2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemtrace.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/kmemleak.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
37 * The slab_lock protects operations on the object of a particular
38 * slab and its metadata in the page struct. If the slab lock
39 * has been taken then no allocations nor frees can be performed
40 * on the objects in the slab nor can the slab be added or removed
41 * from the partial or full lists since this would mean modifying
42 * the page_struct of the slab.
44 * The list_lock protects the partial and full list on each node and
45 * the partial slab counter. If taken then no new slabs may be added or
46 * removed from the lists nor make the number of partial slabs be modified.
47 * (Note that the total number of slabs is an atomic value that may be
48 * modified without taking the list lock).
50 * The list_lock is a centralized lock and thus we avoid taking it as
51 * much as possible. As long as SLUB does not have to handle partial
52 * slabs, operations can continue without any centralized lock. F.e.
53 * allocating a long series of objects that fill up slabs does not require
56 * The lock order is sometimes inverted when we are trying to get a slab
57 * off a list. We take the list_lock and then look for a page on the list
58 * to use. While we do that objects in the slabs may be freed. We can
59 * only operate on the slab if we have also taken the slab_lock. So we use
60 * a slab_trylock() on the slab. If trylock was successful then no frees
61 * can occur anymore and we can use the slab for allocations etc. If the
62 * slab_trylock() does not succeed then frees are in progress in the slab and
63 * we must stay away from it for a while since we may cause a bouncing
64 * cacheline if we try to acquire the lock. So go onto the next slab.
65 * If all pages are busy then we may allocate a new slab instead of reusing
66 * a partial slab. A new slab has noone operating on it and thus there is
67 * no danger of cacheline contention.
69 * Interrupts are disabled during allocation and deallocation in order to
70 * make the slab allocator safe to use in the context of an irq. In addition
71 * interrupts are disabled to ensure that the processor does not change
72 * while handling per_cpu slabs, due to kernel preemption.
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
77 * Slabs with free elements are kept on a partial list and during regular
78 * operations no list for full slabs is used. If an object in a full slab is
79 * freed then the slab will show up again on the partial lists.
80 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
83 * Slabs are freed when they become empty. Teardown and setup is
84 * minimal so we rely on the page allocators per cpu caches for
85 * fast frees and allocs.
87 * Overloading of page flags that are otherwise used for LRU management.
89 * PageActive The slab is frozen and exempt from list processing.
90 * This means that the slab is dedicated to a purpose
91 * such as satisfying allocations for a specific
92 * processor. Objects may be freed in the slab while
93 * it is frozen but slab_free will then skip the usual
94 * list operations. It is up to the processor holding
95 * the slab to integrate the slab into the slab lists
96 * when the slab is no longer needed.
98 * One use of this flag is to mark slabs that are
99 * used for allocations. Then such a slab becomes a cpu
100 * slab. The cpu slab may be equipped with an additional
101 * freelist that allows lockless access to
102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
107 * the fast path and disables lockless freelists.
110 #ifdef CONFIG_SLUB_DEBUG
117 * Issues still to be resolved:
119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
121 * - Variable sizing of the per node arrays
124 /* Enable to test recovery from slab corruption on boot */
125 #undef SLUB_RESILIENCY_TEST
128 * Mininum number of partial slabs. These will be left on the partial
129 * lists even if they are empty. kmem_cache_shrink may reclaim them.
131 #define MIN_PARTIAL 5
134 * Maximum number of desirable partial slabs.
135 * The existence of more partial slabs makes kmem_cache_shrink
136 * sort the partial list by the number of objects in the.
138 #define MAX_PARTIAL 10
140 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
141 SLAB_POISON | SLAB_STORE_USER)
144 * Set of flags that will prevent slab merging
146 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
147 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
149 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
152 #ifndef ARCH_KMALLOC_MINALIGN
153 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
156 #ifndef ARCH_SLAB_MINALIGN
157 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
161 #define OO_MASK ((1 << OO_SHIFT) - 1)
162 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
164 /* Internal SLUB flags */
165 #define __OBJECT_POISON 0x80000000 /* Poison object */
166 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
168 static int kmem_size = sizeof(struct kmem_cache);
171 static struct notifier_block slab_notifier;
175 DOWN, /* No slab functionality available */
176 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
177 UP, /* Everything works but does not show up in sysfs */
181 /* A list of all slab caches on the system */
182 static DECLARE_RWSEM(slub_lock);
183 static LIST_HEAD(slab_caches);
186 * Tracking user of a slab.
189 unsigned long addr; /* Called from address */
190 int cpu; /* Was running on cpu */
191 int pid; /* Pid context */
192 unsigned long when; /* When did the operation occur */
195 enum track_item { TRACK_ALLOC, TRACK_FREE };
197 #ifdef CONFIG_SLUB_DEBUG
198 static int sysfs_slab_add(struct kmem_cache *);
199 static int sysfs_slab_alias(struct kmem_cache *, const char *);
200 static void sysfs_slab_remove(struct kmem_cache *);
203 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
204 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
206 static inline void sysfs_slab_remove(struct kmem_cache *s)
213 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
215 #ifdef CONFIG_SLUB_STATS
220 /********************************************************************
221 * Core slab cache functions
222 *******************************************************************/
224 int slab_is_available(void)
226 return slab_state >= UP;
229 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
232 return s->node[node];
234 return &s->local_node;
238 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
241 return s->cpu_slab[cpu];
247 /* Verify that a pointer has an address that is valid within a slab page */
248 static inline int check_valid_pointer(struct kmem_cache *s,
249 struct page *page, const void *object)
256 base = page_address(page);
257 if (object < base || object >= base + page->objects * s->size ||
258 (object - base) % s->size) {
266 * Slow version of get and set free pointer.
268 * This version requires touching the cache lines of kmem_cache which
269 * we avoid to do in the fast alloc free paths. There we obtain the offset
270 * from the page struct.
272 static inline void *get_freepointer(struct kmem_cache *s, void *object)
274 return *(void **)(object + s->offset);
277 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
279 *(void **)(object + s->offset) = fp;
282 /* Loop over all objects in a slab */
283 #define for_each_object(__p, __s, __addr, __objects) \
284 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
288 #define for_each_free_object(__p, __s, __free) \
289 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
294 return (p - addr) / s->size;
297 static inline struct kmem_cache_order_objects oo_make(int order,
300 struct kmem_cache_order_objects x = {
301 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
307 static inline int oo_order(struct kmem_cache_order_objects x)
309 return x.x >> OO_SHIFT;
312 static inline int oo_objects(struct kmem_cache_order_objects x)
314 return x.x & OO_MASK;
317 #ifdef CONFIG_SLUB_DEBUG
321 #ifdef CONFIG_SLUB_DEBUG_ON
322 static int slub_debug = DEBUG_DEFAULT_FLAGS;
324 static int slub_debug;
327 static char *slub_debug_slabs;
332 static void print_section(char *text, u8 *addr, unsigned int length)
340 for (i = 0; i < length; i++) {
342 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
345 printk(KERN_CONT " %02x", addr[i]);
347 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
349 printk(KERN_CONT " %s\n", ascii);
356 printk(KERN_CONT " ");
360 printk(KERN_CONT " %s\n", ascii);
364 static struct track *get_track(struct kmem_cache *s, void *object,
365 enum track_item alloc)
370 p = object + s->offset + sizeof(void *);
372 p = object + s->inuse;
377 static void set_track(struct kmem_cache *s, void *object,
378 enum track_item alloc, unsigned long addr)
380 struct track *p = get_track(s, object, alloc);
384 p->cpu = smp_processor_id();
385 p->pid = current->pid;
388 memset(p, 0, sizeof(struct track));
391 static void init_tracking(struct kmem_cache *s, void *object)
393 if (!(s->flags & SLAB_STORE_USER))
396 set_track(s, object, TRACK_FREE, 0UL);
397 set_track(s, object, TRACK_ALLOC, 0UL);
400 static void print_track(const char *s, struct track *t)
405 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
406 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
409 static void print_tracking(struct kmem_cache *s, void *object)
411 if (!(s->flags & SLAB_STORE_USER))
414 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
415 print_track("Freed", get_track(s, object, TRACK_FREE));
418 static void print_page_info(struct page *page)
420 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
421 page, page->objects, page->inuse, page->freelist, page->flags);
425 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
431 vsnprintf(buf, sizeof(buf), fmt, args);
433 printk(KERN_ERR "========================================"
434 "=====================================\n");
435 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
436 printk(KERN_ERR "----------------------------------------"
437 "-------------------------------------\n\n");
440 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
446 vsnprintf(buf, sizeof(buf), fmt, args);
448 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
451 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
453 unsigned int off; /* Offset of last byte */
454 u8 *addr = page_address(page);
456 print_tracking(s, p);
458 print_page_info(page);
460 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
461 p, p - addr, get_freepointer(s, p));
464 print_section("Bytes b4", p - 16, 16);
466 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
468 if (s->flags & SLAB_RED_ZONE)
469 print_section("Redzone", p + s->objsize,
470 s->inuse - s->objsize);
473 off = s->offset + sizeof(void *);
477 if (s->flags & SLAB_STORE_USER)
478 off += 2 * sizeof(struct track);
481 /* Beginning of the filler is the free pointer */
482 print_section("Padding", p + off, s->size - off);
487 static void object_err(struct kmem_cache *s, struct page *page,
488 u8 *object, char *reason)
490 slab_bug(s, "%s", reason);
491 print_trailer(s, page, object);
494 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
500 vsnprintf(buf, sizeof(buf), fmt, args);
502 slab_bug(s, "%s", buf);
503 print_page_info(page);
507 static void init_object(struct kmem_cache *s, void *object, int active)
511 if (s->flags & __OBJECT_POISON) {
512 memset(p, POISON_FREE, s->objsize - 1);
513 p[s->objsize - 1] = POISON_END;
516 if (s->flags & SLAB_RED_ZONE)
517 memset(p + s->objsize,
518 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
519 s->inuse - s->objsize);
522 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
525 if (*start != (u8)value)
533 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
534 void *from, void *to)
536 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
537 memset(from, data, to - from);
540 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
541 u8 *object, char *what,
542 u8 *start, unsigned int value, unsigned int bytes)
547 fault = check_bytes(start, value, bytes);
552 while (end > fault && end[-1] == value)
555 slab_bug(s, "%s overwritten", what);
556 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
557 fault, end - 1, fault[0], value);
558 print_trailer(s, page, object);
560 restore_bytes(s, what, value, fault, end);
568 * Bytes of the object to be managed.
569 * If the freepointer may overlay the object then the free
570 * pointer is the first word of the object.
572 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
575 * object + s->objsize
576 * Padding to reach word boundary. This is also used for Redzoning.
577 * Padding is extended by another word if Redzoning is enabled and
580 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
581 * 0xcc (RED_ACTIVE) for objects in use.
584 * Meta data starts here.
586 * A. Free pointer (if we cannot overwrite object on free)
587 * B. Tracking data for SLAB_STORE_USER
588 * C. Padding to reach required alignment boundary or at mininum
589 * one word if debugging is on to be able to detect writes
590 * before the word boundary.
592 * Padding is done using 0x5a (POISON_INUSE)
595 * Nothing is used beyond s->size.
597 * If slabcaches are merged then the objsize and inuse boundaries are mostly
598 * ignored. And therefore no slab options that rely on these boundaries
599 * may be used with merged slabcaches.
602 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
604 unsigned long off = s->inuse; /* The end of info */
607 /* Freepointer is placed after the object. */
608 off += sizeof(void *);
610 if (s->flags & SLAB_STORE_USER)
611 /* We also have user information there */
612 off += 2 * sizeof(struct track);
617 return check_bytes_and_report(s, page, p, "Object padding",
618 p + off, POISON_INUSE, s->size - off);
621 /* Check the pad bytes at the end of a slab page */
622 static int slab_pad_check(struct kmem_cache *s, struct page *page)
630 if (!(s->flags & SLAB_POISON))
633 start = page_address(page);
634 length = (PAGE_SIZE << compound_order(page));
635 end = start + length;
636 remainder = length % s->size;
640 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
643 while (end > fault && end[-1] == POISON_INUSE)
646 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
647 print_section("Padding", end - remainder, remainder);
649 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
653 static int check_object(struct kmem_cache *s, struct page *page,
654 void *object, int active)
657 u8 *endobject = object + s->objsize;
659 if (s->flags & SLAB_RED_ZONE) {
661 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
663 if (!check_bytes_and_report(s, page, object, "Redzone",
664 endobject, red, s->inuse - s->objsize))
667 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
668 check_bytes_and_report(s, page, p, "Alignment padding",
669 endobject, POISON_INUSE, s->inuse - s->objsize);
673 if (s->flags & SLAB_POISON) {
674 if (!active && (s->flags & __OBJECT_POISON) &&
675 (!check_bytes_and_report(s, page, p, "Poison", p,
676 POISON_FREE, s->objsize - 1) ||
677 !check_bytes_and_report(s, page, p, "Poison",
678 p + s->objsize - 1, POISON_END, 1)))
681 * check_pad_bytes cleans up on its own.
683 check_pad_bytes(s, page, p);
686 if (!s->offset && active)
688 * Object and freepointer overlap. Cannot check
689 * freepointer while object is allocated.
693 /* Check free pointer validity */
694 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
695 object_err(s, page, p, "Freepointer corrupt");
697 * No choice but to zap it and thus lose the remainder
698 * of the free objects in this slab. May cause
699 * another error because the object count is now wrong.
701 set_freepointer(s, p, NULL);
707 static int check_slab(struct kmem_cache *s, struct page *page)
711 VM_BUG_ON(!irqs_disabled());
713 if (!PageSlab(page)) {
714 slab_err(s, page, "Not a valid slab page");
718 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
719 if (page->objects > maxobj) {
720 slab_err(s, page, "objects %u > max %u",
721 s->name, page->objects, maxobj);
724 if (page->inuse > page->objects) {
725 slab_err(s, page, "inuse %u > max %u",
726 s->name, page->inuse, page->objects);
729 /* Slab_pad_check fixes things up after itself */
730 slab_pad_check(s, page);
735 * Determine if a certain object on a page is on the freelist. Must hold the
736 * slab lock to guarantee that the chains are in a consistent state.
738 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
741 void *fp = page->freelist;
743 unsigned long max_objects;
745 while (fp && nr <= page->objects) {
748 if (!check_valid_pointer(s, page, fp)) {
750 object_err(s, page, object,
751 "Freechain corrupt");
752 set_freepointer(s, object, NULL);
755 slab_err(s, page, "Freepointer corrupt");
756 page->freelist = NULL;
757 page->inuse = page->objects;
758 slab_fix(s, "Freelist cleared");
764 fp = get_freepointer(s, object);
768 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
769 if (max_objects > MAX_OBJS_PER_PAGE)
770 max_objects = MAX_OBJS_PER_PAGE;
772 if (page->objects != max_objects) {
773 slab_err(s, page, "Wrong number of objects. Found %d but "
774 "should be %d", page->objects, max_objects);
775 page->objects = max_objects;
776 slab_fix(s, "Number of objects adjusted.");
778 if (page->inuse != page->objects - nr) {
779 slab_err(s, page, "Wrong object count. Counter is %d but "
780 "counted were %d", page->inuse, page->objects - nr);
781 page->inuse = page->objects - nr;
782 slab_fix(s, "Object count adjusted.");
784 return search == NULL;
787 static void trace(struct kmem_cache *s, struct page *page, void *object,
790 if (s->flags & SLAB_TRACE) {
791 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
793 alloc ? "alloc" : "free",
798 print_section("Object", (void *)object, s->objsize);
805 * Tracking of fully allocated slabs for debugging purposes.
807 static void add_full(struct kmem_cache_node *n, struct page *page)
809 spin_lock(&n->list_lock);
810 list_add(&page->lru, &n->full);
811 spin_unlock(&n->list_lock);
814 static void remove_full(struct kmem_cache *s, struct page *page)
816 struct kmem_cache_node *n;
818 if (!(s->flags & SLAB_STORE_USER))
821 n = get_node(s, page_to_nid(page));
823 spin_lock(&n->list_lock);
824 list_del(&page->lru);
825 spin_unlock(&n->list_lock);
828 /* Tracking of the number of slabs for debugging purposes */
829 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
831 struct kmem_cache_node *n = get_node(s, node);
833 return atomic_long_read(&n->nr_slabs);
836 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
838 struct kmem_cache_node *n = get_node(s, node);
841 * May be called early in order to allocate a slab for the
842 * kmem_cache_node structure. Solve the chicken-egg
843 * dilemma by deferring the increment of the count during
844 * bootstrap (see early_kmem_cache_node_alloc).
846 if (!NUMA_BUILD || n) {
847 atomic_long_inc(&n->nr_slabs);
848 atomic_long_add(objects, &n->total_objects);
851 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
853 struct kmem_cache_node *n = get_node(s, node);
855 atomic_long_dec(&n->nr_slabs);
856 atomic_long_sub(objects, &n->total_objects);
859 /* Object debug checks for alloc/free paths */
860 static void setup_object_debug(struct kmem_cache *s, struct page *page,
863 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
866 init_object(s, object, 0);
867 init_tracking(s, object);
870 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
871 void *object, unsigned long addr)
873 if (!check_slab(s, page))
876 if (!on_freelist(s, page, object)) {
877 object_err(s, page, object, "Object already allocated");
881 if (!check_valid_pointer(s, page, object)) {
882 object_err(s, page, object, "Freelist Pointer check fails");
886 if (!check_object(s, page, object, 0))
889 /* Success perform special debug activities for allocs */
890 if (s->flags & SLAB_STORE_USER)
891 set_track(s, object, TRACK_ALLOC, addr);
892 trace(s, page, object, 1);
893 init_object(s, object, 1);
897 if (PageSlab(page)) {
899 * If this is a slab page then lets do the best we can
900 * to avoid issues in the future. Marking all objects
901 * as used avoids touching the remaining objects.
903 slab_fix(s, "Marking all objects used");
904 page->inuse = page->objects;
905 page->freelist = NULL;
910 static int free_debug_processing(struct kmem_cache *s, struct page *page,
911 void *object, unsigned long addr)
913 if (!check_slab(s, page))
916 if (!check_valid_pointer(s, page, object)) {
917 slab_err(s, page, "Invalid object pointer 0x%p", object);
921 if (on_freelist(s, page, object)) {
922 object_err(s, page, object, "Object already free");
926 if (!check_object(s, page, object, 1))
929 if (unlikely(s != page->slab)) {
930 if (!PageSlab(page)) {
931 slab_err(s, page, "Attempt to free object(0x%p) "
932 "outside of slab", object);
933 } else if (!page->slab) {
935 "SLUB <none>: no slab for object 0x%p.\n",
939 object_err(s, page, object,
940 "page slab pointer corrupt.");
944 /* Special debug activities for freeing objects */
945 if (!PageSlubFrozen(page) && !page->freelist)
946 remove_full(s, page);
947 if (s->flags & SLAB_STORE_USER)
948 set_track(s, object, TRACK_FREE, addr);
949 trace(s, page, object, 0);
950 init_object(s, object, 0);
954 slab_fix(s, "Object at 0x%p not freed", object);
958 static int __init setup_slub_debug(char *str)
960 slub_debug = DEBUG_DEFAULT_FLAGS;
961 if (*str++ != '=' || !*str)
963 * No options specified. Switch on full debugging.
969 * No options but restriction on slabs. This means full
970 * debugging for slabs matching a pattern.
977 * Switch off all debugging measures.
982 * Determine which debug features should be switched on
984 for (; *str && *str != ','; str++) {
985 switch (tolower(*str)) {
987 slub_debug |= SLAB_DEBUG_FREE;
990 slub_debug |= SLAB_RED_ZONE;
993 slub_debug |= SLAB_POISON;
996 slub_debug |= SLAB_STORE_USER;
999 slub_debug |= SLAB_TRACE;
1002 printk(KERN_ERR "slub_debug option '%c' "
1003 "unknown. skipped\n", *str);
1009 slub_debug_slabs = str + 1;
1014 __setup("slub_debug", setup_slub_debug);
1016 static unsigned long kmem_cache_flags(unsigned long objsize,
1017 unsigned long flags, const char *name,
1018 void (*ctor)(void *))
1021 * Enable debugging if selected on the kernel commandline.
1023 if (slub_debug && (!slub_debug_slabs ||
1024 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1025 flags |= slub_debug;
1030 static inline void setup_object_debug(struct kmem_cache *s,
1031 struct page *page, void *object) {}
1033 static inline int alloc_debug_processing(struct kmem_cache *s,
1034 struct page *page, void *object, unsigned long addr) { return 0; }
1036 static inline int free_debug_processing(struct kmem_cache *s,
1037 struct page *page, void *object, unsigned long addr) { return 0; }
1039 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1041 static inline int check_object(struct kmem_cache *s, struct page *page,
1042 void *object, int active) { return 1; }
1043 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1044 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1045 unsigned long flags, const char *name,
1046 void (*ctor)(void *))
1050 #define slub_debug 0
1052 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1054 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1056 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1061 * Slab allocation and freeing
1063 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1064 struct kmem_cache_order_objects oo)
1066 int order = oo_order(oo);
1069 return alloc_pages(flags, order);
1071 return alloc_pages_node(node, flags, order);
1074 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1077 struct kmem_cache_order_objects oo = s->oo;
1079 flags |= s->allocflags;
1081 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1083 if (unlikely(!page)) {
1086 * Allocation may have failed due to fragmentation.
1087 * Try a lower order alloc if possible
1089 page = alloc_slab_page(flags, node, oo);
1093 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1095 page->objects = oo_objects(oo);
1096 mod_zone_page_state(page_zone(page),
1097 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1098 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1104 static void setup_object(struct kmem_cache *s, struct page *page,
1107 setup_object_debug(s, page, object);
1108 if (unlikely(s->ctor))
1112 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1119 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1121 page = allocate_slab(s,
1122 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1126 inc_slabs_node(s, page_to_nid(page), page->objects);
1128 page->flags |= 1 << PG_slab;
1129 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1130 SLAB_STORE_USER | SLAB_TRACE))
1131 __SetPageSlubDebug(page);
1133 start = page_address(page);
1135 if (unlikely(s->flags & SLAB_POISON))
1136 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1139 for_each_object(p, s, start, page->objects) {
1140 setup_object(s, page, last);
1141 set_freepointer(s, last, p);
1144 setup_object(s, page, last);
1145 set_freepointer(s, last, NULL);
1147 page->freelist = start;
1153 static void __free_slab(struct kmem_cache *s, struct page *page)
1155 int order = compound_order(page);
1156 int pages = 1 << order;
1158 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1161 slab_pad_check(s, page);
1162 for_each_object(p, s, page_address(page),
1164 check_object(s, page, p, 0);
1165 __ClearPageSlubDebug(page);
1168 mod_zone_page_state(page_zone(page),
1169 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1170 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1173 __ClearPageSlab(page);
1174 reset_page_mapcount(page);
1175 if (current->reclaim_state)
1176 current->reclaim_state->reclaimed_slab += pages;
1177 __free_pages(page, order);
1180 static void rcu_free_slab(struct rcu_head *h)
1184 page = container_of((struct list_head *)h, struct page, lru);
1185 __free_slab(page->slab, page);
1188 static void free_slab(struct kmem_cache *s, struct page *page)
1190 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1192 * RCU free overloads the RCU head over the LRU
1194 struct rcu_head *head = (void *)&page->lru;
1196 call_rcu(head, rcu_free_slab);
1198 __free_slab(s, page);
1201 static void discard_slab(struct kmem_cache *s, struct page *page)
1203 dec_slabs_node(s, page_to_nid(page), page->objects);
1208 * Per slab locking using the pagelock
1210 static __always_inline void slab_lock(struct page *page)
1212 bit_spin_lock(PG_locked, &page->flags);
1215 static __always_inline void slab_unlock(struct page *page)
1217 __bit_spin_unlock(PG_locked, &page->flags);
1220 static __always_inline int slab_trylock(struct page *page)
1224 rc = bit_spin_trylock(PG_locked, &page->flags);
1229 * Management of partially allocated slabs
1231 static void add_partial(struct kmem_cache_node *n,
1232 struct page *page, int tail)
1234 spin_lock(&n->list_lock);
1237 list_add_tail(&page->lru, &n->partial);
1239 list_add(&page->lru, &n->partial);
1240 spin_unlock(&n->list_lock);
1243 static void remove_partial(struct kmem_cache *s, struct page *page)
1245 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1247 spin_lock(&n->list_lock);
1248 list_del(&page->lru);
1250 spin_unlock(&n->list_lock);
1254 * Lock slab and remove from the partial list.
1256 * Must hold list_lock.
1258 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1261 if (slab_trylock(page)) {
1262 list_del(&page->lru);
1264 __SetPageSlubFrozen(page);
1271 * Try to allocate a partial slab from a specific node.
1273 static struct page *get_partial_node(struct kmem_cache_node *n)
1278 * Racy check. If we mistakenly see no partial slabs then we
1279 * just allocate an empty slab. If we mistakenly try to get a
1280 * partial slab and there is none available then get_partials()
1283 if (!n || !n->nr_partial)
1286 spin_lock(&n->list_lock);
1287 list_for_each_entry(page, &n->partial, lru)
1288 if (lock_and_freeze_slab(n, page))
1292 spin_unlock(&n->list_lock);
1297 * Get a page from somewhere. Search in increasing NUMA distances.
1299 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1302 struct zonelist *zonelist;
1305 enum zone_type high_zoneidx = gfp_zone(flags);
1309 * The defrag ratio allows a configuration of the tradeoffs between
1310 * inter node defragmentation and node local allocations. A lower
1311 * defrag_ratio increases the tendency to do local allocations
1312 * instead of attempting to obtain partial slabs from other nodes.
1314 * If the defrag_ratio is set to 0 then kmalloc() always
1315 * returns node local objects. If the ratio is higher then kmalloc()
1316 * may return off node objects because partial slabs are obtained
1317 * from other nodes and filled up.
1319 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1320 * defrag_ratio = 1000) then every (well almost) allocation will
1321 * first attempt to defrag slab caches on other nodes. This means
1322 * scanning over all nodes to look for partial slabs which may be
1323 * expensive if we do it every time we are trying to find a slab
1324 * with available objects.
1326 if (!s->remote_node_defrag_ratio ||
1327 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1330 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1331 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1332 struct kmem_cache_node *n;
1334 n = get_node(s, zone_to_nid(zone));
1336 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1337 n->nr_partial > s->min_partial) {
1338 page = get_partial_node(n);
1348 * Get a partial page, lock it and return it.
1350 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1353 int searchnode = (node == -1) ? numa_node_id() : node;
1355 page = get_partial_node(get_node(s, searchnode));
1356 if (page || (flags & __GFP_THISNODE))
1359 return get_any_partial(s, flags);
1363 * Move a page back to the lists.
1365 * Must be called with the slab lock held.
1367 * On exit the slab lock will have been dropped.
1369 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1371 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1372 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1374 __ClearPageSlubFrozen(page);
1377 if (page->freelist) {
1378 add_partial(n, page, tail);
1379 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1381 stat(c, DEACTIVATE_FULL);
1382 if (SLABDEBUG && PageSlubDebug(page) &&
1383 (s->flags & SLAB_STORE_USER))
1388 stat(c, DEACTIVATE_EMPTY);
1389 if (n->nr_partial < s->min_partial) {
1391 * Adding an empty slab to the partial slabs in order
1392 * to avoid page allocator overhead. This slab needs
1393 * to come after the other slabs with objects in
1394 * so that the others get filled first. That way the
1395 * size of the partial list stays small.
1397 * kmem_cache_shrink can reclaim any empty slabs from
1400 add_partial(n, page, 1);
1404 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1405 discard_slab(s, page);
1411 * Remove the cpu slab
1413 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1415 struct page *page = c->page;
1419 stat(c, DEACTIVATE_REMOTE_FREES);
1421 * Merge cpu freelist into slab freelist. Typically we get here
1422 * because both freelists are empty. So this is unlikely
1425 while (unlikely(c->freelist)) {
1428 tail = 0; /* Hot objects. Put the slab first */
1430 /* Retrieve object from cpu_freelist */
1431 object = c->freelist;
1432 c->freelist = c->freelist[c->offset];
1434 /* And put onto the regular freelist */
1435 object[c->offset] = page->freelist;
1436 page->freelist = object;
1440 unfreeze_slab(s, page, tail);
1443 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1445 stat(c, CPUSLAB_FLUSH);
1447 deactivate_slab(s, c);
1453 * Called from IPI handler with interrupts disabled.
1455 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1457 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1459 if (likely(c && c->page))
1463 static void flush_cpu_slab(void *d)
1465 struct kmem_cache *s = d;
1467 __flush_cpu_slab(s, smp_processor_id());
1470 static void flush_all(struct kmem_cache *s)
1472 on_each_cpu(flush_cpu_slab, s, 1);
1476 * Check if the objects in a per cpu structure fit numa
1477 * locality expectations.
1479 static inline int node_match(struct kmem_cache_cpu *c, int node)
1482 if (node != -1 && c->node != node)
1489 * Slow path. The lockless freelist is empty or we need to perform
1492 * Interrupts are disabled.
1494 * Processing is still very fast if new objects have been freed to the
1495 * regular freelist. In that case we simply take over the regular freelist
1496 * as the lockless freelist and zap the regular freelist.
1498 * If that is not working then we fall back to the partial lists. We take the
1499 * first element of the freelist as the object to allocate now and move the
1500 * rest of the freelist to the lockless freelist.
1502 * And if we were unable to get a new slab from the partial slab lists then
1503 * we need to allocate a new slab. This is the slowest path since it involves
1504 * a call to the page allocator and the setup of a new slab.
1506 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1507 unsigned long addr, struct kmem_cache_cpu *c)
1512 /* We handle __GFP_ZERO in the caller */
1513 gfpflags &= ~__GFP_ZERO;
1519 if (unlikely(!node_match(c, node)))
1522 stat(c, ALLOC_REFILL);
1525 object = c->page->freelist;
1526 if (unlikely(!object))
1528 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1531 c->freelist = object[c->offset];
1532 c->page->inuse = c->page->objects;
1533 c->page->freelist = NULL;
1534 c->node = page_to_nid(c->page);
1536 slab_unlock(c->page);
1537 stat(c, ALLOC_SLOWPATH);
1541 deactivate_slab(s, c);
1544 new = get_partial(s, gfpflags, node);
1547 stat(c, ALLOC_FROM_PARTIAL);
1551 if (gfpflags & __GFP_WAIT)
1554 new = new_slab(s, gfpflags, node);
1556 if (gfpflags & __GFP_WAIT)
1557 local_irq_disable();
1560 c = get_cpu_slab(s, smp_processor_id());
1561 stat(c, ALLOC_SLAB);
1565 __SetPageSlubFrozen(new);
1571 if (!alloc_debug_processing(s, c->page, object, addr))
1575 c->page->freelist = object[c->offset];
1581 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1582 * have the fastpath folded into their functions. So no function call
1583 * overhead for requests that can be satisfied on the fastpath.
1585 * The fastpath works by first checking if the lockless freelist can be used.
1586 * If not then __slab_alloc is called for slow processing.
1588 * Otherwise we can simply pick the next object from the lockless free list.
1590 static __always_inline void *slab_alloc(struct kmem_cache *s,
1591 gfp_t gfpflags, int node, unsigned long addr)
1594 struct kmem_cache_cpu *c;
1595 unsigned long flags;
1596 unsigned int objsize;
1598 lockdep_trace_alloc(gfpflags);
1599 might_sleep_if(gfpflags & __GFP_WAIT);
1601 if (should_failslab(s->objsize, gfpflags))
1604 local_irq_save(flags);
1605 c = get_cpu_slab(s, smp_processor_id());
1606 objsize = c->objsize;
1607 if (unlikely(!c->freelist || !node_match(c, node)))
1609 object = __slab_alloc(s, gfpflags, node, addr, c);
1612 object = c->freelist;
1613 c->freelist = object[c->offset];
1614 stat(c, ALLOC_FASTPATH);
1616 local_irq_restore(flags);
1618 if (unlikely((gfpflags & __GFP_ZERO) && object))
1619 memset(object, 0, objsize);
1621 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
1625 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1627 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1629 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1633 EXPORT_SYMBOL(kmem_cache_alloc);
1635 #ifdef CONFIG_KMEMTRACE
1636 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1638 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1640 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1644 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1646 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1648 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1649 s->objsize, s->size, gfpflags, node);
1653 EXPORT_SYMBOL(kmem_cache_alloc_node);
1656 #ifdef CONFIG_KMEMTRACE
1657 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1661 return slab_alloc(s, gfpflags, node, _RET_IP_);
1663 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1667 * Slow patch handling. This may still be called frequently since objects
1668 * have a longer lifetime than the cpu slabs in most processing loads.
1670 * So we still attempt to reduce cache line usage. Just take the slab
1671 * lock and free the item. If there is no additional partial page
1672 * handling required then we can return immediately.
1674 static void __slab_free(struct kmem_cache *s, struct page *page,
1675 void *x, unsigned long addr, unsigned int offset)
1678 void **object = (void *)x;
1679 struct kmem_cache_cpu *c;
1681 c = get_cpu_slab(s, raw_smp_processor_id());
1682 stat(c, FREE_SLOWPATH);
1685 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1689 prior = object[offset] = page->freelist;
1690 page->freelist = object;
1693 if (unlikely(PageSlubFrozen(page))) {
1694 stat(c, FREE_FROZEN);
1698 if (unlikely(!page->inuse))
1702 * Objects left in the slab. If it was not on the partial list before
1705 if (unlikely(!prior)) {
1706 add_partial(get_node(s, page_to_nid(page)), page, 1);
1707 stat(c, FREE_ADD_PARTIAL);
1717 * Slab still on the partial list.
1719 remove_partial(s, page);
1720 stat(c, FREE_REMOVE_PARTIAL);
1724 discard_slab(s, page);
1728 if (!free_debug_processing(s, page, x, addr))
1734 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1735 * can perform fastpath freeing without additional function calls.
1737 * The fastpath is only possible if we are freeing to the current cpu slab
1738 * of this processor. This typically the case if we have just allocated
1741 * If fastpath is not possible then fall back to __slab_free where we deal
1742 * with all sorts of special processing.
1744 static __always_inline void slab_free(struct kmem_cache *s,
1745 struct page *page, void *x, unsigned long addr)
1747 void **object = (void *)x;
1748 struct kmem_cache_cpu *c;
1749 unsigned long flags;
1751 kmemleak_free_recursive(x, s->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, c->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 trace_kmem_cache_free(_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)
1881 * Attempt to find best configuration for a slab. This
1882 * works by first attempting to generate a layout with
1883 * the best configuration and backing off gradually.
1885 * First we reduce the acceptable waste in a slab. Then
1886 * we reduce the minimum objects required in a slab.
1888 min_objects = slub_min_objects;
1890 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1891 max_objects = (PAGE_SIZE << slub_max_order)/size;
1892 min_objects = min(min_objects, max_objects);
1894 while (min_objects > 1) {
1896 while (fraction >= 4) {
1897 order = slab_order(size, min_objects,
1898 slub_max_order, fraction);
1899 if (order <= slub_max_order)
1907 * We were unable to place multiple objects in a slab. Now
1908 * lets see if we can place a single object there.
1910 order = slab_order(size, 1, slub_max_order, 1);
1911 if (order <= slub_max_order)
1915 * Doh this slab cannot be placed using slub_max_order.
1917 order = slab_order(size, 1, MAX_ORDER, 1);
1918 if (order < MAX_ORDER)
1924 * Figure out what the alignment of the objects will be.
1926 static unsigned long calculate_alignment(unsigned long flags,
1927 unsigned long align, unsigned long size)
1930 * If the user wants hardware cache aligned objects then follow that
1931 * suggestion if the object is sufficiently large.
1933 * The hardware cache alignment cannot override the specified
1934 * alignment though. If that is greater then use it.
1936 if (flags & SLAB_HWCACHE_ALIGN) {
1937 unsigned long ralign = cache_line_size();
1938 while (size <= ralign / 2)
1940 align = max(align, ralign);
1943 if (align < ARCH_SLAB_MINALIGN)
1944 align = ARCH_SLAB_MINALIGN;
1946 return ALIGN(align, sizeof(void *));
1949 static void init_kmem_cache_cpu(struct kmem_cache *s,
1950 struct kmem_cache_cpu *c)
1955 c->offset = s->offset / sizeof(void *);
1956 c->objsize = s->objsize;
1957 #ifdef CONFIG_SLUB_STATS
1958 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1963 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1966 spin_lock_init(&n->list_lock);
1967 INIT_LIST_HEAD(&n->partial);
1968 #ifdef CONFIG_SLUB_DEBUG
1969 atomic_long_set(&n->nr_slabs, 0);
1970 atomic_long_set(&n->total_objects, 0);
1971 INIT_LIST_HEAD(&n->full);
1977 * Per cpu array for per cpu structures.
1979 * The per cpu array places all kmem_cache_cpu structures from one processor
1980 * close together meaning that it becomes possible that multiple per cpu
1981 * structures are contained in one cacheline. This may be particularly
1982 * beneficial for the kmalloc caches.
1984 * A desktop system typically has around 60-80 slabs. With 100 here we are
1985 * likely able to get per cpu structures for all caches from the array defined
1986 * here. We must be able to cover all kmalloc caches during bootstrap.
1988 * If the per cpu array is exhausted then fall back to kmalloc
1989 * of individual cachelines. No sharing is possible then.
1991 #define NR_KMEM_CACHE_CPU 100
1993 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1994 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1996 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1997 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
1999 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2000 int cpu, gfp_t flags)
2002 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2005 per_cpu(kmem_cache_cpu_free, cpu) =
2006 (void *)c->freelist;
2008 /* Table overflow: So allocate ourselves */
2010 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2011 flags, cpu_to_node(cpu));
2016 init_kmem_cache_cpu(s, c);
2020 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2022 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2023 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2027 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2028 per_cpu(kmem_cache_cpu_free, cpu) = c;
2031 static void free_kmem_cache_cpus(struct kmem_cache *s)
2035 for_each_online_cpu(cpu) {
2036 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2039 s->cpu_slab[cpu] = NULL;
2040 free_kmem_cache_cpu(c, cpu);
2045 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2049 for_each_online_cpu(cpu) {
2050 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2055 c = alloc_kmem_cache_cpu(s, cpu, flags);
2057 free_kmem_cache_cpus(s);
2060 s->cpu_slab[cpu] = c;
2066 * Initialize the per cpu array.
2068 static void init_alloc_cpu_cpu(int cpu)
2072 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2075 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2076 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2078 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2081 static void __init init_alloc_cpu(void)
2085 for_each_online_cpu(cpu)
2086 init_alloc_cpu_cpu(cpu);
2090 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2091 static inline void init_alloc_cpu(void) {}
2093 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2095 init_kmem_cache_cpu(s, &s->cpu_slab);
2102 * No kmalloc_node yet so do it by hand. We know that this is the first
2103 * slab on the node for this slabcache. There are no concurrent accesses
2106 * Note that this function only works on the kmalloc_node_cache
2107 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2108 * memory on a fresh node that has no slab structures yet.
2110 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2113 struct kmem_cache_node *n;
2114 unsigned long flags;
2116 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2118 page = new_slab(kmalloc_caches, gfpflags, node);
2121 if (page_to_nid(page) != node) {
2122 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2124 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2125 "in order to be able to continue\n");
2130 page->freelist = get_freepointer(kmalloc_caches, n);
2132 kmalloc_caches->node[node] = n;
2133 #ifdef CONFIG_SLUB_DEBUG
2134 init_object(kmalloc_caches, n, 1);
2135 init_tracking(kmalloc_caches, n);
2137 init_kmem_cache_node(n, kmalloc_caches);
2138 inc_slabs_node(kmalloc_caches, node, page->objects);
2141 * lockdep requires consistent irq usage for each lock
2142 * so even though there cannot be a race this early in
2143 * the boot sequence, we still disable irqs.
2145 local_irq_save(flags);
2146 add_partial(n, page, 0);
2147 local_irq_restore(flags);
2150 static void free_kmem_cache_nodes(struct kmem_cache *s)
2154 for_each_node_state(node, N_NORMAL_MEMORY) {
2155 struct kmem_cache_node *n = s->node[node];
2156 if (n && n != &s->local_node)
2157 kmem_cache_free(kmalloc_caches, n);
2158 s->node[node] = NULL;
2162 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2167 if (slab_state >= UP)
2168 local_node = page_to_nid(virt_to_page(s));
2172 for_each_node_state(node, N_NORMAL_MEMORY) {
2173 struct kmem_cache_node *n;
2175 if (local_node == node)
2178 if (slab_state == DOWN) {
2179 early_kmem_cache_node_alloc(gfpflags, node);
2182 n = kmem_cache_alloc_node(kmalloc_caches,
2186 free_kmem_cache_nodes(s);
2192 init_kmem_cache_node(n, s);
2197 static void free_kmem_cache_nodes(struct kmem_cache *s)
2201 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2203 init_kmem_cache_node(&s->local_node, s);
2208 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2210 if (min < MIN_PARTIAL)
2212 else if (min > MAX_PARTIAL)
2214 s->min_partial = min;
2218 * calculate_sizes() determines the order and the distribution of data within
2221 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2223 unsigned long flags = s->flags;
2224 unsigned long size = s->objsize;
2225 unsigned long align = s->align;
2229 * Round up object size to the next word boundary. We can only
2230 * place the free pointer at word boundaries and this determines
2231 * the possible location of the free pointer.
2233 size = ALIGN(size, sizeof(void *));
2235 #ifdef CONFIG_SLUB_DEBUG
2237 * Determine if we can poison the object itself. If the user of
2238 * the slab may touch the object after free or before allocation
2239 * then we should never poison the object itself.
2241 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2243 s->flags |= __OBJECT_POISON;
2245 s->flags &= ~__OBJECT_POISON;
2249 * If we are Redzoning then check if there is some space between the
2250 * end of the object and the free pointer. If not then add an
2251 * additional word to have some bytes to store Redzone information.
2253 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2254 size += sizeof(void *);
2258 * With that we have determined the number of bytes in actual use
2259 * by the object. This is the potential offset to the free pointer.
2263 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2266 * Relocate free pointer after the object if it is not
2267 * permitted to overwrite the first word of the object on
2270 * This is the case if we do RCU, have a constructor or
2271 * destructor or are poisoning the objects.
2274 size += sizeof(void *);
2277 #ifdef CONFIG_SLUB_DEBUG
2278 if (flags & SLAB_STORE_USER)
2280 * Need to store information about allocs and frees after
2283 size += 2 * sizeof(struct track);
2285 if (flags & SLAB_RED_ZONE)
2287 * Add some empty padding so that we can catch
2288 * overwrites from earlier objects rather than let
2289 * tracking information or the free pointer be
2290 * corrupted if a user writes before the start
2293 size += sizeof(void *);
2297 * Determine the alignment based on various parameters that the
2298 * user specified and the dynamic determination of cache line size
2301 align = calculate_alignment(flags, align, s->objsize);
2304 * SLUB stores one object immediately after another beginning from
2305 * offset 0. In order to align the objects we have to simply size
2306 * each object to conform to the alignment.
2308 size = ALIGN(size, align);
2310 if (forced_order >= 0)
2311 order = forced_order;
2313 order = calculate_order(size);
2320 s->allocflags |= __GFP_COMP;
2322 if (s->flags & SLAB_CACHE_DMA)
2323 s->allocflags |= SLUB_DMA;
2325 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2326 s->allocflags |= __GFP_RECLAIMABLE;
2329 * Determine the number of objects per slab
2331 s->oo = oo_make(order, size);
2332 s->min = oo_make(get_order(size), size);
2333 if (oo_objects(s->oo) > oo_objects(s->max))
2336 return !!oo_objects(s->oo);
2340 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2341 const char *name, size_t size,
2342 size_t align, unsigned long flags,
2343 void (*ctor)(void *))
2345 memset(s, 0, kmem_size);
2350 s->flags = kmem_cache_flags(size, flags, name, ctor);
2352 if (!calculate_sizes(s, -1))
2356 * The larger the object size is, the more pages we want on the partial
2357 * list to avoid pounding the page allocator excessively.
2359 set_min_partial(s, ilog2(s->size));
2362 s->remote_node_defrag_ratio = 1000;
2364 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2367 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2369 free_kmem_cache_nodes(s);
2371 if (flags & SLAB_PANIC)
2372 panic("Cannot create slab %s size=%lu realsize=%u "
2373 "order=%u offset=%u flags=%lx\n",
2374 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2380 * Check if a given pointer is valid
2382 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2386 page = get_object_page(object);
2388 if (!page || s != page->slab)
2389 /* No slab or wrong slab */
2392 if (!check_valid_pointer(s, page, object))
2396 * We could also check if the object is on the slabs freelist.
2397 * But this would be too expensive and it seems that the main
2398 * purpose of kmem_ptr_valid() is to check if the object belongs
2399 * to a certain slab.
2403 EXPORT_SYMBOL(kmem_ptr_validate);
2406 * Determine the size of a slab object
2408 unsigned int kmem_cache_size(struct kmem_cache *s)
2412 EXPORT_SYMBOL(kmem_cache_size);
2414 const char *kmem_cache_name(struct kmem_cache *s)
2418 EXPORT_SYMBOL(kmem_cache_name);
2420 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2423 #ifdef CONFIG_SLUB_DEBUG
2424 void *addr = page_address(page);
2426 DECLARE_BITMAP(map, page->objects);
2428 bitmap_zero(map, page->objects);
2429 slab_err(s, page, "%s", text);
2431 for_each_free_object(p, s, page->freelist)
2432 set_bit(slab_index(p, s, addr), map);
2434 for_each_object(p, s, addr, page->objects) {
2436 if (!test_bit(slab_index(p, s, addr), map)) {
2437 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2439 print_tracking(s, p);
2447 * Attempt to free all partial slabs on a node.
2449 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2451 unsigned long flags;
2452 struct page *page, *h;
2454 spin_lock_irqsave(&n->list_lock, flags);
2455 list_for_each_entry_safe(page, h, &n->partial, lru) {
2457 list_del(&page->lru);
2458 discard_slab(s, page);
2461 list_slab_objects(s, page,
2462 "Objects remaining on kmem_cache_close()");
2465 spin_unlock_irqrestore(&n->list_lock, flags);
2469 * Release all resources used by a slab cache.
2471 static inline int kmem_cache_close(struct kmem_cache *s)
2477 /* Attempt to free all objects */
2478 free_kmem_cache_cpus(s);
2479 for_each_node_state(node, N_NORMAL_MEMORY) {
2480 struct kmem_cache_node *n = get_node(s, node);
2483 if (n->nr_partial || slabs_node(s, node))
2486 free_kmem_cache_nodes(s);
2491 * Close a cache and release the kmem_cache structure
2492 * (must be used for caches created using kmem_cache_create)
2494 void kmem_cache_destroy(struct kmem_cache *s)
2496 down_write(&slub_lock);
2500 up_write(&slub_lock);
2501 if (kmem_cache_close(s)) {
2502 printk(KERN_ERR "SLUB %s: %s called for cache that "
2503 "still has objects.\n", s->name, __func__);
2506 sysfs_slab_remove(s);
2508 up_write(&slub_lock);
2510 EXPORT_SYMBOL(kmem_cache_destroy);
2512 /********************************************************************
2514 *******************************************************************/
2516 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2517 EXPORT_SYMBOL(kmalloc_caches);
2519 static int __init setup_slub_min_order(char *str)
2521 get_option(&str, &slub_min_order);
2526 __setup("slub_min_order=", setup_slub_min_order);
2528 static int __init setup_slub_max_order(char *str)
2530 get_option(&str, &slub_max_order);
2531 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2536 __setup("slub_max_order=", setup_slub_max_order);
2538 static int __init setup_slub_min_objects(char *str)
2540 get_option(&str, &slub_min_objects);
2545 __setup("slub_min_objects=", setup_slub_min_objects);
2547 static int __init setup_slub_nomerge(char *str)
2553 __setup("slub_nomerge", setup_slub_nomerge);
2555 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2556 const char *name, int size, gfp_t gfp_flags)
2558 unsigned int flags = 0;
2560 if (gfp_flags & SLUB_DMA)
2561 flags = SLAB_CACHE_DMA;
2563 down_write(&slub_lock);
2564 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2568 list_add(&s->list, &slab_caches);
2569 up_write(&slub_lock);
2570 if (sysfs_slab_add(s))
2575 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2578 #ifdef CONFIG_ZONE_DMA
2579 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2581 static void sysfs_add_func(struct work_struct *w)
2583 struct kmem_cache *s;
2585 down_write(&slub_lock);
2586 list_for_each_entry(s, &slab_caches, list) {
2587 if (s->flags & __SYSFS_ADD_DEFERRED) {
2588 s->flags &= ~__SYSFS_ADD_DEFERRED;
2592 up_write(&slub_lock);
2595 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2597 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2599 struct kmem_cache *s;
2603 s = kmalloc_caches_dma[index];
2607 /* Dynamically create dma cache */
2608 if (flags & __GFP_WAIT)
2609 down_write(&slub_lock);
2611 if (!down_write_trylock(&slub_lock))
2615 if (kmalloc_caches_dma[index])
2618 realsize = kmalloc_caches[index].objsize;
2619 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2620 (unsigned int)realsize);
2621 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2623 if (!s || !text || !kmem_cache_open(s, flags, text,
2624 realsize, ARCH_KMALLOC_MINALIGN,
2625 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2631 list_add(&s->list, &slab_caches);
2632 kmalloc_caches_dma[index] = s;
2634 schedule_work(&sysfs_add_work);
2637 up_write(&slub_lock);
2639 return kmalloc_caches_dma[index];
2644 * Conversion table for small slabs sizes / 8 to the index in the
2645 * kmalloc array. This is necessary for slabs < 192 since we have non power
2646 * of two cache sizes there. The size of larger slabs can be determined using
2649 static s8 size_index[24] = {
2676 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2682 return ZERO_SIZE_PTR;
2684 index = size_index[(size - 1) / 8];
2686 index = fls(size - 1);
2688 #ifdef CONFIG_ZONE_DMA
2689 if (unlikely((flags & SLUB_DMA)))
2690 return dma_kmalloc_cache(index, flags);
2693 return &kmalloc_caches[index];
2696 void *__kmalloc(size_t size, gfp_t flags)
2698 struct kmem_cache *s;
2701 if (unlikely(size > SLUB_MAX_SIZE))
2702 return kmalloc_large(size, flags);
2704 s = get_slab(size, flags);
2706 if (unlikely(ZERO_OR_NULL_PTR(s)))
2709 ret = slab_alloc(s, flags, -1, _RET_IP_);
2711 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2715 EXPORT_SYMBOL(__kmalloc);
2717 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2719 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2723 return page_address(page);
2729 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2731 struct kmem_cache *s;
2734 if (unlikely(size > SLUB_MAX_SIZE)) {
2735 ret = kmalloc_large_node(size, flags, node);
2737 trace_kmalloc_node(_RET_IP_, ret,
2738 size, PAGE_SIZE << get_order(size),
2744 s = get_slab(size, flags);
2746 if (unlikely(ZERO_OR_NULL_PTR(s)))
2749 ret = slab_alloc(s, flags, node, _RET_IP_);
2751 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2755 EXPORT_SYMBOL(__kmalloc_node);
2758 size_t ksize(const void *object)
2761 struct kmem_cache *s;
2763 if (unlikely(object == ZERO_SIZE_PTR))
2766 page = virt_to_head_page(object);
2768 if (unlikely(!PageSlab(page))) {
2769 WARN_ON(!PageCompound(page));
2770 return PAGE_SIZE << compound_order(page);
2774 #ifdef CONFIG_SLUB_DEBUG
2776 * Debugging requires use of the padding between object
2777 * and whatever may come after it.
2779 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2784 * If we have the need to store the freelist pointer
2785 * back there or track user information then we can
2786 * only use the space before that information.
2788 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2791 * Else we can use all the padding etc for the allocation
2795 EXPORT_SYMBOL(ksize);
2797 void kfree(const void *x)
2800 void *object = (void *)x;
2802 trace_kfree(_RET_IP_, x);
2804 if (unlikely(ZERO_OR_NULL_PTR(x)))
2807 page = virt_to_head_page(x);
2808 if (unlikely(!PageSlab(page))) {
2809 BUG_ON(!PageCompound(page));
2813 slab_free(page->slab, page, object, _RET_IP_);
2815 EXPORT_SYMBOL(kfree);
2818 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2819 * the remaining slabs by the number of items in use. The slabs with the
2820 * most items in use come first. New allocations will then fill those up
2821 * and thus they can be removed from the partial lists.
2823 * The slabs with the least items are placed last. This results in them
2824 * being allocated from last increasing the chance that the last objects
2825 * are freed in them.
2827 int kmem_cache_shrink(struct kmem_cache *s)
2831 struct kmem_cache_node *n;
2834 int objects = oo_objects(s->max);
2835 struct list_head *slabs_by_inuse =
2836 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2837 unsigned long flags;
2839 if (!slabs_by_inuse)
2843 for_each_node_state(node, N_NORMAL_MEMORY) {
2844 n = get_node(s, node);
2849 for (i = 0; i < objects; i++)
2850 INIT_LIST_HEAD(slabs_by_inuse + i);
2852 spin_lock_irqsave(&n->list_lock, flags);
2855 * Build lists indexed by the items in use in each slab.
2857 * Note that concurrent frees may occur while we hold the
2858 * list_lock. page->inuse here is the upper limit.
2860 list_for_each_entry_safe(page, t, &n->partial, lru) {
2861 if (!page->inuse && slab_trylock(page)) {
2863 * Must hold slab lock here because slab_free
2864 * may have freed the last object and be
2865 * waiting to release the slab.
2867 list_del(&page->lru);
2870 discard_slab(s, page);
2872 list_move(&page->lru,
2873 slabs_by_inuse + page->inuse);
2878 * Rebuild the partial list with the slabs filled up most
2879 * first and the least used slabs at the end.
2881 for (i = objects - 1; i >= 0; i--)
2882 list_splice(slabs_by_inuse + i, n->partial.prev);
2884 spin_unlock_irqrestore(&n->list_lock, flags);
2887 kfree(slabs_by_inuse);
2890 EXPORT_SYMBOL(kmem_cache_shrink);
2892 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2893 static int slab_mem_going_offline_callback(void *arg)
2895 struct kmem_cache *s;
2897 down_read(&slub_lock);
2898 list_for_each_entry(s, &slab_caches, list)
2899 kmem_cache_shrink(s);
2900 up_read(&slub_lock);
2905 static void slab_mem_offline_callback(void *arg)
2907 struct kmem_cache_node *n;
2908 struct kmem_cache *s;
2909 struct memory_notify *marg = arg;
2912 offline_node = marg->status_change_nid;
2915 * If the node still has available memory. we need kmem_cache_node
2918 if (offline_node < 0)
2921 down_read(&slub_lock);
2922 list_for_each_entry(s, &slab_caches, list) {
2923 n = get_node(s, offline_node);
2926 * if n->nr_slabs > 0, slabs still exist on the node
2927 * that is going down. We were unable to free them,
2928 * and offline_pages() function shoudn't call this
2929 * callback. So, we must fail.
2931 BUG_ON(slabs_node(s, offline_node));
2933 s->node[offline_node] = NULL;
2934 kmem_cache_free(kmalloc_caches, n);
2937 up_read(&slub_lock);
2940 static int slab_mem_going_online_callback(void *arg)
2942 struct kmem_cache_node *n;
2943 struct kmem_cache *s;
2944 struct memory_notify *marg = arg;
2945 int nid = marg->status_change_nid;
2949 * If the node's memory is already available, then kmem_cache_node is
2950 * already created. Nothing to do.
2956 * We are bringing a node online. No memory is available yet. We must
2957 * allocate a kmem_cache_node structure in order to bring the node
2960 down_read(&slub_lock);
2961 list_for_each_entry(s, &slab_caches, list) {
2963 * XXX: kmem_cache_alloc_node will fallback to other nodes
2964 * since memory is not yet available from the node that
2967 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2972 init_kmem_cache_node(n, s);
2976 up_read(&slub_lock);
2980 static int slab_memory_callback(struct notifier_block *self,
2981 unsigned long action, void *arg)
2986 case MEM_GOING_ONLINE:
2987 ret = slab_mem_going_online_callback(arg);
2989 case MEM_GOING_OFFLINE:
2990 ret = slab_mem_going_offline_callback(arg);
2993 case MEM_CANCEL_ONLINE:
2994 slab_mem_offline_callback(arg);
2997 case MEM_CANCEL_OFFLINE:
3001 ret = notifier_from_errno(ret);
3007 #endif /* CONFIG_MEMORY_HOTPLUG */
3009 /********************************************************************
3010 * Basic setup of slabs
3011 *******************************************************************/
3013 void __init kmem_cache_init(void)
3022 * Must first have the slab cache available for the allocations of the
3023 * struct kmem_cache_node's. There is special bootstrap code in
3024 * kmem_cache_open for slab_state == DOWN.
3026 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3027 sizeof(struct kmem_cache_node), GFP_KERNEL);
3028 kmalloc_caches[0].refcount = -1;
3031 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3034 /* Able to allocate the per node structures */
3035 slab_state = PARTIAL;
3037 /* Caches that are not of the two-to-the-power-of size */
3038 if (KMALLOC_MIN_SIZE <= 64) {
3039 create_kmalloc_cache(&kmalloc_caches[1],
3040 "kmalloc-96", 96, GFP_KERNEL);
3042 create_kmalloc_cache(&kmalloc_caches[2],
3043 "kmalloc-192", 192, GFP_KERNEL);
3047 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3048 create_kmalloc_cache(&kmalloc_caches[i],
3049 "kmalloc", 1 << i, GFP_KERNEL);
3055 * Patch up the size_index table if we have strange large alignment
3056 * requirements for the kmalloc array. This is only the case for
3057 * MIPS it seems. The standard arches will not generate any code here.
3059 * Largest permitted alignment is 256 bytes due to the way we
3060 * handle the index determination for the smaller caches.
3062 * Make sure that nothing crazy happens if someone starts tinkering
3063 * around with ARCH_KMALLOC_MINALIGN
3065 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3066 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3068 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3069 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3071 if (KMALLOC_MIN_SIZE == 128) {
3073 * The 192 byte sized cache is not used if the alignment
3074 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3077 for (i = 128 + 8; i <= 192; i += 8)
3078 size_index[(i - 1) / 8] = 8;
3083 /* Provide the correct kmalloc names now that the caches are up */
3084 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3085 kmalloc_caches[i]. name =
3086 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3089 register_cpu_notifier(&slab_notifier);
3090 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3091 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3093 kmem_size = sizeof(struct kmem_cache);
3097 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3098 " CPUs=%d, Nodes=%d\n",
3099 caches, cache_line_size(),
3100 slub_min_order, slub_max_order, slub_min_objects,
3101 nr_cpu_ids, nr_node_ids);
3105 * Find a mergeable slab cache
3107 static int slab_unmergeable(struct kmem_cache *s)
3109 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3116 * We may have set a slab to be unmergeable during bootstrap.
3118 if (s->refcount < 0)
3124 static struct kmem_cache *find_mergeable(size_t size,
3125 size_t align, unsigned long flags, const char *name,
3126 void (*ctor)(void *))
3128 struct kmem_cache *s;
3130 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3136 size = ALIGN(size, sizeof(void *));
3137 align = calculate_alignment(flags, align, size);
3138 size = ALIGN(size, align);
3139 flags = kmem_cache_flags(size, flags, name, NULL);
3141 list_for_each_entry(s, &slab_caches, list) {
3142 if (slab_unmergeable(s))
3148 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3151 * Check if alignment is compatible.
3152 * Courtesy of Adrian Drzewiecki
3154 if ((s->size & ~(align - 1)) != s->size)
3157 if (s->size - size >= sizeof(void *))
3165 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3166 size_t align, unsigned long flags, void (*ctor)(void *))
3168 struct kmem_cache *s;
3170 down_write(&slub_lock);
3171 s = find_mergeable(size, align, flags, name, ctor);
3177 * Adjust the object sizes so that we clear
3178 * the complete object on kzalloc.
3180 s->objsize = max(s->objsize, (int)size);
3183 * And then we need to update the object size in the
3184 * per cpu structures
3186 for_each_online_cpu(cpu)
3187 get_cpu_slab(s, cpu)->objsize = s->objsize;
3189 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3190 up_write(&slub_lock);
3192 if (sysfs_slab_alias(s, name)) {
3193 down_write(&slub_lock);
3195 up_write(&slub_lock);
3201 s = kmalloc(kmem_size, GFP_KERNEL);
3203 if (kmem_cache_open(s, GFP_KERNEL, name,
3204 size, align, flags, ctor)) {
3205 list_add(&s->list, &slab_caches);
3206 up_write(&slub_lock);
3207 if (sysfs_slab_add(s)) {
3208 down_write(&slub_lock);
3210 up_write(&slub_lock);
3218 up_write(&slub_lock);
3221 if (flags & SLAB_PANIC)
3222 panic("Cannot create slabcache %s\n", name);
3227 EXPORT_SYMBOL(kmem_cache_create);
3231 * Use the cpu notifier to insure that the cpu slabs are flushed when
3234 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3235 unsigned long action, void *hcpu)
3237 long cpu = (long)hcpu;
3238 struct kmem_cache *s;
3239 unsigned long flags;
3242 case CPU_UP_PREPARE:
3243 case CPU_UP_PREPARE_FROZEN:
3244 init_alloc_cpu_cpu(cpu);
3245 down_read(&slub_lock);
3246 list_for_each_entry(s, &slab_caches, list)
3247 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3249 up_read(&slub_lock);
3252 case CPU_UP_CANCELED:
3253 case CPU_UP_CANCELED_FROZEN:
3255 case CPU_DEAD_FROZEN:
3256 down_read(&slub_lock);
3257 list_for_each_entry(s, &slab_caches, list) {
3258 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3260 local_irq_save(flags);
3261 __flush_cpu_slab(s, cpu);
3262 local_irq_restore(flags);
3263 free_kmem_cache_cpu(c, cpu);
3264 s->cpu_slab[cpu] = NULL;
3266 up_read(&slub_lock);
3274 static struct notifier_block __cpuinitdata slab_notifier = {
3275 .notifier_call = slab_cpuup_callback
3280 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3282 struct kmem_cache *s;
3285 if (unlikely(size > SLUB_MAX_SIZE))
3286 return kmalloc_large(size, gfpflags);
3288 s = get_slab(size, gfpflags);
3290 if (unlikely(ZERO_OR_NULL_PTR(s)))
3293 ret = slab_alloc(s, gfpflags, -1, caller);
3295 /* Honor the call site pointer we recieved. */
3296 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3301 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3302 int node, unsigned long caller)
3304 struct kmem_cache *s;
3307 if (unlikely(size > SLUB_MAX_SIZE))
3308 return kmalloc_large_node(size, gfpflags, node);
3310 s = get_slab(size, gfpflags);
3312 if (unlikely(ZERO_OR_NULL_PTR(s)))
3315 ret = slab_alloc(s, gfpflags, node, caller);
3317 /* Honor the call site pointer we recieved. */
3318 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3323 #ifdef CONFIG_SLUB_DEBUG
3324 static unsigned long count_partial(struct kmem_cache_node *n,
3325 int (*get_count)(struct page *))
3327 unsigned long flags;
3328 unsigned long x = 0;
3331 spin_lock_irqsave(&n->list_lock, flags);
3332 list_for_each_entry(page, &n->partial, lru)
3333 x += get_count(page);
3334 spin_unlock_irqrestore(&n->list_lock, flags);
3338 static int count_inuse(struct page *page)
3343 static int count_total(struct page *page)
3345 return page->objects;
3348 static int count_free(struct page *page)
3350 return page->objects - page->inuse;
3353 static int validate_slab(struct kmem_cache *s, struct page *page,
3357 void *addr = page_address(page);
3359 if (!check_slab(s, page) ||
3360 !on_freelist(s, page, NULL))
3363 /* Now we know that a valid freelist exists */
3364 bitmap_zero(map, page->objects);
3366 for_each_free_object(p, s, page->freelist) {
3367 set_bit(slab_index(p, s, addr), map);
3368 if (!check_object(s, page, p, 0))
3372 for_each_object(p, s, addr, page->objects)
3373 if (!test_bit(slab_index(p, s, addr), map))
3374 if (!check_object(s, page, p, 1))
3379 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3382 if (slab_trylock(page)) {
3383 validate_slab(s, page, map);
3386 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3389 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3390 if (!PageSlubDebug(page))
3391 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3392 "on slab 0x%p\n", s->name, page);
3394 if (PageSlubDebug(page))
3395 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3396 "slab 0x%p\n", s->name, page);
3400 static int validate_slab_node(struct kmem_cache *s,
3401 struct kmem_cache_node *n, unsigned long *map)
3403 unsigned long count = 0;
3405 unsigned long flags;
3407 spin_lock_irqsave(&n->list_lock, flags);
3409 list_for_each_entry(page, &n->partial, lru) {
3410 validate_slab_slab(s, page, map);
3413 if (count != n->nr_partial)
3414 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3415 "counter=%ld\n", s->name, count, n->nr_partial);
3417 if (!(s->flags & SLAB_STORE_USER))
3420 list_for_each_entry(page, &n->full, lru) {
3421 validate_slab_slab(s, page, map);
3424 if (count != atomic_long_read(&n->nr_slabs))
3425 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3426 "counter=%ld\n", s->name, count,
3427 atomic_long_read(&n->nr_slabs));
3430 spin_unlock_irqrestore(&n->list_lock, flags);
3434 static long validate_slab_cache(struct kmem_cache *s)
3437 unsigned long count = 0;
3438 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3439 sizeof(unsigned long), GFP_KERNEL);
3445 for_each_node_state(node, N_NORMAL_MEMORY) {
3446 struct kmem_cache_node *n = get_node(s, node);
3448 count += validate_slab_node(s, n, map);
3454 #ifdef SLUB_RESILIENCY_TEST
3455 static void resiliency_test(void)
3459 printk(KERN_ERR "SLUB resiliency testing\n");
3460 printk(KERN_ERR "-----------------------\n");
3461 printk(KERN_ERR "A. Corruption after allocation\n");
3463 p = kzalloc(16, GFP_KERNEL);
3465 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3466 " 0x12->0x%p\n\n", p + 16);
3468 validate_slab_cache(kmalloc_caches + 4);
3470 /* Hmmm... The next two are dangerous */
3471 p = kzalloc(32, GFP_KERNEL);
3472 p[32 + sizeof(void *)] = 0x34;
3473 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3474 " 0x34 -> -0x%p\n", p);
3476 "If allocated object is overwritten then not detectable\n\n");
3478 validate_slab_cache(kmalloc_caches + 5);
3479 p = kzalloc(64, GFP_KERNEL);
3480 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3482 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3485 "If allocated object is overwritten then not detectable\n\n");
3486 validate_slab_cache(kmalloc_caches + 6);
3488 printk(KERN_ERR "\nB. Corruption after free\n");
3489 p = kzalloc(128, GFP_KERNEL);
3492 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3493 validate_slab_cache(kmalloc_caches + 7);
3495 p = kzalloc(256, GFP_KERNEL);
3498 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3500 validate_slab_cache(kmalloc_caches + 8);
3502 p = kzalloc(512, GFP_KERNEL);
3505 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3506 validate_slab_cache(kmalloc_caches + 9);
3509 static void resiliency_test(void) {};
3513 * Generate lists of code addresses where slabcache objects are allocated
3518 unsigned long count;
3525 DECLARE_BITMAP(cpus, NR_CPUS);
3531 unsigned long count;
3532 struct location *loc;
3535 static void free_loc_track(struct loc_track *t)
3538 free_pages((unsigned long)t->loc,
3539 get_order(sizeof(struct location) * t->max));
3542 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3547 order = get_order(sizeof(struct location) * max);
3549 l = (void *)__get_free_pages(flags, order);
3554 memcpy(l, t->loc, sizeof(struct location) * t->count);
3562 static int add_location(struct loc_track *t, struct kmem_cache *s,
3563 const struct track *track)
3565 long start, end, pos;
3567 unsigned long caddr;
3568 unsigned long age = jiffies - track->when;
3574 pos = start + (end - start + 1) / 2;
3577 * There is nothing at "end". If we end up there
3578 * we need to add something to before end.
3583 caddr = t->loc[pos].addr;
3584 if (track->addr == caddr) {
3590 if (age < l->min_time)
3592 if (age > l->max_time)
3595 if (track->pid < l->min_pid)
3596 l->min_pid = track->pid;
3597 if (track->pid > l->max_pid)
3598 l->max_pid = track->pid;
3600 cpumask_set_cpu(track->cpu,
3601 to_cpumask(l->cpus));
3603 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3607 if (track->addr < caddr)
3614 * Not found. Insert new tracking element.
3616 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3622 (t->count - pos) * sizeof(struct location));
3625 l->addr = track->addr;
3629 l->min_pid = track->pid;
3630 l->max_pid = track->pid;
3631 cpumask_clear(to_cpumask(l->cpus));
3632 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3633 nodes_clear(l->nodes);
3634 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3638 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3639 struct page *page, enum track_item alloc)
3641 void *addr = page_address(page);
3642 DECLARE_BITMAP(map, page->objects);
3645 bitmap_zero(map, page->objects);
3646 for_each_free_object(p, s, page->freelist)
3647 set_bit(slab_index(p, s, addr), map);
3649 for_each_object(p, s, addr, page->objects)
3650 if (!test_bit(slab_index(p, s, addr), map))
3651 add_location(t, s, get_track(s, p, alloc));
3654 static int list_locations(struct kmem_cache *s, char *buf,
3655 enum track_item alloc)
3659 struct loc_track t = { 0, 0, NULL };
3662 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3664 return sprintf(buf, "Out of memory\n");
3666 /* Push back cpu slabs */
3669 for_each_node_state(node, N_NORMAL_MEMORY) {
3670 struct kmem_cache_node *n = get_node(s, node);
3671 unsigned long flags;
3674 if (!atomic_long_read(&n->nr_slabs))
3677 spin_lock_irqsave(&n->list_lock, flags);
3678 list_for_each_entry(page, &n->partial, lru)
3679 process_slab(&t, s, page, alloc);
3680 list_for_each_entry(page, &n->full, lru)
3681 process_slab(&t, s, page, alloc);
3682 spin_unlock_irqrestore(&n->list_lock, flags);
3685 for (i = 0; i < t.count; i++) {
3686 struct location *l = &t.loc[i];
3688 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3690 len += sprintf(buf + len, "%7ld ", l->count);
3693 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3695 len += sprintf(buf + len, "<not-available>");
3697 if (l->sum_time != l->min_time) {
3698 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3700 (long)div_u64(l->sum_time, l->count),
3703 len += sprintf(buf + len, " age=%ld",
3706 if (l->min_pid != l->max_pid)
3707 len += sprintf(buf + len, " pid=%ld-%ld",
3708 l->min_pid, l->max_pid);
3710 len += sprintf(buf + len, " pid=%ld",
3713 if (num_online_cpus() > 1 &&
3714 !cpumask_empty(to_cpumask(l->cpus)) &&
3715 len < PAGE_SIZE - 60) {
3716 len += sprintf(buf + len, " cpus=");
3717 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3718 to_cpumask(l->cpus));
3721 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3722 len < PAGE_SIZE - 60) {
3723 len += sprintf(buf + len, " nodes=");
3724 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3728 len += sprintf(buf + len, "\n");
3733 len += sprintf(buf, "No data\n");
3737 enum slab_stat_type {
3738 SL_ALL, /* All slabs */
3739 SL_PARTIAL, /* Only partially allocated slabs */
3740 SL_CPU, /* Only slabs used for cpu caches */
3741 SL_OBJECTS, /* Determine allocated objects not slabs */
3742 SL_TOTAL /* Determine object capacity not slabs */
3745 #define SO_ALL (1 << SL_ALL)
3746 #define SO_PARTIAL (1 << SL_PARTIAL)
3747 #define SO_CPU (1 << SL_CPU)
3748 #define SO_OBJECTS (1 << SL_OBJECTS)
3749 #define SO_TOTAL (1 << SL_TOTAL)
3751 static ssize_t show_slab_objects(struct kmem_cache *s,
3752 char *buf, unsigned long flags)
3754 unsigned long total = 0;
3757 unsigned long *nodes;
3758 unsigned long *per_cpu;
3760 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3763 per_cpu = nodes + nr_node_ids;
3765 if (flags & SO_CPU) {
3768 for_each_possible_cpu(cpu) {
3769 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3771 if (!c || c->node < 0)
3775 if (flags & SO_TOTAL)
3776 x = c->page->objects;
3777 else if (flags & SO_OBJECTS)
3783 nodes[c->node] += x;
3789 if (flags & SO_ALL) {
3790 for_each_node_state(node, N_NORMAL_MEMORY) {
3791 struct kmem_cache_node *n = get_node(s, node);
3793 if (flags & SO_TOTAL)
3794 x = atomic_long_read(&n->total_objects);
3795 else if (flags & SO_OBJECTS)
3796 x = atomic_long_read(&n->total_objects) -
3797 count_partial(n, count_free);
3800 x = atomic_long_read(&n->nr_slabs);
3805 } else if (flags & SO_PARTIAL) {
3806 for_each_node_state(node, N_NORMAL_MEMORY) {
3807 struct kmem_cache_node *n = get_node(s, node);
3809 if (flags & SO_TOTAL)
3810 x = count_partial(n, count_total);
3811 else if (flags & SO_OBJECTS)
3812 x = count_partial(n, count_inuse);
3819 x = sprintf(buf, "%lu", total);
3821 for_each_node_state(node, N_NORMAL_MEMORY)
3823 x += sprintf(buf + x, " N%d=%lu",
3827 return x + sprintf(buf + x, "\n");
3830 static int any_slab_objects(struct kmem_cache *s)
3834 for_each_online_node(node) {
3835 struct kmem_cache_node *n = get_node(s, node);
3840 if (atomic_long_read(&n->total_objects))
3846 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3847 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3849 struct slab_attribute {
3850 struct attribute attr;
3851 ssize_t (*show)(struct kmem_cache *s, char *buf);
3852 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3855 #define SLAB_ATTR_RO(_name) \
3856 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3858 #define SLAB_ATTR(_name) \
3859 static struct slab_attribute _name##_attr = \
3860 __ATTR(_name, 0644, _name##_show, _name##_store)
3862 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3864 return sprintf(buf, "%d\n", s->size);
3866 SLAB_ATTR_RO(slab_size);
3868 static ssize_t align_show(struct kmem_cache *s, char *buf)
3870 return sprintf(buf, "%d\n", s->align);
3872 SLAB_ATTR_RO(align);
3874 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3876 return sprintf(buf, "%d\n", s->objsize);
3878 SLAB_ATTR_RO(object_size);
3880 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3882 return sprintf(buf, "%d\n", oo_objects(s->oo));
3884 SLAB_ATTR_RO(objs_per_slab);
3886 static ssize_t order_store(struct kmem_cache *s,
3887 const char *buf, size_t length)
3889 unsigned long order;
3892 err = strict_strtoul(buf, 10, &order);
3896 if (order > slub_max_order || order < slub_min_order)
3899 calculate_sizes(s, order);
3903 static ssize_t order_show(struct kmem_cache *s, char *buf)
3905 return sprintf(buf, "%d\n", oo_order(s->oo));
3909 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3911 return sprintf(buf, "%lu\n", s->min_partial);
3914 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3920 err = strict_strtoul(buf, 10, &min);
3924 set_min_partial(s, min);
3927 SLAB_ATTR(min_partial);
3929 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3932 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3934 return n + sprintf(buf + n, "\n");
3940 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3942 return sprintf(buf, "%d\n", s->refcount - 1);
3944 SLAB_ATTR_RO(aliases);
3946 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3948 return show_slab_objects(s, buf, SO_ALL);
3950 SLAB_ATTR_RO(slabs);
3952 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3954 return show_slab_objects(s, buf, SO_PARTIAL);
3956 SLAB_ATTR_RO(partial);
3958 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3960 return show_slab_objects(s, buf, SO_CPU);
3962 SLAB_ATTR_RO(cpu_slabs);
3964 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3966 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3968 SLAB_ATTR_RO(objects);
3970 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3972 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3974 SLAB_ATTR_RO(objects_partial);
3976 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3978 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3980 SLAB_ATTR_RO(total_objects);
3982 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3984 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3987 static ssize_t sanity_checks_store(struct kmem_cache *s,
3988 const char *buf, size_t length)
3990 s->flags &= ~SLAB_DEBUG_FREE;
3992 s->flags |= SLAB_DEBUG_FREE;
3995 SLAB_ATTR(sanity_checks);
3997 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3999 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4002 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4005 s->flags &= ~SLAB_TRACE;
4007 s->flags |= SLAB_TRACE;
4012 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4014 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4017 static ssize_t reclaim_account_store(struct kmem_cache *s,
4018 const char *buf, size_t length)
4020 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4022 s->flags |= SLAB_RECLAIM_ACCOUNT;
4025 SLAB_ATTR(reclaim_account);
4027 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4029 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4031 SLAB_ATTR_RO(hwcache_align);
4033 #ifdef CONFIG_ZONE_DMA
4034 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4036 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4038 SLAB_ATTR_RO(cache_dma);
4041 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4043 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4045 SLAB_ATTR_RO(destroy_by_rcu);
4047 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4049 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4052 static ssize_t red_zone_store(struct kmem_cache *s,
4053 const char *buf, size_t length)
4055 if (any_slab_objects(s))
4058 s->flags &= ~SLAB_RED_ZONE;
4060 s->flags |= SLAB_RED_ZONE;
4061 calculate_sizes(s, -1);
4064 SLAB_ATTR(red_zone);
4066 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4068 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4071 static ssize_t poison_store(struct kmem_cache *s,
4072 const char *buf, size_t length)
4074 if (any_slab_objects(s))
4077 s->flags &= ~SLAB_POISON;
4079 s->flags |= SLAB_POISON;
4080 calculate_sizes(s, -1);
4085 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4087 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4090 static ssize_t store_user_store(struct kmem_cache *s,
4091 const char *buf, size_t length)
4093 if (any_slab_objects(s))
4096 s->flags &= ~SLAB_STORE_USER;
4098 s->flags |= SLAB_STORE_USER;
4099 calculate_sizes(s, -1);
4102 SLAB_ATTR(store_user);
4104 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4109 static ssize_t validate_store(struct kmem_cache *s,
4110 const char *buf, size_t length)
4114 if (buf[0] == '1') {
4115 ret = validate_slab_cache(s);
4121 SLAB_ATTR(validate);
4123 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4128 static ssize_t shrink_store(struct kmem_cache *s,
4129 const char *buf, size_t length)
4131 if (buf[0] == '1') {
4132 int rc = kmem_cache_shrink(s);
4142 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4144 if (!(s->flags & SLAB_STORE_USER))
4146 return list_locations(s, buf, TRACK_ALLOC);
4148 SLAB_ATTR_RO(alloc_calls);
4150 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4152 if (!(s->flags & SLAB_STORE_USER))
4154 return list_locations(s, buf, TRACK_FREE);
4156 SLAB_ATTR_RO(free_calls);
4159 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4161 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4164 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4165 const char *buf, size_t length)
4167 unsigned long ratio;
4170 err = strict_strtoul(buf, 10, &ratio);
4175 s->remote_node_defrag_ratio = ratio * 10;
4179 SLAB_ATTR(remote_node_defrag_ratio);
4182 #ifdef CONFIG_SLUB_STATS
4183 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4185 unsigned long sum = 0;
4188 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4193 for_each_online_cpu(cpu) {
4194 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4200 len = sprintf(buf, "%lu", sum);
4203 for_each_online_cpu(cpu) {
4204 if (data[cpu] && len < PAGE_SIZE - 20)
4205 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4209 return len + sprintf(buf + len, "\n");
4212 #define STAT_ATTR(si, text) \
4213 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4215 return show_stat(s, buf, si); \
4217 SLAB_ATTR_RO(text); \
4219 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4220 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4221 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4222 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4223 STAT_ATTR(FREE_FROZEN, free_frozen);
4224 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4225 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4226 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4227 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4228 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4229 STAT_ATTR(FREE_SLAB, free_slab);
4230 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4231 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4232 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4233 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4234 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4235 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4236 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4239 static struct attribute *slab_attrs[] = {
4240 &slab_size_attr.attr,
4241 &object_size_attr.attr,
4242 &objs_per_slab_attr.attr,
4244 &min_partial_attr.attr,
4246 &objects_partial_attr.attr,
4247 &total_objects_attr.attr,
4250 &cpu_slabs_attr.attr,
4254 &sanity_checks_attr.attr,
4256 &hwcache_align_attr.attr,
4257 &reclaim_account_attr.attr,
4258 &destroy_by_rcu_attr.attr,
4259 &red_zone_attr.attr,
4261 &store_user_attr.attr,
4262 &validate_attr.attr,
4264 &alloc_calls_attr.attr,
4265 &free_calls_attr.attr,
4266 #ifdef CONFIG_ZONE_DMA
4267 &cache_dma_attr.attr,
4270 &remote_node_defrag_ratio_attr.attr,
4272 #ifdef CONFIG_SLUB_STATS
4273 &alloc_fastpath_attr.attr,
4274 &alloc_slowpath_attr.attr,
4275 &free_fastpath_attr.attr,
4276 &free_slowpath_attr.attr,
4277 &free_frozen_attr.attr,
4278 &free_add_partial_attr.attr,
4279 &free_remove_partial_attr.attr,
4280 &alloc_from_partial_attr.attr,
4281 &alloc_slab_attr.attr,
4282 &alloc_refill_attr.attr,
4283 &free_slab_attr.attr,
4284 &cpuslab_flush_attr.attr,
4285 &deactivate_full_attr.attr,
4286 &deactivate_empty_attr.attr,
4287 &deactivate_to_head_attr.attr,
4288 &deactivate_to_tail_attr.attr,
4289 &deactivate_remote_frees_attr.attr,
4290 &order_fallback_attr.attr,
4295 static struct attribute_group slab_attr_group = {
4296 .attrs = slab_attrs,
4299 static ssize_t slab_attr_show(struct kobject *kobj,
4300 struct attribute *attr,
4303 struct slab_attribute *attribute;
4304 struct kmem_cache *s;
4307 attribute = to_slab_attr(attr);
4310 if (!attribute->show)
4313 err = attribute->show(s, buf);
4318 static ssize_t slab_attr_store(struct kobject *kobj,
4319 struct attribute *attr,
4320 const char *buf, size_t len)
4322 struct slab_attribute *attribute;
4323 struct kmem_cache *s;
4326 attribute = to_slab_attr(attr);
4329 if (!attribute->store)
4332 err = attribute->store(s, buf, len);
4337 static void kmem_cache_release(struct kobject *kobj)
4339 struct kmem_cache *s = to_slab(kobj);
4344 static struct sysfs_ops slab_sysfs_ops = {
4345 .show = slab_attr_show,
4346 .store = slab_attr_store,
4349 static struct kobj_type slab_ktype = {
4350 .sysfs_ops = &slab_sysfs_ops,
4351 .release = kmem_cache_release
4354 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4356 struct kobj_type *ktype = get_ktype(kobj);
4358 if (ktype == &slab_ktype)
4363 static struct kset_uevent_ops slab_uevent_ops = {
4364 .filter = uevent_filter,
4367 static struct kset *slab_kset;
4369 #define ID_STR_LENGTH 64
4371 /* Create a unique string id for a slab cache:
4373 * Format :[flags-]size
4375 static char *create_unique_id(struct kmem_cache *s)
4377 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4384 * First flags affecting slabcache operations. We will only
4385 * get here for aliasable slabs so we do not need to support
4386 * too many flags. The flags here must cover all flags that
4387 * are matched during merging to guarantee that the id is
4390 if (s->flags & SLAB_CACHE_DMA)
4392 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4394 if (s->flags & SLAB_DEBUG_FREE)
4398 p += sprintf(p, "%07d", s->size);
4399 BUG_ON(p > name + ID_STR_LENGTH - 1);
4403 static int sysfs_slab_add(struct kmem_cache *s)
4409 if (slab_state < SYSFS)
4410 /* Defer until later */
4413 unmergeable = slab_unmergeable(s);
4416 * Slabcache can never be merged so we can use the name proper.
4417 * This is typically the case for debug situations. In that
4418 * case we can catch duplicate names easily.
4420 sysfs_remove_link(&slab_kset->kobj, s->name);
4424 * Create a unique name for the slab as a target
4427 name = create_unique_id(s);
4430 s->kobj.kset = slab_kset;
4431 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4433 kobject_put(&s->kobj);
4437 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4440 kobject_uevent(&s->kobj, KOBJ_ADD);
4442 /* Setup first alias */
4443 sysfs_slab_alias(s, s->name);
4449 static void sysfs_slab_remove(struct kmem_cache *s)
4451 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4452 kobject_del(&s->kobj);
4453 kobject_put(&s->kobj);
4457 * Need to buffer aliases during bootup until sysfs becomes
4458 * available lest we lose that information.
4460 struct saved_alias {
4461 struct kmem_cache *s;
4463 struct saved_alias *next;
4466 static struct saved_alias *alias_list;
4468 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4470 struct saved_alias *al;
4472 if (slab_state == SYSFS) {
4474 * If we have a leftover link then remove it.
4476 sysfs_remove_link(&slab_kset->kobj, name);
4477 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4480 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4486 al->next = alias_list;
4491 static int __init slab_sysfs_init(void)
4493 struct kmem_cache *s;
4496 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4498 printk(KERN_ERR "Cannot register slab subsystem.\n");
4504 list_for_each_entry(s, &slab_caches, list) {
4505 err = sysfs_slab_add(s);
4507 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4508 " to sysfs\n", s->name);
4511 while (alias_list) {
4512 struct saved_alias *al = alias_list;
4514 alias_list = alias_list->next;
4515 err = sysfs_slab_alias(al->s, al->name);
4517 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4518 " %s to sysfs\n", s->name);
4526 __initcall(slab_sysfs_init);
4530 * The /proc/slabinfo ABI
4532 #ifdef CONFIG_SLABINFO
4533 static void print_slabinfo_header(struct seq_file *m)
4535 seq_puts(m, "slabinfo - version: 2.1\n");
4536 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4537 "<objperslab> <pagesperslab>");
4538 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4539 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4543 static void *s_start(struct seq_file *m, loff_t *pos)
4547 down_read(&slub_lock);
4549 print_slabinfo_header(m);
4551 return seq_list_start(&slab_caches, *pos);
4554 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4556 return seq_list_next(p, &slab_caches, pos);
4559 static void s_stop(struct seq_file *m, void *p)
4561 up_read(&slub_lock);
4564 static int s_show(struct seq_file *m, void *p)
4566 unsigned long nr_partials = 0;
4567 unsigned long nr_slabs = 0;
4568 unsigned long nr_inuse = 0;
4569 unsigned long nr_objs = 0;
4570 unsigned long nr_free = 0;
4571 struct kmem_cache *s;
4574 s = list_entry(p, struct kmem_cache, list);
4576 for_each_online_node(node) {
4577 struct kmem_cache_node *n = get_node(s, node);
4582 nr_partials += n->nr_partial;
4583 nr_slabs += atomic_long_read(&n->nr_slabs);
4584 nr_objs += atomic_long_read(&n->total_objects);
4585 nr_free += count_partial(n, count_free);
4588 nr_inuse = nr_objs - nr_free;
4590 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4591 nr_objs, s->size, oo_objects(s->oo),
4592 (1 << oo_order(s->oo)));
4593 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4594 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4600 static const struct seq_operations slabinfo_op = {
4607 static int slabinfo_open(struct inode *inode, struct file *file)
4609 return seq_open(file, &slabinfo_op);
4612 static const struct file_operations proc_slabinfo_operations = {
4613 .open = slabinfo_open,
4615 .llseek = seq_lseek,
4616 .release = seq_release,
4619 static int __init slab_proc_init(void)
4621 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4624 module_init(slab_proc_init);
4625 #endif /* CONFIG_SLABINFO */