2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemtrace.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/kmemleak.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has noone operating on it and thus there is
68 * no danger of cacheline contention.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #ifdef CONFIG_SLUB_DEBUG
118 * Issues still to be resolved:
120 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
122 * - Variable sizing of the per node arrays
125 /* Enable to test recovery from slab corruption on boot */
126 #undef SLUB_RESILIENCY_TEST
129 * Mininum number of partial slabs. These will be left on the partial
130 * lists even if they are empty. kmem_cache_shrink may reclaim them.
132 #define MIN_PARTIAL 5
135 * Maximum number of desirable partial slabs.
136 * The existence of more partial slabs makes kmem_cache_shrink
137 * sort the partial list by the number of objects in the.
139 #define MAX_PARTIAL 10
141 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
142 SLAB_POISON | SLAB_STORE_USER)
145 * Debugging flags that require metadata to be stored in the slab, up to
146 * DEBUG_SIZE in size.
148 #define DEBUG_SIZE_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
149 #define DEBUG_SIZE (3 * sizeof(void *) + 2 * sizeof(struct track))
152 * Set of flags that will prevent slab merging
154 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
155 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
157 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
158 SLAB_CACHE_DMA | SLAB_NOTRACK)
160 #ifndef ARCH_KMALLOC_MINALIGN
161 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
164 #ifndef ARCH_SLAB_MINALIGN
165 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
169 #define OO_MASK ((1 << OO_SHIFT) - 1)
170 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
172 /* Internal SLUB flags */
173 #define __OBJECT_POISON 0x80000000 /* Poison object */
174 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
176 static int kmem_size = sizeof(struct kmem_cache);
179 static struct notifier_block slab_notifier;
183 DOWN, /* No slab functionality available */
184 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
185 UP, /* Everything works but does not show up in sysfs */
189 /* A list of all slab caches on the system */
190 static DECLARE_RWSEM(slub_lock);
191 static LIST_HEAD(slab_caches);
194 * Tracking user of a slab.
197 unsigned long addr; /* Called from address */
198 int cpu; /* Was running on cpu */
199 int pid; /* Pid context */
200 unsigned long when; /* When did the operation occur */
203 enum track_item { TRACK_ALLOC, TRACK_FREE };
205 #ifdef CONFIG_SLUB_DEBUG
206 static int sysfs_slab_add(struct kmem_cache *);
207 static int sysfs_slab_alias(struct kmem_cache *, const char *);
208 static void sysfs_slab_remove(struct kmem_cache *);
211 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
212 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
214 static inline void sysfs_slab_remove(struct kmem_cache *s)
221 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
228 /********************************************************************
229 * Core slab cache functions
230 *******************************************************************/
232 int slab_is_available(void)
234 return slab_state >= UP;
237 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
240 return s->node[node];
242 return &s->local_node;
246 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
249 return s->cpu_slab[cpu];
255 /* Verify that a pointer has an address that is valid within a slab page */
256 static inline int check_valid_pointer(struct kmem_cache *s,
257 struct page *page, const void *object)
264 base = page_address(page);
265 if (object < base || object >= base + page->objects * s->size ||
266 (object - base) % s->size) {
274 * Slow version of get and set free pointer.
276 * This version requires touching the cache lines of kmem_cache which
277 * we avoid to do in the fast alloc free paths. There we obtain the offset
278 * from the page struct.
280 static inline void *get_freepointer(struct kmem_cache *s, void *object)
282 return *(void **)(object + s->offset);
285 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
287 *(void **)(object + s->offset) = fp;
290 /* Loop over all objects in a slab */
291 #define for_each_object(__p, __s, __addr, __objects) \
292 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
296 #define for_each_free_object(__p, __s, __free) \
297 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
299 /* Determine object index from a given position */
300 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
302 return (p - addr) / s->size;
305 static inline struct kmem_cache_order_objects oo_make(int order,
308 struct kmem_cache_order_objects x = {
309 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
315 static inline int oo_order(struct kmem_cache_order_objects x)
317 return x.x >> OO_SHIFT;
320 static inline int oo_objects(struct kmem_cache_order_objects x)
322 return x.x & OO_MASK;
325 #ifdef CONFIG_SLUB_DEBUG
329 #ifdef CONFIG_SLUB_DEBUG_ON
330 static int slub_debug = DEBUG_DEFAULT_FLAGS;
332 static int slub_debug;
335 static char *slub_debug_slabs;
336 static int disable_higher_order_debug;
341 static void print_section(char *text, u8 *addr, unsigned int length)
349 for (i = 0; i < length; i++) {
351 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
354 printk(KERN_CONT " %02x", addr[i]);
356 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
358 printk(KERN_CONT " %s\n", ascii);
365 printk(KERN_CONT " ");
369 printk(KERN_CONT " %s\n", ascii);
373 static struct track *get_track(struct kmem_cache *s, void *object,
374 enum track_item alloc)
379 p = object + s->offset + sizeof(void *);
381 p = object + s->inuse;
386 static void set_track(struct kmem_cache *s, void *object,
387 enum track_item alloc, unsigned long addr)
389 struct track *p = get_track(s, object, alloc);
393 p->cpu = smp_processor_id();
394 p->pid = current->pid;
397 memset(p, 0, sizeof(struct track));
400 static void init_tracking(struct kmem_cache *s, void *object)
402 if (!(s->flags & SLAB_STORE_USER))
405 set_track(s, object, TRACK_FREE, 0UL);
406 set_track(s, object, TRACK_ALLOC, 0UL);
409 static void print_track(const char *s, struct track *t)
414 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
415 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
418 static void print_tracking(struct kmem_cache *s, void *object)
420 if (!(s->flags & SLAB_STORE_USER))
423 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
424 print_track("Freed", get_track(s, object, TRACK_FREE));
427 static void print_page_info(struct page *page)
429 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
430 page, page->objects, page->inuse, page->freelist, page->flags);
434 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
440 vsnprintf(buf, sizeof(buf), fmt, args);
442 printk(KERN_ERR "========================================"
443 "=====================================\n");
444 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
445 printk(KERN_ERR "----------------------------------------"
446 "-------------------------------------\n\n");
449 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
455 vsnprintf(buf, sizeof(buf), fmt, args);
457 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
460 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
462 unsigned int off; /* Offset of last byte */
463 u8 *addr = page_address(page);
465 print_tracking(s, p);
467 print_page_info(page);
469 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
470 p, p - addr, get_freepointer(s, p));
473 print_section("Bytes b4", p - 16, 16);
475 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
477 if (s->flags & SLAB_RED_ZONE)
478 print_section("Redzone", p + s->objsize,
479 s->inuse - s->objsize);
482 off = s->offset + sizeof(void *);
486 if (s->flags & SLAB_STORE_USER)
487 off += 2 * sizeof(struct track);
490 /* Beginning of the filler is the free pointer */
491 print_section("Padding", p + off, s->size - off);
496 static void object_err(struct kmem_cache *s, struct page *page,
497 u8 *object, char *reason)
499 slab_bug(s, "%s", reason);
500 print_trailer(s, page, object);
503 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
509 vsnprintf(buf, sizeof(buf), fmt, args);
511 slab_bug(s, "%s", buf);
512 print_page_info(page);
516 static void init_object(struct kmem_cache *s, void *object, int active)
520 if (s->flags & __OBJECT_POISON) {
521 memset(p, POISON_FREE, s->objsize - 1);
522 p[s->objsize - 1] = POISON_END;
525 if (s->flags & SLAB_RED_ZONE)
526 memset(p + s->objsize,
527 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
528 s->inuse - s->objsize);
531 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
534 if (*start != (u8)value)
542 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
543 void *from, void *to)
545 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
546 memset(from, data, to - from);
549 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
550 u8 *object, char *what,
551 u8 *start, unsigned int value, unsigned int bytes)
556 fault = check_bytes(start, value, bytes);
561 while (end > fault && end[-1] == value)
564 slab_bug(s, "%s overwritten", what);
565 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
566 fault, end - 1, fault[0], value);
567 print_trailer(s, page, object);
569 restore_bytes(s, what, value, fault, end);
577 * Bytes of the object to be managed.
578 * If the freepointer may overlay the object then the free
579 * pointer is the first word of the object.
581 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
584 * object + s->objsize
585 * Padding to reach word boundary. This is also used for Redzoning.
586 * Padding is extended by another word if Redzoning is enabled and
589 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
590 * 0xcc (RED_ACTIVE) for objects in use.
593 * Meta data starts here.
595 * A. Free pointer (if we cannot overwrite object on free)
596 * B. Tracking data for SLAB_STORE_USER
597 * C. Padding to reach required alignment boundary or at mininum
598 * one word if debugging is on to be able to detect writes
599 * before the word boundary.
601 * Padding is done using 0x5a (POISON_INUSE)
604 * Nothing is used beyond s->size.
606 * If slabcaches are merged then the objsize and inuse boundaries are mostly
607 * ignored. And therefore no slab options that rely on these boundaries
608 * may be used with merged slabcaches.
611 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
613 unsigned long off = s->inuse; /* The end of info */
616 /* Freepointer is placed after the object. */
617 off += sizeof(void *);
619 if (s->flags & SLAB_STORE_USER)
620 /* We also have user information there */
621 off += 2 * sizeof(struct track);
626 return check_bytes_and_report(s, page, p, "Object padding",
627 p + off, POISON_INUSE, s->size - off);
630 /* Check the pad bytes at the end of a slab page */
631 static int slab_pad_check(struct kmem_cache *s, struct page *page)
639 if (!(s->flags & SLAB_POISON))
642 start = page_address(page);
643 length = (PAGE_SIZE << compound_order(page));
644 end = start + length;
645 remainder = length % s->size;
649 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
652 while (end > fault && end[-1] == POISON_INUSE)
655 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
656 print_section("Padding", end - remainder, remainder);
658 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
662 static int check_object(struct kmem_cache *s, struct page *page,
663 void *object, int active)
666 u8 *endobject = object + s->objsize;
668 if (s->flags & SLAB_RED_ZONE) {
670 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
672 if (!check_bytes_and_report(s, page, object, "Redzone",
673 endobject, red, s->inuse - s->objsize))
676 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
677 check_bytes_and_report(s, page, p, "Alignment padding",
678 endobject, POISON_INUSE, s->inuse - s->objsize);
682 if (s->flags & SLAB_POISON) {
683 if (!active && (s->flags & __OBJECT_POISON) &&
684 (!check_bytes_and_report(s, page, p, "Poison", p,
685 POISON_FREE, s->objsize - 1) ||
686 !check_bytes_and_report(s, page, p, "Poison",
687 p + s->objsize - 1, POISON_END, 1)))
690 * check_pad_bytes cleans up on its own.
692 check_pad_bytes(s, page, p);
695 if (!s->offset && active)
697 * Object and freepointer overlap. Cannot check
698 * freepointer while object is allocated.
702 /* Check free pointer validity */
703 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
704 object_err(s, page, p, "Freepointer corrupt");
706 * No choice but to zap it and thus lose the remainder
707 * of the free objects in this slab. May cause
708 * another error because the object count is now wrong.
710 set_freepointer(s, p, NULL);
716 static int check_slab(struct kmem_cache *s, struct page *page)
720 VM_BUG_ON(!irqs_disabled());
722 if (!PageSlab(page)) {
723 slab_err(s, page, "Not a valid slab page");
727 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
728 if (page->objects > maxobj) {
729 slab_err(s, page, "objects %u > max %u",
730 s->name, page->objects, maxobj);
733 if (page->inuse > page->objects) {
734 slab_err(s, page, "inuse %u > max %u",
735 s->name, page->inuse, page->objects);
738 /* Slab_pad_check fixes things up after itself */
739 slab_pad_check(s, page);
744 * Determine if a certain object on a page is on the freelist. Must hold the
745 * slab lock to guarantee that the chains are in a consistent state.
747 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
750 void *fp = page->freelist;
752 unsigned long max_objects;
754 while (fp && nr <= page->objects) {
757 if (!check_valid_pointer(s, page, fp)) {
759 object_err(s, page, object,
760 "Freechain corrupt");
761 set_freepointer(s, object, NULL);
764 slab_err(s, page, "Freepointer corrupt");
765 page->freelist = NULL;
766 page->inuse = page->objects;
767 slab_fix(s, "Freelist cleared");
773 fp = get_freepointer(s, object);
777 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
778 if (max_objects > MAX_OBJS_PER_PAGE)
779 max_objects = MAX_OBJS_PER_PAGE;
781 if (page->objects != max_objects) {
782 slab_err(s, page, "Wrong number of objects. Found %d but "
783 "should be %d", page->objects, max_objects);
784 page->objects = max_objects;
785 slab_fix(s, "Number of objects adjusted.");
787 if (page->inuse != page->objects - nr) {
788 slab_err(s, page, "Wrong object count. Counter is %d but "
789 "counted were %d", page->inuse, page->objects - nr);
790 page->inuse = page->objects - nr;
791 slab_fix(s, "Object count adjusted.");
793 return search == NULL;
796 static void trace(struct kmem_cache *s, struct page *page, void *object,
799 if (s->flags & SLAB_TRACE) {
800 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
802 alloc ? "alloc" : "free",
807 print_section("Object", (void *)object, s->objsize);
814 * Tracking of fully allocated slabs for debugging purposes.
816 static void add_full(struct kmem_cache_node *n, struct page *page)
818 spin_lock(&n->list_lock);
819 list_add(&page->lru, &n->full);
820 spin_unlock(&n->list_lock);
823 static void remove_full(struct kmem_cache *s, struct page *page)
825 struct kmem_cache_node *n;
827 if (!(s->flags & SLAB_STORE_USER))
830 n = get_node(s, page_to_nid(page));
832 spin_lock(&n->list_lock);
833 list_del(&page->lru);
834 spin_unlock(&n->list_lock);
837 /* Tracking of the number of slabs for debugging purposes */
838 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
840 struct kmem_cache_node *n = get_node(s, node);
842 return atomic_long_read(&n->nr_slabs);
845 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
847 return atomic_long_read(&n->nr_slabs);
850 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
852 struct kmem_cache_node *n = get_node(s, node);
855 * May be called early in order to allocate a slab for the
856 * kmem_cache_node structure. Solve the chicken-egg
857 * dilemma by deferring the increment of the count during
858 * bootstrap (see early_kmem_cache_node_alloc).
860 if (!NUMA_BUILD || n) {
861 atomic_long_inc(&n->nr_slabs);
862 atomic_long_add(objects, &n->total_objects);
865 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
867 struct kmem_cache_node *n = get_node(s, node);
869 atomic_long_dec(&n->nr_slabs);
870 atomic_long_sub(objects, &n->total_objects);
873 /* Object debug checks for alloc/free paths */
874 static void setup_object_debug(struct kmem_cache *s, struct page *page,
877 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
880 init_object(s, object, 0);
881 init_tracking(s, object);
884 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
885 void *object, unsigned long addr)
887 if (!check_slab(s, page))
890 if (!on_freelist(s, page, object)) {
891 object_err(s, page, object, "Object already allocated");
895 if (!check_valid_pointer(s, page, object)) {
896 object_err(s, page, object, "Freelist Pointer check fails");
900 if (!check_object(s, page, object, 0))
903 /* Success perform special debug activities for allocs */
904 if (s->flags & SLAB_STORE_USER)
905 set_track(s, object, TRACK_ALLOC, addr);
906 trace(s, page, object, 1);
907 init_object(s, object, 1);
911 if (PageSlab(page)) {
913 * If this is a slab page then lets do the best we can
914 * to avoid issues in the future. Marking all objects
915 * as used avoids touching the remaining objects.
917 slab_fix(s, "Marking all objects used");
918 page->inuse = page->objects;
919 page->freelist = NULL;
924 static int free_debug_processing(struct kmem_cache *s, struct page *page,
925 void *object, unsigned long addr)
927 if (!check_slab(s, page))
930 if (!check_valid_pointer(s, page, object)) {
931 slab_err(s, page, "Invalid object pointer 0x%p", object);
935 if (on_freelist(s, page, object)) {
936 object_err(s, page, object, "Object already free");
940 if (!check_object(s, page, object, 1))
943 if (unlikely(s != page->slab)) {
944 if (!PageSlab(page)) {
945 slab_err(s, page, "Attempt to free object(0x%p) "
946 "outside of slab", object);
947 } else if (!page->slab) {
949 "SLUB <none>: no slab for object 0x%p.\n",
953 object_err(s, page, object,
954 "page slab pointer corrupt.");
958 /* Special debug activities for freeing objects */
959 if (!PageSlubFrozen(page) && !page->freelist)
960 remove_full(s, page);
961 if (s->flags & SLAB_STORE_USER)
962 set_track(s, object, TRACK_FREE, addr);
963 trace(s, page, object, 0);
964 init_object(s, object, 0);
968 slab_fix(s, "Object at 0x%p not freed", object);
972 static int __init setup_slub_debug(char *str)
974 slub_debug = DEBUG_DEFAULT_FLAGS;
975 if (*str++ != '=' || !*str)
977 * No options specified. Switch on full debugging.
983 * No options but restriction on slabs. This means full
984 * debugging for slabs matching a pattern.
988 if (tolower(*str) == 'o') {
990 * Avoid enabling debugging on caches if its minimum order
991 * would increase as a result.
993 disable_higher_order_debug = 1;
1000 * Switch off all debugging measures.
1005 * Determine which debug features should be switched on
1007 for (; *str && *str != ','; str++) {
1008 switch (tolower(*str)) {
1010 slub_debug |= SLAB_DEBUG_FREE;
1013 slub_debug |= SLAB_RED_ZONE;
1016 slub_debug |= SLAB_POISON;
1019 slub_debug |= SLAB_STORE_USER;
1022 slub_debug |= SLAB_TRACE;
1025 printk(KERN_ERR "slub_debug option '%c' "
1026 "unknown. skipped\n", *str);
1032 slub_debug_slabs = str + 1;
1037 __setup("slub_debug", setup_slub_debug);
1039 static unsigned long kmem_cache_flags(unsigned long objsize,
1040 unsigned long flags, const char *name,
1041 void (*ctor)(void *))
1043 int debug_flags = slub_debug;
1046 * Enable debugging if selected on the kernel commandline.
1049 if (slub_debug_slabs &&
1050 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))
1054 * Disable debugging that increases slab size if the minimum
1055 * slab order would have increased as a result.
1057 if (disable_higher_order_debug &&
1058 get_order(objsize + DEBUG_SIZE) > get_order(objsize))
1059 debug_flags &= ~DEBUG_SIZE_FLAGS;
1061 flags |= debug_flags;
1067 static inline void setup_object_debug(struct kmem_cache *s,
1068 struct page *page, void *object) {}
1070 static inline int alloc_debug_processing(struct kmem_cache *s,
1071 struct page *page, void *object, unsigned long addr) { return 0; }
1073 static inline int free_debug_processing(struct kmem_cache *s,
1074 struct page *page, void *object, unsigned long addr) { return 0; }
1076 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1078 static inline int check_object(struct kmem_cache *s, struct page *page,
1079 void *object, int active) { return 1; }
1080 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1081 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1082 unsigned long flags, const char *name,
1083 void (*ctor)(void *))
1087 #define slub_debug 0
1089 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1091 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1093 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1095 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1100 * Slab allocation and freeing
1102 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1103 struct kmem_cache_order_objects oo)
1105 int order = oo_order(oo);
1107 flags |= __GFP_NOTRACK;
1110 return alloc_pages(flags, order);
1112 return alloc_pages_node(node, flags, order);
1115 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1118 struct kmem_cache_order_objects oo = s->oo;
1121 flags |= s->allocflags;
1124 * Let the initial higher-order allocation fail under memory pressure
1125 * so we fall-back to the minimum order allocation.
1127 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1129 page = alloc_slab_page(alloc_gfp, node, oo);
1130 if (unlikely(!page)) {
1133 * Allocation may have failed due to fragmentation.
1134 * Try a lower order alloc if possible
1136 page = alloc_slab_page(flags, node, oo);
1140 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1143 if (kmemcheck_enabled
1144 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS)))
1146 int pages = 1 << oo_order(oo);
1148 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1151 * Objects from caches that have a constructor don't get
1152 * cleared when they're allocated, so we need to do it here.
1155 kmemcheck_mark_uninitialized_pages(page, pages);
1157 kmemcheck_mark_unallocated_pages(page, pages);
1160 page->objects = oo_objects(oo);
1161 mod_zone_page_state(page_zone(page),
1162 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1163 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1169 static void setup_object(struct kmem_cache *s, struct page *page,
1172 setup_object_debug(s, page, object);
1173 if (unlikely(s->ctor))
1177 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1184 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1186 page = allocate_slab(s,
1187 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1191 inc_slabs_node(s, page_to_nid(page), page->objects);
1193 page->flags |= 1 << PG_slab;
1194 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1195 SLAB_STORE_USER | SLAB_TRACE))
1196 __SetPageSlubDebug(page);
1198 start = page_address(page);
1200 if (unlikely(s->flags & SLAB_POISON))
1201 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1204 for_each_object(p, s, start, page->objects) {
1205 setup_object(s, page, last);
1206 set_freepointer(s, last, p);
1209 setup_object(s, page, last);
1210 set_freepointer(s, last, NULL);
1212 page->freelist = start;
1218 static void __free_slab(struct kmem_cache *s, struct page *page)
1220 int order = compound_order(page);
1221 int pages = 1 << order;
1223 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1226 slab_pad_check(s, page);
1227 for_each_object(p, s, page_address(page),
1229 check_object(s, page, p, 0);
1230 __ClearPageSlubDebug(page);
1233 kmemcheck_free_shadow(page, compound_order(page));
1235 mod_zone_page_state(page_zone(page),
1236 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1237 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1240 __ClearPageSlab(page);
1241 reset_page_mapcount(page);
1242 if (current->reclaim_state)
1243 current->reclaim_state->reclaimed_slab += pages;
1244 __free_pages(page, order);
1247 static void rcu_free_slab(struct rcu_head *h)
1251 page = container_of((struct list_head *)h, struct page, lru);
1252 __free_slab(page->slab, page);
1255 static void free_slab(struct kmem_cache *s, struct page *page)
1257 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1259 * RCU free overloads the RCU head over the LRU
1261 struct rcu_head *head = (void *)&page->lru;
1263 call_rcu(head, rcu_free_slab);
1265 __free_slab(s, page);
1268 static void discard_slab(struct kmem_cache *s, struct page *page)
1270 dec_slabs_node(s, page_to_nid(page), page->objects);
1275 * Per slab locking using the pagelock
1277 static __always_inline void slab_lock(struct page *page)
1279 bit_spin_lock(PG_locked, &page->flags);
1282 static __always_inline void slab_unlock(struct page *page)
1284 __bit_spin_unlock(PG_locked, &page->flags);
1287 static __always_inline int slab_trylock(struct page *page)
1291 rc = bit_spin_trylock(PG_locked, &page->flags);
1296 * Management of partially allocated slabs
1298 static void add_partial(struct kmem_cache_node *n,
1299 struct page *page, int tail)
1301 spin_lock(&n->list_lock);
1304 list_add_tail(&page->lru, &n->partial);
1306 list_add(&page->lru, &n->partial);
1307 spin_unlock(&n->list_lock);
1310 static void remove_partial(struct kmem_cache *s, struct page *page)
1312 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1314 spin_lock(&n->list_lock);
1315 list_del(&page->lru);
1317 spin_unlock(&n->list_lock);
1321 * Lock slab and remove from the partial list.
1323 * Must hold list_lock.
1325 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1328 if (slab_trylock(page)) {
1329 list_del(&page->lru);
1331 __SetPageSlubFrozen(page);
1338 * Try to allocate a partial slab from a specific node.
1340 static struct page *get_partial_node(struct kmem_cache_node *n)
1345 * Racy check. If we mistakenly see no partial slabs then we
1346 * just allocate an empty slab. If we mistakenly try to get a
1347 * partial slab and there is none available then get_partials()
1350 if (!n || !n->nr_partial)
1353 spin_lock(&n->list_lock);
1354 list_for_each_entry(page, &n->partial, lru)
1355 if (lock_and_freeze_slab(n, page))
1359 spin_unlock(&n->list_lock);
1364 * Get a page from somewhere. Search in increasing NUMA distances.
1366 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1369 struct zonelist *zonelist;
1372 enum zone_type high_zoneidx = gfp_zone(flags);
1376 * The defrag ratio allows a configuration of the tradeoffs between
1377 * inter node defragmentation and node local allocations. A lower
1378 * defrag_ratio increases the tendency to do local allocations
1379 * instead of attempting to obtain partial slabs from other nodes.
1381 * If the defrag_ratio is set to 0 then kmalloc() always
1382 * returns node local objects. If the ratio is higher then kmalloc()
1383 * may return off node objects because partial slabs are obtained
1384 * from other nodes and filled up.
1386 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1387 * defrag_ratio = 1000) then every (well almost) allocation will
1388 * first attempt to defrag slab caches on other nodes. This means
1389 * scanning over all nodes to look for partial slabs which may be
1390 * expensive if we do it every time we are trying to find a slab
1391 * with available objects.
1393 if (!s->remote_node_defrag_ratio ||
1394 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1397 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1398 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1399 struct kmem_cache_node *n;
1401 n = get_node(s, zone_to_nid(zone));
1403 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1404 n->nr_partial > s->min_partial) {
1405 page = get_partial_node(n);
1415 * Get a partial page, lock it and return it.
1417 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1420 int searchnode = (node == -1) ? numa_node_id() : node;
1422 page = get_partial_node(get_node(s, searchnode));
1423 if (page || (flags & __GFP_THISNODE))
1426 return get_any_partial(s, flags);
1430 * Move a page back to the lists.
1432 * Must be called with the slab lock held.
1434 * On exit the slab lock will have been dropped.
1436 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1438 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1439 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1441 __ClearPageSlubFrozen(page);
1444 if (page->freelist) {
1445 add_partial(n, page, tail);
1446 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1448 stat(c, DEACTIVATE_FULL);
1449 if (SLABDEBUG && PageSlubDebug(page) &&
1450 (s->flags & SLAB_STORE_USER))
1455 stat(c, DEACTIVATE_EMPTY);
1456 if (n->nr_partial < s->min_partial) {
1458 * Adding an empty slab to the partial slabs in order
1459 * to avoid page allocator overhead. This slab needs
1460 * to come after the other slabs with objects in
1461 * so that the others get filled first. That way the
1462 * size of the partial list stays small.
1464 * kmem_cache_shrink can reclaim any empty slabs from
1467 add_partial(n, page, 1);
1471 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1472 discard_slab(s, page);
1478 * Remove the cpu slab
1480 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1482 struct page *page = c->page;
1486 stat(c, DEACTIVATE_REMOTE_FREES);
1488 * Merge cpu freelist into slab freelist. Typically we get here
1489 * because both freelists are empty. So this is unlikely
1492 while (unlikely(c->freelist)) {
1495 tail = 0; /* Hot objects. Put the slab first */
1497 /* Retrieve object from cpu_freelist */
1498 object = c->freelist;
1499 c->freelist = c->freelist[c->offset];
1501 /* And put onto the regular freelist */
1502 object[c->offset] = page->freelist;
1503 page->freelist = object;
1507 unfreeze_slab(s, page, tail);
1510 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1512 stat(c, CPUSLAB_FLUSH);
1514 deactivate_slab(s, c);
1520 * Called from IPI handler with interrupts disabled.
1522 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1524 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1526 if (likely(c && c->page))
1530 static void flush_cpu_slab(void *d)
1532 struct kmem_cache *s = d;
1534 __flush_cpu_slab(s, smp_processor_id());
1537 static void flush_all(struct kmem_cache *s)
1539 on_each_cpu(flush_cpu_slab, s, 1);
1543 * Check if the objects in a per cpu structure fit numa
1544 * locality expectations.
1546 static inline int node_match(struct kmem_cache_cpu *c, int node)
1549 if (node != -1 && c->node != node)
1555 static int count_free(struct page *page)
1557 return page->objects - page->inuse;
1560 static unsigned long count_partial(struct kmem_cache_node *n,
1561 int (*get_count)(struct page *))
1563 unsigned long flags;
1564 unsigned long x = 0;
1567 spin_lock_irqsave(&n->list_lock, flags);
1568 list_for_each_entry(page, &n->partial, lru)
1569 x += get_count(page);
1570 spin_unlock_irqrestore(&n->list_lock, flags);
1574 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1576 #ifdef CONFIG_SLUB_DEBUG
1577 return atomic_long_read(&n->total_objects);
1583 static noinline void
1584 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1589 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1591 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1592 "default order: %d, min order: %d\n", s->name, s->objsize,
1593 s->size, oo_order(s->oo), oo_order(s->min));
1595 if (oo_order(s->min) > get_order(s->objsize))
1596 printk(KERN_WARNING " %s debugging increased min order, use "
1597 "slub_debug=O to disable.\n", s->name);
1599 for_each_online_node(node) {
1600 struct kmem_cache_node *n = get_node(s, node);
1601 unsigned long nr_slabs;
1602 unsigned long nr_objs;
1603 unsigned long nr_free;
1608 nr_free = count_partial(n, count_free);
1609 nr_slabs = node_nr_slabs(n);
1610 nr_objs = node_nr_objs(n);
1613 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1614 node, nr_slabs, nr_objs, nr_free);
1619 * Slow path. The lockless freelist is empty or we need to perform
1622 * Interrupts are disabled.
1624 * Processing is still very fast if new objects have been freed to the
1625 * regular freelist. In that case we simply take over the regular freelist
1626 * as the lockless freelist and zap the regular freelist.
1628 * If that is not working then we fall back to the partial lists. We take the
1629 * first element of the freelist as the object to allocate now and move the
1630 * rest of the freelist to the lockless freelist.
1632 * And if we were unable to get a new slab from the partial slab lists then
1633 * we need to allocate a new slab. This is the slowest path since it involves
1634 * a call to the page allocator and the setup of a new slab.
1636 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1637 unsigned long addr, struct kmem_cache_cpu *c)
1642 /* We handle __GFP_ZERO in the caller */
1643 gfpflags &= ~__GFP_ZERO;
1649 if (unlikely(!node_match(c, node)))
1652 stat(c, ALLOC_REFILL);
1655 object = c->page->freelist;
1656 if (unlikely(!object))
1658 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1661 c->freelist = object[c->offset];
1662 c->page->inuse = c->page->objects;
1663 c->page->freelist = NULL;
1664 c->node = page_to_nid(c->page);
1666 slab_unlock(c->page);
1667 stat(c, ALLOC_SLOWPATH);
1671 deactivate_slab(s, c);
1674 new = get_partial(s, gfpflags, node);
1677 stat(c, ALLOC_FROM_PARTIAL);
1681 if (gfpflags & __GFP_WAIT)
1684 new = new_slab(s, gfpflags, node);
1686 if (gfpflags & __GFP_WAIT)
1687 local_irq_disable();
1690 c = get_cpu_slab(s, smp_processor_id());
1691 stat(c, ALLOC_SLAB);
1695 __SetPageSlubFrozen(new);
1699 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1700 slab_out_of_memory(s, gfpflags, node);
1703 if (!alloc_debug_processing(s, c->page, object, addr))
1707 c->page->freelist = object[c->offset];
1713 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1714 * have the fastpath folded into their functions. So no function call
1715 * overhead for requests that can be satisfied on the fastpath.
1717 * The fastpath works by first checking if the lockless freelist can be used.
1718 * If not then __slab_alloc is called for slow processing.
1720 * Otherwise we can simply pick the next object from the lockless free list.
1722 static __always_inline void *slab_alloc(struct kmem_cache *s,
1723 gfp_t gfpflags, int node, unsigned long addr)
1726 struct kmem_cache_cpu *c;
1727 unsigned long flags;
1728 unsigned int objsize;
1730 gfpflags &= gfp_allowed_mask;
1732 lockdep_trace_alloc(gfpflags);
1733 might_sleep_if(gfpflags & __GFP_WAIT);
1735 if (should_failslab(s->objsize, gfpflags))
1738 local_irq_save(flags);
1739 c = get_cpu_slab(s, smp_processor_id());
1740 objsize = c->objsize;
1741 if (unlikely(!c->freelist || !node_match(c, node)))
1743 object = __slab_alloc(s, gfpflags, node, addr, c);
1746 object = c->freelist;
1747 c->freelist = object[c->offset];
1748 stat(c, ALLOC_FASTPATH);
1750 local_irq_restore(flags);
1752 if (unlikely((gfpflags & __GFP_ZERO) && object))
1753 memset(object, 0, objsize);
1755 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize);
1756 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
1761 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1763 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1765 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1769 EXPORT_SYMBOL(kmem_cache_alloc);
1771 #ifdef CONFIG_KMEMTRACE
1772 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1774 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1776 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1780 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1782 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1784 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1785 s->objsize, s->size, gfpflags, node);
1789 EXPORT_SYMBOL(kmem_cache_alloc_node);
1792 #ifdef CONFIG_KMEMTRACE
1793 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1797 return slab_alloc(s, gfpflags, node, _RET_IP_);
1799 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1803 * Slow patch handling. This may still be called frequently since objects
1804 * have a longer lifetime than the cpu slabs in most processing loads.
1806 * So we still attempt to reduce cache line usage. Just take the slab
1807 * lock and free the item. If there is no additional partial page
1808 * handling required then we can return immediately.
1810 static void __slab_free(struct kmem_cache *s, struct page *page,
1811 void *x, unsigned long addr, unsigned int offset)
1814 void **object = (void *)x;
1815 struct kmem_cache_cpu *c;
1817 c = get_cpu_slab(s, raw_smp_processor_id());
1818 stat(c, FREE_SLOWPATH);
1821 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1825 prior = object[offset] = page->freelist;
1826 page->freelist = object;
1829 if (unlikely(PageSlubFrozen(page))) {
1830 stat(c, FREE_FROZEN);
1834 if (unlikely(!page->inuse))
1838 * Objects left in the slab. If it was not on the partial list before
1841 if (unlikely(!prior)) {
1842 add_partial(get_node(s, page_to_nid(page)), page, 1);
1843 stat(c, FREE_ADD_PARTIAL);
1853 * Slab still on the partial list.
1855 remove_partial(s, page);
1856 stat(c, FREE_REMOVE_PARTIAL);
1860 discard_slab(s, page);
1864 if (!free_debug_processing(s, page, x, addr))
1870 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1871 * can perform fastpath freeing without additional function calls.
1873 * The fastpath is only possible if we are freeing to the current cpu slab
1874 * of this processor. This typically the case if we have just allocated
1877 * If fastpath is not possible then fall back to __slab_free where we deal
1878 * with all sorts of special processing.
1880 static __always_inline void slab_free(struct kmem_cache *s,
1881 struct page *page, void *x, unsigned long addr)
1883 void **object = (void *)x;
1884 struct kmem_cache_cpu *c;
1885 unsigned long flags;
1887 kmemleak_free_recursive(x, s->flags);
1888 local_irq_save(flags);
1889 c = get_cpu_slab(s, smp_processor_id());
1890 kmemcheck_slab_free(s, object, c->objsize);
1891 debug_check_no_locks_freed(object, c->objsize);
1892 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1893 debug_check_no_obj_freed(object, c->objsize);
1894 if (likely(page == c->page && c->node >= 0)) {
1895 object[c->offset] = c->freelist;
1896 c->freelist = object;
1897 stat(c, FREE_FASTPATH);
1899 __slab_free(s, page, x, addr, c->offset);
1901 local_irq_restore(flags);
1904 void kmem_cache_free(struct kmem_cache *s, void *x)
1908 page = virt_to_head_page(x);
1910 slab_free(s, page, x, _RET_IP_);
1912 trace_kmem_cache_free(_RET_IP_, x);
1914 EXPORT_SYMBOL(kmem_cache_free);
1916 /* Figure out on which slab page the object resides */
1917 static struct page *get_object_page(const void *x)
1919 struct page *page = virt_to_head_page(x);
1921 if (!PageSlab(page))
1928 * Object placement in a slab is made very easy because we always start at
1929 * offset 0. If we tune the size of the object to the alignment then we can
1930 * get the required alignment by putting one properly sized object after
1933 * Notice that the allocation order determines the sizes of the per cpu
1934 * caches. Each processor has always one slab available for allocations.
1935 * Increasing the allocation order reduces the number of times that slabs
1936 * must be moved on and off the partial lists and is therefore a factor in
1941 * Mininum / Maximum order of slab pages. This influences locking overhead
1942 * and slab fragmentation. A higher order reduces the number of partial slabs
1943 * and increases the number of allocations possible without having to
1944 * take the list_lock.
1946 static int slub_min_order;
1947 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1948 static int slub_min_objects;
1951 * Merge control. If this is set then no merging of slab caches will occur.
1952 * (Could be removed. This was introduced to pacify the merge skeptics.)
1954 static int slub_nomerge;
1957 * Calculate the order of allocation given an slab object size.
1959 * The order of allocation has significant impact on performance and other
1960 * system components. Generally order 0 allocations should be preferred since
1961 * order 0 does not cause fragmentation in the page allocator. Larger objects
1962 * be problematic to put into order 0 slabs because there may be too much
1963 * unused space left. We go to a higher order if more than 1/16th of the slab
1966 * In order to reach satisfactory performance we must ensure that a minimum
1967 * number of objects is in one slab. Otherwise we may generate too much
1968 * activity on the partial lists which requires taking the list_lock. This is
1969 * less a concern for large slabs though which are rarely used.
1971 * slub_max_order specifies the order where we begin to stop considering the
1972 * number of objects in a slab as critical. If we reach slub_max_order then
1973 * we try to keep the page order as low as possible. So we accept more waste
1974 * of space in favor of a small page order.
1976 * Higher order allocations also allow the placement of more objects in a
1977 * slab and thereby reduce object handling overhead. If the user has
1978 * requested a higher mininum order then we start with that one instead of
1979 * the smallest order which will fit the object.
1981 static inline int slab_order(int size, int min_objects,
1982 int max_order, int fract_leftover)
1986 int min_order = slub_min_order;
1988 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1989 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1991 for (order = max(min_order,
1992 fls(min_objects * size - 1) - PAGE_SHIFT);
1993 order <= max_order; order++) {
1995 unsigned long slab_size = PAGE_SIZE << order;
1997 if (slab_size < min_objects * size)
2000 rem = slab_size % size;
2002 if (rem <= slab_size / fract_leftover)
2010 static inline int calculate_order(int size)
2018 * Attempt to find best configuration for a slab. This
2019 * works by first attempting to generate a layout with
2020 * the best configuration and backing off gradually.
2022 * First we reduce the acceptable waste in a slab. Then
2023 * we reduce the minimum objects required in a slab.
2025 min_objects = slub_min_objects;
2027 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2028 max_objects = (PAGE_SIZE << slub_max_order)/size;
2029 min_objects = min(min_objects, max_objects);
2031 while (min_objects > 1) {
2033 while (fraction >= 4) {
2034 order = slab_order(size, min_objects,
2035 slub_max_order, fraction);
2036 if (order <= slub_max_order)
2044 * We were unable to place multiple objects in a slab. Now
2045 * lets see if we can place a single object there.
2047 order = slab_order(size, 1, slub_max_order, 1);
2048 if (order <= slub_max_order)
2052 * Doh this slab cannot be placed using slub_max_order.
2054 order = slab_order(size, 1, MAX_ORDER, 1);
2055 if (order < MAX_ORDER)
2061 * Figure out what the alignment of the objects will be.
2063 static unsigned long calculate_alignment(unsigned long flags,
2064 unsigned long align, unsigned long size)
2067 * If the user wants hardware cache aligned objects then follow that
2068 * suggestion if the object is sufficiently large.
2070 * The hardware cache alignment cannot override the specified
2071 * alignment though. If that is greater then use it.
2073 if (flags & SLAB_HWCACHE_ALIGN) {
2074 unsigned long ralign = cache_line_size();
2075 while (size <= ralign / 2)
2077 align = max(align, ralign);
2080 if (align < ARCH_SLAB_MINALIGN)
2081 align = ARCH_SLAB_MINALIGN;
2083 return ALIGN(align, sizeof(void *));
2086 static void init_kmem_cache_cpu(struct kmem_cache *s,
2087 struct kmem_cache_cpu *c)
2092 c->offset = s->offset / sizeof(void *);
2093 c->objsize = s->objsize;
2094 #ifdef CONFIG_SLUB_STATS
2095 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
2100 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2103 spin_lock_init(&n->list_lock);
2104 INIT_LIST_HEAD(&n->partial);
2105 #ifdef CONFIG_SLUB_DEBUG
2106 atomic_long_set(&n->nr_slabs, 0);
2107 atomic_long_set(&n->total_objects, 0);
2108 INIT_LIST_HEAD(&n->full);
2114 * Per cpu array for per cpu structures.
2116 * The per cpu array places all kmem_cache_cpu structures from one processor
2117 * close together meaning that it becomes possible that multiple per cpu
2118 * structures are contained in one cacheline. This may be particularly
2119 * beneficial for the kmalloc caches.
2121 * A desktop system typically has around 60-80 slabs. With 100 here we are
2122 * likely able to get per cpu structures for all caches from the array defined
2123 * here. We must be able to cover all kmalloc caches during bootstrap.
2125 * If the per cpu array is exhausted then fall back to kmalloc
2126 * of individual cachelines. No sharing is possible then.
2128 #define NR_KMEM_CACHE_CPU 100
2130 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2131 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2133 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2134 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
2136 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2137 int cpu, gfp_t flags)
2139 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2142 per_cpu(kmem_cache_cpu_free, cpu) =
2143 (void *)c->freelist;
2145 /* Table overflow: So allocate ourselves */
2147 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2148 flags, cpu_to_node(cpu));
2153 init_kmem_cache_cpu(s, c);
2157 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2159 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2160 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2164 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2165 per_cpu(kmem_cache_cpu_free, cpu) = c;
2168 static void free_kmem_cache_cpus(struct kmem_cache *s)
2172 for_each_online_cpu(cpu) {
2173 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2176 s->cpu_slab[cpu] = NULL;
2177 free_kmem_cache_cpu(c, cpu);
2182 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2186 for_each_online_cpu(cpu) {
2187 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2192 c = alloc_kmem_cache_cpu(s, cpu, flags);
2194 free_kmem_cache_cpus(s);
2197 s->cpu_slab[cpu] = c;
2203 * Initialize the per cpu array.
2205 static void init_alloc_cpu_cpu(int cpu)
2209 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2212 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2213 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2215 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2218 static void __init init_alloc_cpu(void)
2222 for_each_online_cpu(cpu)
2223 init_alloc_cpu_cpu(cpu);
2227 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2228 static inline void init_alloc_cpu(void) {}
2230 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2232 init_kmem_cache_cpu(s, &s->cpu_slab);
2239 * No kmalloc_node yet so do it by hand. We know that this is the first
2240 * slab on the node for this slabcache. There are no concurrent accesses
2243 * Note that this function only works on the kmalloc_node_cache
2244 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2245 * memory on a fresh node that has no slab structures yet.
2247 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2250 struct kmem_cache_node *n;
2251 unsigned long flags;
2253 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2255 page = new_slab(kmalloc_caches, gfpflags, node);
2258 if (page_to_nid(page) != node) {
2259 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2261 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2262 "in order to be able to continue\n");
2267 page->freelist = get_freepointer(kmalloc_caches, n);
2269 kmalloc_caches->node[node] = n;
2270 #ifdef CONFIG_SLUB_DEBUG
2271 init_object(kmalloc_caches, n, 1);
2272 init_tracking(kmalloc_caches, n);
2274 init_kmem_cache_node(n, kmalloc_caches);
2275 inc_slabs_node(kmalloc_caches, node, page->objects);
2278 * lockdep requires consistent irq usage for each lock
2279 * so even though there cannot be a race this early in
2280 * the boot sequence, we still disable irqs.
2282 local_irq_save(flags);
2283 add_partial(n, page, 0);
2284 local_irq_restore(flags);
2287 static void free_kmem_cache_nodes(struct kmem_cache *s)
2291 for_each_node_state(node, N_NORMAL_MEMORY) {
2292 struct kmem_cache_node *n = s->node[node];
2293 if (n && n != &s->local_node)
2294 kmem_cache_free(kmalloc_caches, n);
2295 s->node[node] = NULL;
2299 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2304 if (slab_state >= UP)
2305 local_node = page_to_nid(virt_to_page(s));
2309 for_each_node_state(node, N_NORMAL_MEMORY) {
2310 struct kmem_cache_node *n;
2312 if (local_node == node)
2315 if (slab_state == DOWN) {
2316 early_kmem_cache_node_alloc(gfpflags, node);
2319 n = kmem_cache_alloc_node(kmalloc_caches,
2323 free_kmem_cache_nodes(s);
2329 init_kmem_cache_node(n, s);
2334 static void free_kmem_cache_nodes(struct kmem_cache *s)
2338 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2340 init_kmem_cache_node(&s->local_node, s);
2345 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2347 if (min < MIN_PARTIAL)
2349 else if (min > MAX_PARTIAL)
2351 s->min_partial = min;
2355 * calculate_sizes() determines the order and the distribution of data within
2358 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2360 unsigned long flags = s->flags;
2361 unsigned long size = s->objsize;
2362 unsigned long align = s->align;
2366 * Round up object size to the next word boundary. We can only
2367 * place the free pointer at word boundaries and this determines
2368 * the possible location of the free pointer.
2370 size = ALIGN(size, sizeof(void *));
2372 #ifdef CONFIG_SLUB_DEBUG
2374 * Determine if we can poison the object itself. If the user of
2375 * the slab may touch the object after free or before allocation
2376 * then we should never poison the object itself.
2378 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2380 s->flags |= __OBJECT_POISON;
2382 s->flags &= ~__OBJECT_POISON;
2386 * If we are Redzoning then check if there is some space between the
2387 * end of the object and the free pointer. If not then add an
2388 * additional word to have some bytes to store Redzone information.
2390 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2391 size += sizeof(void *);
2395 * With that we have determined the number of bytes in actual use
2396 * by the object. This is the potential offset to the free pointer.
2400 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2403 * Relocate free pointer after the object if it is not
2404 * permitted to overwrite the first word of the object on
2407 * This is the case if we do RCU, have a constructor or
2408 * destructor or are poisoning the objects.
2411 size += sizeof(void *);
2414 #ifdef CONFIG_SLUB_DEBUG
2415 if (flags & SLAB_STORE_USER)
2417 * Need to store information about allocs and frees after
2420 size += 2 * sizeof(struct track);
2422 if (flags & SLAB_RED_ZONE)
2424 * Add some empty padding so that we can catch
2425 * overwrites from earlier objects rather than let
2426 * tracking information or the free pointer be
2427 * corrupted if a user writes before the start
2430 size += sizeof(void *);
2434 * Determine the alignment based on various parameters that the
2435 * user specified and the dynamic determination of cache line size
2438 align = calculate_alignment(flags, align, s->objsize);
2441 * SLUB stores one object immediately after another beginning from
2442 * offset 0. In order to align the objects we have to simply size
2443 * each object to conform to the alignment.
2445 size = ALIGN(size, align);
2447 if (forced_order >= 0)
2448 order = forced_order;
2450 order = calculate_order(size);
2457 s->allocflags |= __GFP_COMP;
2459 if (s->flags & SLAB_CACHE_DMA)
2460 s->allocflags |= SLUB_DMA;
2462 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2463 s->allocflags |= __GFP_RECLAIMABLE;
2466 * Determine the number of objects per slab
2468 s->oo = oo_make(order, size);
2469 s->min = oo_make(get_order(size), size);
2470 if (oo_objects(s->oo) > oo_objects(s->max))
2473 return !!oo_objects(s->oo);
2477 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2478 const char *name, size_t size,
2479 size_t align, unsigned long flags,
2480 void (*ctor)(void *))
2482 memset(s, 0, kmem_size);
2487 s->flags = kmem_cache_flags(size, flags, name, ctor);
2489 if (!calculate_sizes(s, -1))
2493 * The larger the object size is, the more pages we want on the partial
2494 * list to avoid pounding the page allocator excessively.
2496 set_min_partial(s, ilog2(s->size));
2499 s->remote_node_defrag_ratio = 1000;
2501 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2504 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2506 free_kmem_cache_nodes(s);
2508 if (flags & SLAB_PANIC)
2509 panic("Cannot create slab %s size=%lu realsize=%u "
2510 "order=%u offset=%u flags=%lx\n",
2511 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2517 * Check if a given pointer is valid
2519 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2523 page = get_object_page(object);
2525 if (!page || s != page->slab)
2526 /* No slab or wrong slab */
2529 if (!check_valid_pointer(s, page, object))
2533 * We could also check if the object is on the slabs freelist.
2534 * But this would be too expensive and it seems that the main
2535 * purpose of kmem_ptr_valid() is to check if the object belongs
2536 * to a certain slab.
2540 EXPORT_SYMBOL(kmem_ptr_validate);
2543 * Determine the size of a slab object
2545 unsigned int kmem_cache_size(struct kmem_cache *s)
2549 EXPORT_SYMBOL(kmem_cache_size);
2551 const char *kmem_cache_name(struct kmem_cache *s)
2555 EXPORT_SYMBOL(kmem_cache_name);
2557 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2560 #ifdef CONFIG_SLUB_DEBUG
2561 void *addr = page_address(page);
2563 DECLARE_BITMAP(map, page->objects);
2565 bitmap_zero(map, page->objects);
2566 slab_err(s, page, "%s", text);
2568 for_each_free_object(p, s, page->freelist)
2569 set_bit(slab_index(p, s, addr), map);
2571 for_each_object(p, s, addr, page->objects) {
2573 if (!test_bit(slab_index(p, s, addr), map)) {
2574 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2576 print_tracking(s, p);
2584 * Attempt to free all partial slabs on a node.
2586 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2588 unsigned long flags;
2589 struct page *page, *h;
2591 spin_lock_irqsave(&n->list_lock, flags);
2592 list_for_each_entry_safe(page, h, &n->partial, lru) {
2594 list_del(&page->lru);
2595 discard_slab(s, page);
2598 list_slab_objects(s, page,
2599 "Objects remaining on kmem_cache_close()");
2602 spin_unlock_irqrestore(&n->list_lock, flags);
2606 * Release all resources used by a slab cache.
2608 static inline int kmem_cache_close(struct kmem_cache *s)
2614 /* Attempt to free all objects */
2615 free_kmem_cache_cpus(s);
2616 for_each_node_state(node, N_NORMAL_MEMORY) {
2617 struct kmem_cache_node *n = get_node(s, node);
2620 if (n->nr_partial || slabs_node(s, node))
2623 free_kmem_cache_nodes(s);
2628 * Close a cache and release the kmem_cache structure
2629 * (must be used for caches created using kmem_cache_create)
2631 void kmem_cache_destroy(struct kmem_cache *s)
2633 if (s->flags & SLAB_DESTROY_BY_RCU)
2635 down_write(&slub_lock);
2639 up_write(&slub_lock);
2640 if (kmem_cache_close(s)) {
2641 printk(KERN_ERR "SLUB %s: %s called for cache that "
2642 "still has objects.\n", s->name, __func__);
2645 sysfs_slab_remove(s);
2647 up_write(&slub_lock);
2649 EXPORT_SYMBOL(kmem_cache_destroy);
2651 /********************************************************************
2653 *******************************************************************/
2655 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2656 EXPORT_SYMBOL(kmalloc_caches);
2658 static int __init setup_slub_min_order(char *str)
2660 get_option(&str, &slub_min_order);
2665 __setup("slub_min_order=", setup_slub_min_order);
2667 static int __init setup_slub_max_order(char *str)
2669 get_option(&str, &slub_max_order);
2670 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2675 __setup("slub_max_order=", setup_slub_max_order);
2677 static int __init setup_slub_min_objects(char *str)
2679 get_option(&str, &slub_min_objects);
2684 __setup("slub_min_objects=", setup_slub_min_objects);
2686 static int __init setup_slub_nomerge(char *str)
2692 __setup("slub_nomerge", setup_slub_nomerge);
2694 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2695 const char *name, int size, gfp_t gfp_flags)
2697 unsigned int flags = 0;
2699 if (gfp_flags & SLUB_DMA)
2700 flags = SLAB_CACHE_DMA;
2703 * This function is called with IRQs disabled during early-boot on
2704 * single CPU so there's no need to take slub_lock here.
2706 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2710 list_add(&s->list, &slab_caches);
2712 if (sysfs_slab_add(s))
2717 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2720 #ifdef CONFIG_ZONE_DMA
2721 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2723 static void sysfs_add_func(struct work_struct *w)
2725 struct kmem_cache *s;
2727 down_write(&slub_lock);
2728 list_for_each_entry(s, &slab_caches, list) {
2729 if (s->flags & __SYSFS_ADD_DEFERRED) {
2730 s->flags &= ~__SYSFS_ADD_DEFERRED;
2734 up_write(&slub_lock);
2737 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2739 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2741 struct kmem_cache *s;
2744 unsigned long slabflags;
2746 s = kmalloc_caches_dma[index];
2750 /* Dynamically create dma cache */
2751 if (flags & __GFP_WAIT)
2752 down_write(&slub_lock);
2754 if (!down_write_trylock(&slub_lock))
2758 if (kmalloc_caches_dma[index])
2761 realsize = kmalloc_caches[index].objsize;
2762 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2763 (unsigned int)realsize);
2764 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2767 * Must defer sysfs creation to a workqueue because we don't know
2768 * what context we are called from. Before sysfs comes up, we don't
2769 * need to do anything because our sysfs initcall will start by
2770 * adding all existing slabs to sysfs.
2772 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
2773 if (slab_state >= SYSFS)
2774 slabflags |= __SYSFS_ADD_DEFERRED;
2776 if (!s || !text || !kmem_cache_open(s, flags, text,
2777 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
2783 list_add(&s->list, &slab_caches);
2784 kmalloc_caches_dma[index] = s;
2786 if (slab_state >= SYSFS)
2787 schedule_work(&sysfs_add_work);
2790 up_write(&slub_lock);
2792 return kmalloc_caches_dma[index];
2797 * Conversion table for small slabs sizes / 8 to the index in the
2798 * kmalloc array. This is necessary for slabs < 192 since we have non power
2799 * of two cache sizes there. The size of larger slabs can be determined using
2802 static s8 size_index[24] = {
2829 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2835 return ZERO_SIZE_PTR;
2837 index = size_index[(size - 1) / 8];
2839 index = fls(size - 1);
2841 #ifdef CONFIG_ZONE_DMA
2842 if (unlikely((flags & SLUB_DMA)))
2843 return dma_kmalloc_cache(index, flags);
2846 return &kmalloc_caches[index];
2849 void *__kmalloc(size_t size, gfp_t flags)
2851 struct kmem_cache *s;
2854 if (unlikely(size > SLUB_MAX_SIZE))
2855 return kmalloc_large(size, flags);
2857 s = get_slab(size, flags);
2859 if (unlikely(ZERO_OR_NULL_PTR(s)))
2862 ret = slab_alloc(s, flags, -1, _RET_IP_);
2864 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2868 EXPORT_SYMBOL(__kmalloc);
2870 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2874 flags |= __GFP_COMP | __GFP_NOTRACK;
2875 page = alloc_pages_node(node, flags, get_order(size));
2877 return page_address(page);
2883 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2885 struct kmem_cache *s;
2888 if (unlikely(size > SLUB_MAX_SIZE)) {
2889 ret = kmalloc_large_node(size, flags, node);
2891 trace_kmalloc_node(_RET_IP_, ret,
2892 size, PAGE_SIZE << get_order(size),
2898 s = get_slab(size, flags);
2900 if (unlikely(ZERO_OR_NULL_PTR(s)))
2903 ret = slab_alloc(s, flags, node, _RET_IP_);
2905 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2909 EXPORT_SYMBOL(__kmalloc_node);
2912 size_t ksize(const void *object)
2915 struct kmem_cache *s;
2917 if (unlikely(object == ZERO_SIZE_PTR))
2920 page = virt_to_head_page(object);
2922 if (unlikely(!PageSlab(page))) {
2923 WARN_ON(!PageCompound(page));
2924 return PAGE_SIZE << compound_order(page);
2928 #ifdef CONFIG_SLUB_DEBUG
2930 * Debugging requires use of the padding between object
2931 * and whatever may come after it.
2933 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2938 * If we have the need to store the freelist pointer
2939 * back there or track user information then we can
2940 * only use the space before that information.
2942 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2945 * Else we can use all the padding etc for the allocation
2949 EXPORT_SYMBOL(ksize);
2951 void kfree(const void *x)
2954 void *object = (void *)x;
2956 trace_kfree(_RET_IP_, x);
2958 if (unlikely(ZERO_OR_NULL_PTR(x)))
2961 page = virt_to_head_page(x);
2962 if (unlikely(!PageSlab(page))) {
2963 BUG_ON(!PageCompound(page));
2967 slab_free(page->slab, page, object, _RET_IP_);
2969 EXPORT_SYMBOL(kfree);
2972 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2973 * the remaining slabs by the number of items in use. The slabs with the
2974 * most items in use come first. New allocations will then fill those up
2975 * and thus they can be removed from the partial lists.
2977 * The slabs with the least items are placed last. This results in them
2978 * being allocated from last increasing the chance that the last objects
2979 * are freed in them.
2981 int kmem_cache_shrink(struct kmem_cache *s)
2985 struct kmem_cache_node *n;
2988 int objects = oo_objects(s->max);
2989 struct list_head *slabs_by_inuse =
2990 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2991 unsigned long flags;
2993 if (!slabs_by_inuse)
2997 for_each_node_state(node, N_NORMAL_MEMORY) {
2998 n = get_node(s, node);
3003 for (i = 0; i < objects; i++)
3004 INIT_LIST_HEAD(slabs_by_inuse + i);
3006 spin_lock_irqsave(&n->list_lock, flags);
3009 * Build lists indexed by the items in use in each slab.
3011 * Note that concurrent frees may occur while we hold the
3012 * list_lock. page->inuse here is the upper limit.
3014 list_for_each_entry_safe(page, t, &n->partial, lru) {
3015 if (!page->inuse && slab_trylock(page)) {
3017 * Must hold slab lock here because slab_free
3018 * may have freed the last object and be
3019 * waiting to release the slab.
3021 list_del(&page->lru);
3024 discard_slab(s, page);
3026 list_move(&page->lru,
3027 slabs_by_inuse + page->inuse);
3032 * Rebuild the partial list with the slabs filled up most
3033 * first and the least used slabs at the end.
3035 for (i = objects - 1; i >= 0; i--)
3036 list_splice(slabs_by_inuse + i, n->partial.prev);
3038 spin_unlock_irqrestore(&n->list_lock, flags);
3041 kfree(slabs_by_inuse);
3044 EXPORT_SYMBOL(kmem_cache_shrink);
3046 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
3047 static int slab_mem_going_offline_callback(void *arg)
3049 struct kmem_cache *s;
3051 down_read(&slub_lock);
3052 list_for_each_entry(s, &slab_caches, list)
3053 kmem_cache_shrink(s);
3054 up_read(&slub_lock);
3059 static void slab_mem_offline_callback(void *arg)
3061 struct kmem_cache_node *n;
3062 struct kmem_cache *s;
3063 struct memory_notify *marg = arg;
3066 offline_node = marg->status_change_nid;
3069 * If the node still has available memory. we need kmem_cache_node
3072 if (offline_node < 0)
3075 down_read(&slub_lock);
3076 list_for_each_entry(s, &slab_caches, list) {
3077 n = get_node(s, offline_node);
3080 * if n->nr_slabs > 0, slabs still exist on the node
3081 * that is going down. We were unable to free them,
3082 * and offline_pages() function shoudn't call this
3083 * callback. So, we must fail.
3085 BUG_ON(slabs_node(s, offline_node));
3087 s->node[offline_node] = NULL;
3088 kmem_cache_free(kmalloc_caches, n);
3091 up_read(&slub_lock);
3094 static int slab_mem_going_online_callback(void *arg)
3096 struct kmem_cache_node *n;
3097 struct kmem_cache *s;
3098 struct memory_notify *marg = arg;
3099 int nid = marg->status_change_nid;
3103 * If the node's memory is already available, then kmem_cache_node is
3104 * already created. Nothing to do.
3110 * We are bringing a node online. No memory is available yet. We must
3111 * allocate a kmem_cache_node structure in order to bring the node
3114 down_read(&slub_lock);
3115 list_for_each_entry(s, &slab_caches, list) {
3117 * XXX: kmem_cache_alloc_node will fallback to other nodes
3118 * since memory is not yet available from the node that
3121 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3126 init_kmem_cache_node(n, s);
3130 up_read(&slub_lock);
3134 static int slab_memory_callback(struct notifier_block *self,
3135 unsigned long action, void *arg)
3140 case MEM_GOING_ONLINE:
3141 ret = slab_mem_going_online_callback(arg);
3143 case MEM_GOING_OFFLINE:
3144 ret = slab_mem_going_offline_callback(arg);
3147 case MEM_CANCEL_ONLINE:
3148 slab_mem_offline_callback(arg);
3151 case MEM_CANCEL_OFFLINE:
3155 ret = notifier_from_errno(ret);
3161 #endif /* CONFIG_MEMORY_HOTPLUG */
3163 /********************************************************************
3164 * Basic setup of slabs
3165 *******************************************************************/
3167 void __init kmem_cache_init(void)
3176 * Must first have the slab cache available for the allocations of the
3177 * struct kmem_cache_node's. There is special bootstrap code in
3178 * kmem_cache_open for slab_state == DOWN.
3180 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3181 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3182 kmalloc_caches[0].refcount = -1;
3185 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3188 /* Able to allocate the per node structures */
3189 slab_state = PARTIAL;
3191 /* Caches that are not of the two-to-the-power-of size */
3192 if (KMALLOC_MIN_SIZE <= 64) {
3193 create_kmalloc_cache(&kmalloc_caches[1],
3194 "kmalloc-96", 96, GFP_NOWAIT);
3196 create_kmalloc_cache(&kmalloc_caches[2],
3197 "kmalloc-192", 192, GFP_NOWAIT);
3201 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3202 create_kmalloc_cache(&kmalloc_caches[i],
3203 "kmalloc", 1 << i, GFP_NOWAIT);
3209 * Patch up the size_index table if we have strange large alignment
3210 * requirements for the kmalloc array. This is only the case for
3211 * MIPS it seems. The standard arches will not generate any code here.
3213 * Largest permitted alignment is 256 bytes due to the way we
3214 * handle the index determination for the smaller caches.
3216 * Make sure that nothing crazy happens if someone starts tinkering
3217 * around with ARCH_KMALLOC_MINALIGN
3219 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3220 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3222 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3223 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3225 if (KMALLOC_MIN_SIZE == 128) {
3227 * The 192 byte sized cache is not used if the alignment
3228 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3231 for (i = 128 + 8; i <= 192; i += 8)
3232 size_index[(i - 1) / 8] = 8;
3237 /* Provide the correct kmalloc names now that the caches are up */
3238 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3239 kmalloc_caches[i]. name =
3240 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3243 register_cpu_notifier(&slab_notifier);
3244 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3245 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3247 kmem_size = sizeof(struct kmem_cache);
3251 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3252 " CPUs=%d, Nodes=%d\n",
3253 caches, cache_line_size(),
3254 slub_min_order, slub_max_order, slub_min_objects,
3255 nr_cpu_ids, nr_node_ids);
3258 void __init kmem_cache_init_late(void)
3263 * Find a mergeable slab cache
3265 static int slab_unmergeable(struct kmem_cache *s)
3267 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3274 * We may have set a slab to be unmergeable during bootstrap.
3276 if (s->refcount < 0)
3282 static struct kmem_cache *find_mergeable(size_t size,
3283 size_t align, unsigned long flags, const char *name,
3284 void (*ctor)(void *))
3286 struct kmem_cache *s;
3288 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3294 size = ALIGN(size, sizeof(void *));
3295 align = calculate_alignment(flags, align, size);
3296 size = ALIGN(size, align);
3297 flags = kmem_cache_flags(size, flags, name, NULL);
3299 list_for_each_entry(s, &slab_caches, list) {
3300 if (slab_unmergeable(s))
3306 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3309 * Check if alignment is compatible.
3310 * Courtesy of Adrian Drzewiecki
3312 if ((s->size & ~(align - 1)) != s->size)
3315 if (s->size - size >= sizeof(void *))
3323 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3324 size_t align, unsigned long flags, void (*ctor)(void *))
3326 struct kmem_cache *s;
3328 down_write(&slub_lock);
3329 s = find_mergeable(size, align, flags, name, ctor);
3335 * Adjust the object sizes so that we clear
3336 * the complete object on kzalloc.
3338 s->objsize = max(s->objsize, (int)size);
3341 * And then we need to update the object size in the
3342 * per cpu structures
3344 for_each_online_cpu(cpu)
3345 get_cpu_slab(s, cpu)->objsize = s->objsize;
3347 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3348 up_write(&slub_lock);
3350 if (sysfs_slab_alias(s, name)) {
3351 down_write(&slub_lock);
3353 up_write(&slub_lock);
3359 s = kmalloc(kmem_size, GFP_KERNEL);
3361 if (kmem_cache_open(s, GFP_KERNEL, name,
3362 size, align, flags, ctor)) {
3363 list_add(&s->list, &slab_caches);
3364 up_write(&slub_lock);
3365 if (sysfs_slab_add(s)) {
3366 down_write(&slub_lock);
3368 up_write(&slub_lock);
3376 up_write(&slub_lock);
3379 if (flags & SLAB_PANIC)
3380 panic("Cannot create slabcache %s\n", name);
3385 EXPORT_SYMBOL(kmem_cache_create);
3389 * Use the cpu notifier to insure that the cpu slabs are flushed when
3392 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3393 unsigned long action, void *hcpu)
3395 long cpu = (long)hcpu;
3396 struct kmem_cache *s;
3397 unsigned long flags;
3400 case CPU_UP_PREPARE:
3401 case CPU_UP_PREPARE_FROZEN:
3402 init_alloc_cpu_cpu(cpu);
3403 down_read(&slub_lock);
3404 list_for_each_entry(s, &slab_caches, list)
3405 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3407 up_read(&slub_lock);
3410 case CPU_UP_CANCELED:
3411 case CPU_UP_CANCELED_FROZEN:
3413 case CPU_DEAD_FROZEN:
3414 down_read(&slub_lock);
3415 list_for_each_entry(s, &slab_caches, list) {
3416 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3418 local_irq_save(flags);
3419 __flush_cpu_slab(s, cpu);
3420 local_irq_restore(flags);
3421 free_kmem_cache_cpu(c, cpu);
3422 s->cpu_slab[cpu] = NULL;
3424 up_read(&slub_lock);
3432 static struct notifier_block __cpuinitdata slab_notifier = {
3433 .notifier_call = slab_cpuup_callback
3438 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3440 struct kmem_cache *s;
3443 if (unlikely(size > SLUB_MAX_SIZE))
3444 return kmalloc_large(size, gfpflags);
3446 s = get_slab(size, gfpflags);
3448 if (unlikely(ZERO_OR_NULL_PTR(s)))
3451 ret = slab_alloc(s, gfpflags, -1, caller);
3453 /* Honor the call site pointer we recieved. */
3454 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3459 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3460 int node, unsigned long caller)
3462 struct kmem_cache *s;
3465 if (unlikely(size > SLUB_MAX_SIZE))
3466 return kmalloc_large_node(size, gfpflags, node);
3468 s = get_slab(size, gfpflags);
3470 if (unlikely(ZERO_OR_NULL_PTR(s)))
3473 ret = slab_alloc(s, gfpflags, node, caller);
3475 /* Honor the call site pointer we recieved. */
3476 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3481 #ifdef CONFIG_SLUB_DEBUG
3482 static int count_inuse(struct page *page)
3487 static int count_total(struct page *page)
3489 return page->objects;
3492 static int validate_slab(struct kmem_cache *s, struct page *page,
3496 void *addr = page_address(page);
3498 if (!check_slab(s, page) ||
3499 !on_freelist(s, page, NULL))
3502 /* Now we know that a valid freelist exists */
3503 bitmap_zero(map, page->objects);
3505 for_each_free_object(p, s, page->freelist) {
3506 set_bit(slab_index(p, s, addr), map);
3507 if (!check_object(s, page, p, 0))
3511 for_each_object(p, s, addr, page->objects)
3512 if (!test_bit(slab_index(p, s, addr), map))
3513 if (!check_object(s, page, p, 1))
3518 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3521 if (slab_trylock(page)) {
3522 validate_slab(s, page, map);
3525 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3528 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3529 if (!PageSlubDebug(page))
3530 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3531 "on slab 0x%p\n", s->name, page);
3533 if (PageSlubDebug(page))
3534 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3535 "slab 0x%p\n", s->name, page);
3539 static int validate_slab_node(struct kmem_cache *s,
3540 struct kmem_cache_node *n, unsigned long *map)
3542 unsigned long count = 0;
3544 unsigned long flags;
3546 spin_lock_irqsave(&n->list_lock, flags);
3548 list_for_each_entry(page, &n->partial, lru) {
3549 validate_slab_slab(s, page, map);
3552 if (count != n->nr_partial)
3553 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3554 "counter=%ld\n", s->name, count, n->nr_partial);
3556 if (!(s->flags & SLAB_STORE_USER))
3559 list_for_each_entry(page, &n->full, lru) {
3560 validate_slab_slab(s, page, map);
3563 if (count != atomic_long_read(&n->nr_slabs))
3564 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3565 "counter=%ld\n", s->name, count,
3566 atomic_long_read(&n->nr_slabs));
3569 spin_unlock_irqrestore(&n->list_lock, flags);
3573 static long validate_slab_cache(struct kmem_cache *s)
3576 unsigned long count = 0;
3577 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3578 sizeof(unsigned long), GFP_KERNEL);
3584 for_each_node_state(node, N_NORMAL_MEMORY) {
3585 struct kmem_cache_node *n = get_node(s, node);
3587 count += validate_slab_node(s, n, map);
3593 #ifdef SLUB_RESILIENCY_TEST
3594 static void resiliency_test(void)
3598 printk(KERN_ERR "SLUB resiliency testing\n");
3599 printk(KERN_ERR "-----------------------\n");
3600 printk(KERN_ERR "A. Corruption after allocation\n");
3602 p = kzalloc(16, GFP_KERNEL);
3604 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3605 " 0x12->0x%p\n\n", p + 16);
3607 validate_slab_cache(kmalloc_caches + 4);
3609 /* Hmmm... The next two are dangerous */
3610 p = kzalloc(32, GFP_KERNEL);
3611 p[32 + sizeof(void *)] = 0x34;
3612 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3613 " 0x34 -> -0x%p\n", p);
3615 "If allocated object is overwritten then not detectable\n\n");
3617 validate_slab_cache(kmalloc_caches + 5);
3618 p = kzalloc(64, GFP_KERNEL);
3619 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3621 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3624 "If allocated object is overwritten then not detectable\n\n");
3625 validate_slab_cache(kmalloc_caches + 6);
3627 printk(KERN_ERR "\nB. Corruption after free\n");
3628 p = kzalloc(128, GFP_KERNEL);
3631 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3632 validate_slab_cache(kmalloc_caches + 7);
3634 p = kzalloc(256, GFP_KERNEL);
3637 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3639 validate_slab_cache(kmalloc_caches + 8);
3641 p = kzalloc(512, GFP_KERNEL);
3644 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3645 validate_slab_cache(kmalloc_caches + 9);
3648 static void resiliency_test(void) {};
3652 * Generate lists of code addresses where slabcache objects are allocated
3657 unsigned long count;
3664 DECLARE_BITMAP(cpus, NR_CPUS);
3670 unsigned long count;
3671 struct location *loc;
3674 static void free_loc_track(struct loc_track *t)
3677 free_pages((unsigned long)t->loc,
3678 get_order(sizeof(struct location) * t->max));
3681 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3686 order = get_order(sizeof(struct location) * max);
3688 l = (void *)__get_free_pages(flags, order);
3693 memcpy(l, t->loc, sizeof(struct location) * t->count);
3701 static int add_location(struct loc_track *t, struct kmem_cache *s,
3702 const struct track *track)
3704 long start, end, pos;
3706 unsigned long caddr;
3707 unsigned long age = jiffies - track->when;
3713 pos = start + (end - start + 1) / 2;
3716 * There is nothing at "end". If we end up there
3717 * we need to add something to before end.
3722 caddr = t->loc[pos].addr;
3723 if (track->addr == caddr) {
3729 if (age < l->min_time)
3731 if (age > l->max_time)
3734 if (track->pid < l->min_pid)
3735 l->min_pid = track->pid;
3736 if (track->pid > l->max_pid)
3737 l->max_pid = track->pid;
3739 cpumask_set_cpu(track->cpu,
3740 to_cpumask(l->cpus));
3742 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3746 if (track->addr < caddr)
3753 * Not found. Insert new tracking element.
3755 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3761 (t->count - pos) * sizeof(struct location));
3764 l->addr = track->addr;
3768 l->min_pid = track->pid;
3769 l->max_pid = track->pid;
3770 cpumask_clear(to_cpumask(l->cpus));
3771 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3772 nodes_clear(l->nodes);
3773 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3777 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3778 struct page *page, enum track_item alloc)
3780 void *addr = page_address(page);
3781 DECLARE_BITMAP(map, page->objects);
3784 bitmap_zero(map, page->objects);
3785 for_each_free_object(p, s, page->freelist)
3786 set_bit(slab_index(p, s, addr), map);
3788 for_each_object(p, s, addr, page->objects)
3789 if (!test_bit(slab_index(p, s, addr), map))
3790 add_location(t, s, get_track(s, p, alloc));
3793 static int list_locations(struct kmem_cache *s, char *buf,
3794 enum track_item alloc)
3798 struct loc_track t = { 0, 0, NULL };
3801 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3803 return sprintf(buf, "Out of memory\n");
3805 /* Push back cpu slabs */
3808 for_each_node_state(node, N_NORMAL_MEMORY) {
3809 struct kmem_cache_node *n = get_node(s, node);
3810 unsigned long flags;
3813 if (!atomic_long_read(&n->nr_slabs))
3816 spin_lock_irqsave(&n->list_lock, flags);
3817 list_for_each_entry(page, &n->partial, lru)
3818 process_slab(&t, s, page, alloc);
3819 list_for_each_entry(page, &n->full, lru)
3820 process_slab(&t, s, page, alloc);
3821 spin_unlock_irqrestore(&n->list_lock, flags);
3824 for (i = 0; i < t.count; i++) {
3825 struct location *l = &t.loc[i];
3827 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3829 len += sprintf(buf + len, "%7ld ", l->count);
3832 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3834 len += sprintf(buf + len, "<not-available>");
3836 if (l->sum_time != l->min_time) {
3837 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3839 (long)div_u64(l->sum_time, l->count),
3842 len += sprintf(buf + len, " age=%ld",
3845 if (l->min_pid != l->max_pid)
3846 len += sprintf(buf + len, " pid=%ld-%ld",
3847 l->min_pid, l->max_pid);
3849 len += sprintf(buf + len, " pid=%ld",
3852 if (num_online_cpus() > 1 &&
3853 !cpumask_empty(to_cpumask(l->cpus)) &&
3854 len < PAGE_SIZE - 60) {
3855 len += sprintf(buf + len, " cpus=");
3856 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3857 to_cpumask(l->cpus));
3860 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3861 len < PAGE_SIZE - 60) {
3862 len += sprintf(buf + len, " nodes=");
3863 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3867 len += sprintf(buf + len, "\n");
3872 len += sprintf(buf, "No data\n");
3876 enum slab_stat_type {
3877 SL_ALL, /* All slabs */
3878 SL_PARTIAL, /* Only partially allocated slabs */
3879 SL_CPU, /* Only slabs used for cpu caches */
3880 SL_OBJECTS, /* Determine allocated objects not slabs */
3881 SL_TOTAL /* Determine object capacity not slabs */
3884 #define SO_ALL (1 << SL_ALL)
3885 #define SO_PARTIAL (1 << SL_PARTIAL)
3886 #define SO_CPU (1 << SL_CPU)
3887 #define SO_OBJECTS (1 << SL_OBJECTS)
3888 #define SO_TOTAL (1 << SL_TOTAL)
3890 static ssize_t show_slab_objects(struct kmem_cache *s,
3891 char *buf, unsigned long flags)
3893 unsigned long total = 0;
3896 unsigned long *nodes;
3897 unsigned long *per_cpu;
3899 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3902 per_cpu = nodes + nr_node_ids;
3904 if (flags & SO_CPU) {
3907 for_each_possible_cpu(cpu) {
3908 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3910 if (!c || c->node < 0)
3914 if (flags & SO_TOTAL)
3915 x = c->page->objects;
3916 else if (flags & SO_OBJECTS)
3922 nodes[c->node] += x;
3928 if (flags & SO_ALL) {
3929 for_each_node_state(node, N_NORMAL_MEMORY) {
3930 struct kmem_cache_node *n = get_node(s, node);
3932 if (flags & SO_TOTAL)
3933 x = atomic_long_read(&n->total_objects);
3934 else if (flags & SO_OBJECTS)
3935 x = atomic_long_read(&n->total_objects) -
3936 count_partial(n, count_free);
3939 x = atomic_long_read(&n->nr_slabs);
3944 } else if (flags & SO_PARTIAL) {
3945 for_each_node_state(node, N_NORMAL_MEMORY) {
3946 struct kmem_cache_node *n = get_node(s, node);
3948 if (flags & SO_TOTAL)
3949 x = count_partial(n, count_total);
3950 else if (flags & SO_OBJECTS)
3951 x = count_partial(n, count_inuse);
3958 x = sprintf(buf, "%lu", total);
3960 for_each_node_state(node, N_NORMAL_MEMORY)
3962 x += sprintf(buf + x, " N%d=%lu",
3966 return x + sprintf(buf + x, "\n");
3969 static int any_slab_objects(struct kmem_cache *s)
3973 for_each_online_node(node) {
3974 struct kmem_cache_node *n = get_node(s, node);
3979 if (atomic_long_read(&n->total_objects))
3985 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3986 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3988 struct slab_attribute {
3989 struct attribute attr;
3990 ssize_t (*show)(struct kmem_cache *s, char *buf);
3991 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3994 #define SLAB_ATTR_RO(_name) \
3995 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3997 #define SLAB_ATTR(_name) \
3998 static struct slab_attribute _name##_attr = \
3999 __ATTR(_name, 0644, _name##_show, _name##_store)
4001 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4003 return sprintf(buf, "%d\n", s->size);
4005 SLAB_ATTR_RO(slab_size);
4007 static ssize_t align_show(struct kmem_cache *s, char *buf)
4009 return sprintf(buf, "%d\n", s->align);
4011 SLAB_ATTR_RO(align);
4013 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4015 return sprintf(buf, "%d\n", s->objsize);
4017 SLAB_ATTR_RO(object_size);
4019 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4021 return sprintf(buf, "%d\n", oo_objects(s->oo));
4023 SLAB_ATTR_RO(objs_per_slab);
4025 static ssize_t order_store(struct kmem_cache *s,
4026 const char *buf, size_t length)
4028 unsigned long order;
4031 err = strict_strtoul(buf, 10, &order);
4035 if (order > slub_max_order || order < slub_min_order)
4038 calculate_sizes(s, order);
4042 static ssize_t order_show(struct kmem_cache *s, char *buf)
4044 return sprintf(buf, "%d\n", oo_order(s->oo));
4048 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4050 return sprintf(buf, "%lu\n", s->min_partial);
4053 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4059 err = strict_strtoul(buf, 10, &min);
4063 set_min_partial(s, min);
4066 SLAB_ATTR(min_partial);
4068 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4071 int n = sprint_symbol(buf, (unsigned long)s->ctor);
4073 return n + sprintf(buf + n, "\n");
4079 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4081 return sprintf(buf, "%d\n", s->refcount - 1);
4083 SLAB_ATTR_RO(aliases);
4085 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4087 return show_slab_objects(s, buf, SO_ALL);
4089 SLAB_ATTR_RO(slabs);
4091 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4093 return show_slab_objects(s, buf, SO_PARTIAL);
4095 SLAB_ATTR_RO(partial);
4097 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4099 return show_slab_objects(s, buf, SO_CPU);
4101 SLAB_ATTR_RO(cpu_slabs);
4103 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4105 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4107 SLAB_ATTR_RO(objects);
4109 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4111 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4113 SLAB_ATTR_RO(objects_partial);
4115 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4117 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4119 SLAB_ATTR_RO(total_objects);
4121 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4123 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4126 static ssize_t sanity_checks_store(struct kmem_cache *s,
4127 const char *buf, size_t length)
4129 s->flags &= ~SLAB_DEBUG_FREE;
4131 s->flags |= SLAB_DEBUG_FREE;
4134 SLAB_ATTR(sanity_checks);
4136 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4138 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4141 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4144 s->flags &= ~SLAB_TRACE;
4146 s->flags |= SLAB_TRACE;
4151 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4153 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4156 static ssize_t reclaim_account_store(struct kmem_cache *s,
4157 const char *buf, size_t length)
4159 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4161 s->flags |= SLAB_RECLAIM_ACCOUNT;
4164 SLAB_ATTR(reclaim_account);
4166 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4168 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4170 SLAB_ATTR_RO(hwcache_align);
4172 #ifdef CONFIG_ZONE_DMA
4173 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4175 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4177 SLAB_ATTR_RO(cache_dma);
4180 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4182 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4184 SLAB_ATTR_RO(destroy_by_rcu);
4186 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4188 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4191 static ssize_t red_zone_store(struct kmem_cache *s,
4192 const char *buf, size_t length)
4194 if (any_slab_objects(s))
4197 s->flags &= ~SLAB_RED_ZONE;
4199 s->flags |= SLAB_RED_ZONE;
4200 calculate_sizes(s, -1);
4203 SLAB_ATTR(red_zone);
4205 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4207 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4210 static ssize_t poison_store(struct kmem_cache *s,
4211 const char *buf, size_t length)
4213 if (any_slab_objects(s))
4216 s->flags &= ~SLAB_POISON;
4218 s->flags |= SLAB_POISON;
4219 calculate_sizes(s, -1);
4224 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4226 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4229 static ssize_t store_user_store(struct kmem_cache *s,
4230 const char *buf, size_t length)
4232 if (any_slab_objects(s))
4235 s->flags &= ~SLAB_STORE_USER;
4237 s->flags |= SLAB_STORE_USER;
4238 calculate_sizes(s, -1);
4241 SLAB_ATTR(store_user);
4243 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4248 static ssize_t validate_store(struct kmem_cache *s,
4249 const char *buf, size_t length)
4253 if (buf[0] == '1') {
4254 ret = validate_slab_cache(s);
4260 SLAB_ATTR(validate);
4262 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4267 static ssize_t shrink_store(struct kmem_cache *s,
4268 const char *buf, size_t length)
4270 if (buf[0] == '1') {
4271 int rc = kmem_cache_shrink(s);
4281 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4283 if (!(s->flags & SLAB_STORE_USER))
4285 return list_locations(s, buf, TRACK_ALLOC);
4287 SLAB_ATTR_RO(alloc_calls);
4289 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4291 if (!(s->flags & SLAB_STORE_USER))
4293 return list_locations(s, buf, TRACK_FREE);
4295 SLAB_ATTR_RO(free_calls);
4298 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4300 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4303 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4304 const char *buf, size_t length)
4306 unsigned long ratio;
4309 err = strict_strtoul(buf, 10, &ratio);
4314 s->remote_node_defrag_ratio = ratio * 10;
4318 SLAB_ATTR(remote_node_defrag_ratio);
4321 #ifdef CONFIG_SLUB_STATS
4322 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4324 unsigned long sum = 0;
4327 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4332 for_each_online_cpu(cpu) {
4333 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4339 len = sprintf(buf, "%lu", sum);
4342 for_each_online_cpu(cpu) {
4343 if (data[cpu] && len < PAGE_SIZE - 20)
4344 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4348 return len + sprintf(buf + len, "\n");
4351 #define STAT_ATTR(si, text) \
4352 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4354 return show_stat(s, buf, si); \
4356 SLAB_ATTR_RO(text); \
4358 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4359 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4360 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4361 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4362 STAT_ATTR(FREE_FROZEN, free_frozen);
4363 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4364 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4365 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4366 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4367 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4368 STAT_ATTR(FREE_SLAB, free_slab);
4369 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4370 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4371 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4372 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4373 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4374 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4375 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4378 static struct attribute *slab_attrs[] = {
4379 &slab_size_attr.attr,
4380 &object_size_attr.attr,
4381 &objs_per_slab_attr.attr,
4383 &min_partial_attr.attr,
4385 &objects_partial_attr.attr,
4386 &total_objects_attr.attr,
4389 &cpu_slabs_attr.attr,
4393 &sanity_checks_attr.attr,
4395 &hwcache_align_attr.attr,
4396 &reclaim_account_attr.attr,
4397 &destroy_by_rcu_attr.attr,
4398 &red_zone_attr.attr,
4400 &store_user_attr.attr,
4401 &validate_attr.attr,
4403 &alloc_calls_attr.attr,
4404 &free_calls_attr.attr,
4405 #ifdef CONFIG_ZONE_DMA
4406 &cache_dma_attr.attr,
4409 &remote_node_defrag_ratio_attr.attr,
4411 #ifdef CONFIG_SLUB_STATS
4412 &alloc_fastpath_attr.attr,
4413 &alloc_slowpath_attr.attr,
4414 &free_fastpath_attr.attr,
4415 &free_slowpath_attr.attr,
4416 &free_frozen_attr.attr,
4417 &free_add_partial_attr.attr,
4418 &free_remove_partial_attr.attr,
4419 &alloc_from_partial_attr.attr,
4420 &alloc_slab_attr.attr,
4421 &alloc_refill_attr.attr,
4422 &free_slab_attr.attr,
4423 &cpuslab_flush_attr.attr,
4424 &deactivate_full_attr.attr,
4425 &deactivate_empty_attr.attr,
4426 &deactivate_to_head_attr.attr,
4427 &deactivate_to_tail_attr.attr,
4428 &deactivate_remote_frees_attr.attr,
4429 &order_fallback_attr.attr,
4434 static struct attribute_group slab_attr_group = {
4435 .attrs = slab_attrs,
4438 static ssize_t slab_attr_show(struct kobject *kobj,
4439 struct attribute *attr,
4442 struct slab_attribute *attribute;
4443 struct kmem_cache *s;
4446 attribute = to_slab_attr(attr);
4449 if (!attribute->show)
4452 err = attribute->show(s, buf);
4457 static ssize_t slab_attr_store(struct kobject *kobj,
4458 struct attribute *attr,
4459 const char *buf, size_t len)
4461 struct slab_attribute *attribute;
4462 struct kmem_cache *s;
4465 attribute = to_slab_attr(attr);
4468 if (!attribute->store)
4471 err = attribute->store(s, buf, len);
4476 static void kmem_cache_release(struct kobject *kobj)
4478 struct kmem_cache *s = to_slab(kobj);
4483 static struct sysfs_ops slab_sysfs_ops = {
4484 .show = slab_attr_show,
4485 .store = slab_attr_store,
4488 static struct kobj_type slab_ktype = {
4489 .sysfs_ops = &slab_sysfs_ops,
4490 .release = kmem_cache_release
4493 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4495 struct kobj_type *ktype = get_ktype(kobj);
4497 if (ktype == &slab_ktype)
4502 static struct kset_uevent_ops slab_uevent_ops = {
4503 .filter = uevent_filter,
4506 static struct kset *slab_kset;
4508 #define ID_STR_LENGTH 64
4510 /* Create a unique string id for a slab cache:
4512 * Format :[flags-]size
4514 static char *create_unique_id(struct kmem_cache *s)
4516 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4523 * First flags affecting slabcache operations. We will only
4524 * get here for aliasable slabs so we do not need to support
4525 * too many flags. The flags here must cover all flags that
4526 * are matched during merging to guarantee that the id is
4529 if (s->flags & SLAB_CACHE_DMA)
4531 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4533 if (s->flags & SLAB_DEBUG_FREE)
4535 if (!(s->flags & SLAB_NOTRACK))
4539 p += sprintf(p, "%07d", s->size);
4540 BUG_ON(p > name + ID_STR_LENGTH - 1);
4544 static int sysfs_slab_add(struct kmem_cache *s)
4550 if (slab_state < SYSFS)
4551 /* Defer until later */
4554 unmergeable = slab_unmergeable(s);
4557 * Slabcache can never be merged so we can use the name proper.
4558 * This is typically the case for debug situations. In that
4559 * case we can catch duplicate names easily.
4561 sysfs_remove_link(&slab_kset->kobj, s->name);
4565 * Create a unique name for the slab as a target
4568 name = create_unique_id(s);
4571 s->kobj.kset = slab_kset;
4572 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4574 kobject_put(&s->kobj);
4578 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4581 kobject_uevent(&s->kobj, KOBJ_ADD);
4583 /* Setup first alias */
4584 sysfs_slab_alias(s, s->name);
4590 static void sysfs_slab_remove(struct kmem_cache *s)
4592 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4593 kobject_del(&s->kobj);
4594 kobject_put(&s->kobj);
4598 * Need to buffer aliases during bootup until sysfs becomes
4599 * available lest we lose that information.
4601 struct saved_alias {
4602 struct kmem_cache *s;
4604 struct saved_alias *next;
4607 static struct saved_alias *alias_list;
4609 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4611 struct saved_alias *al;
4613 if (slab_state == SYSFS) {
4615 * If we have a leftover link then remove it.
4617 sysfs_remove_link(&slab_kset->kobj, name);
4618 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4621 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4627 al->next = alias_list;
4632 static int __init slab_sysfs_init(void)
4634 struct kmem_cache *s;
4637 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4639 printk(KERN_ERR "Cannot register slab subsystem.\n");
4645 list_for_each_entry(s, &slab_caches, list) {
4646 err = sysfs_slab_add(s);
4648 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4649 " to sysfs\n", s->name);
4652 while (alias_list) {
4653 struct saved_alias *al = alias_list;
4655 alias_list = alias_list->next;
4656 err = sysfs_slab_alias(al->s, al->name);
4658 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4659 " %s to sysfs\n", s->name);
4667 __initcall(slab_sysfs_init);
4671 * The /proc/slabinfo ABI
4673 #ifdef CONFIG_SLABINFO
4674 static void print_slabinfo_header(struct seq_file *m)
4676 seq_puts(m, "slabinfo - version: 2.1\n");
4677 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4678 "<objperslab> <pagesperslab>");
4679 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4680 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4684 static void *s_start(struct seq_file *m, loff_t *pos)
4688 down_read(&slub_lock);
4690 print_slabinfo_header(m);
4692 return seq_list_start(&slab_caches, *pos);
4695 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4697 return seq_list_next(p, &slab_caches, pos);
4700 static void s_stop(struct seq_file *m, void *p)
4702 up_read(&slub_lock);
4705 static int s_show(struct seq_file *m, void *p)
4707 unsigned long nr_partials = 0;
4708 unsigned long nr_slabs = 0;
4709 unsigned long nr_inuse = 0;
4710 unsigned long nr_objs = 0;
4711 unsigned long nr_free = 0;
4712 struct kmem_cache *s;
4715 s = list_entry(p, struct kmem_cache, list);
4717 for_each_online_node(node) {
4718 struct kmem_cache_node *n = get_node(s, node);
4723 nr_partials += n->nr_partial;
4724 nr_slabs += atomic_long_read(&n->nr_slabs);
4725 nr_objs += atomic_long_read(&n->total_objects);
4726 nr_free += count_partial(n, count_free);
4729 nr_inuse = nr_objs - nr_free;
4731 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4732 nr_objs, s->size, oo_objects(s->oo),
4733 (1 << oo_order(s->oo)));
4734 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4735 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4741 static const struct seq_operations slabinfo_op = {
4748 static int slabinfo_open(struct inode *inode, struct file *file)
4750 return seq_open(file, &slabinfo_op);
4753 static const struct file_operations proc_slabinfo_operations = {
4754 .open = slabinfo_open,
4756 .llseek = seq_lseek,
4757 .release = seq_release,
4760 static int __init slab_proc_init(void)
4762 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4765 module_init(slab_proc_init);
4766 #endif /* CONFIG_SLABINFO */