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/kmemcheck.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/mempolicy.h>
24 #include <linux/ctype.h>
25 #include <linux/debugobjects.h>
26 #include <linux/kallsyms.h>
27 #include <linux/memory.h>
28 #include <linux/math64.h>
29 #include <linux/fault-inject.h>
36 * The slab_lock protects operations on the object of a particular
37 * slab and its metadata in the page struct. If the slab lock
38 * has been taken then no allocations nor frees can be performed
39 * on the objects in the slab nor can the slab be added or removed
40 * from the partial or full lists since this would mean modifying
41 * the page_struct of the slab.
43 * The list_lock protects the partial and full list on each node and
44 * the partial slab counter. If taken then no new slabs may be added or
45 * removed from the lists nor make the number of partial slabs be modified.
46 * (Note that the total number of slabs is an atomic value that may be
47 * modified without taking the list lock).
49 * The list_lock is a centralized lock and thus we avoid taking it as
50 * much as possible. As long as SLUB does not have to handle partial
51 * slabs, operations can continue without any centralized lock. F.e.
52 * allocating a long series of objects that fill up slabs does not require
55 * The lock order is sometimes inverted when we are trying to get a slab
56 * off a list. We take the list_lock and then look for a page on the list
57 * to use. While we do that objects in the slabs may be freed. We can
58 * only operate on the slab if we have also taken the slab_lock. So we use
59 * a slab_trylock() on the slab. If trylock was successful then no frees
60 * can occur anymore and we can use the slab for allocations etc. If the
61 * slab_trylock() does not succeed then frees are in progress in the slab and
62 * we must stay away from it for a while since we may cause a bouncing
63 * cacheline if we try to acquire the lock. So go onto the next slab.
64 * If all pages are busy then we may allocate a new slab instead of reusing
65 * a partial slab. A new slab has noone operating on it and thus there is
66 * no danger of cacheline contention.
68 * Interrupts are disabled during allocation and deallocation in order to
69 * make the slab allocator safe to use in the context of an irq. In addition
70 * interrupts are disabled to ensure that the processor does not change
71 * while handling per_cpu slabs, due to kernel preemption.
73 * SLUB assigns one slab for allocation to each processor.
74 * Allocations only occur from these slabs called cpu slabs.
76 * Slabs with free elements are kept on a partial list and during regular
77 * operations no list for full slabs is used. If an object in a full slab is
78 * freed then the slab will show up again on the partial lists.
79 * We track full slabs for debugging purposes though because otherwise we
80 * cannot scan all objects.
82 * Slabs are freed when they become empty. Teardown and setup is
83 * minimal so we rely on the page allocators per cpu caches for
84 * fast frees and allocs.
86 * Overloading of page flags that are otherwise used for LRU management.
88 * PageActive The slab is frozen and exempt from list processing.
89 * This means that the slab is dedicated to a purpose
90 * such as satisfying allocations for a specific
91 * processor. Objects may be freed in the slab while
92 * it is frozen but slab_free will then skip the usual
93 * list operations. It is up to the processor holding
94 * the slab to integrate the slab into the slab lists
95 * when the slab is no longer needed.
97 * One use of this flag is to mark slabs that are
98 * used for allocations. Then such a slab becomes a cpu
99 * slab. The cpu slab may be equipped with an additional
100 * freelist that allows lockless access to
101 * free objects in addition to the regular freelist
102 * that requires the slab lock.
104 * PageError Slab requires special handling due to debug
105 * options set. This moves slab handling out of
106 * the fast path and disables lockless freelists.
109 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
110 SLAB_TRACE | SLAB_DEBUG_FREE)
112 static inline int kmem_cache_debug(struct kmem_cache *s)
114 #ifdef CONFIG_SLUB_DEBUG
115 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
122 * Issues still to be resolved:
124 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
126 * - Variable sizing of the per node arrays
129 /* Enable to test recovery from slab corruption on boot */
130 #undef SLUB_RESILIENCY_TEST
133 * Mininum number of partial slabs. These will be left on the partial
134 * lists even if they are empty. kmem_cache_shrink may reclaim them.
136 #define MIN_PARTIAL 5
139 * Maximum number of desirable partial slabs.
140 * The existence of more partial slabs makes kmem_cache_shrink
141 * sort the partial list by the number of objects in the.
143 #define MAX_PARTIAL 10
145 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
146 SLAB_POISON | SLAB_STORE_USER)
149 * Debugging flags that require metadata to be stored in the slab. These get
150 * disabled when slub_debug=O is used and a cache's min order increases with
153 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
156 * Set of flags that will prevent slab merging
158 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
159 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
162 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
163 SLAB_CACHE_DMA | SLAB_NOTRACK)
166 #define OO_MASK ((1 << OO_SHIFT) - 1)
167 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
169 /* Internal SLUB flags */
170 #define __OBJECT_POISON 0x80000000UL /* Poison object */
172 static int kmem_size = sizeof(struct kmem_cache);
175 static struct notifier_block slab_notifier;
179 DOWN, /* No slab functionality available */
180 PARTIAL, /* Kmem_cache_node works */
181 UP, /* Everything works but does not show up in sysfs */
185 /* A list of all slab caches on the system */
186 static DECLARE_RWSEM(slub_lock);
187 static LIST_HEAD(slab_caches);
190 * Tracking user of a slab.
193 unsigned long addr; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
201 #ifdef CONFIG_SLUB_DEBUG
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void sysfs_slab_remove(struct kmem_cache *);
207 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
208 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
210 static inline void sysfs_slab_remove(struct kmem_cache *s)
217 static inline void stat(struct kmem_cache *s, enum stat_item si)
219 #ifdef CONFIG_SLUB_STATS
220 __this_cpu_inc(s->cpu_slab->stat[si]);
224 /********************************************************************
225 * Core slab cache functions
226 *******************************************************************/
228 int slab_is_available(void)
230 return slab_state >= UP;
233 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
236 return s->node[node];
238 return &s->local_node;
242 /* Verify that a pointer has an address that is valid within a slab page */
243 static inline int check_valid_pointer(struct kmem_cache *s,
244 struct page *page, const void *object)
251 base = page_address(page);
252 if (object < base || object >= base + page->objects * s->size ||
253 (object - base) % s->size) {
260 static inline void *get_freepointer(struct kmem_cache *s, void *object)
262 return *(void **)(object + s->offset);
265 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
267 *(void **)(object + s->offset) = fp;
270 /* Loop over all objects in a slab */
271 #define for_each_object(__p, __s, __addr, __objects) \
272 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
276 #define for_each_free_object(__p, __s, __free) \
277 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
279 /* Determine object index from a given position */
280 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282 return (p - addr) / s->size;
285 static inline struct kmem_cache_order_objects oo_make(int order,
288 struct kmem_cache_order_objects x = {
289 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
295 static inline int oo_order(struct kmem_cache_order_objects x)
297 return x.x >> OO_SHIFT;
300 static inline int oo_objects(struct kmem_cache_order_objects x)
302 return x.x & OO_MASK;
305 #ifdef CONFIG_SLUB_DEBUG
309 #ifdef CONFIG_SLUB_DEBUG_ON
310 static int slub_debug = DEBUG_DEFAULT_FLAGS;
312 static int slub_debug;
315 static char *slub_debug_slabs;
316 static int disable_higher_order_debug;
321 static void print_section(char *text, u8 *addr, unsigned int length)
329 for (i = 0; i < length; i++) {
331 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
334 printk(KERN_CONT " %02x", addr[i]);
336 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
338 printk(KERN_CONT " %s\n", ascii);
345 printk(KERN_CONT " ");
349 printk(KERN_CONT " %s\n", ascii);
353 static struct track *get_track(struct kmem_cache *s, void *object,
354 enum track_item alloc)
359 p = object + s->offset + sizeof(void *);
361 p = object + s->inuse;
366 static void set_track(struct kmem_cache *s, void *object,
367 enum track_item alloc, unsigned long addr)
369 struct track *p = get_track(s, object, alloc);
373 p->cpu = smp_processor_id();
374 p->pid = current->pid;
377 memset(p, 0, sizeof(struct track));
380 static void init_tracking(struct kmem_cache *s, void *object)
382 if (!(s->flags & SLAB_STORE_USER))
385 set_track(s, object, TRACK_FREE, 0UL);
386 set_track(s, object, TRACK_ALLOC, 0UL);
389 static void print_track(const char *s, struct track *t)
394 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
395 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
398 static void print_tracking(struct kmem_cache *s, void *object)
400 if (!(s->flags & SLAB_STORE_USER))
403 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
404 print_track("Freed", get_track(s, object, TRACK_FREE));
407 static void print_page_info(struct page *page)
409 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
410 page, page->objects, page->inuse, page->freelist, page->flags);
414 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
420 vsnprintf(buf, sizeof(buf), fmt, args);
422 printk(KERN_ERR "========================================"
423 "=====================================\n");
424 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
425 printk(KERN_ERR "----------------------------------------"
426 "-------------------------------------\n\n");
429 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
435 vsnprintf(buf, sizeof(buf), fmt, args);
437 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
440 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
442 unsigned int off; /* Offset of last byte */
443 u8 *addr = page_address(page);
445 print_tracking(s, p);
447 print_page_info(page);
449 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
450 p, p - addr, get_freepointer(s, p));
453 print_section("Bytes b4", p - 16, 16);
455 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
457 if (s->flags & SLAB_RED_ZONE)
458 print_section("Redzone", p + s->objsize,
459 s->inuse - s->objsize);
462 off = s->offset + sizeof(void *);
466 if (s->flags & SLAB_STORE_USER)
467 off += 2 * sizeof(struct track);
470 /* Beginning of the filler is the free pointer */
471 print_section("Padding", p + off, s->size - off);
476 static void object_err(struct kmem_cache *s, struct page *page,
477 u8 *object, char *reason)
479 slab_bug(s, "%s", reason);
480 print_trailer(s, page, object);
483 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
489 vsnprintf(buf, sizeof(buf), fmt, args);
491 slab_bug(s, "%s", buf);
492 print_page_info(page);
496 static void init_object(struct kmem_cache *s, void *object, int active)
500 if (s->flags & __OBJECT_POISON) {
501 memset(p, POISON_FREE, s->objsize - 1);
502 p[s->objsize - 1] = POISON_END;
505 if (s->flags & SLAB_RED_ZONE)
506 memset(p + s->objsize,
507 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
508 s->inuse - s->objsize);
511 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
514 if (*start != (u8)value)
522 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
523 void *from, void *to)
525 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
526 memset(from, data, to - from);
529 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
530 u8 *object, char *what,
531 u8 *start, unsigned int value, unsigned int bytes)
536 fault = check_bytes(start, value, bytes);
541 while (end > fault && end[-1] == value)
544 slab_bug(s, "%s overwritten", what);
545 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
546 fault, end - 1, fault[0], value);
547 print_trailer(s, page, object);
549 restore_bytes(s, what, value, fault, end);
557 * Bytes of the object to be managed.
558 * If the freepointer may overlay the object then the free
559 * pointer is the first word of the object.
561 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
564 * object + s->objsize
565 * Padding to reach word boundary. This is also used for Redzoning.
566 * Padding is extended by another word if Redzoning is enabled and
569 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
570 * 0xcc (RED_ACTIVE) for objects in use.
573 * Meta data starts here.
575 * A. Free pointer (if we cannot overwrite object on free)
576 * B. Tracking data for SLAB_STORE_USER
577 * C. Padding to reach required alignment boundary or at mininum
578 * one word if debugging is on to be able to detect writes
579 * before the word boundary.
581 * Padding is done using 0x5a (POISON_INUSE)
584 * Nothing is used beyond s->size.
586 * If slabcaches are merged then the objsize and inuse boundaries are mostly
587 * ignored. And therefore no slab options that rely on these boundaries
588 * may be used with merged slabcaches.
591 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
593 unsigned long off = s->inuse; /* The end of info */
596 /* Freepointer is placed after the object. */
597 off += sizeof(void *);
599 if (s->flags & SLAB_STORE_USER)
600 /* We also have user information there */
601 off += 2 * sizeof(struct track);
606 return check_bytes_and_report(s, page, p, "Object padding",
607 p + off, POISON_INUSE, s->size - off);
610 /* Check the pad bytes at the end of a slab page */
611 static int slab_pad_check(struct kmem_cache *s, struct page *page)
619 if (!(s->flags & SLAB_POISON))
622 start = page_address(page);
623 length = (PAGE_SIZE << compound_order(page));
624 end = start + length;
625 remainder = length % s->size;
629 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
632 while (end > fault && end[-1] == POISON_INUSE)
635 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
636 print_section("Padding", end - remainder, remainder);
638 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
642 static int check_object(struct kmem_cache *s, struct page *page,
643 void *object, int active)
646 u8 *endobject = object + s->objsize;
648 if (s->flags & SLAB_RED_ZONE) {
650 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
652 if (!check_bytes_and_report(s, page, object, "Redzone",
653 endobject, red, s->inuse - s->objsize))
656 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
657 check_bytes_and_report(s, page, p, "Alignment padding",
658 endobject, POISON_INUSE, s->inuse - s->objsize);
662 if (s->flags & SLAB_POISON) {
663 if (!active && (s->flags & __OBJECT_POISON) &&
664 (!check_bytes_and_report(s, page, p, "Poison", p,
665 POISON_FREE, s->objsize - 1) ||
666 !check_bytes_and_report(s, page, p, "Poison",
667 p + s->objsize - 1, POISON_END, 1)))
670 * check_pad_bytes cleans up on its own.
672 check_pad_bytes(s, page, p);
675 if (!s->offset && active)
677 * Object and freepointer overlap. Cannot check
678 * freepointer while object is allocated.
682 /* Check free pointer validity */
683 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
684 object_err(s, page, p, "Freepointer corrupt");
686 * No choice but to zap it and thus lose the remainder
687 * of the free objects in this slab. May cause
688 * another error because the object count is now wrong.
690 set_freepointer(s, p, NULL);
696 static int check_slab(struct kmem_cache *s, struct page *page)
700 VM_BUG_ON(!irqs_disabled());
702 if (!PageSlab(page)) {
703 slab_err(s, page, "Not a valid slab page");
707 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
708 if (page->objects > maxobj) {
709 slab_err(s, page, "objects %u > max %u",
710 s->name, page->objects, maxobj);
713 if (page->inuse > page->objects) {
714 slab_err(s, page, "inuse %u > max %u",
715 s->name, page->inuse, page->objects);
718 /* Slab_pad_check fixes things up after itself */
719 slab_pad_check(s, page);
724 * Determine if a certain object on a page is on the freelist. Must hold the
725 * slab lock to guarantee that the chains are in a consistent state.
727 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
730 void *fp = page->freelist;
732 unsigned long max_objects;
734 while (fp && nr <= page->objects) {
737 if (!check_valid_pointer(s, page, fp)) {
739 object_err(s, page, object,
740 "Freechain corrupt");
741 set_freepointer(s, object, NULL);
744 slab_err(s, page, "Freepointer corrupt");
745 page->freelist = NULL;
746 page->inuse = page->objects;
747 slab_fix(s, "Freelist cleared");
753 fp = get_freepointer(s, object);
757 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
758 if (max_objects > MAX_OBJS_PER_PAGE)
759 max_objects = MAX_OBJS_PER_PAGE;
761 if (page->objects != max_objects) {
762 slab_err(s, page, "Wrong number of objects. Found %d but "
763 "should be %d", page->objects, max_objects);
764 page->objects = max_objects;
765 slab_fix(s, "Number of objects adjusted.");
767 if (page->inuse != page->objects - nr) {
768 slab_err(s, page, "Wrong object count. Counter is %d but "
769 "counted were %d", page->inuse, page->objects - nr);
770 page->inuse = page->objects - nr;
771 slab_fix(s, "Object count adjusted.");
773 return search == NULL;
776 static void trace(struct kmem_cache *s, struct page *page, void *object,
779 if (s->flags & SLAB_TRACE) {
780 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
782 alloc ? "alloc" : "free",
787 print_section("Object", (void *)object, s->objsize);
794 * Tracking of fully allocated slabs for debugging purposes.
796 static void add_full(struct kmem_cache_node *n, struct page *page)
798 spin_lock(&n->list_lock);
799 list_add(&page->lru, &n->full);
800 spin_unlock(&n->list_lock);
803 static void remove_full(struct kmem_cache *s, struct page *page)
805 struct kmem_cache_node *n;
807 if (!(s->flags & SLAB_STORE_USER))
810 n = get_node(s, page_to_nid(page));
812 spin_lock(&n->list_lock);
813 list_del(&page->lru);
814 spin_unlock(&n->list_lock);
817 /* Tracking of the number of slabs for debugging purposes */
818 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
820 struct kmem_cache_node *n = get_node(s, node);
822 return atomic_long_read(&n->nr_slabs);
825 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
827 return atomic_long_read(&n->nr_slabs);
830 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
832 struct kmem_cache_node *n = get_node(s, node);
835 * May be called early in order to allocate a slab for the
836 * kmem_cache_node structure. Solve the chicken-egg
837 * dilemma by deferring the increment of the count during
838 * bootstrap (see early_kmem_cache_node_alloc).
840 if (!NUMA_BUILD || n) {
841 atomic_long_inc(&n->nr_slabs);
842 atomic_long_add(objects, &n->total_objects);
845 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
847 struct kmem_cache_node *n = get_node(s, node);
849 atomic_long_dec(&n->nr_slabs);
850 atomic_long_sub(objects, &n->total_objects);
853 /* Object debug checks for alloc/free paths */
854 static void setup_object_debug(struct kmem_cache *s, struct page *page,
857 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
860 init_object(s, object, 0);
861 init_tracking(s, object);
864 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
865 void *object, unsigned long addr)
867 if (!check_slab(s, page))
870 if (!on_freelist(s, page, object)) {
871 object_err(s, page, object, "Object already allocated");
875 if (!check_valid_pointer(s, page, object)) {
876 object_err(s, page, object, "Freelist Pointer check fails");
880 if (!check_object(s, page, object, 0))
883 /* Success perform special debug activities for allocs */
884 if (s->flags & SLAB_STORE_USER)
885 set_track(s, object, TRACK_ALLOC, addr);
886 trace(s, page, object, 1);
887 init_object(s, object, 1);
891 if (PageSlab(page)) {
893 * If this is a slab page then lets do the best we can
894 * to avoid issues in the future. Marking all objects
895 * as used avoids touching the remaining objects.
897 slab_fix(s, "Marking all objects used");
898 page->inuse = page->objects;
899 page->freelist = NULL;
904 static noinline int free_debug_processing(struct kmem_cache *s,
905 struct page *page, void *object, unsigned long addr)
907 if (!check_slab(s, page))
910 if (!check_valid_pointer(s, page, object)) {
911 slab_err(s, page, "Invalid object pointer 0x%p", object);
915 if (on_freelist(s, page, object)) {
916 object_err(s, page, object, "Object already free");
920 if (!check_object(s, page, object, 1))
923 if (unlikely(s != page->slab)) {
924 if (!PageSlab(page)) {
925 slab_err(s, page, "Attempt to free object(0x%p) "
926 "outside of slab", object);
927 } else if (!page->slab) {
929 "SLUB <none>: no slab for object 0x%p.\n",
933 object_err(s, page, object,
934 "page slab pointer corrupt.");
938 /* Special debug activities for freeing objects */
939 if (!PageSlubFrozen(page) && !page->freelist)
940 remove_full(s, page);
941 if (s->flags & SLAB_STORE_USER)
942 set_track(s, object, TRACK_FREE, addr);
943 trace(s, page, object, 0);
944 init_object(s, object, 0);
948 slab_fix(s, "Object at 0x%p not freed", object);
952 static int __init setup_slub_debug(char *str)
954 slub_debug = DEBUG_DEFAULT_FLAGS;
955 if (*str++ != '=' || !*str)
957 * No options specified. Switch on full debugging.
963 * No options but restriction on slabs. This means full
964 * debugging for slabs matching a pattern.
968 if (tolower(*str) == 'o') {
970 * Avoid enabling debugging on caches if its minimum order
971 * would increase as a result.
973 disable_higher_order_debug = 1;
980 * Switch off all debugging measures.
985 * Determine which debug features should be switched on
987 for (; *str && *str != ','; str++) {
988 switch (tolower(*str)) {
990 slub_debug |= SLAB_DEBUG_FREE;
993 slub_debug |= SLAB_RED_ZONE;
996 slub_debug |= SLAB_POISON;
999 slub_debug |= SLAB_STORE_USER;
1002 slub_debug |= SLAB_TRACE;
1005 slub_debug |= SLAB_FAILSLAB;
1008 printk(KERN_ERR "slub_debug option '%c' "
1009 "unknown. skipped\n", *str);
1015 slub_debug_slabs = str + 1;
1020 __setup("slub_debug", setup_slub_debug);
1022 static unsigned long kmem_cache_flags(unsigned long objsize,
1023 unsigned long flags, const char *name,
1024 void (*ctor)(void *))
1027 * Enable debugging if selected on the kernel commandline.
1029 if (slub_debug && (!slub_debug_slabs ||
1030 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1031 flags |= slub_debug;
1036 static inline void setup_object_debug(struct kmem_cache *s,
1037 struct page *page, void *object) {}
1039 static inline int alloc_debug_processing(struct kmem_cache *s,
1040 struct page *page, void *object, unsigned long addr) { return 0; }
1042 static inline int free_debug_processing(struct kmem_cache *s,
1043 struct page *page, void *object, unsigned long addr) { return 0; }
1045 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1047 static inline int check_object(struct kmem_cache *s, struct page *page,
1048 void *object, int active) { return 1; }
1049 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1050 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1051 unsigned long flags, const char *name,
1052 void (*ctor)(void *))
1056 #define slub_debug 0
1058 #define disable_higher_order_debug 0
1060 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1062 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1064 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1066 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1071 * Slab allocation and freeing
1073 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1074 struct kmem_cache_order_objects oo)
1076 int order = oo_order(oo);
1078 flags |= __GFP_NOTRACK;
1080 if (node == NUMA_NO_NODE)
1081 return alloc_pages(flags, order);
1083 return alloc_pages_exact_node(node, flags, order);
1086 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1089 struct kmem_cache_order_objects oo = s->oo;
1092 flags |= s->allocflags;
1095 * Let the initial higher-order allocation fail under memory pressure
1096 * so we fall-back to the minimum order allocation.
1098 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1100 page = alloc_slab_page(alloc_gfp, node, oo);
1101 if (unlikely(!page)) {
1104 * Allocation may have failed due to fragmentation.
1105 * Try a lower order alloc if possible
1107 page = alloc_slab_page(flags, node, oo);
1111 stat(s, ORDER_FALLBACK);
1114 if (kmemcheck_enabled
1115 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1116 int pages = 1 << oo_order(oo);
1118 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1121 * Objects from caches that have a constructor don't get
1122 * cleared when they're allocated, so we need to do it here.
1125 kmemcheck_mark_uninitialized_pages(page, pages);
1127 kmemcheck_mark_unallocated_pages(page, pages);
1130 page->objects = oo_objects(oo);
1131 mod_zone_page_state(page_zone(page),
1132 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1133 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1139 static void setup_object(struct kmem_cache *s, struct page *page,
1142 setup_object_debug(s, page, object);
1143 if (unlikely(s->ctor))
1147 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1154 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1156 page = allocate_slab(s,
1157 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1161 inc_slabs_node(s, page_to_nid(page), page->objects);
1163 page->flags |= 1 << PG_slab;
1165 start = page_address(page);
1167 if (unlikely(s->flags & SLAB_POISON))
1168 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1171 for_each_object(p, s, start, page->objects) {
1172 setup_object(s, page, last);
1173 set_freepointer(s, last, p);
1176 setup_object(s, page, last);
1177 set_freepointer(s, last, NULL);
1179 page->freelist = start;
1185 static void __free_slab(struct kmem_cache *s, struct page *page)
1187 int order = compound_order(page);
1188 int pages = 1 << order;
1190 if (kmem_cache_debug(s)) {
1193 slab_pad_check(s, page);
1194 for_each_object(p, s, page_address(page),
1196 check_object(s, page, p, 0);
1199 kmemcheck_free_shadow(page, compound_order(page));
1201 mod_zone_page_state(page_zone(page),
1202 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1203 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1206 __ClearPageSlab(page);
1207 reset_page_mapcount(page);
1208 if (current->reclaim_state)
1209 current->reclaim_state->reclaimed_slab += pages;
1210 __free_pages(page, order);
1213 static void rcu_free_slab(struct rcu_head *h)
1217 page = container_of((struct list_head *)h, struct page, lru);
1218 __free_slab(page->slab, page);
1221 static void free_slab(struct kmem_cache *s, struct page *page)
1223 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1225 * RCU free overloads the RCU head over the LRU
1227 struct rcu_head *head = (void *)&page->lru;
1229 call_rcu(head, rcu_free_slab);
1231 __free_slab(s, page);
1234 static void discard_slab(struct kmem_cache *s, struct page *page)
1236 dec_slabs_node(s, page_to_nid(page), page->objects);
1241 * Per slab locking using the pagelock
1243 static __always_inline void slab_lock(struct page *page)
1245 bit_spin_lock(PG_locked, &page->flags);
1248 static __always_inline void slab_unlock(struct page *page)
1250 __bit_spin_unlock(PG_locked, &page->flags);
1253 static __always_inline int slab_trylock(struct page *page)
1257 rc = bit_spin_trylock(PG_locked, &page->flags);
1262 * Management of partially allocated slabs
1264 static void add_partial(struct kmem_cache_node *n,
1265 struct page *page, int tail)
1267 spin_lock(&n->list_lock);
1270 list_add_tail(&page->lru, &n->partial);
1272 list_add(&page->lru, &n->partial);
1273 spin_unlock(&n->list_lock);
1276 static void remove_partial(struct kmem_cache *s, struct page *page)
1278 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1280 spin_lock(&n->list_lock);
1281 list_del(&page->lru);
1283 spin_unlock(&n->list_lock);
1287 * Lock slab and remove from the partial list.
1289 * Must hold list_lock.
1291 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1294 if (slab_trylock(page)) {
1295 list_del(&page->lru);
1297 __SetPageSlubFrozen(page);
1304 * Try to allocate a partial slab from a specific node.
1306 static struct page *get_partial_node(struct kmem_cache_node *n)
1311 * Racy check. If we mistakenly see no partial slabs then we
1312 * just allocate an empty slab. If we mistakenly try to get a
1313 * partial slab and there is none available then get_partials()
1316 if (!n || !n->nr_partial)
1319 spin_lock(&n->list_lock);
1320 list_for_each_entry(page, &n->partial, lru)
1321 if (lock_and_freeze_slab(n, page))
1325 spin_unlock(&n->list_lock);
1330 * Get a page from somewhere. Search in increasing NUMA distances.
1332 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1335 struct zonelist *zonelist;
1338 enum zone_type high_zoneidx = gfp_zone(flags);
1342 * The defrag ratio allows a configuration of the tradeoffs between
1343 * inter node defragmentation and node local allocations. A lower
1344 * defrag_ratio increases the tendency to do local allocations
1345 * instead of attempting to obtain partial slabs from other nodes.
1347 * If the defrag_ratio is set to 0 then kmalloc() always
1348 * returns node local objects. If the ratio is higher then kmalloc()
1349 * may return off node objects because partial slabs are obtained
1350 * from other nodes and filled up.
1352 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1353 * defrag_ratio = 1000) then every (well almost) allocation will
1354 * first attempt to defrag slab caches on other nodes. This means
1355 * scanning over all nodes to look for partial slabs which may be
1356 * expensive if we do it every time we are trying to find a slab
1357 * with available objects.
1359 if (!s->remote_node_defrag_ratio ||
1360 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1364 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1365 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1366 struct kmem_cache_node *n;
1368 n = get_node(s, zone_to_nid(zone));
1370 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1371 n->nr_partial > s->min_partial) {
1372 page = get_partial_node(n);
1385 * Get a partial page, lock it and return it.
1387 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1390 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1392 page = get_partial_node(get_node(s, searchnode));
1393 if (page || node != -1)
1396 return get_any_partial(s, flags);
1400 * Move a page back to the lists.
1402 * Must be called with the slab lock held.
1404 * On exit the slab lock will have been dropped.
1406 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1408 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1410 __ClearPageSlubFrozen(page);
1413 if (page->freelist) {
1414 add_partial(n, page, tail);
1415 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1417 stat(s, DEACTIVATE_FULL);
1418 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1423 stat(s, DEACTIVATE_EMPTY);
1424 if (n->nr_partial < s->min_partial) {
1426 * Adding an empty slab to the partial slabs in order
1427 * to avoid page allocator overhead. This slab needs
1428 * to come after the other slabs with objects in
1429 * so that the others get filled first. That way the
1430 * size of the partial list stays small.
1432 * kmem_cache_shrink can reclaim any empty slabs from
1435 add_partial(n, page, 1);
1440 discard_slab(s, page);
1446 * Remove the cpu slab
1448 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1450 struct page *page = c->page;
1454 stat(s, DEACTIVATE_REMOTE_FREES);
1456 * Merge cpu freelist into slab freelist. Typically we get here
1457 * because both freelists are empty. So this is unlikely
1460 while (unlikely(c->freelist)) {
1463 tail = 0; /* Hot objects. Put the slab first */
1465 /* Retrieve object from cpu_freelist */
1466 object = c->freelist;
1467 c->freelist = get_freepointer(s, c->freelist);
1469 /* And put onto the regular freelist */
1470 set_freepointer(s, object, page->freelist);
1471 page->freelist = object;
1475 unfreeze_slab(s, page, tail);
1478 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1480 stat(s, CPUSLAB_FLUSH);
1482 deactivate_slab(s, c);
1488 * Called from IPI handler with interrupts disabled.
1490 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1492 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1494 if (likely(c && c->page))
1498 static void flush_cpu_slab(void *d)
1500 struct kmem_cache *s = d;
1502 __flush_cpu_slab(s, smp_processor_id());
1505 static void flush_all(struct kmem_cache *s)
1507 on_each_cpu(flush_cpu_slab, s, 1);
1511 * Check if the objects in a per cpu structure fit numa
1512 * locality expectations.
1514 static inline int node_match(struct kmem_cache_cpu *c, int node)
1517 if (node != NUMA_NO_NODE && c->node != node)
1523 static int count_free(struct page *page)
1525 return page->objects - page->inuse;
1528 static unsigned long count_partial(struct kmem_cache_node *n,
1529 int (*get_count)(struct page *))
1531 unsigned long flags;
1532 unsigned long x = 0;
1535 spin_lock_irqsave(&n->list_lock, flags);
1536 list_for_each_entry(page, &n->partial, lru)
1537 x += get_count(page);
1538 spin_unlock_irqrestore(&n->list_lock, flags);
1542 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1544 #ifdef CONFIG_SLUB_DEBUG
1545 return atomic_long_read(&n->total_objects);
1551 static noinline void
1552 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1557 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1559 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1560 "default order: %d, min order: %d\n", s->name, s->objsize,
1561 s->size, oo_order(s->oo), oo_order(s->min));
1563 if (oo_order(s->min) > get_order(s->objsize))
1564 printk(KERN_WARNING " %s debugging increased min order, use "
1565 "slub_debug=O to disable.\n", s->name);
1567 for_each_online_node(node) {
1568 struct kmem_cache_node *n = get_node(s, node);
1569 unsigned long nr_slabs;
1570 unsigned long nr_objs;
1571 unsigned long nr_free;
1576 nr_free = count_partial(n, count_free);
1577 nr_slabs = node_nr_slabs(n);
1578 nr_objs = node_nr_objs(n);
1581 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1582 node, nr_slabs, nr_objs, nr_free);
1587 * Slow path. The lockless freelist is empty or we need to perform
1590 * Interrupts are disabled.
1592 * Processing is still very fast if new objects have been freed to the
1593 * regular freelist. In that case we simply take over the regular freelist
1594 * as the lockless freelist and zap the regular freelist.
1596 * If that is not working then we fall back to the partial lists. We take the
1597 * first element of the freelist as the object to allocate now and move the
1598 * rest of the freelist to the lockless freelist.
1600 * And if we were unable to get a new slab from the partial slab lists then
1601 * we need to allocate a new slab. This is the slowest path since it involves
1602 * a call to the page allocator and the setup of a new slab.
1604 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1605 unsigned long addr, struct kmem_cache_cpu *c)
1610 /* We handle __GFP_ZERO in the caller */
1611 gfpflags &= ~__GFP_ZERO;
1617 if (unlikely(!node_match(c, node)))
1620 stat(s, ALLOC_REFILL);
1623 object = c->page->freelist;
1624 if (unlikely(!object))
1626 if (kmem_cache_debug(s))
1629 c->freelist = get_freepointer(s, object);
1630 c->page->inuse = c->page->objects;
1631 c->page->freelist = NULL;
1632 c->node = page_to_nid(c->page);
1634 slab_unlock(c->page);
1635 stat(s, ALLOC_SLOWPATH);
1639 deactivate_slab(s, c);
1642 new = get_partial(s, gfpflags, node);
1645 stat(s, ALLOC_FROM_PARTIAL);
1649 if (gfpflags & __GFP_WAIT)
1652 new = new_slab(s, gfpflags, node);
1654 if (gfpflags & __GFP_WAIT)
1655 local_irq_disable();
1658 c = __this_cpu_ptr(s->cpu_slab);
1659 stat(s, ALLOC_SLAB);
1663 __SetPageSlubFrozen(new);
1667 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1668 slab_out_of_memory(s, gfpflags, node);
1671 if (!alloc_debug_processing(s, c->page, object, addr))
1675 c->page->freelist = get_freepointer(s, object);
1681 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1682 * have the fastpath folded into their functions. So no function call
1683 * overhead for requests that can be satisfied on the fastpath.
1685 * The fastpath works by first checking if the lockless freelist can be used.
1686 * If not then __slab_alloc is called for slow processing.
1688 * Otherwise we can simply pick the next object from the lockless free list.
1690 static __always_inline void *slab_alloc(struct kmem_cache *s,
1691 gfp_t gfpflags, int node, unsigned long addr)
1694 struct kmem_cache_cpu *c;
1695 unsigned long flags;
1697 gfpflags &= gfp_allowed_mask;
1699 lockdep_trace_alloc(gfpflags);
1700 might_sleep_if(gfpflags & __GFP_WAIT);
1702 if (should_failslab(s->objsize, gfpflags, s->flags))
1705 local_irq_save(flags);
1706 c = __this_cpu_ptr(s->cpu_slab);
1707 object = c->freelist;
1708 if (unlikely(!object || !node_match(c, node)))
1710 object = __slab_alloc(s, gfpflags, node, addr, c);
1713 c->freelist = get_freepointer(s, object);
1714 stat(s, ALLOC_FASTPATH);
1716 local_irq_restore(flags);
1718 if (unlikely(gfpflags & __GFP_ZERO) && object)
1719 memset(object, 0, s->objsize);
1721 kmemcheck_slab_alloc(s, gfpflags, object, s->objsize);
1722 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, gfpflags);
1727 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1729 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1731 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1735 EXPORT_SYMBOL(kmem_cache_alloc);
1737 #ifdef CONFIG_TRACING
1738 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1740 return slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1742 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1746 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1748 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1750 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1751 s->objsize, s->size, gfpflags, node);
1755 EXPORT_SYMBOL(kmem_cache_alloc_node);
1758 #ifdef CONFIG_TRACING
1759 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1763 return slab_alloc(s, gfpflags, node, _RET_IP_);
1765 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1769 * Slow patch handling. This may still be called frequently since objects
1770 * have a longer lifetime than the cpu slabs in most processing loads.
1772 * So we still attempt to reduce cache line usage. Just take the slab
1773 * lock and free the item. If there is no additional partial page
1774 * handling required then we can return immediately.
1776 static void __slab_free(struct kmem_cache *s, struct page *page,
1777 void *x, unsigned long addr)
1780 void **object = (void *)x;
1782 stat(s, FREE_SLOWPATH);
1785 if (kmem_cache_debug(s))
1789 prior = page->freelist;
1790 set_freepointer(s, object, prior);
1791 page->freelist = object;
1794 if (unlikely(PageSlubFrozen(page))) {
1795 stat(s, FREE_FROZEN);
1799 if (unlikely(!page->inuse))
1803 * Objects left in the slab. If it was not on the partial list before
1806 if (unlikely(!prior)) {
1807 add_partial(get_node(s, page_to_nid(page)), page, 1);
1808 stat(s, FREE_ADD_PARTIAL);
1818 * Slab still on the partial list.
1820 remove_partial(s, page);
1821 stat(s, FREE_REMOVE_PARTIAL);
1825 discard_slab(s, page);
1829 if (!free_debug_processing(s, page, x, addr))
1835 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1836 * can perform fastpath freeing without additional function calls.
1838 * The fastpath is only possible if we are freeing to the current cpu slab
1839 * of this processor. This typically the case if we have just allocated
1842 * If fastpath is not possible then fall back to __slab_free where we deal
1843 * with all sorts of special processing.
1845 static __always_inline void slab_free(struct kmem_cache *s,
1846 struct page *page, void *x, unsigned long addr)
1848 void **object = (void *)x;
1849 struct kmem_cache_cpu *c;
1850 unsigned long flags;
1852 kmemleak_free_recursive(x, s->flags);
1853 local_irq_save(flags);
1854 c = __this_cpu_ptr(s->cpu_slab);
1855 kmemcheck_slab_free(s, object, s->objsize);
1856 debug_check_no_locks_freed(object, s->objsize);
1857 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1858 debug_check_no_obj_freed(object, s->objsize);
1859 if (likely(page == c->page && c->node >= 0)) {
1860 set_freepointer(s, object, c->freelist);
1861 c->freelist = object;
1862 stat(s, FREE_FASTPATH);
1864 __slab_free(s, page, x, addr);
1866 local_irq_restore(flags);
1869 void kmem_cache_free(struct kmem_cache *s, void *x)
1873 page = virt_to_head_page(x);
1875 slab_free(s, page, x, _RET_IP_);
1877 trace_kmem_cache_free(_RET_IP_, x);
1879 EXPORT_SYMBOL(kmem_cache_free);
1881 /* Figure out on which slab page the object resides */
1882 static struct page *get_object_page(const void *x)
1884 struct page *page = virt_to_head_page(x);
1886 if (!PageSlab(page))
1893 * Object placement in a slab is made very easy because we always start at
1894 * offset 0. If we tune the size of the object to the alignment then we can
1895 * get the required alignment by putting one properly sized object after
1898 * Notice that the allocation order determines the sizes of the per cpu
1899 * caches. Each processor has always one slab available for allocations.
1900 * Increasing the allocation order reduces the number of times that slabs
1901 * must be moved on and off the partial lists and is therefore a factor in
1906 * Mininum / Maximum order of slab pages. This influences locking overhead
1907 * and slab fragmentation. A higher order reduces the number of partial slabs
1908 * and increases the number of allocations possible without having to
1909 * take the list_lock.
1911 static int slub_min_order;
1912 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1913 static int slub_min_objects;
1916 * Merge control. If this is set then no merging of slab caches will occur.
1917 * (Could be removed. This was introduced to pacify the merge skeptics.)
1919 static int slub_nomerge;
1922 * Calculate the order of allocation given an slab object size.
1924 * The order of allocation has significant impact on performance and other
1925 * system components. Generally order 0 allocations should be preferred since
1926 * order 0 does not cause fragmentation in the page allocator. Larger objects
1927 * be problematic to put into order 0 slabs because there may be too much
1928 * unused space left. We go to a higher order if more than 1/16th of the slab
1931 * In order to reach satisfactory performance we must ensure that a minimum
1932 * number of objects is in one slab. Otherwise we may generate too much
1933 * activity on the partial lists which requires taking the list_lock. This is
1934 * less a concern for large slabs though which are rarely used.
1936 * slub_max_order specifies the order where we begin to stop considering the
1937 * number of objects in a slab as critical. If we reach slub_max_order then
1938 * we try to keep the page order as low as possible. So we accept more waste
1939 * of space in favor of a small page order.
1941 * Higher order allocations also allow the placement of more objects in a
1942 * slab and thereby reduce object handling overhead. If the user has
1943 * requested a higher mininum order then we start with that one instead of
1944 * the smallest order which will fit the object.
1946 static inline int slab_order(int size, int min_objects,
1947 int max_order, int fract_leftover)
1951 int min_order = slub_min_order;
1953 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1954 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1956 for (order = max(min_order,
1957 fls(min_objects * size - 1) - PAGE_SHIFT);
1958 order <= max_order; order++) {
1960 unsigned long slab_size = PAGE_SIZE << order;
1962 if (slab_size < min_objects * size)
1965 rem = slab_size % size;
1967 if (rem <= slab_size / fract_leftover)
1975 static inline int calculate_order(int size)
1983 * Attempt to find best configuration for a slab. This
1984 * works by first attempting to generate a layout with
1985 * the best configuration and backing off gradually.
1987 * First we reduce the acceptable waste in a slab. Then
1988 * we reduce the minimum objects required in a slab.
1990 min_objects = slub_min_objects;
1992 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1993 max_objects = (PAGE_SIZE << slub_max_order)/size;
1994 min_objects = min(min_objects, max_objects);
1996 while (min_objects > 1) {
1998 while (fraction >= 4) {
1999 order = slab_order(size, min_objects,
2000 slub_max_order, fraction);
2001 if (order <= slub_max_order)
2009 * We were unable to place multiple objects in a slab. Now
2010 * lets see if we can place a single object there.
2012 order = slab_order(size, 1, slub_max_order, 1);
2013 if (order <= slub_max_order)
2017 * Doh this slab cannot be placed using slub_max_order.
2019 order = slab_order(size, 1, MAX_ORDER, 1);
2020 if (order < MAX_ORDER)
2026 * Figure out what the alignment of the objects will be.
2028 static unsigned long calculate_alignment(unsigned long flags,
2029 unsigned long align, unsigned long size)
2032 * If the user wants hardware cache aligned objects then follow that
2033 * suggestion if the object is sufficiently large.
2035 * The hardware cache alignment cannot override the specified
2036 * alignment though. If that is greater then use it.
2038 if (flags & SLAB_HWCACHE_ALIGN) {
2039 unsigned long ralign = cache_line_size();
2040 while (size <= ralign / 2)
2042 align = max(align, ralign);
2045 if (align < ARCH_SLAB_MINALIGN)
2046 align = ARCH_SLAB_MINALIGN;
2048 return ALIGN(align, sizeof(void *));
2052 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2055 spin_lock_init(&n->list_lock);
2056 INIT_LIST_HEAD(&n->partial);
2057 #ifdef CONFIG_SLUB_DEBUG
2058 atomic_long_set(&n->nr_slabs, 0);
2059 atomic_long_set(&n->total_objects, 0);
2060 INIT_LIST_HEAD(&n->full);
2064 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2066 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2067 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2069 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2071 return s->cpu_slab != NULL;
2075 static struct kmem_cache *kmem_cache_node;
2078 * No kmalloc_node yet so do it by hand. We know that this is the first
2079 * slab on the node for this slabcache. There are no concurrent accesses
2082 * Note that this function only works on the kmalloc_node_cache
2083 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2084 * memory on a fresh node that has no slab structures yet.
2086 static void early_kmem_cache_node_alloc(int node)
2089 struct kmem_cache_node *n;
2090 unsigned long flags;
2092 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2094 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2097 if (page_to_nid(page) != node) {
2098 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2100 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2101 "in order to be able to continue\n");
2106 page->freelist = get_freepointer(kmem_cache_node, n);
2108 kmem_cache_node->node[node] = n;
2109 #ifdef CONFIG_SLUB_DEBUG
2110 init_object(kmem_cache_node, n, 1);
2111 init_tracking(kmem_cache_node, n);
2113 init_kmem_cache_node(n, kmem_cache_node);
2114 inc_slabs_node(kmem_cache_node, node, page->objects);
2117 * lockdep requires consistent irq usage for each lock
2118 * so even though there cannot be a race this early in
2119 * the boot sequence, we still disable irqs.
2121 local_irq_save(flags);
2122 add_partial(n, page, 0);
2123 local_irq_restore(flags);
2126 static void free_kmem_cache_nodes(struct kmem_cache *s)
2130 for_each_node_state(node, N_NORMAL_MEMORY) {
2131 struct kmem_cache_node *n = s->node[node];
2134 kmem_cache_free(kmem_cache_node, n);
2136 s->node[node] = NULL;
2140 static int init_kmem_cache_nodes(struct kmem_cache *s)
2144 for_each_node_state(node, N_NORMAL_MEMORY) {
2145 struct kmem_cache_node *n;
2147 if (slab_state == DOWN) {
2148 early_kmem_cache_node_alloc(node);
2151 n = kmem_cache_alloc_node(kmem_cache_node,
2155 free_kmem_cache_nodes(s);
2160 init_kmem_cache_node(n, s);
2165 static void free_kmem_cache_nodes(struct kmem_cache *s)
2169 static int init_kmem_cache_nodes(struct kmem_cache *s)
2171 init_kmem_cache_node(&s->local_node, s);
2176 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2178 if (min < MIN_PARTIAL)
2180 else if (min > MAX_PARTIAL)
2182 s->min_partial = min;
2186 * calculate_sizes() determines the order and the distribution of data within
2189 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2191 unsigned long flags = s->flags;
2192 unsigned long size = s->objsize;
2193 unsigned long align = s->align;
2197 * Round up object size to the next word boundary. We can only
2198 * place the free pointer at word boundaries and this determines
2199 * the possible location of the free pointer.
2201 size = ALIGN(size, sizeof(void *));
2203 #ifdef CONFIG_SLUB_DEBUG
2205 * Determine if we can poison the object itself. If the user of
2206 * the slab may touch the object after free or before allocation
2207 * then we should never poison the object itself.
2209 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2211 s->flags |= __OBJECT_POISON;
2213 s->flags &= ~__OBJECT_POISON;
2217 * If we are Redzoning then check if there is some space between the
2218 * end of the object and the free pointer. If not then add an
2219 * additional word to have some bytes to store Redzone information.
2221 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2222 size += sizeof(void *);
2226 * With that we have determined the number of bytes in actual use
2227 * by the object. This is the potential offset to the free pointer.
2231 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2234 * Relocate free pointer after the object if it is not
2235 * permitted to overwrite the first word of the object on
2238 * This is the case if we do RCU, have a constructor or
2239 * destructor or are poisoning the objects.
2242 size += sizeof(void *);
2245 #ifdef CONFIG_SLUB_DEBUG
2246 if (flags & SLAB_STORE_USER)
2248 * Need to store information about allocs and frees after
2251 size += 2 * sizeof(struct track);
2253 if (flags & SLAB_RED_ZONE)
2255 * Add some empty padding so that we can catch
2256 * overwrites from earlier objects rather than let
2257 * tracking information or the free pointer be
2258 * corrupted if a user writes before the start
2261 size += sizeof(void *);
2265 * Determine the alignment based on various parameters that the
2266 * user specified and the dynamic determination of cache line size
2269 align = calculate_alignment(flags, align, s->objsize);
2273 * SLUB stores one object immediately after another beginning from
2274 * offset 0. In order to align the objects we have to simply size
2275 * each object to conform to the alignment.
2277 size = ALIGN(size, align);
2279 if (forced_order >= 0)
2280 order = forced_order;
2282 order = calculate_order(size);
2289 s->allocflags |= __GFP_COMP;
2291 if (s->flags & SLAB_CACHE_DMA)
2292 s->allocflags |= SLUB_DMA;
2294 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2295 s->allocflags |= __GFP_RECLAIMABLE;
2298 * Determine the number of objects per slab
2300 s->oo = oo_make(order, size);
2301 s->min = oo_make(get_order(size), size);
2302 if (oo_objects(s->oo) > oo_objects(s->max))
2305 return !!oo_objects(s->oo);
2309 static int kmem_cache_open(struct kmem_cache *s,
2310 const char *name, size_t size,
2311 size_t align, unsigned long flags,
2312 void (*ctor)(void *))
2314 memset(s, 0, kmem_size);
2319 s->flags = kmem_cache_flags(size, flags, name, ctor);
2321 if (!calculate_sizes(s, -1))
2323 if (disable_higher_order_debug) {
2325 * Disable debugging flags that store metadata if the min slab
2328 if (get_order(s->size) > get_order(s->objsize)) {
2329 s->flags &= ~DEBUG_METADATA_FLAGS;
2331 if (!calculate_sizes(s, -1))
2337 * The larger the object size is, the more pages we want on the partial
2338 * list to avoid pounding the page allocator excessively.
2340 set_min_partial(s, ilog2(s->size));
2343 s->remote_node_defrag_ratio = 1000;
2345 if (!init_kmem_cache_nodes(s))
2348 if (alloc_kmem_cache_cpus(s))
2351 free_kmem_cache_nodes(s);
2353 if (flags & SLAB_PANIC)
2354 panic("Cannot create slab %s size=%lu realsize=%u "
2355 "order=%u offset=%u flags=%lx\n",
2356 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2362 * Check if a given pointer is valid
2364 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2368 if (!kern_ptr_validate(object, s->size))
2371 page = get_object_page(object);
2373 if (!page || s != page->slab)
2374 /* No slab or wrong slab */
2377 if (!check_valid_pointer(s, page, object))
2381 * We could also check if the object is on the slabs freelist.
2382 * But this would be too expensive and it seems that the main
2383 * purpose of kmem_ptr_valid() is to check if the object belongs
2384 * to a certain slab.
2388 EXPORT_SYMBOL(kmem_ptr_validate);
2391 * Determine the size of a slab object
2393 unsigned int kmem_cache_size(struct kmem_cache *s)
2397 EXPORT_SYMBOL(kmem_cache_size);
2399 const char *kmem_cache_name(struct kmem_cache *s)
2403 EXPORT_SYMBOL(kmem_cache_name);
2405 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2408 #ifdef CONFIG_SLUB_DEBUG
2409 void *addr = page_address(page);
2411 long *map = kzalloc(BITS_TO_LONGS(page->objects) * sizeof(long),
2416 slab_err(s, page, "%s", text);
2418 for_each_free_object(p, s, page->freelist)
2419 set_bit(slab_index(p, s, addr), map);
2421 for_each_object(p, s, addr, page->objects) {
2423 if (!test_bit(slab_index(p, s, addr), map)) {
2424 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2426 print_tracking(s, p);
2435 * Attempt to free all partial slabs on a node.
2437 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2439 unsigned long flags;
2440 struct page *page, *h;
2442 spin_lock_irqsave(&n->list_lock, flags);
2443 list_for_each_entry_safe(page, h, &n->partial, lru) {
2445 list_del(&page->lru);
2446 discard_slab(s, page);
2449 list_slab_objects(s, page,
2450 "Objects remaining on kmem_cache_close()");
2453 spin_unlock_irqrestore(&n->list_lock, flags);
2457 * Release all resources used by a slab cache.
2459 static inline int kmem_cache_close(struct kmem_cache *s)
2464 free_percpu(s->cpu_slab);
2465 /* Attempt to free all objects */
2466 for_each_node_state(node, N_NORMAL_MEMORY) {
2467 struct kmem_cache_node *n = get_node(s, node);
2470 if (n->nr_partial || slabs_node(s, node))
2473 free_kmem_cache_nodes(s);
2478 * Close a cache and release the kmem_cache structure
2479 * (must be used for caches created using kmem_cache_create)
2481 void kmem_cache_destroy(struct kmem_cache *s)
2483 down_write(&slub_lock);
2487 if (kmem_cache_close(s)) {
2488 printk(KERN_ERR "SLUB %s: %s called for cache that "
2489 "still has objects.\n", s->name, __func__);
2492 if (s->flags & SLAB_DESTROY_BY_RCU)
2494 sysfs_slab_remove(s);
2496 up_write(&slub_lock);
2498 EXPORT_SYMBOL(kmem_cache_destroy);
2500 /********************************************************************
2502 *******************************************************************/
2504 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2505 EXPORT_SYMBOL(kmalloc_caches);
2507 static struct kmem_cache *kmem_cache;
2509 #ifdef CONFIG_ZONE_DMA
2510 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2513 static int __init setup_slub_min_order(char *str)
2515 get_option(&str, &slub_min_order);
2520 __setup("slub_min_order=", setup_slub_min_order);
2522 static int __init setup_slub_max_order(char *str)
2524 get_option(&str, &slub_max_order);
2525 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2530 __setup("slub_max_order=", setup_slub_max_order);
2532 static int __init setup_slub_min_objects(char *str)
2534 get_option(&str, &slub_min_objects);
2539 __setup("slub_min_objects=", setup_slub_min_objects);
2541 static int __init setup_slub_nomerge(char *str)
2547 __setup("slub_nomerge", setup_slub_nomerge);
2549 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2550 int size, unsigned int flags)
2552 struct kmem_cache *s;
2554 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2557 * This function is called with IRQs disabled during early-boot on
2558 * single CPU so there's no need to take slub_lock here.
2560 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2564 list_add(&s->list, &slab_caches);
2568 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2573 * Conversion table for small slabs sizes / 8 to the index in the
2574 * kmalloc array. This is necessary for slabs < 192 since we have non power
2575 * of two cache sizes there. The size of larger slabs can be determined using
2578 static s8 size_index[24] = {
2605 static inline int size_index_elem(size_t bytes)
2607 return (bytes - 1) / 8;
2610 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2616 return ZERO_SIZE_PTR;
2618 index = size_index[size_index_elem(size)];
2620 index = fls(size - 1);
2622 #ifdef CONFIG_ZONE_DMA
2623 if (unlikely((flags & SLUB_DMA)))
2624 return kmalloc_dma_caches[index];
2627 return kmalloc_caches[index];
2630 void *__kmalloc(size_t size, gfp_t flags)
2632 struct kmem_cache *s;
2635 if (unlikely(size > SLUB_MAX_SIZE))
2636 return kmalloc_large(size, flags);
2638 s = get_slab(size, flags);
2640 if (unlikely(ZERO_OR_NULL_PTR(s)))
2643 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2645 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2649 EXPORT_SYMBOL(__kmalloc);
2651 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2656 flags |= __GFP_COMP | __GFP_NOTRACK;
2657 page = alloc_pages_node(node, flags, get_order(size));
2659 ptr = page_address(page);
2661 kmemleak_alloc(ptr, size, 1, flags);
2666 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2668 struct kmem_cache *s;
2671 if (unlikely(size > SLUB_MAX_SIZE)) {
2672 ret = kmalloc_large_node(size, flags, node);
2674 trace_kmalloc_node(_RET_IP_, ret,
2675 size, PAGE_SIZE << get_order(size),
2681 s = get_slab(size, flags);
2683 if (unlikely(ZERO_OR_NULL_PTR(s)))
2686 ret = slab_alloc(s, flags, node, _RET_IP_);
2688 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2692 EXPORT_SYMBOL(__kmalloc_node);
2695 size_t ksize(const void *object)
2698 struct kmem_cache *s;
2700 if (unlikely(object == ZERO_SIZE_PTR))
2703 page = virt_to_head_page(object);
2705 if (unlikely(!PageSlab(page))) {
2706 WARN_ON(!PageCompound(page));
2707 return PAGE_SIZE << compound_order(page);
2711 #ifdef CONFIG_SLUB_DEBUG
2713 * Debugging requires use of the padding between object
2714 * and whatever may come after it.
2716 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2721 * If we have the need to store the freelist pointer
2722 * back there or track user information then we can
2723 * only use the space before that information.
2725 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2728 * Else we can use all the padding etc for the allocation
2732 EXPORT_SYMBOL(ksize);
2734 void kfree(const void *x)
2737 void *object = (void *)x;
2739 trace_kfree(_RET_IP_, x);
2741 if (unlikely(ZERO_OR_NULL_PTR(x)))
2744 page = virt_to_head_page(x);
2745 if (unlikely(!PageSlab(page))) {
2746 BUG_ON(!PageCompound(page));
2751 slab_free(page->slab, page, object, _RET_IP_);
2753 EXPORT_SYMBOL(kfree);
2756 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2757 * the remaining slabs by the number of items in use. The slabs with the
2758 * most items in use come first. New allocations will then fill those up
2759 * and thus they can be removed from the partial lists.
2761 * The slabs with the least items are placed last. This results in them
2762 * being allocated from last increasing the chance that the last objects
2763 * are freed in them.
2765 int kmem_cache_shrink(struct kmem_cache *s)
2769 struct kmem_cache_node *n;
2772 int objects = oo_objects(s->max);
2773 struct list_head *slabs_by_inuse =
2774 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2775 unsigned long flags;
2777 if (!slabs_by_inuse)
2781 for_each_node_state(node, N_NORMAL_MEMORY) {
2782 n = get_node(s, node);
2787 for (i = 0; i < objects; i++)
2788 INIT_LIST_HEAD(slabs_by_inuse + i);
2790 spin_lock_irqsave(&n->list_lock, flags);
2793 * Build lists indexed by the items in use in each slab.
2795 * Note that concurrent frees may occur while we hold the
2796 * list_lock. page->inuse here is the upper limit.
2798 list_for_each_entry_safe(page, t, &n->partial, lru) {
2799 if (!page->inuse && slab_trylock(page)) {
2801 * Must hold slab lock here because slab_free
2802 * may have freed the last object and be
2803 * waiting to release the slab.
2805 list_del(&page->lru);
2808 discard_slab(s, page);
2810 list_move(&page->lru,
2811 slabs_by_inuse + page->inuse);
2816 * Rebuild the partial list with the slabs filled up most
2817 * first and the least used slabs at the end.
2819 for (i = objects - 1; i >= 0; i--)
2820 list_splice(slabs_by_inuse + i, n->partial.prev);
2822 spin_unlock_irqrestore(&n->list_lock, flags);
2825 kfree(slabs_by_inuse);
2828 EXPORT_SYMBOL(kmem_cache_shrink);
2830 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2831 static int slab_mem_going_offline_callback(void *arg)
2833 struct kmem_cache *s;
2835 down_read(&slub_lock);
2836 list_for_each_entry(s, &slab_caches, list)
2837 kmem_cache_shrink(s);
2838 up_read(&slub_lock);
2843 static void slab_mem_offline_callback(void *arg)
2845 struct kmem_cache_node *n;
2846 struct kmem_cache *s;
2847 struct memory_notify *marg = arg;
2850 offline_node = marg->status_change_nid;
2853 * If the node still has available memory. we need kmem_cache_node
2856 if (offline_node < 0)
2859 down_read(&slub_lock);
2860 list_for_each_entry(s, &slab_caches, list) {
2861 n = get_node(s, offline_node);
2864 * if n->nr_slabs > 0, slabs still exist on the node
2865 * that is going down. We were unable to free them,
2866 * and offline_pages() function shouldn't call this
2867 * callback. So, we must fail.
2869 BUG_ON(slabs_node(s, offline_node));
2871 s->node[offline_node] = NULL;
2872 kmem_cache_free(kmalloc_caches, n);
2875 up_read(&slub_lock);
2878 static int slab_mem_going_online_callback(void *arg)
2880 struct kmem_cache_node *n;
2881 struct kmem_cache *s;
2882 struct memory_notify *marg = arg;
2883 int nid = marg->status_change_nid;
2887 * If the node's memory is already available, then kmem_cache_node is
2888 * already created. Nothing to do.
2894 * We are bringing a node online. No memory is available yet. We must
2895 * allocate a kmem_cache_node structure in order to bring the node
2898 down_read(&slub_lock);
2899 list_for_each_entry(s, &slab_caches, list) {
2901 * XXX: kmem_cache_alloc_node will fallback to other nodes
2902 * since memory is not yet available from the node that
2905 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2910 init_kmem_cache_node(n, s);
2914 up_read(&slub_lock);
2918 static int slab_memory_callback(struct notifier_block *self,
2919 unsigned long action, void *arg)
2924 case MEM_GOING_ONLINE:
2925 ret = slab_mem_going_online_callback(arg);
2927 case MEM_GOING_OFFLINE:
2928 ret = slab_mem_going_offline_callback(arg);
2931 case MEM_CANCEL_ONLINE:
2932 slab_mem_offline_callback(arg);
2935 case MEM_CANCEL_OFFLINE:
2939 ret = notifier_from_errno(ret);
2945 #endif /* CONFIG_MEMORY_HOTPLUG */
2947 /********************************************************************
2948 * Basic setup of slabs
2949 *******************************************************************/
2952 * Used for early kmem_cache structures that were allocated using
2953 * the page allocator
2956 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
2960 list_add(&s->list, &slab_caches);
2963 for_each_node_state(node, N_NORMAL_MEMORY) {
2964 struct kmem_cache_node *n = get_node(s, node);
2968 list_for_each_entry(p, &n->partial, lru)
2971 #ifdef CONFIG_SLAB_DEBUG
2972 list_for_each_entry(p, &n->full, lru)
2979 void __init kmem_cache_init(void)
2983 struct kmem_cache *temp_kmem_cache;
2987 struct kmem_cache *temp_kmem_cache_node;
2988 unsigned long kmalloc_size;
2990 kmem_size = offsetof(struct kmem_cache, node) +
2991 nr_node_ids * sizeof(struct kmem_cache_node *);
2993 /* Allocate two kmem_caches from the page allocator */
2994 kmalloc_size = ALIGN(kmem_size, cache_line_size());
2995 order = get_order(2 * kmalloc_size);
2996 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
2999 * Must first have the slab cache available for the allocations of the
3000 * struct kmem_cache_node's. There is special bootstrap code in
3001 * kmem_cache_open for slab_state == DOWN.
3003 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3005 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3006 sizeof(struct kmem_cache_node),
3007 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3009 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3011 /* Allocate a single kmem_cache from the page allocator */
3012 kmem_size = sizeof(struct kmem_cache);
3013 order = get_order(kmem_size);
3014 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3017 /* Able to allocate the per node structures */
3018 slab_state = PARTIAL;
3020 temp_kmem_cache = kmem_cache;
3021 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3022 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3023 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3024 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3028 * Allocate kmem_cache_node properly from the kmem_cache slab.
3029 * kmem_cache_node is separately allocated so no need to
3030 * update any list pointers.
3032 temp_kmem_cache_node = kmem_cache_node;
3034 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3035 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3037 kmem_cache_bootstrap_fixup(kmem_cache_node);
3042 * kmem_cache has kmem_cache_node embedded and we moved it!
3043 * Update the list heads
3045 INIT_LIST_HEAD(&kmem_cache->local_node.partial);
3046 list_splice(&temp_kmem_cache->local_node.partial, &kmem_cache->local_node.partial);
3047 #ifdef CONFIG_SLUB_DEBUG
3048 INIT_LIST_HEAD(&kmem_cache->local_node.full);
3049 list_splice(&temp_kmem_cache->local_node.full, &kmem_cache->local_node.full);
3052 kmem_cache_bootstrap_fixup(kmem_cache);
3054 /* Free temporary boot structure */
3055 free_pages((unsigned long)temp_kmem_cache, order);
3057 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3060 * Patch up the size_index table if we have strange large alignment
3061 * requirements for the kmalloc array. This is only the case for
3062 * MIPS it seems. The standard arches will not generate any code here.
3064 * Largest permitted alignment is 256 bytes due to the way we
3065 * handle the index determination for the smaller caches.
3067 * Make sure that nothing crazy happens if someone starts tinkering
3068 * around with ARCH_KMALLOC_MINALIGN
3070 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3071 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3073 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3074 int elem = size_index_elem(i);
3075 if (elem >= ARRAY_SIZE(size_index))
3077 size_index[elem] = KMALLOC_SHIFT_LOW;
3080 if (KMALLOC_MIN_SIZE == 64) {
3082 * The 96 byte size cache is not used if the alignment
3085 for (i = 64 + 8; i <= 96; i += 8)
3086 size_index[size_index_elem(i)] = 7;
3087 } else if (KMALLOC_MIN_SIZE == 128) {
3089 * The 192 byte sized cache is not used if the alignment
3090 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3093 for (i = 128 + 8; i <= 192; i += 8)
3094 size_index[size_index_elem(i)] = 8;
3097 /* Caches that are not of the two-to-the-power-of size */
3098 if (KMALLOC_MIN_SIZE <= 32) {
3099 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3103 if (KMALLOC_MIN_SIZE <= 64) {
3104 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3108 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3109 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3115 /* Provide the correct kmalloc names now that the caches are up */
3116 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3117 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3120 kmalloc_caches[i]->name = s;
3124 register_cpu_notifier(&slab_notifier);
3127 #ifdef CONFIG_ZONE_DMA
3128 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3129 struct kmem_cache *s = kmalloc_caches[i];
3132 char *name = kasprintf(GFP_NOWAIT,
3133 "dma-kmalloc-%d", s->objsize);
3136 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3137 s->objsize, SLAB_CACHE_DMA);
3142 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3143 " CPUs=%d, Nodes=%d\n",
3144 caches, cache_line_size(),
3145 slub_min_order, slub_max_order, slub_min_objects,
3146 nr_cpu_ids, nr_node_ids);
3149 void __init kmem_cache_init_late(void)
3154 * Find a mergeable slab cache
3156 static int slab_unmergeable(struct kmem_cache *s)
3158 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3165 * We may have set a slab to be unmergeable during bootstrap.
3167 if (s->refcount < 0)
3173 static struct kmem_cache *find_mergeable(size_t size,
3174 size_t align, unsigned long flags, const char *name,
3175 void (*ctor)(void *))
3177 struct kmem_cache *s;
3179 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3185 size = ALIGN(size, sizeof(void *));
3186 align = calculate_alignment(flags, align, size);
3187 size = ALIGN(size, align);
3188 flags = kmem_cache_flags(size, flags, name, NULL);
3190 list_for_each_entry(s, &slab_caches, list) {
3191 if (slab_unmergeable(s))
3197 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3200 * Check if alignment is compatible.
3201 * Courtesy of Adrian Drzewiecki
3203 if ((s->size & ~(align - 1)) != s->size)
3206 if (s->size - size >= sizeof(void *))
3214 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3215 size_t align, unsigned long flags, void (*ctor)(void *))
3217 struct kmem_cache *s;
3222 down_write(&slub_lock);
3223 s = find_mergeable(size, align, flags, name, ctor);
3227 * Adjust the object sizes so that we clear
3228 * the complete object on kzalloc.
3230 s->objsize = max(s->objsize, (int)size);
3231 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3233 if (sysfs_slab_alias(s, name)) {
3237 up_write(&slub_lock);
3241 s = kmalloc(kmem_size, GFP_KERNEL);
3243 if (kmem_cache_open(s, name,
3244 size, align, flags, ctor)) {
3245 list_add(&s->list, &slab_caches);
3246 if (sysfs_slab_add(s)) {
3251 up_write(&slub_lock);
3256 up_write(&slub_lock);
3259 if (flags & SLAB_PANIC)
3260 panic("Cannot create slabcache %s\n", name);
3265 EXPORT_SYMBOL(kmem_cache_create);
3269 * Use the cpu notifier to insure that the cpu slabs are flushed when
3272 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3273 unsigned long action, void *hcpu)
3275 long cpu = (long)hcpu;
3276 struct kmem_cache *s;
3277 unsigned long flags;
3280 case CPU_UP_CANCELED:
3281 case CPU_UP_CANCELED_FROZEN:
3283 case CPU_DEAD_FROZEN:
3284 down_read(&slub_lock);
3285 list_for_each_entry(s, &slab_caches, list) {
3286 local_irq_save(flags);
3287 __flush_cpu_slab(s, cpu);
3288 local_irq_restore(flags);
3290 up_read(&slub_lock);
3298 static struct notifier_block __cpuinitdata slab_notifier = {
3299 .notifier_call = slab_cpuup_callback
3304 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3306 struct kmem_cache *s;
3309 if (unlikely(size > SLUB_MAX_SIZE))
3310 return kmalloc_large(size, gfpflags);
3312 s = get_slab(size, gfpflags);
3314 if (unlikely(ZERO_OR_NULL_PTR(s)))
3317 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3319 /* Honor the call site pointer we recieved. */
3320 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3325 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3326 int node, unsigned long caller)
3328 struct kmem_cache *s;
3331 if (unlikely(size > SLUB_MAX_SIZE)) {
3332 ret = kmalloc_large_node(size, gfpflags, node);
3334 trace_kmalloc_node(caller, ret,
3335 size, PAGE_SIZE << get_order(size),
3341 s = get_slab(size, gfpflags);
3343 if (unlikely(ZERO_OR_NULL_PTR(s)))
3346 ret = slab_alloc(s, gfpflags, node, caller);
3348 /* Honor the call site pointer we recieved. */
3349 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3354 #ifdef CONFIG_SLUB_DEBUG
3355 static int count_inuse(struct page *page)
3360 static int count_total(struct page *page)
3362 return page->objects;
3365 static int validate_slab(struct kmem_cache *s, struct page *page,
3369 void *addr = page_address(page);
3371 if (!check_slab(s, page) ||
3372 !on_freelist(s, page, NULL))
3375 /* Now we know that a valid freelist exists */
3376 bitmap_zero(map, page->objects);
3378 for_each_free_object(p, s, page->freelist) {
3379 set_bit(slab_index(p, s, addr), map);
3380 if (!check_object(s, page, p, 0))
3384 for_each_object(p, s, addr, page->objects)
3385 if (!test_bit(slab_index(p, s, addr), map))
3386 if (!check_object(s, page, p, 1))
3391 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3394 if (slab_trylock(page)) {
3395 validate_slab(s, page, map);
3398 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3402 static int validate_slab_node(struct kmem_cache *s,
3403 struct kmem_cache_node *n, unsigned long *map)
3405 unsigned long count = 0;
3407 unsigned long flags;
3409 spin_lock_irqsave(&n->list_lock, flags);
3411 list_for_each_entry(page, &n->partial, lru) {
3412 validate_slab_slab(s, page, map);
3415 if (count != n->nr_partial)
3416 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3417 "counter=%ld\n", s->name, count, n->nr_partial);
3419 if (!(s->flags & SLAB_STORE_USER))
3422 list_for_each_entry(page, &n->full, lru) {
3423 validate_slab_slab(s, page, map);
3426 if (count != atomic_long_read(&n->nr_slabs))
3427 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3428 "counter=%ld\n", s->name, count,
3429 atomic_long_read(&n->nr_slabs));
3432 spin_unlock_irqrestore(&n->list_lock, flags);
3436 static long validate_slab_cache(struct kmem_cache *s)
3439 unsigned long count = 0;
3440 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3441 sizeof(unsigned long), GFP_KERNEL);
3447 for_each_node_state(node, N_NORMAL_MEMORY) {
3448 struct kmem_cache_node *n = get_node(s, node);
3450 count += validate_slab_node(s, n, map);
3456 #ifdef SLUB_RESILIENCY_TEST
3457 static void resiliency_test(void)
3461 printk(KERN_ERR "SLUB resiliency testing\n");
3462 printk(KERN_ERR "-----------------------\n");
3463 printk(KERN_ERR "A. Corruption after allocation\n");
3465 p = kzalloc(16, GFP_KERNEL);
3467 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3468 " 0x12->0x%p\n\n", p + 16);
3470 validate_slab_cache(kmalloc_caches + 4);
3472 /* Hmmm... The next two are dangerous */
3473 p = kzalloc(32, GFP_KERNEL);
3474 p[32 + sizeof(void *)] = 0x34;
3475 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3476 " 0x34 -> -0x%p\n", p);
3478 "If allocated object is overwritten then not detectable\n\n");
3480 validate_slab_cache(kmalloc_caches + 5);
3481 p = kzalloc(64, GFP_KERNEL);
3482 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3484 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3487 "If allocated object is overwritten then not detectable\n\n");
3488 validate_slab_cache(kmalloc_caches + 6);
3490 printk(KERN_ERR "\nB. Corruption after free\n");
3491 p = kzalloc(128, GFP_KERNEL);
3494 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3495 validate_slab_cache(kmalloc_caches + 7);
3497 p = kzalloc(256, GFP_KERNEL);
3500 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3502 validate_slab_cache(kmalloc_caches + 8);
3504 p = kzalloc(512, GFP_KERNEL);
3507 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3508 validate_slab_cache(kmalloc_caches + 9);
3511 static void resiliency_test(void) {};
3515 * Generate lists of code addresses where slabcache objects are allocated
3520 unsigned long count;
3527 DECLARE_BITMAP(cpus, NR_CPUS);
3533 unsigned long count;
3534 struct location *loc;
3537 static void free_loc_track(struct loc_track *t)
3540 free_pages((unsigned long)t->loc,
3541 get_order(sizeof(struct location) * t->max));
3544 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3549 order = get_order(sizeof(struct location) * max);
3551 l = (void *)__get_free_pages(flags, order);
3556 memcpy(l, t->loc, sizeof(struct location) * t->count);
3564 static int add_location(struct loc_track *t, struct kmem_cache *s,
3565 const struct track *track)
3567 long start, end, pos;
3569 unsigned long caddr;
3570 unsigned long age = jiffies - track->when;
3576 pos = start + (end - start + 1) / 2;
3579 * There is nothing at "end". If we end up there
3580 * we need to add something to before end.
3585 caddr = t->loc[pos].addr;
3586 if (track->addr == caddr) {
3592 if (age < l->min_time)
3594 if (age > l->max_time)
3597 if (track->pid < l->min_pid)
3598 l->min_pid = track->pid;
3599 if (track->pid > l->max_pid)
3600 l->max_pid = track->pid;
3602 cpumask_set_cpu(track->cpu,
3603 to_cpumask(l->cpus));
3605 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3609 if (track->addr < caddr)
3616 * Not found. Insert new tracking element.
3618 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3624 (t->count - pos) * sizeof(struct location));
3627 l->addr = track->addr;
3631 l->min_pid = track->pid;
3632 l->max_pid = track->pid;
3633 cpumask_clear(to_cpumask(l->cpus));
3634 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3635 nodes_clear(l->nodes);
3636 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3640 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3641 struct page *page, enum track_item alloc,
3644 void *addr = page_address(page);
3647 bitmap_zero(map, page->objects);
3648 for_each_free_object(p, s, page->freelist)
3649 set_bit(slab_index(p, s, addr), map);
3651 for_each_object(p, s, addr, page->objects)
3652 if (!test_bit(slab_index(p, s, addr), map))
3653 add_location(t, s, get_track(s, p, alloc));
3656 static int list_locations(struct kmem_cache *s, char *buf,
3657 enum track_item alloc)
3661 struct loc_track t = { 0, 0, NULL };
3663 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3664 sizeof(unsigned long), GFP_KERNEL);
3666 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3669 return sprintf(buf, "Out of memory\n");
3671 /* Push back cpu slabs */
3674 for_each_node_state(node, N_NORMAL_MEMORY) {
3675 struct kmem_cache_node *n = get_node(s, node);
3676 unsigned long flags;
3679 if (!atomic_long_read(&n->nr_slabs))
3682 spin_lock_irqsave(&n->list_lock, flags);
3683 list_for_each_entry(page, &n->partial, lru)
3684 process_slab(&t, s, page, alloc, map);
3685 list_for_each_entry(page, &n->full, lru)
3686 process_slab(&t, s, page, alloc, map);
3687 spin_unlock_irqrestore(&n->list_lock, flags);
3690 for (i = 0; i < t.count; i++) {
3691 struct location *l = &t.loc[i];
3693 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3695 len += sprintf(buf + len, "%7ld ", l->count);
3698 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3700 len += sprintf(buf + len, "<not-available>");
3702 if (l->sum_time != l->min_time) {
3703 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3705 (long)div_u64(l->sum_time, l->count),
3708 len += sprintf(buf + len, " age=%ld",
3711 if (l->min_pid != l->max_pid)
3712 len += sprintf(buf + len, " pid=%ld-%ld",
3713 l->min_pid, l->max_pid);
3715 len += sprintf(buf + len, " pid=%ld",
3718 if (num_online_cpus() > 1 &&
3719 !cpumask_empty(to_cpumask(l->cpus)) &&
3720 len < PAGE_SIZE - 60) {
3721 len += sprintf(buf + len, " cpus=");
3722 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3723 to_cpumask(l->cpus));
3726 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3727 len < PAGE_SIZE - 60) {
3728 len += sprintf(buf + len, " nodes=");
3729 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3733 len += sprintf(buf + len, "\n");
3739 len += sprintf(buf, "No data\n");
3743 enum slab_stat_type {
3744 SL_ALL, /* All slabs */
3745 SL_PARTIAL, /* Only partially allocated slabs */
3746 SL_CPU, /* Only slabs used for cpu caches */
3747 SL_OBJECTS, /* Determine allocated objects not slabs */
3748 SL_TOTAL /* Determine object capacity not slabs */
3751 #define SO_ALL (1 << SL_ALL)
3752 #define SO_PARTIAL (1 << SL_PARTIAL)
3753 #define SO_CPU (1 << SL_CPU)
3754 #define SO_OBJECTS (1 << SL_OBJECTS)
3755 #define SO_TOTAL (1 << SL_TOTAL)
3757 static ssize_t show_slab_objects(struct kmem_cache *s,
3758 char *buf, unsigned long flags)
3760 unsigned long total = 0;
3763 unsigned long *nodes;
3764 unsigned long *per_cpu;
3766 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3769 per_cpu = nodes + nr_node_ids;
3771 if (flags & SO_CPU) {
3774 for_each_possible_cpu(cpu) {
3775 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3777 if (!c || c->node < 0)
3781 if (flags & SO_TOTAL)
3782 x = c->page->objects;
3783 else if (flags & SO_OBJECTS)
3789 nodes[c->node] += x;
3795 if (flags & SO_ALL) {
3796 for_each_node_state(node, N_NORMAL_MEMORY) {
3797 struct kmem_cache_node *n = get_node(s, node);
3799 if (flags & SO_TOTAL)
3800 x = atomic_long_read(&n->total_objects);
3801 else if (flags & SO_OBJECTS)
3802 x = atomic_long_read(&n->total_objects) -
3803 count_partial(n, count_free);
3806 x = atomic_long_read(&n->nr_slabs);
3811 } else if (flags & SO_PARTIAL) {
3812 for_each_node_state(node, N_NORMAL_MEMORY) {
3813 struct kmem_cache_node *n = get_node(s, node);
3815 if (flags & SO_TOTAL)
3816 x = count_partial(n, count_total);
3817 else if (flags & SO_OBJECTS)
3818 x = count_partial(n, count_inuse);
3825 x = sprintf(buf, "%lu", total);
3827 for_each_node_state(node, N_NORMAL_MEMORY)
3829 x += sprintf(buf + x, " N%d=%lu",
3833 return x + sprintf(buf + x, "\n");
3836 static int any_slab_objects(struct kmem_cache *s)
3840 for_each_online_node(node) {
3841 struct kmem_cache_node *n = get_node(s, node);
3846 if (atomic_long_read(&n->total_objects))
3852 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3853 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3855 struct slab_attribute {
3856 struct attribute attr;
3857 ssize_t (*show)(struct kmem_cache *s, char *buf);
3858 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3861 #define SLAB_ATTR_RO(_name) \
3862 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3864 #define SLAB_ATTR(_name) \
3865 static struct slab_attribute _name##_attr = \
3866 __ATTR(_name, 0644, _name##_show, _name##_store)
3868 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3870 return sprintf(buf, "%d\n", s->size);
3872 SLAB_ATTR_RO(slab_size);
3874 static ssize_t align_show(struct kmem_cache *s, char *buf)
3876 return sprintf(buf, "%d\n", s->align);
3878 SLAB_ATTR_RO(align);
3880 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3882 return sprintf(buf, "%d\n", s->objsize);
3884 SLAB_ATTR_RO(object_size);
3886 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3888 return sprintf(buf, "%d\n", oo_objects(s->oo));
3890 SLAB_ATTR_RO(objs_per_slab);
3892 static ssize_t order_store(struct kmem_cache *s,
3893 const char *buf, size_t length)
3895 unsigned long order;
3898 err = strict_strtoul(buf, 10, &order);
3902 if (order > slub_max_order || order < slub_min_order)
3905 calculate_sizes(s, order);
3909 static ssize_t order_show(struct kmem_cache *s, char *buf)
3911 return sprintf(buf, "%d\n", oo_order(s->oo));
3915 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3917 return sprintf(buf, "%lu\n", s->min_partial);
3920 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3926 err = strict_strtoul(buf, 10, &min);
3930 set_min_partial(s, min);
3933 SLAB_ATTR(min_partial);
3935 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3938 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3940 return n + sprintf(buf + n, "\n");
3946 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3948 return sprintf(buf, "%d\n", s->refcount - 1);
3950 SLAB_ATTR_RO(aliases);
3952 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3954 return show_slab_objects(s, buf, SO_ALL);
3956 SLAB_ATTR_RO(slabs);
3958 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3960 return show_slab_objects(s, buf, SO_PARTIAL);
3962 SLAB_ATTR_RO(partial);
3964 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3966 return show_slab_objects(s, buf, SO_CPU);
3968 SLAB_ATTR_RO(cpu_slabs);
3970 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3972 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3974 SLAB_ATTR_RO(objects);
3976 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3978 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3980 SLAB_ATTR_RO(objects_partial);
3982 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3984 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3986 SLAB_ATTR_RO(total_objects);
3988 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3990 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3993 static ssize_t sanity_checks_store(struct kmem_cache *s,
3994 const char *buf, size_t length)
3996 s->flags &= ~SLAB_DEBUG_FREE;
3998 s->flags |= SLAB_DEBUG_FREE;
4001 SLAB_ATTR(sanity_checks);
4003 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4005 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4008 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4011 s->flags &= ~SLAB_TRACE;
4013 s->flags |= SLAB_TRACE;
4018 #ifdef CONFIG_FAILSLAB
4019 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4021 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4024 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4027 s->flags &= ~SLAB_FAILSLAB;
4029 s->flags |= SLAB_FAILSLAB;
4032 SLAB_ATTR(failslab);
4035 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4037 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4040 static ssize_t reclaim_account_store(struct kmem_cache *s,
4041 const char *buf, size_t length)
4043 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4045 s->flags |= SLAB_RECLAIM_ACCOUNT;
4048 SLAB_ATTR(reclaim_account);
4050 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4052 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4054 SLAB_ATTR_RO(hwcache_align);
4056 #ifdef CONFIG_ZONE_DMA
4057 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4059 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4061 SLAB_ATTR_RO(cache_dma);
4064 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4066 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4068 SLAB_ATTR_RO(destroy_by_rcu);
4070 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4072 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4075 static ssize_t red_zone_store(struct kmem_cache *s,
4076 const char *buf, size_t length)
4078 if (any_slab_objects(s))
4081 s->flags &= ~SLAB_RED_ZONE;
4083 s->flags |= SLAB_RED_ZONE;
4084 calculate_sizes(s, -1);
4087 SLAB_ATTR(red_zone);
4089 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4091 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4094 static ssize_t poison_store(struct kmem_cache *s,
4095 const char *buf, size_t length)
4097 if (any_slab_objects(s))
4100 s->flags &= ~SLAB_POISON;
4102 s->flags |= SLAB_POISON;
4103 calculate_sizes(s, -1);
4108 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4110 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4113 static ssize_t store_user_store(struct kmem_cache *s,
4114 const char *buf, size_t length)
4116 if (any_slab_objects(s))
4119 s->flags &= ~SLAB_STORE_USER;
4121 s->flags |= SLAB_STORE_USER;
4122 calculate_sizes(s, -1);
4125 SLAB_ATTR(store_user);
4127 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4132 static ssize_t validate_store(struct kmem_cache *s,
4133 const char *buf, size_t length)
4137 if (buf[0] == '1') {
4138 ret = validate_slab_cache(s);
4144 SLAB_ATTR(validate);
4146 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4151 static ssize_t shrink_store(struct kmem_cache *s,
4152 const char *buf, size_t length)
4154 if (buf[0] == '1') {
4155 int rc = kmem_cache_shrink(s);
4165 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4167 if (!(s->flags & SLAB_STORE_USER))
4169 return list_locations(s, buf, TRACK_ALLOC);
4171 SLAB_ATTR_RO(alloc_calls);
4173 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4175 if (!(s->flags & SLAB_STORE_USER))
4177 return list_locations(s, buf, TRACK_FREE);
4179 SLAB_ATTR_RO(free_calls);
4182 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4184 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4187 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4188 const char *buf, size_t length)
4190 unsigned long ratio;
4193 err = strict_strtoul(buf, 10, &ratio);
4198 s->remote_node_defrag_ratio = ratio * 10;
4202 SLAB_ATTR(remote_node_defrag_ratio);
4205 #ifdef CONFIG_SLUB_STATS
4206 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4208 unsigned long sum = 0;
4211 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4216 for_each_online_cpu(cpu) {
4217 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4223 len = sprintf(buf, "%lu", sum);
4226 for_each_online_cpu(cpu) {
4227 if (data[cpu] && len < PAGE_SIZE - 20)
4228 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4232 return len + sprintf(buf + len, "\n");
4235 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4239 for_each_online_cpu(cpu)
4240 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4243 #define STAT_ATTR(si, text) \
4244 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4246 return show_stat(s, buf, si); \
4248 static ssize_t text##_store(struct kmem_cache *s, \
4249 const char *buf, size_t length) \
4251 if (buf[0] != '0') \
4253 clear_stat(s, si); \
4258 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4259 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4260 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4261 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4262 STAT_ATTR(FREE_FROZEN, free_frozen);
4263 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4264 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4265 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4266 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4267 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4268 STAT_ATTR(FREE_SLAB, free_slab);
4269 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4270 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4271 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4272 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4273 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4274 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4275 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4278 static struct attribute *slab_attrs[] = {
4279 &slab_size_attr.attr,
4280 &object_size_attr.attr,
4281 &objs_per_slab_attr.attr,
4283 &min_partial_attr.attr,
4285 &objects_partial_attr.attr,
4286 &total_objects_attr.attr,
4289 &cpu_slabs_attr.attr,
4293 &sanity_checks_attr.attr,
4295 &hwcache_align_attr.attr,
4296 &reclaim_account_attr.attr,
4297 &destroy_by_rcu_attr.attr,
4298 &red_zone_attr.attr,
4300 &store_user_attr.attr,
4301 &validate_attr.attr,
4303 &alloc_calls_attr.attr,
4304 &free_calls_attr.attr,
4305 #ifdef CONFIG_ZONE_DMA
4306 &cache_dma_attr.attr,
4309 &remote_node_defrag_ratio_attr.attr,
4311 #ifdef CONFIG_SLUB_STATS
4312 &alloc_fastpath_attr.attr,
4313 &alloc_slowpath_attr.attr,
4314 &free_fastpath_attr.attr,
4315 &free_slowpath_attr.attr,
4316 &free_frozen_attr.attr,
4317 &free_add_partial_attr.attr,
4318 &free_remove_partial_attr.attr,
4319 &alloc_from_partial_attr.attr,
4320 &alloc_slab_attr.attr,
4321 &alloc_refill_attr.attr,
4322 &free_slab_attr.attr,
4323 &cpuslab_flush_attr.attr,
4324 &deactivate_full_attr.attr,
4325 &deactivate_empty_attr.attr,
4326 &deactivate_to_head_attr.attr,
4327 &deactivate_to_tail_attr.attr,
4328 &deactivate_remote_frees_attr.attr,
4329 &order_fallback_attr.attr,
4331 #ifdef CONFIG_FAILSLAB
4332 &failslab_attr.attr,
4338 static struct attribute_group slab_attr_group = {
4339 .attrs = slab_attrs,
4342 static ssize_t slab_attr_show(struct kobject *kobj,
4343 struct attribute *attr,
4346 struct slab_attribute *attribute;
4347 struct kmem_cache *s;
4350 attribute = to_slab_attr(attr);
4353 if (!attribute->show)
4356 err = attribute->show(s, buf);
4361 static ssize_t slab_attr_store(struct kobject *kobj,
4362 struct attribute *attr,
4363 const char *buf, size_t len)
4365 struct slab_attribute *attribute;
4366 struct kmem_cache *s;
4369 attribute = to_slab_attr(attr);
4372 if (!attribute->store)
4375 err = attribute->store(s, buf, len);
4380 static void kmem_cache_release(struct kobject *kobj)
4382 struct kmem_cache *s = to_slab(kobj);
4387 static const struct sysfs_ops slab_sysfs_ops = {
4388 .show = slab_attr_show,
4389 .store = slab_attr_store,
4392 static struct kobj_type slab_ktype = {
4393 .sysfs_ops = &slab_sysfs_ops,
4394 .release = kmem_cache_release
4397 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4399 struct kobj_type *ktype = get_ktype(kobj);
4401 if (ktype == &slab_ktype)
4406 static const struct kset_uevent_ops slab_uevent_ops = {
4407 .filter = uevent_filter,
4410 static struct kset *slab_kset;
4412 #define ID_STR_LENGTH 64
4414 /* Create a unique string id for a slab cache:
4416 * Format :[flags-]size
4418 static char *create_unique_id(struct kmem_cache *s)
4420 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4427 * First flags affecting slabcache operations. We will only
4428 * get here for aliasable slabs so we do not need to support
4429 * too many flags. The flags here must cover all flags that
4430 * are matched during merging to guarantee that the id is
4433 if (s->flags & SLAB_CACHE_DMA)
4435 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4437 if (s->flags & SLAB_DEBUG_FREE)
4439 if (!(s->flags & SLAB_NOTRACK))
4443 p += sprintf(p, "%07d", s->size);
4444 BUG_ON(p > name + ID_STR_LENGTH - 1);
4448 static int sysfs_slab_add(struct kmem_cache *s)
4454 if (slab_state < SYSFS)
4455 /* Defer until later */
4458 unmergeable = slab_unmergeable(s);
4461 * Slabcache can never be merged so we can use the name proper.
4462 * This is typically the case for debug situations. In that
4463 * case we can catch duplicate names easily.
4465 sysfs_remove_link(&slab_kset->kobj, s->name);
4469 * Create a unique name for the slab as a target
4472 name = create_unique_id(s);
4475 s->kobj.kset = slab_kset;
4476 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4478 kobject_put(&s->kobj);
4482 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4484 kobject_del(&s->kobj);
4485 kobject_put(&s->kobj);
4488 kobject_uevent(&s->kobj, KOBJ_ADD);
4490 /* Setup first alias */
4491 sysfs_slab_alias(s, s->name);
4497 static void sysfs_slab_remove(struct kmem_cache *s)
4499 if (slab_state < SYSFS)
4501 * Sysfs has not been setup yet so no need to remove the
4506 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4507 kobject_del(&s->kobj);
4508 kobject_put(&s->kobj);
4512 * Need to buffer aliases during bootup until sysfs becomes
4513 * available lest we lose that information.
4515 struct saved_alias {
4516 struct kmem_cache *s;
4518 struct saved_alias *next;
4521 static struct saved_alias *alias_list;
4523 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4525 struct saved_alias *al;
4527 if (slab_state == SYSFS) {
4529 * If we have a leftover link then remove it.
4531 sysfs_remove_link(&slab_kset->kobj, name);
4532 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4535 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4541 al->next = alias_list;
4546 static int __init slab_sysfs_init(void)
4548 struct kmem_cache *s;
4551 down_write(&slub_lock);
4553 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4555 up_write(&slub_lock);
4556 printk(KERN_ERR "Cannot register slab subsystem.\n");
4562 list_for_each_entry(s, &slab_caches, list) {
4563 err = sysfs_slab_add(s);
4565 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4566 " to sysfs\n", s->name);
4569 while (alias_list) {
4570 struct saved_alias *al = alias_list;
4572 alias_list = alias_list->next;
4573 err = sysfs_slab_alias(al->s, al->name);
4575 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4576 " %s to sysfs\n", s->name);
4580 up_write(&slub_lock);
4585 __initcall(slab_sysfs_init);
4589 * The /proc/slabinfo ABI
4591 #ifdef CONFIG_SLABINFO
4592 static void print_slabinfo_header(struct seq_file *m)
4594 seq_puts(m, "slabinfo - version: 2.1\n");
4595 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4596 "<objperslab> <pagesperslab>");
4597 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4598 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4602 static void *s_start(struct seq_file *m, loff_t *pos)
4606 down_read(&slub_lock);
4608 print_slabinfo_header(m);
4610 return seq_list_start(&slab_caches, *pos);
4613 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4615 return seq_list_next(p, &slab_caches, pos);
4618 static void s_stop(struct seq_file *m, void *p)
4620 up_read(&slub_lock);
4623 static int s_show(struct seq_file *m, void *p)
4625 unsigned long nr_partials = 0;
4626 unsigned long nr_slabs = 0;
4627 unsigned long nr_inuse = 0;
4628 unsigned long nr_objs = 0;
4629 unsigned long nr_free = 0;
4630 struct kmem_cache *s;
4633 s = list_entry(p, struct kmem_cache, list);
4635 for_each_online_node(node) {
4636 struct kmem_cache_node *n = get_node(s, node);
4641 nr_partials += n->nr_partial;
4642 nr_slabs += atomic_long_read(&n->nr_slabs);
4643 nr_objs += atomic_long_read(&n->total_objects);
4644 nr_free += count_partial(n, count_free);
4647 nr_inuse = nr_objs - nr_free;
4649 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4650 nr_objs, s->size, oo_objects(s->oo),
4651 (1 << oo_order(s->oo)));
4652 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4653 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4659 static const struct seq_operations slabinfo_op = {
4666 static int slabinfo_open(struct inode *inode, struct file *file)
4668 return seq_open(file, &slabinfo_op);
4671 static const struct file_operations proc_slabinfo_operations = {
4672 .open = slabinfo_open,
4674 .llseek = seq_lseek,
4675 .release = seq_release,
4678 static int __init slab_proc_init(void)
4680 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4683 module_init(slab_proc_init);
4684 #endif /* CONFIG_SLABINFO */