if (unlikely(!c->freelist || !node_match(c, node)))
1610 object = __slab_alloc(s, gfpflags, node, addr, c);
1613 object = c->freelist;
1614 c->freelist = object[c->offset];
1615 stat(c, ALLOC_FASTPATH);
1617 local_irq_restore(flags);
1619 if (unlikely((gfpflags & __GFP_ZERO) && object))
1620 memset(object, 0, objsize);
1625 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1627 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1629 kmemtrace_mark_alloc(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
1630 s->objsize, s->size, gfpflags);
1634 EXPORT_SYMBOL(kmem_cache_alloc);
1636 #ifdef CONFIG_KMEMTRACE
1637 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1639 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1641 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1645 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1647 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1649 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
1650 s->objsize, s->size, gfpflags, node);
1654 EXPORT_SYMBOL(kmem_cache_alloc_node);
1657 #ifdef CONFIG_KMEMTRACE
1658 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1662 return slab_alloc(s, gfpflags, node, _RET_IP_);
1664 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1668 * Slow patch handling. This may still be called frequently since objects
1669 * have a longer lifetime than the cpu slabs in most processing loads.
1671 * So we still attempt to reduce cache line usage. Just take the slab
1672 * lock and free the item. If there is no additional partial page
1673 * handling required then we can return immediately.
1675 static void __slab_free(struct kmem_cache *s, struct page *page,
1676 void *x, unsigned long addr, unsigned int offset)
1679 void **object = (void *)x;
1680 struct kmem_cache_cpu *c;
1682 c = get_cpu_slab(s, raw_smp_processor_id());
1683 stat(c, FREE_SLOWPATH);
1686 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1690 prior = object[offset] = page->freelist;
1691 page->freelist = object;
1694 if (unlikely(PageSlubFrozen(page))) {
1695 stat(c, FREE_FROZEN);
1699 if (unlikely(!page->inuse))
1703 * Objects left in the slab. If it was not on the partial list before
1706 if (unlikely(!prior)) {
1707 add_partial(get_node(s, page_to_nid(page)), page, 1);
1708 stat(c, FREE_ADD_PARTIAL);
1718 * Slab still on the partial list.
1720 remove_partial(s, page);
1721 stat(c, FREE_REMOVE_PARTIAL);
1725 discard_slab(s, page);
1729 if (!free_debug_processing(s, page, x, addr))
1735 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1736 * can perform fastpath freeing without additional function calls.
1738 * The fastpath is only possible if we are freeing to the current cpu slab
1739 * of this processor. This typically the case if we have just allocated
1742 * If fastpath is not possible then fall back to __slab_free where we deal
1743 * with all sorts of special processing.
1745 static __always_inline void slab_free(struct kmem_cache *s,
1746 struct page *page, void *x, unsigned long addr)
1748 void **object = (void *)x;
1749 struct kmem_cache_cpu *c;
1750 unsigned long flags;
1752 local_irq_save(flags);
1753 c = get_cpu_slab(s, smp_processor_id());
1754 debug_check_no_locks_freed(object, c->objsize);
1755 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1756 debug_check_no_obj_freed(object, s->objsize);
1757 if (likely(page == c->page && c->node >= 0)) {
1758 object[c->offset] = c->freelist;
1759 c->freelist = object;
1760 stat(c, FREE_FASTPATH);
1762 __slab_free(s, page, x, addr, c->offset);
1764 local_irq_restore(flags);
1767 void kmem_cache_free(struct kmem_cache *s, void *x)
1771 page = virt_to_head_page(x);
1773 slab_free(s, page, x, _RET_IP_);
1775 kmemtrace_mark_free(KMEMTRACE_TYPE_CACHE, _RET_IP_, x);
1777 EXPORT_SYMBOL(kmem_cache_free);
1779 /* Figure out on which slab page the object resides */
1780 static struct page *get_object_page(const void *x)
1782 struct page *page = virt_to_head_page(x);
1784 if (!PageSlab(page))
1791 * Object placement in a slab is made very easy because we always start at
1792 * offset 0. If we tune the size of the object to the alignment then we can
1793 * get the required alignment by putting one properly sized object after
1796 * Notice that the allocation order determines the sizes of the per cpu
1797 * caches. Each processor has always one slab available for allocations.
1798 * Increasing the allocation order reduces the number of times that slabs
1799 * must be moved on and off the partial lists and is therefore a factor in
1804 * Mininum / Maximum order of slab pages. This influences locking overhead
1805 * and slab fragmentation. A higher order reduces the number of partial slabs
1806 * and increases the number of allocations possible without having to
1807 * take the list_lock.
1809 static int slub_min_order;
1810 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1811 static int slub_min_objects;
1814 * Merge control. If this is set then no merging of slab caches will occur.
1815 * (Could be removed. This was introduced to pacify the merge skeptics.)
1817 static int slub_nomerge;
1820 * Calculate the order of allocation given an slab object size.
1822 * The order of allocation has significant impact on performance and other
1823 * system components. Generally order 0 allocations should be preferred since
1824 * order 0 does not cause fragmentation in the page allocator. Larger objects
1825 * be problematic to put into order 0 slabs because there may be too much
1826 * unused space left. We go to a higher order if more than 1/16th of the slab
1829 * In order to reach satisfactory performance we must ensure that a minimum
1830 * number of objects is in one slab. Otherwise we may generate too much
1831 * activity on the partial lists which requires taking the list_lock. This is
1832 * less a concern for large slabs though which are rarely used.
1834 * slub_max_order specifies the order where we begin to stop considering the
1835 * number of objects in a slab as critical. If we reach slub_max_order then
1836 * we try to keep the page order as low as possible. So we accept more waste
1837 * of space in favor of a small page order.
1839 * Higher order allocations also allow the placement of more objects in a
1840 * slab and thereby reduce object handling overhead. If the user has
1841 * requested a higher mininum order then we start with that one instead of
1842 * the smallest order which will fit the object.
1844 static inline int slab_order(int size, int min_objects,
1845 int max_order, int fract_leftover)
1849 int min_order = slub_min_order;
1851 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1852 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1854 for (order = max(min_order,
1855 fls(min_objects * size - 1) - PAGE_SHIFT);
1856 order <= max_order; order++) {
1858 unsigned long slab_size = PAGE_SIZE << order;
1860 if (slab_size < min_objects * size)
1863 rem = slab_size % size;
1865 if (rem <= slab_size / fract_leftover)
1873 static inline int calculate_order(int size)
1880 * Attempt to find best configuration for a slab. This
1881 * works by first attempting to generate a layout with
1882 * the best configuration and backing off gradually.
1884 * First we reduce the acceptable waste in a slab. Then
1885 * we reduce the minimum objects required in a slab.
1887 min_objects = slub_min_objects;
1889 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1890 while (min_objects > 1) {
1892 while (fraction >= 4) {
1893 order = slab_order(size, min_objects,
1894 slub_max_order, fraction);
1895 if (order <= slub_max_order)
1903 * We were unable to place multiple objects in a slab. Now
1904 * lets see if we can place a single object there.
1906 order = slab_order(size, 1, slub_max_order, 1);
1907 if (order <= slub_max_order)
1911 * Doh this slab cannot be placed using slub_max_order.
1913 order = slab_order(size, 1, MAX_ORDER, 1);
1914 if (order <= MAX_ORDER)
1920 * Figure out what the alignment of the objects will be.
1922 static unsigned long calculate_alignment(unsigned long flags,
1923 unsigned long align, unsigned long size)
1926 * If the user wants hardware cache aligned objects then follow that
1927 * suggestion if the object is sufficiently large.
1929 * The hardware cache alignment cannot override the specified
1930 * alignment though. If that is greater then use it.
1932 if (flags & SLAB_HWCACHE_ALIGN) {
1933 unsigned long ralign = cache_line_size();
1934 while (size <= ralign / 2)
1936 align = max(align, ralign);
1939 if (align < ARCH_SLAB_MINALIGN)
1940 align = ARCH_SLAB_MINALIGN;
1942 return ALIGN(align, sizeof(void *));
1945 static void init_kmem_cache_cpu(struct kmem_cache *s,
1946 struct kmem_cache_cpu *c)
1951 c->offset = s->offset / sizeof(void *);
1952 c->objsize = s->objsize;
1953 #ifdef CONFIG_SLUB_STATS
1954 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1959 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1964 * The larger the object size is, the more pages we want on the partial
1965 * list to avoid pounding the page allocator excessively.
1967 n->min_partial = ilog2(s->size);
1968 if (n->min_partial < MIN_PARTIAL)
1969 n->min_partial = MIN_PARTIAL;
1970 else if (n->min_partial > MAX_PARTIAL)
1971 n->min_partial = MAX_PARTIAL;
1973 spin_lock_init(&n->list_lock);
1974 INIT_LIST_HEAD(&n->partial);
1975 #ifdef CONFIG_SLUB_DEBUG
1976 atomic_long_set(&n->nr_slabs, 0);
1977 atomic_long_set(&n->total_objects, 0);
1978 INIT_LIST_HEAD(&n->full);
1984 * Per cpu array for per cpu structures.
1986 * The per cpu array places all kmem_cache_cpu structures from one processor
1987 * close together meaning that it becomes possible that multiple per cpu
1988 * structures are contained in one cacheline. This may be particularly
1989 * beneficial for the kmalloc caches.
1991 * A desktop system typically has around 60-80 slabs. With 100 here we are
1992 * likely able to get per cpu structures for all caches from the array defined
1993 * here. We must be able to cover all kmalloc caches during bootstrap.
1995 * If the per cpu array is exhausted then fall back to kmalloc
1996 * of individual cachelines. No sharing is possible then.
1998 #define NR_KMEM_CACHE_CPU 100
2000 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2001 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2003 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2004 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
2006 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2007 int cpu, gfp_t flags)
2009 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2012 per_cpu(kmem_cache_cpu_free, cpu) =
2013 (void *)c->freelist;
2015 /* Table overflow: So allocate ourselves */
2017 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2018 flags, cpu_to_node(cpu));
2023 init_kmem_cache_cpu(s, c);
2027 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2029 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2030 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2034 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2035 per_cpu(kmem_cache_cpu_free, cpu) = c;
2038 static void free_kmem_cache_cpus(struct kmem_cache *s)
2042 for_each_online_cpu(cpu) {
2043 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2046 s->cpu_slab[cpu] = NULL;
2047 free_kmem_cache_cpu(c, cpu);
2052 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2056 for_each_online_cpu(cpu) {
2057 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2062 c = alloc_kmem_cache_cpu(s, cpu, flags);
2064 free_kmem_cache_cpus(s);
2067 s->cpu_slab[cpu] = c;
2073 * Initialize the per cpu array.
2075 static void init_alloc_cpu_cpu(int cpu)
2079 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2082 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2083 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2085 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2088 static void __init init_alloc_cpu(void)
2092 for_each_online_cpu(cpu)
2093 init_alloc_cpu_cpu(cpu);
2097 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2098 static inline void init_alloc_cpu(void) {}
2100 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2102 init_kmem_cache_cpu(s, &s->cpu_slab);
2109 * No kmalloc_node yet so do it by hand. We know that this is the first
2110 * slab on the node for this slabcache. There are no concurrent accesses
2113 * Note that this function only works on the kmalloc_node_cache
2114 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2115 * memory on a fresh node that has no slab structures yet.
2117 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2120 struct kmem_cache_node *n;
2121 unsigned long flags;
2123 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2125 page = new_slab(kmalloc_caches, gfpflags, node);
2128 if (page_to_nid(page) != node) {
2129 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2131 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2132 "in order to be able to continue\n");
2137 page->freelist = get_freepointer(kmalloc_caches, n);
2139 kmalloc_caches->node[node] = n;
2140 #ifdef CONFIG_SLUB_DEBUG
2141 init_object(kmalloc_caches, n, 1);
2142 init_tracking(kmalloc_caches, n);
2144 init_kmem_cache_node(n, kmalloc_caches);
2145 inc_slabs_node(kmalloc_caches, node, page->objects);
2148 * lockdep requires consistent irq usage for each lock
2149 * so even though there cannot be a race this early in
2150 * the boot sequence, we still disable irqs.
2152 local_irq_save(flags);
2153 add_partial(n, page, 0);
2154 local_irq_restore(flags);
2157 static void free_kmem_cache_nodes(struct kmem_cache *s)
2161 for_each_node_state(node, N_NORMAL_MEMORY) {
2162 struct kmem_cache_node *n = s->node[node];
2163 if (n && n != &s->local_node)
2164 kmem_cache_free(kmalloc_caches, n);
2165 s->node[node] = NULL;
2169 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2174 if (slab_state >= UP)
2175 local_node = page_to_nid(virt_to_page(s));
2179 for_each_node_state(node, N_NORMAL_MEMORY) {
2180 struct kmem_cache_node *n;
2182 if (local_node == node)
2185 if (slab_state == DOWN) {
2186 early_kmem_cache_node_alloc(gfpflags, node);
2189 n = kmem_cache_alloc_node(kmalloc_caches,
2193 free_kmem_cache_nodes(s);
2199 init_kmem_cache_node(n, s);
2204 static void free_kmem_cache_nodes(struct kmem_cache *s)
2208 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2210 init_kmem_cache_node(&s->local_node, s);
2216 * calculate_sizes() determines the order and the distribution of data within
2219 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2221 unsigned long flags = s->flags;
2222 unsigned long size = s->objsize;
2223 unsigned long align = s->align;
2227 * Round up object size to the next word boundary. We can only
2228 * place the free pointer at word boundaries and this determines
2229 * the possible location of the free pointer.
2231 size = ALIGN(size, sizeof(void *));
2233 #ifdef CONFIG_SLUB_DEBUG
2235 * Determine if we can poison the object itself. If the user of
2236 * the slab may touch the object after free or before allocation
2237 * then we should never poison the object itself.
2239 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2241 s->flags |= __OBJECT_POISON;
2243 s->flags &= ~__OBJECT_POISON;
2247 * If we are Redzoning then check if there is some space between the
2248 * end of the object and the free pointer. If not then add an
2249 * additional word to have some bytes to store Redzone information.
2251 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2252 size += sizeof(void *);
2256 * With that we have determined the number of bytes in actual use
2257 * by the object. This is the potential offset to the free pointer.
2261 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2264 * Relocate free pointer after the object if it is not
2265 * permitted to overwrite the first word of the object on
2268 * This is the case if we do RCU, have a constructor or
2269 * destructor or are poisoning the objects.
2272 size += sizeof(void *);
2275 #ifdef CONFIG_SLUB_DEBUG
2276 if (flags & SLAB_STORE_USER)
2278 * Need to store information about allocs and frees after
2281 size += 2 * sizeof(struct track);
2283 if (flags & SLAB_RED_ZONE)
2285 * Add some empty padding so that we can catch
2286 * overwrites from earlier objects rather than let
2287 * tracking information or the free pointer be
2288 * corrupted if a user writes before the start
2291 size += sizeof(void *);
2295 * Determine the alignment based on various parameters that the
2296 * user specified and the dynamic determination of cache line size
2299 align = calculate_alignment(flags, align, s->objsize);
2302 * SLUB stores one object immediately after another beginning from
2303 * offset 0. In order to align the objects we have to simply size
2304 * each object to conform to the alignment.
2306 size = ALIGN(size, align);
2308 if (forced_order >= 0)
2309 order = forced_order;
2311 order = calculate_order(size);
2318 s->allocflags |= __GFP_COMP;
2320 if (s->flags & SLAB_CACHE_DMA)
2321 s->allocflags |= SLUB_DMA;
2323 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2324 s->allocflags |= __GFP_RECLAIMABLE;
2327 * Determine the number of objects per slab
2329 s->oo = oo_make(order, size);
2330 s->min = oo_make(get_order(size), size);
2331 if (oo_objects(s->oo) > oo_objects(s->max))
2334 return !!oo_objects(s->oo);
2338 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2339 const char *name, size_t size,
2340 size_t align, unsigned long flags,
2341 void (*ctor)(void *))
2343 memset(s, 0, kmem_size);
2348 s->flags = kmem_cache_flags(size, flags, name, ctor);
2350 if (!calculate_sizes(s, -1))
2355 s->remote_node_defrag_ratio = 1000;
2357 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2360 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2362 free_kmem_cache_nodes(s);
2364 if (flags & SLAB_PANIC)
2365 panic("Cannot create slab %s size=%lu realsize=%u "
2366 "order=%u offset=%u flags=%lx\n",
2367 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2373 * Check if a given pointer is valid
2375 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2379 page = get_object_page(object);
2381 if (!page || s != page->slab)
2382 /* No slab or wrong slab */
2385 if (!check_valid_pointer(s, page, object))
2389 * We could also check if the object is on the slabs freelist.
2390 * But this would be too expensive and it seems that the main
2391 * purpose of kmem_ptr_valid() is to check if the object belongs
2392 * to a certain slab.
2396 EXPORT_SYMBOL(kmem_ptr_validate);
2399 * Determine the size of a slab object
2401 unsigned int kmem_cache_size(struct kmem_cache *s)
2405 EXPORT_SYMBOL(kmem_cache_size);
2407 const char *kmem_cache_name(struct kmem_cache *s)
2411 EXPORT_SYMBOL(kmem_cache_name);
2413 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2416 #ifdef CONFIG_SLUB_DEBUG
2417 void *addr = page_address(page);
2419 DECLARE_BITMAP(map, page->objects);
2421 bitmap_zero(map, page->objects);
2422 slab_err(s, page, "%s", text);
2424 for_each_free_object(p, s, page->freelist)
2425 set_bit(slab_index(p, s, addr), map);
2427 for_each_object(p, s, addr, page->objects) {
2429 if (!test_bit(slab_index(p, s, addr), map)) {
2430 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2432 print_tracking(s, p);
2440 * Attempt to free all partial slabs on a node.
2442 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2444 unsigned long flags;
2445 struct page *page, *h;
2447 spin_lock_irqsave(&n->list_lock, flags);
2448 list_for_each_entry_safe(page, h, &n->partial, lru) {
2450 list_del(&page->lru);
2451 discard_slab(s, page);
2454 list_slab_objects(s, page,
2455 "Objects remaining on kmem_cache_close()");
2458 spin_unlock_irqrestore(&n->list_lock, flags);
2462 * Release all resources used by a slab cache.
2464 static inline int kmem_cache_close(struct kmem_cache *s)
2470 /* Attempt to free all objects */
2471 free_kmem_cache_cpus(s);
2472 for_each_node_state(node, N_NORMAL_MEMORY) {
2473 struct kmem_cache_node *n = get_node(s, node);
2476 if (n->nr_partial || slabs_node(s, node))
2479 free_kmem_cache_nodes(s);
2484 * Close a cache and release the kmem_cache structure
2485 * (must be used for caches created using kmem_cache_create)
2487 void kmem_cache_destroy(struct kmem_cache *s)
2489 down_write(&slub_lock);
2493 up_write(&slub_lock);
2494 if (kmem_cache_close(s)) {
2495 printk(KERN_ERR "SLUB %s: %s called for cache that "
2496 "still has objects.\n", s->name, __func__);
2499 sysfs_slab_remove(s);
2501 up_write(&slub_lock);
2503 EXPORT_SYMBOL(kmem_cache_destroy);
2505 /********************************************************************
2507 *******************************************************************/
2509 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2510 EXPORT_SYMBOL(kmalloc_caches);
2512 static int __init setup_slub_min_order(char *str)
2514 get_option(&str, &slub_min_order);
2519 __setup("slub_min_order=", setup_slub_min_order);
2521 static int __init setup_slub_max_order(char *str)
2523 get_option(&str, &slub_max_order);
2528 __setup("slub_max_order=", setup_slub_max_order);
2530 static int __init setup_slub_min_objects(char *str)
2532 get_option(&str, &slub_min_objects);
2537 __setup("slub_min_objects=", setup_slub_min_objects);
2539 static int __init setup_slub_nomerge(char *str)
2545 __setup("slub_nomerge", setup_slub_nomerge);
2547 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2548 const char *name, int size, gfp_t gfp_flags)
2550 unsigned int flags = 0;
2552 if (gfp_flags & SLUB_DMA)
2553 flags = SLAB_CACHE_DMA;
2555 down_write(&slub_lock);
2556 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2560 list_add(&s->list, &slab_caches);
2561 up_write(&slub_lock);
2562 if (sysfs_slab_add(s))
2567 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2570 #ifdef CONFIG_ZONE_DMA
2571 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2573 static void sysfs_add_func(struct work_struct *w)
2575 struct kmem_cache *s;
2577 down_write(&slub_lock);
2578 list_for_each_entry(s, &slab_caches, list) {
2579 if (s->flags & __SYSFS_ADD_DEFERRED) {
2580 s->flags &= ~__SYSFS_ADD_DEFERRED;
2584 up_write(&slub_lock);
2587 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2589 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2591 struct kmem_cache *s;
2595 s = kmalloc_caches_dma[index];
2599 /* Dynamically create dma cache */
2600 if (flags & __GFP_WAIT)
2601 down_write(&slub_lock);
2603 if (!down_write_trylock(&slub_lock))
2607 if (kmalloc_caches_dma[index])
2610 realsize = kmalloc_caches[index].objsize;
2611 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2612 (unsigned int)realsize);
2613 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2615 if (!s || !text || !kmem_cache_open(s, flags, text,
2616 realsize, ARCH_KMALLOC_MINALIGN,
2617 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2623 list_add(&s->list, &slab_caches);
2624 kmalloc_caches_dma[index] = s;
2626 schedule_work(&sysfs_add_work);
2629 up_write(&slub_lock);
2631 return kmalloc_caches_dma[index];
2636 * Conversion table for small slabs sizes / 8 to the index in the
2637 * kmalloc array. This is necessary for slabs < 192 since we have non power
2638 * of two cache sizes there. The size of larger slabs can be determined using
2641 static s8 size_index[24] = {
2668 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2674 return ZERO_SIZE_PTR;
2676 index = size_index[(size - 1) / 8];
2678 index = fls(size - 1);
2680 #ifdef CONFIG_ZONE_DMA
2681 if (unlikely((flags & SLUB_DMA)))
2682 return dma_kmalloc_cache(index, flags);
2685 return &kmalloc_caches[index];
2688 void *__kmalloc(size_t size, gfp_t flags)
2690 struct kmem_cache *s;
2693 if (unlikely(size > SLUB_MAX_SIZE))
2694 return kmalloc_large(size, flags);
2696 s = get_slab(size, flags);
2698 if (unlikely(ZERO_OR_NULL_PTR(s)))
2701 ret = slab_alloc(s, flags, -1, _RET_IP_);
2703 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, ret,
2704 size, s->size, flags);
2708 EXPORT_SYMBOL(__kmalloc);
2710 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2712 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2716 return page_address(page);
2722 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2724 struct kmem_cache *s;
2727 if (unlikely(size > SLUB_MAX_SIZE)) {
2728 ret = kmalloc_large_node(size, flags, node);
2730 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC,
2732 size, PAGE_SIZE << get_order(size),
2738 s = get_slab(size, flags);
2740 if (unlikely(ZERO_OR_NULL_PTR(s)))
2743 ret = slab_alloc(s, flags, node, _RET_IP_);
2745 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, ret,
2746 size, s->size, flags, node);
2750 EXPORT_SYMBOL(__kmalloc_node);
2753 size_t ksize(const void *object)
2756 struct kmem_cache *s;
2758 if (unlikely(object == ZERO_SIZE_PTR))
2761 page = virt_to_head_page(object);
2763 if (unlikely(!PageSlab(page))) {
2764 WARN_ON(!PageCompound(page));
2765 return PAGE_SIZE << compound_order(page);
2769 #ifdef CONFIG_SLUB_DEBUG
2771 * Debugging requires use of the padding between object
2772 * and whatever may come after it.
2774 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2779 * If we have the need to store the freelist pointer
2780 * back there or track user information then we can
2781 * only use the space before that information.
2783 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2786 * Else we can use all the padding etc for the allocation
2790 EXPORT_SYMBOL(ksize);
2792 void kfree(const void *x)
2795 void *object = (void *)x;
2797 if (unlikely(ZERO_OR_NULL_PTR(x)))
2800 page = virt_to_head_page(x);
2801 if (unlikely(!PageSlab(page))) {
2802 BUG_ON(!PageCompound(page));
2806 slab_free(page->slab, page, object, _RET_IP_);
2808 kmemtrace_mark_free(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, x);
2810 EXPORT_SYMBOL(kfree);
2813 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2814 * the remaining slabs by the number of items in use. The slabs with the
2815 * most items in use come first. New allocations will then fill those up
2816 * and thus they can be removed from the partial lists.
2818 * The slabs with the least items are placed last. This results in them
2819 * being allocated from last increasing the chance that the last objects
2820 * are freed in them.
2822 int kmem_cache_shrink(struct kmem_cache *s)
2826 struct kmem_cache_node *n;
2829 int objects = oo_objects(s->max);
2830 struct list_head *slabs_by_inuse =
2831 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2832 unsigned long flags;
2834 if (!slabs_by_inuse)
2838 for_each_node_state(node, N_NORMAL_MEMORY) {
2839 n = get_node(s, node);
2844 for (i = 0; i < objects; i++)
2845 INIT_LIST_HEAD(slabs_by_inuse + i);
2847 spin_lock_irqsave(&n->list_lock, flags);
2850 * Build lists indexed by the items in use in each slab.
2852 * Note that concurrent frees may occur while we hold the
2853 * list_lock. page->inuse here is the upper limit.
2855 list_for_each_entry_safe(page, t, &n->partial, lru) {
2856 if (!page->inuse && slab_trylock(page)) {
2858 * Must hold slab lock here because slab_free
2859 * may have freed the last object and be
2860 * waiting to release the slab.
2862 list_del(&page->lru);
2865 discard_slab(s, page);
2867 list_move(&page->lru,
2868 slabs_by_inuse + page->inuse);
2873 * Rebuild the partial list with the slabs filled up most
2874 * first and the least used slabs at the end.
2876 for (i = objects - 1; i >= 0; i--)
2877 list_splice(slabs_by_inuse + i, n->partial.prev);
2879 spin_unlock_irqrestore(&n->list_lock, flags);
2882 kfree(slabs_by_inuse);
2885 EXPORT_SYMBOL(kmem_cache_shrink);
2887 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2888 static int slab_mem_going_offline_callback(void *arg)
2890 struct kmem_cache *s;
2892 down_read(&slub_lock);
2893 list_for_each_entry(s, &slab_caches, list)
2894 kmem_cache_shrink(s);
2895 up_read(&slub_lock);
2900 static void slab_mem_offline_callback(void *arg)
2902 struct kmem_cache_node *n;
2903 struct kmem_cache *s;
2904 struct memory_notify *marg = arg;
2907 offline_node = marg->status_change_nid;
2910 * If the node still has available memory. we need kmem_cache_node
2913 if (offline_node < 0)
2916 down_read(&slub_lock);
2917 list_for_each_entry(s, &slab_caches, list) {
2918 n = get_node(s, offline_node);
2921 * if n->nr_slabs > 0, slabs still exist on the node
2922 * that is going down. We were unable to free them,
2923 * and offline_pages() function shoudn't call this
2924 * callback. So, we must fail.
2926 BUG_ON(slabs_node(s, offline_node));
2928 s->node[offline_node] = NULL;
2929 kmem_cache_free(kmalloc_caches, n);
2932 up_read(&slub_lock);
2935 static int slab_mem_going_online_callback(void *arg)
2937 struct kmem_cache_node *n;
2938 struct kmem_cache *s;
2939 struct memory_notify *marg = arg;
2940 int nid = marg->status_change_nid;
2944 * If the node's memory is already available, then kmem_cache_node is
2945 * already created. Nothing to do.
2951 * We are bringing a node online. No memory is available yet. We must
2952 * allocate a kmem_cache_node structure in order to bring the node
2955 down_read(&slub_lock);
2956 list_for_each_entry(s, &slab_caches, list) {
2958 * XXX: kmem_cache_alloc_node will fallback to other nodes
2959 * since memory is not yet available from the node that
2962 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2967 init_kmem_cache_node(n, s);
2971 up_read(&slub_lock);
2975 static int slab_memory_callback(struct notifier_block *self,
2976 unsigned long action, void *arg)
2981 case MEM_GOING_ONLINE:
2982 ret = slab_mem_going_online_callback(arg);
2984 case MEM_GOING_OFFLINE:
2985 ret = slab_mem_going_offline_callback(arg);
2988 case MEM_CANCEL_ONLINE:
2989 slab_mem_offline_callback(arg);
2992 case MEM_CANCEL_OFFLINE:
2996 ret = notifier_from_errno(ret);
3002 #endif /* CONFIG_MEMORY_HOTPLUG */
3004 /********************************************************************
3005 * Basic setup of slabs
3006 *******************************************************************/
3008 void __init kmem_cache_init(void)
3017 * Must first have the slab cache available for the allocations of the
3018 * struct kmem_cache_node's. There is special bootstrap code in
3019 * kmem_cache_open for slab_state == DOWN.
3021 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3022 sizeof(struct kmem_cache_node), GFP_KERNEL);
3023 kmalloc_caches[0].refcount = -1;
3026 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3029 /* Able to allocate the per node structures */
3030 slab_state = PARTIAL;
3032 /* Caches that are not of the two-to-the-power-of size */
3033 if (KMALLOC_MIN_SIZE <= 64) {
3034 create_kmalloc_cache(&kmalloc_caches[1],
3035 "kmalloc-96", 96, GFP_KERNEL);
3037 create_kmalloc_cache(&kmalloc_caches[2],
3038 "kmalloc-192", 192, GFP_KERNEL);
3042 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3043 create_kmalloc_cache(&kmalloc_caches[i],
3044 "kmalloc", 1 << i, GFP_KERNEL);
3050 * Patch up the size_index table if we have strange large alignment
3051 * requirements for the kmalloc array. This is only the case for
3052 * MIPS it seems. The standard arches will not generate any code here.
3054 * Largest permitted alignment is 256 bytes due to the way we
3055 * handle the index determination for the smaller caches.
3057 * Make sure that nothing crazy happens if someone starts tinkering
3058 * around with ARCH_KMALLOC_MINALIGN
3060 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3061 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3063 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3064 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3066 if (KMALLOC_MIN_SIZE == 128) {
3068 * The 192 byte sized cache is not used if the alignment
3069 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3072 for (i = 128 + 8; i <= 192; i += 8)
3073 size_index[(i - 1) / 8] = 8;
3078 /* Provide the correct kmalloc names now that the caches are up */
3079 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3080 kmalloc_caches[i]. name =
3081 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3084 register_cpu_notifier(&slab_notifier);
3085 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3086 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3088 kmem_size = sizeof(struct kmem_cache);
3092 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3093 " CPUs=%d, Nodes=%d\n",
3094 caches, cache_line_size(),
3095 slub_min_order, slub_max_order, slub_min_objects,
3096 nr_cpu_ids, nr_node_ids);
3100 * Find a mergeable slab cache
3102 static int slab_unmergeable(struct kmem_cache *s)
3104 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3111 * We may have set a slab to be unmergeable during bootstrap.
3113 if (s->refcount < 0)
3119 static struct kmem_cache *find_mergeable(size_t size,
3120 size_t align, unsigned long flags, const char *name,
3121 void (*ctor)(void *))
3123 struct kmem_cache *s;
3125 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3131 size = ALIGN(size, sizeof(void *));
3132 align = calculate_alignment(flags, align, size);
3133 size = ALIGN(size, align);
3134 flags = kmem_cache_flags(size, flags, name, NULL);
3136 list_for_each_entry(s, &slab_caches, list) {
3137 if (slab_unmergeable(s))
3143 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3146 * Check if alignment is compatible.
3147 * Courtesy of Adrian Drzewiecki
3149 if ((s->size & ~(align - 1)) != s->size)
3152 if (s->size - size >= sizeof(void *))
3160 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3161 size_t align, unsigned long flags, void (*ctor)(void *))
3163 struct kmem_cache *s;
3165 down_write(&slub_lock);
3166 s = find_mergeable(size, align, flags, name, ctor);
3172 * Adjust the object sizes so that we clear
3173 * the complete object on kzalloc.
3175 s->objsize = max(s->objsize, (int)size);
3178 * And then we need to update the object size in the
3179 * per cpu structures
3181 for_each_online_cpu(cpu)
3182 get_cpu_slab(s, cpu)->objsize = s->objsize;
3184 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3185 up_write(&slub_lock);
3187 if (sysfs_slab_alias(s, name)) {
3188 down_write(&slub_lock);
3190 up_write(&slub_lock);
3196 s = kmalloc(kmem_size, GFP_KERNEL);
3198 if (kmem_cache_open(s, GFP_KERNEL, name,
3199 size, align, flags, ctor)) {
3200 list_add(&s->list, &slab_caches);
3201 up_write(&slub_lock);
3202 if (sysfs_slab_add(s)) {
3203 down_write(&slub_lock);
3205 up_write(&slub_lock);
3213 up_write(&slub_lock);
3216 if (flags & SLAB_PANIC)
3217 panic("Cannot create slabcache %s\n", name);
3222 EXPORT_SYMBOL(kmem_cache_create);
3226 * Use the cpu notifier to insure that the cpu slabs are flushed when
3229 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3230 unsigned long action, void *hcpu)
3232 long cpu = (long)hcpu;
3233 struct kmem_cache *s;
3234 unsigned long flags;
3237 case CPU_UP_PREPARE:
3238 case CPU_UP_PREPARE_FROZEN:
3239 init_alloc_cpu_cpu(cpu);
3240 down_read(&slub_lock);
3241 list_for_each_entry(s, &slab_caches, list)
3242 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3244 up_read(&slub_lock);
3247 case CPU_UP_CANCELED:
3248 case CPU_UP_CANCELED_FROZEN:
3250 case CPU_DEAD_FROZEN:
3251 down_read(&slub_lock);
3252 list_for_each_entry(s, &slab_caches, list) {
3253 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3255 local_irq_save(flags);
3256 __flush_cpu_slab(s, cpu);
3257 local_irq_restore(flags);
3258 free_kmem_cache_cpu(c, cpu);
3259 s->cpu_slab[cpu] = NULL;
3261 up_read(&slub_lock);
3269 static struct notifier_block __cpuinitdata slab_notifier = {
3270 .notifier_call = slab_cpuup_callback
3275 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3277 struct kmem_cache *s;
3280 if (unlikely(size > SLUB_MAX_SIZE))
3281 return kmalloc_large(size, gfpflags);
3283 s = get_slab(size, gfpflags);
3285 if (unlikely(ZERO_OR_NULL_PTR(s)))
3288 ret = slab_alloc(s, gfpflags, -1, caller);
3290 /* Honor the call site pointer we recieved. */
3291 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, caller, ret, size,
3297 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3298 int node, unsigned long caller)
3300 struct kmem_cache *s;
3303 if (unlikely(size > SLUB_MAX_SIZE))
3304 return kmalloc_large_node(size, gfpflags, node);
3306 s = get_slab(size, gfpflags);
3308 if (unlikely(ZERO_OR_NULL_PTR(s)))
3311 ret = slab_alloc(s, gfpflags, node, caller);
3313 /* Honor the call site pointer we recieved. */
3314 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC, caller, ret,
3315 size, s->size, gfpflags, node);
3320 #ifdef CONFIG_SLUB_DEBUG
3321 static unsigned long count_partial(struct kmem_cache_node *n,
3322 int (*get_count)(struct page *))
3324 unsigned long flags;
3325 unsigned long x = 0;
3328 spin_lock_irqsave(&n->list_lock, flags);
3329 list_for_each_entry(page, &n->partial, lru)
3330 x += get_count(page);
3331 spin_unlock_irqrestore(&n->list_lock, flags);
3335 static int count_inuse(struct page *page)
3340 static int count_total(struct page *page)
3342 return page->objects;
3345 static int count_free(struct page *page)
3347 return page->objects - page->inuse;
3350 static int validate_slab(struct kmem_cache *s, struct page *page,
3354 void *addr = page_address(page);
3356 if (!check_slab(s, page) ||
3357 !on_freelist(s, page, NULL))
3360 /* Now we know that a valid freelist exists */
3361 bitmap_zero(map, page->objects);
3363 for_each_free_object(p, s, page->freelist) {
3364 set_bit(slab_index(p, s, addr), map);
3365 if (!check_object(s, page, p, 0))
3369 for_each_object(p, s, addr, page->objects)
3370 if (!test_bit(slab_index(p, s, addr), map))
3371 if (!check_object(s, page, p, 1))
3376 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3379 if (slab_trylock(page)) {
3380 validate_slab(s, page, map);
3383 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3386 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3387 if (!PageSlubDebug(page))
3388 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3389 "on slab 0x%p\n", s->name, page);
3391 if (PageSlubDebug(page))
3392 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3393 "slab 0x%p\n", s->name, page);
3397 static int validate_slab_node(struct kmem_cache *s,
3398 struct kmem_cache_node *n, unsigned long *map)
3400 unsigned long count = 0;
3402 unsigned long flags;
3404 spin_lock_irqsave(&n->list_lock, flags);
3406 list_for_each_entry(page, &n->partial, lru) {
3407 validate_slab_slab(s, page, map);
3410 if (count != n->nr_partial)
3411 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3412 "counter=%ld\n", s->name, count, n->nr_partial);
3414 if (!(s->flags & SLAB_STORE_USER))
3417 list_for_each_entry(page, &n->full, lru) {
3418 validate_slab_slab(s, page, map);
3421 if (count != atomic_long_read(&n->nr_slabs))
3422 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3423 "counter=%ld\n", s->name, count,
3424 atomic_long_read(&n->nr_slabs));
3427 spin_unlock_irqrestore(&n->list_lock, flags);
3431 static long validate_slab_cache(struct kmem_cache *s)
3434 unsigned long count = 0;
3435 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3436 sizeof(unsigned long), GFP_KERNEL);
3442 for_each_node_state(node, N_NORMAL_MEMORY) {
3443 struct kmem_cache_node *n = get_node(s, node);
3445 count += validate_slab_node(s, n, map);
3451 #ifdef SLUB_RESILIENCY_TEST
3452 static void resiliency_test(void)
3456 printk(KERN_ERR "SLUB resiliency testing\n");
3457 printk(KERN_ERR "-----------------------\n");
3458 printk(KERN_ERR "A. Corruption after allocation\n");
3460 p = kzalloc(16, GFP_KERNEL);
3462 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3463 " 0x12->0x%p\n\n", p + 16);
3465 validate_slab_cache(kmalloc_caches + 4);
3467 /* Hmmm... The next two are dangerous */
3468 p = kzalloc(32, GFP_KERNEL);
3469 p[32 + sizeof(void *)] = 0x34;
3470 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3471 " 0x34 -> -0x%p\n", p);
3473 "If allocated object is overwritten then not detectable\n\n");
3475 validate_slab_cache(kmalloc_caches + 5);
3476 p = kzalloc(64, GFP_KERNEL);
3477 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3479 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3482 "If allocated object is overwritten then not detectable\n\n");
3483 validate_slab_cache(kmalloc_caches + 6);
3485 printk(KERN_ERR "\nB. Corruption after free\n");
3486 p = kzalloc(128, GFP_KERNEL);
3489 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3490 validate_slab_cache(kmalloc_caches + 7);
3492 p = kzalloc(256, GFP_KERNEL);
3495 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3497 validate_slab_cache(kmalloc_caches + 8);
3499 p = kzalloc(512, GFP_KERNEL);
3502 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3503 validate_slab_cache(kmalloc_caches + 9);
3506 static void resiliency_test(void) {};
3510 * Generate lists of code addresses where slabcache objects are allocated
3515 unsigned long count;
3522 DECLARE_BITMAP(cpus, NR_CPUS);
3528 unsigned long count;
3529 struct location *loc;
3532 static void free_loc_track(struct loc_track *t)
3535 free_pages((unsigned long)t->loc,
3536 get_order(sizeof(struct location) * t->max));
3539 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3544 order = get_order(sizeof(struct location) * max);
3546 l = (void *)__get_free_pages(flags, order);
3551 memcpy(l, t->loc, sizeof(struct location) * t->count);
3559 static int add_location(struct loc_track *t, struct kmem_cache *s,
3560 const struct track *track)
3562 long start, end, pos;
3564 unsigned long caddr;
3565 unsigned long age = jiffies - track->when;
3571 pos = start + (end - start + 1) / 2;
3574 * There is nothing at "end". If we end up there
3575 * we need to add something to before end.
3580 caddr = t->loc[pos].addr;
3581 if (track->addr == caddr) {
3587 if (age < l->min_time)
3589 if (age > l->max_time)
3592 if (track->pid < l->min_pid)
3593 l->min_pid = track->pid;
3594 if (track->pid > l->max_pid)
3595 l->max_pid = track->pid;
3597 cpumask_set_cpu(track->cpu,
3598 to_cpumask(l->cpus));
3600 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3604 if (track->addr < caddr)
3611 * Not found. Insert new tracking element.
3613 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3619 (t->count - pos) * sizeof(struct location));
3622 l->addr = track->addr;
3626 l->min_pid = track->pid;
3627 l->max_pid = track->pid;
3628 cpumask_clear(to_cpumask(l->cpus));
3629 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3630 nodes_clear(l->nodes);
3631 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3635 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3636 struct page *page, enum track_item alloc)
3638 void *addr = page_address(page);
3639 DECLARE_BITMAP(map, page->objects);
3642 bitmap_zero(map, page->objects);
3643 for_each_free_object(p, s, page->freelist)
3644 set_bit(slab_index(p, s, addr), map);
3646 for_each_object(p, s, addr, page->objects)
3647 if (!test_bit(slab_index(p, s, addr), map))
3648 add_location(t, s, get_track(s, p, alloc));
3651 static int list_locations(struct kmem_cache *s, char *buf,
3652 enum track_item alloc)
3656 struct loc_track t = { 0, 0, NULL };
3659 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3661 return sprintf(buf, "Out of memory\n");
3663 /* Push back cpu slabs */
3666 for_each_node_state(node, N_NORMAL_MEMORY) {
3667 struct kmem_cache_node *n = get_node(s, node);
3668 unsigned long flags;
3671 if (!atomic_long_read(&n->nr_slabs))
3674 spin_lock_irqsave(&n->list_lock, flags);
3675 list_for_each_entry(page, &n->partial, lru)
3676 process_slab(&t, s, page, alloc);
3677 list_for_each_entry(page, &n->full, lru)
3678 process_slab(&t, s, page, alloc);
3679 spin_unlock_irqrestore(&n->list_lock, flags);
3682 for (i = 0; i < t.count; i++) {
3683 struct location *l = &t.loc[i];
3685 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3687 len += sprintf(buf + len, "%7ld ", l->count);
3690 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3692 len += sprintf(buf + len, "<not-available>");
3694 if (l->sum_time != l->min_time) {
3695 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3697 (long)div_u64(l->sum_time, l->count),
3700 len += sprintf(buf + len, " age=%ld",
3703 if (l->min_pid != l->max_pid)
3704 len += sprintf(buf + len, " pid=%ld-%ld",
3705 l->min_pid, l->max_pid);
3707 len += sprintf(buf + len, " pid=%ld",
3710 if (num_online_cpus() > 1 &&
3711 !cpumask_empty(to_cpumask(l->cpus)) &&
3712 len < PAGE_SIZE - 60) {
3713 len += sprintf(buf + len, " cpus=");
3714 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3715 to_cpumask(l->cpus));
3718 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3719 len < PAGE_SIZE - 60) {
3720 len += sprintf(buf + len, " nodes=");
3721 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3725 len += sprintf(buf + len, "\n");
3730 len += sprintf(buf, "No data\n");
3734 enum slab_stat_type {
3735 SL_ALL, /* All slabs */
3736 SL_PARTIAL, /* Only partially allocated slabs */
3737 SL_CPU, /* Only slabs used for cpu caches */
3738 SL_OBJECTS, /* Determine allocated objects not slabs */
3739 SL_TOTAL /* Determine object capacity not slabs */
3742 #define SO_ALL (1 << SL_ALL)
3743 #define SO_PARTIAL (1 << SL_PARTIAL)
3744 #define SO_CPU (1 << SL_CPU)
3745 #define SO_OBJECTS (1 << SL_OBJECTS)
3746 #define SO_TOTAL (1 << SL_TOTAL)
3748 static ssize_t show_slab_objects(struct kmem_cache *s,
3749 char *buf, unsigned long flags)
3751 unsigned long total = 0;
3754 unsigned long *nodes;
3755 unsigned long *per_cpu;
3757 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3760 per_cpu = nodes + nr_node_ids;
3762 if (flags & SO_CPU) {
3765 for_each_possible_cpu(cpu) {
3766 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3768 if (!c || c->node < 0)
3772 if (flags & SO_TOTAL)
3773 x = c->page->objects;
3774 else if (flags & SO_OBJECTS)
3780 nodes[c->node] += x;
3786 if (flags & SO_ALL) {
3787 for_each_node_state(node, N_NORMAL_MEMORY) {
3788 struct kmem_cache_node *n = get_node(s, node);
3790 if (flags & SO_TOTAL)
3791 x = atomic_long_read(&n->total_objects);
3792 else if (flags & SO_OBJECTS)
3793 x = atomic_long_read(&n->total_objects) -
3794 count_partial(n, count_free);
3797 x = atomic_long_read(&n->nr_slabs);
3802 } else if (flags & SO_PARTIAL) {
3803 for_each_node_state(node, N_NORMAL_MEMORY) {
3804 struct kmem_cache_node *n = get_node(s, node);
3806 if (flags & SO_TOTAL)
3807 x = count_partial(n, count_total);
3808 else if (flags & SO_OBJECTS)
3809 x = count_partial(n, count_inuse);
3816 x = sprintf(buf, "%lu", total);
3818 for_each_node_state(node, N_NORMAL_MEMORY)
3820 x += sprintf(buf + x, " N%d=%lu",
3824 return x + sprintf(buf + x, "\n");
3827 static int any_slab_objects(struct kmem_cache *s)
3831 for_each_online_node(node) {
3832 struct kmem_cache_node *n = get_node(s, node);
3837 if (atomic_long_read(&n->total_objects))
3843 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3844 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3846 struct slab_attribute {
3847 struct attribute attr;
3848 ssize_t (*show)(struct kmem_cache *s, char *buf);
3849 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3852 #define SLAB_ATTR_RO(_name) \
3853 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3855 #define SLAB_ATTR(_name) \
3856 static struct slab_attribute _name##_attr = \
3857 __ATTR(_name, 0644, _name##_show, _name##_store)
3859 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3861 return sprintf(buf, "%d\n", s->size);
3863 SLAB_ATTR_RO(slab_size);
3865 static ssize_t align_show(struct kmem_cache *s, char *buf)
3867 return sprintf(buf, "%d\n", s->align);
3869 SLAB_ATTR_RO(align);
3871 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3873 return sprintf(buf, "%d\n", s->objsize);
3875 SLAB_ATTR_RO(object_size);
3877 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3879 return sprintf(buf, "%d\n", oo_objects(s->oo));
3881 SLAB_ATTR_RO(objs_per_slab);
3883 static ssize_t order_store(struct kmem_cache *s,
3884 const char *buf, size_t length)
3886 unsigned long order;
3889 err = strict_strtoul(buf, 10, &order);
3893 if (order > slub_max_order || order < slub_min_order)
3896 calculate_sizes(s, order);
3900 static ssize_t order_show(struct kmem_cache *s, char *buf)
3902 return sprintf(buf, "%d\n", oo_order(s->oo));
3906 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3909 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3911 return n + sprintf(buf + n, "\n");
3917 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3919 return sprintf(buf, "%d\n", s->refcount - 1);
3921 SLAB_ATTR_RO(aliases);
3923 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3925 return show_slab_objects(s, buf, SO_ALL);
3927 SLAB_ATTR_RO(slabs);
3929 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3931 return show_slab_objects(s, buf, SO_PARTIAL);
3933 SLAB_ATTR_RO(partial);
3935 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3937 return show_slab_objects(s, buf, SO_CPU);
3939 SLAB_ATTR_RO(cpu_slabs);
3941 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3943 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3945 SLAB_ATTR_RO(objects);
3947 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3949 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3951 SLAB_ATTR_RO(objects_partial);
3953 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3955 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3957 SLAB_ATTR_RO(total_objects);
3959 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3961 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3964 static ssize_t sanity_checks_store(struct kmem_cache *s,
3965 const char *buf, size_t length)
3967 s->flags &= ~SLAB_DEBUG_FREE;
3969 s->flags |= SLAB_DEBUG_FREE;
3972 SLAB_ATTR(sanity_checks);
3974 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3976 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3979 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3982 s->flags &= ~SLAB_TRACE;
3984 s->flags |= SLAB_TRACE;
3989 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3991 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3994 static ssize_t reclaim_account_store(struct kmem_cache *s,
3995 const char *buf, size_t length)
3997 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3999 s->flags |= SLAB_RECLAIM_ACCOUNT;
4002 SLAB_ATTR(reclaim_account);
4004 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4006 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4008 SLAB_ATTR_RO(hwcache_align);
4010 #ifdef CONFIG_ZONE_DMA
4011 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4013 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4015 SLAB_ATTR_RO(cache_dma);
4018 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4020 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4022 SLAB_ATTR_RO(destroy_by_rcu);
4024 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4026 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4029 static ssize_t red_zone_store(struct kmem_cache *s,
4030 const char *buf, size_t length)
4032 if (any_slab_objects(s))
4035 s->flags &= ~SLAB_RED_ZONE;
4037 s->flags |= SLAB_RED_ZONE;
4038 calculate_sizes(s, -1);
4041 SLAB_ATTR(red_zone);
4043 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4045 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4048 static ssize_t poison_store(struct kmem_cache *s,
4049 const char *buf, size_t length)
4051 if (any_slab_objects(s))
4054 s->flags &= ~SLAB_POISON;
4056 s->flags |= SLAB_POISON;
4057 calculate_sizes(s, -1);
4062 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4064 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4067 static ssize_t store_user_store(struct kmem_cache *s,
4068 const char *buf, size_t length)
4070 if (any_slab_objects(s))
4073 s->flags &= ~SLAB_STORE_USER;
4075 s->flags |= SLAB_STORE_USER;
4076 calculate_sizes(s, -1);
4079 SLAB_ATTR(store_user);
4081 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4086 static ssize_t validate_store(struct kmem_cache *s,
4087 const char *buf, size_t length)
4091 if (buf[0] == '1') {
4092 ret = validate_slab_cache(s);
4098 SLAB_ATTR(validate);
4100 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4105 static ssize_t shrink_store(struct kmem_cache *s,
4106 const char *buf, size_t length)
4108 if (buf[0] == '1') {
4109 int rc = kmem_cache_shrink(s);
4119 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4121 if (!(s->flags & SLAB_STORE_USER))
4123 return list_locations(s, buf, TRACK_ALLOC);
4125 SLAB_ATTR_RO(alloc_calls);
4127 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4129 if (!(s->flags & SLAB_STORE_USER))
4131 return list_locations(s, buf, TRACK_FREE);
4133 SLAB_ATTR_RO(free_calls);
4136 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4138 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4141 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4142 const char *buf, size_t length)
4144 unsigned long ratio;
4147 err = strict_strtoul(buf, 10, &ratio);
4152 s->remote_node_defrag_ratio = ratio * 10;
4156 SLAB_ATTR(remote_node_defrag_ratio);
4159 #ifdef CONFIG_SLUB_STATS
4160 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4162 unsigned long sum = 0;
4165 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4170 for_each_online_cpu(cpu) {
4171 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4177 len = sprintf(buf, "%lu", sum);
4180 for_each_online_cpu(cpu) {
4181 if (data[cpu] && len < PAGE_SIZE - 20)
4182 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4186 return len + sprintf(buf + len, "\n");
4189 #define STAT_ATTR(si, text) \
4190 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4192 return show_stat(s, buf, si); \
4194 SLAB_ATTR_RO(text); \
4196 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4197 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4198 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4199 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4200 STAT_ATTR(FREE_FROZEN, free_frozen);
4201 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4202 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4203 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4204 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4205 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4206 STAT_ATTR(FREE_SLAB, free_slab);
4207 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4208 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4209 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4210 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4211 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4212 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4213 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4216 static struct attribute *slab_attrs[] = {
4217 &slab_size_attr.attr,
4218 &object_size_attr.attr,
4219 &objs_per_slab_attr.attr,
4222 &objects_partial_attr.attr,
4223 &total_objects_attr.attr,
4226 &cpu_slabs_attr.attr,
4230 &sanity_checks_attr.attr,
4232 &hwcache_align_attr.attr,
4233 &reclaim_account_attr.attr,
4234 &destroy_by_rcu_attr.attr,
4235 &red_zone_attr.attr,
4237 &store_user_attr.attr,
4238 &validate_attr.attr,
4240 &alloc_calls_attr.attr,
4241 &free_calls_attr.attr,
4242 #ifdef CONFIG_ZONE_DMA
4243 &cache_dma_attr.attr,
4246 &remote_node_defrag_ratio_attr.attr,
4248 #ifdef CONFIG_SLUB_STATS
4249 &alloc_fastpath_attr.attr,
4250 &alloc_slowpath_attr.attr,
4251 &free_fastpath_attr.attr,
4252 &free_slowpath_attr.attr,
4253 &free_frozen_attr.attr,
4254 &free_add_partial_attr.attr,
4255 &free_remove_partial_attr.attr,
4256 &alloc_from_partial_attr.attr,
4257 &alloc_slab_attr.attr,
4258 &alloc_refill_attr.attr,
4259 &free_slab_attr.attr,
4260 &cpuslab_flush_attr.attr,
4261 &deactivate_full_attr.attr,
4262 &deactivate_empty_attr.attr,
4263 &deactivate_to_head_attr.attr,
4264 &deactivate_to_tail_attr.attr,
4265 &deactivate_remote_frees_attr.attr,
4266 &order_fallback_attr.attr,
4271 static struct attribute_group slab_attr_group = {
4272 .attrs = slab_attrs,
4275 static ssize_t slab_attr_show(struct kobject *kobj,
4276 struct attribute *attr,
4279 struct slab_attribute *attribute;
4280 struct kmem_cache *s;
4283 attribute = to_slab_attr(attr);
4286 if (!attribute->show)
4289 err = attribute->show(s, buf);
4294 static ssize_t slab_attr_store(struct kobject *kobj,
4295 struct attribute *attr,
4296 const char *buf, size_t len)
4298 struct slab_attribute *attribute;
4299 struct kmem_cache *s;
4302 attribute = to_slab_attr(attr);
4305 if (!attribute->store)
4308 err = attribute->store(s, buf, len);
4313 static void kmem_cache_release(struct kobject *kobj)
4315 struct kmem_cache *s = to_slab(kobj);
4320 static struct sysfs_ops slab_sysfs_ops = {
4321 .show = slab_attr_show,
4322 .store = slab_attr_store,
4325 static struct kobj_type slab_ktype = {
4326 .sysfs_ops = &slab_sysfs_ops,
4327 .release = kmem_cache_release
4330 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4332 struct kobj_type *ktype = get_ktype(kobj);
4334 if (ktype == &slab_ktype)
4339 static struct kset_uevent_ops slab_uevent_ops = {
4340 .filter = uevent_filter,
4343 static struct kset *slab_kset;
4345 #define ID_STR_LENGTH 64
4347 /* Create a unique string id for a slab cache:
4349 * Format :[flags-]size
4351 static char *create_unique_id(struct kmem_cache *s)
4353 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4360 * First flags affecting slabcache operations. We will only
4361 * get here for aliasable slabs so we do not need to support
4362 * too many flags. The flags here must cover all flags that
4363 * are matched during merging to guarantee that the id is
4366 if (s->flags & SLAB_CACHE_DMA)
4368 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4370 if (s->flags & SLAB_DEBUG_FREE)
4374 p += sprintf(p, "%07d", s->size);
4375 BUG_ON(p > name + ID_STR_LENGTH - 1);
4379 static int sysfs_slab_add(struct kmem_cache *s)
4385 if (slab_state < SYSFS)
4386 /* Defer until later */
4389 unmergeable = slab_unmergeable(s);
4392 * Slabcache can never be merged so we can use the name proper.
4393 * This is typically the case for debug situations. In that
4394 * case we can catch duplicate names easily.
4396 sysfs_remove_link(&slab_kset->kobj, s->name);
4400 * Create a unique name for the slab as a target
4403 name = create_unique_id(s);
4406 s->kobj.kset = slab_kset;
4407 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4409 kobject_put(&s->kobj);
4413 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4416 kobject_uevent(&s->kobj, KOBJ_ADD);
4418 /* Setup first alias */
4419 sysfs_slab_alias(s, s->name);
4425 static void sysfs_slab_remove(struct kmem_cache *s)
4427 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4428 kobject_del(&s->kobj);
4429 kobject_put(&s->kobj);
4433 * Need to buffer aliases during bootup until sysfs becomes
4434 * available lest we lose that information.
4436 struct saved_alias {
4437 struct kmem_cache *s;
4439 struct saved_alias *next;
4442 static struct saved_alias *alias_list;
4444 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4446 struct saved_alias *al;
4448 if (slab_state == SYSFS) {
4450 * If we have a leftover link then remove it.
4452 sysfs_remove_link(&slab_kset->kobj, name);
4453 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4456 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4462 al->next = alias_list;
4467 static int __init slab_sysfs_init(void)
4469 struct kmem_cache *s;
4472 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4474 printk(KERN_ERR "Cannot register slab subsystem.\n");
4480 list_for_each_entry(s, &slab_caches, list) {
4481 err = sysfs_slab_add(s);
4483 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4484 " to sysfs\n", s->name);
4487 while (alias_list) {
4488 struct saved_alias *al = alias_list;
4490 alias_list = alias_list->next;
4491 err = sysfs_slab_alias(al->s, al->name);
4493 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4494 " %s to sysfs\n", s->name);
4502 __initcall(slab_sysfs_init);
4506 * The /proc/slabinfo ABI
4508 #ifdef CONFIG_SLABINFO
4509 static void print_slabinfo_header(struct seq_file *m)
4511 seq_puts(m, "slabinfo - version: 2.1\n");
4512 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4513 "<objperslab> <pagesperslab>");
4514 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4515 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4519 static void *s_start(struct seq_file *m, loff_t *pos)
4523 down_read(&slub_lock);
4525 print_slabinfo_header(m);
4527 return seq_list_start(&slab_caches, *pos);
4530 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4532 return seq_list_next(p, &slab_caches, pos);
4535 static void s_stop(struct seq_file *m, void *p)
4537 up_read(&slub_lock);
4540 static int s_show(struct seq_file *m, void *p)
4542 unsigned long nr_partials = 0;
4543 unsigned long nr_slabs = 0;
4544 unsigned long nr_inuse = 0;
4545 unsigned long nr_objs = 0;
4546 unsigned long nr_free = 0;
4547 struct kmem_cache *s;
4550 s = list_entry(p, struct kmem_cache, list);
4552 for_each_online_node(node) {
4553 struct kmem_cache_node *n = get_node(s, node);
4558 nr_partials += n->nr_partial;
4559 nr_slabs += atomic_long_read(&n->nr_slabs);
4560 nr_objs += atomic_long_read(&n->total_objects);
4561 nr_free += count_partial(n, count_free);
4564 nr_inuse = nr_objs - nr_free;
4566 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4567 nr_objs, s->size, oo_objects(s->oo),
4568 (1 << oo_order(s->oo)));
4569 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4570 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4576 static const struct seq_operations slabinfo_op = {
4583 static int slabinfo_open(struct inode *inode, struct file *file)
4585 return seq_open(file, &slabinfo_op);
4588 static const struct file_operations proc_slabinfo_operations = {
4589 .open = slabinfo_open,
4591 .llseek = seq_lseek,
4592 .release = seq_release,
4595 static int __init slab_proc_init(void)
4597 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4600 module_init(slab_proc_init);
4601 #endif /* CONFIG_SLABINFO */