2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/proc_fs.h>
18 #include <linux/seq_file.h>
19 #include <linux/cpu.h>
20 #include <linux/cpuset.h>
21 #include <linux/mempolicy.h>
22 #include <linux/ctype.h>
23 #include <linux/debugobjects.h>
24 #include <linux/kallsyms.h>
25 #include <linux/memory.h>
26 #include <linux/math64.h>
27 #include <linux/fault-inject.h>
34 * The slab_lock protects operations on the object of a particular
35 * slab and its metadata in the page struct. If the slab lock
36 * has been taken then no allocations nor frees can be performed
37 * on the objects in the slab nor can the slab be added or removed
38 * from the partial or full lists since this would mean modifying
39 * the page_struct of the slab.
41 * The list_lock protects the partial and full list on each node and
42 * the partial slab counter. If taken then no new slabs may be added or
43 * removed from the lists nor make the number of partial slabs be modified.
44 * (Note that the total number of slabs is an atomic value that may be
45 * modified without taking the list lock).
47 * The list_lock is a centralized lock and thus we avoid taking it as
48 * much as possible. As long as SLUB does not have to handle partial
49 * slabs, operations can continue without any centralized lock. F.e.
50 * allocating a long series of objects that fill up slabs does not require
53 * The lock order is sometimes inverted when we are trying to get a slab
54 * off a list. We take the list_lock and then look for a page on the list
55 * to use. While we do that objects in the slabs may be freed. We can
56 * only operate on the slab if we have also taken the slab_lock. So we use
57 * a slab_trylock() on the slab. If trylock was successful then no frees
58 * can occur anymore and we can use the slab for allocations etc. If the
59 * slab_trylock() does not succeed then frees are in progress in the slab and
60 * we must stay away from it for a while since we may cause a bouncing
61 * cacheline if we try to acquire the lock. So go onto the next slab.
62 * If all pages are busy then we may allocate a new slab instead of reusing
63 * a partial slab. A new slab has noone operating on it and thus there is
64 * no danger of cacheline contention.
66 * Interrupts are disabled during allocation and deallocation in order to
67 * make the slab allocator safe to use in the context of an irq. In addition
68 * interrupts are disabled to ensure that the processor does not change
69 * while handling per_cpu slabs, due to kernel preemption.
71 * SLUB assigns one slab for allocation to each processor.
72 * Allocations only occur from these slabs called cpu slabs.
74 * Slabs with free elements are kept on a partial list and during regular
75 * operations no list for full slabs is used. If an object in a full slab is
76 * freed then the slab will show up again on the partial lists.
77 * We track full slabs for debugging purposes though because otherwise we
78 * cannot scan all objects.
80 * Slabs are freed when they become empty. Teardown and setup is
81 * minimal so we rely on the page allocators per cpu caches for
82 * fast frees and allocs.
84 * Overloading of page flags that are otherwise used for LRU management.
86 * PageActive The slab is frozen and exempt from list processing.
87 * This means that the slab is dedicated to a purpose
88 * such as satisfying allocations for a specific
89 * processor. Objects may be freed in the slab while
90 * it is frozen but slab_free will then skip the usual
91 * list operations. It is up to the processor holding
92 * the slab to integrate the slab into the slab lists
93 * when the slab is no longer needed.
95 * One use of this flag is to mark slabs that are
96 * used for allocations. Then such a slab becomes a cpu
97 * slab. The cpu slab may be equipped with an additional
98 * freelist that allows lockless access to
99 * free objects in addition to the regular freelist
100 * that requires the slab lock.
102 * PageError Slab requires special handling due to debug
103 * options set. This moves slab handling out of
104 * the fast path and disables lockless freelists.
107 #ifdef CONFIG_SLUB_DEBUG
114 * Issues still to be resolved:
116 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
118 * - Variable sizing of the per node arrays
121 /* Enable to test recovery from slab corruption on boot */
122 #undef SLUB_RESILIENCY_TEST
125 * Mininum number of partial slabs. These will be left on the partial
126 * lists even if they are empty. kmem_cache_shrink may reclaim them.
128 #define MIN_PARTIAL 5
131 * Maximum number of desirable partial slabs.
132 * The existence of more partial slabs makes kmem_cache_shrink
133 * sort the partial list by the number of objects in the.
135 #define MAX_PARTIAL 10
137 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
138 SLAB_POISON | SLAB_STORE_USER)
141 * Set of flags that will prevent slab merging
143 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
144 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
146 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
149 #ifndef ARCH_KMALLOC_MINALIGN
150 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
153 #ifndef ARCH_SLAB_MINALIGN
154 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
158 #define OO_MASK ((1 << OO_SHIFT) - 1)
159 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
161 /* Internal SLUB flags */
162 #define __OBJECT_POISON 0x80000000 /* Poison object */
163 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
165 static int kmem_size = sizeof(struct kmem_cache);
168 static struct notifier_block slab_notifier;
172 DOWN, /* No slab functionality available */
173 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
174 UP, /* Everything works but does not show up in sysfs */
178 /* A list of all slab caches on the system */
179 static DECLARE_RWSEM(slub_lock);
180 static LIST_HEAD(slab_caches);
183 * Tracking user of a slab.
186 unsigned long addr; /* Called from address */
187 int cpu; /* Was running on cpu */
188 int pid; /* Pid context */
189 unsigned long when; /* When did the operation occur */
192 enum track_item { TRACK_ALLOC, TRACK_FREE };
194 #ifdef CONFIG_SLUB_DEBUG
195 static int sysfs_slab_add(struct kmem_cache *);
196 static int sysfs_slab_alias(struct kmem_cache *, const char *);
197 static void sysfs_slab_remove(struct kmem_cache *);
200 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
201 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
203 static inline void sysfs_slab_remove(struct kmem_cache *s)
210 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
212 #ifdef CONFIG_SLUB_STATS
217 /********************************************************************
218 * Core slab cache functions
219 *******************************************************************/
221 int slab_is_available(void)
223 return slab_state >= UP;
226 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
229 return s->node[node];
231 return &s->local_node;
235 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
238 return s->cpu_slab[cpu];
244 /* Verify that a pointer has an address that is valid within a slab page */
245 static inline int check_valid_pointer(struct kmem_cache *s,
246 struct page *page, const void *object)
253 base = page_address(page);
254 if (object < base || object >= base + page->objects * s->size ||
255 (object - base) % s->size) {
263 * Slow version of get and set free pointer.
265 * This version requires touching the cache lines of kmem_cache which
266 * we avoid to do in the fast alloc free paths. There we obtain the offset
267 * from the page struct.
269 static inline void *get_freepointer(struct kmem_cache *s, void *object)
271 return *(void **)(object + s->offset);
274 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
276 *(void **)(object + s->offset) = fp;
279 /* Loop over all objects in a slab */
280 #define for_each_object(__p, __s, __addr, __objects) \
281 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
285 #define for_each_free_object(__p, __s, __free) \
286 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
288 /* Determine object index from a given position */
289 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
291 return (p - addr) / s->size;
294 static inline struct kmem_cache_order_objects oo_make(int order,
297 struct kmem_cache_order_objects x = {
298 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
304 static inline int oo_order(struct kmem_cache_order_objects x)
306 return x.x >> OO_SHIFT;
309 static inline int oo_objects(struct kmem_cache_order_objects x)
311 return x.x & OO_MASK;
314 #ifdef CONFIG_SLUB_DEBUG
318 #ifdef CONFIG_SLUB_DEBUG_ON
319 static int slub_debug = DEBUG_DEFAULT_FLAGS;
321 static int slub_debug;
324 static char *slub_debug_slabs;
329 static void print_section(char *text, u8 *addr, unsigned int length)
337 for (i = 0; i < length; i++) {
339 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
342 printk(KERN_CONT " %02x", addr[i]);
344 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
346 printk(KERN_CONT " %s\n", ascii);
353 printk(KERN_CONT " ");
357 printk(KERN_CONT " %s\n", ascii);
361 static struct track *get_track(struct kmem_cache *s, void *object,
362 enum track_item alloc)
367 p = object + s->offset + sizeof(void *);
369 p = object + s->inuse;
374 static void set_track(struct kmem_cache *s, void *object,
375 enum track_item alloc, unsigned long addr)
380 p = object + s->offset + sizeof(void *);
382 p = object + s->inuse;
387 p->cpu = smp_processor_id();
388 p->pid = current->pid;
391 memset(p, 0, sizeof(struct track));
394 static void init_tracking(struct kmem_cache *s, void *object)
396 if (!(s->flags & SLAB_STORE_USER))
399 set_track(s, object, TRACK_FREE, 0UL);
400 set_track(s, object, TRACK_ALLOC, 0UL);
403 static void print_track(const char *s, struct track *t)
408 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
409 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
412 static void print_tracking(struct kmem_cache *s, void *object)
414 if (!(s->flags & SLAB_STORE_USER))
417 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
418 print_track("Freed", get_track(s, object, TRACK_FREE));
421 static void print_page_info(struct page *page)
423 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
424 page, page->objects, page->inuse, page->freelist, page->flags);
428 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
434 vsnprintf(buf, sizeof(buf), fmt, args);
436 printk(KERN_ERR "========================================"
437 "=====================================\n");
438 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
439 printk(KERN_ERR "----------------------------------------"
440 "-------------------------------------\n\n");
443 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
449 vsnprintf(buf, sizeof(buf), fmt, args);
451 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
454 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
456 unsigned int off; /* Offset of last byte */
457 u8 *addr = page_address(page);
459 print_tracking(s, p);
461 print_page_info(page);
463 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
464 p, p - addr, get_freepointer(s, p));
467 print_section("Bytes b4", p - 16, 16);
469 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
471 if (s->flags & SLAB_RED_ZONE)
472 print_section("Redzone", p + s->objsize,
473 s->inuse - s->objsize);
476 off = s->offset + sizeof(void *);
480 if (s->flags & SLAB_STORE_USER)
481 off += 2 * sizeof(struct track);
484 /* Beginning of the filler is the free pointer */
485 print_section("Padding", p + off, s->size - off);
490 static void object_err(struct kmem_cache *s, struct page *page,
491 u8 *object, char *reason)
493 slab_bug(s, "%s", reason);
494 print_trailer(s, page, object);
497 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
503 vsnprintf(buf, sizeof(buf), fmt, args);
505 slab_bug(s, "%s", buf);
506 print_page_info(page);
510 static void init_object(struct kmem_cache *s, void *object, int active)
514 if (s->flags & __OBJECT_POISON) {
515 memset(p, POISON_FREE, s->objsize - 1);
516 p[s->objsize - 1] = POISON_END;
519 if (s->flags & SLAB_RED_ZONE)
520 memset(p + s->objsize,
521 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
522 s->inuse - s->objsize);
525 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
528 if (*start != (u8)value)
536 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
537 void *from, void *to)
539 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
540 memset(from, data, to - from);
543 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
544 u8 *object, char *what,
545 u8 *start, unsigned int value, unsigned int bytes)
550 fault = check_bytes(start, value, bytes);
555 while (end > fault && end[-1] == value)
558 slab_bug(s, "%s overwritten", what);
559 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
560 fault, end - 1, fault[0], value);
561 print_trailer(s, page, object);
563 restore_bytes(s, what, value, fault, end);
571 * Bytes of the object to be managed.
572 * If the freepointer may overlay the object then the free
573 * pointer is the first word of the object.
575 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
578 * object + s->objsize
579 * Padding to reach word boundary. This is also used for Redzoning.
580 * Padding is extended by another word if Redzoning is enabled and
583 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
584 * 0xcc (RED_ACTIVE) for objects in use.
587 * Meta data starts here.
589 * A. Free pointer (if we cannot overwrite object on free)
590 * B. Tracking data for SLAB_STORE_USER
591 * C. Padding to reach required alignment boundary or at mininum
592 * one word if debugging is on to be able to detect writes
593 * before the word boundary.
595 * Padding is done using 0x5a (POISON_INUSE)
598 * Nothing is used beyond s->size.
600 * If slabcaches are merged then the objsize and inuse boundaries are mostly
601 * ignored. And therefore no slab options that rely on these boundaries
602 * may be used with merged slabcaches.
605 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
607 unsigned long off = s->inuse; /* The end of info */
610 /* Freepointer is placed after the object. */
611 off += sizeof(void *);
613 if (s->flags & SLAB_STORE_USER)
614 /* We also have user information there */
615 off += 2 * sizeof(struct track);
620 return check_bytes_and_report(s, page, p, "Object padding",
621 p + off, POISON_INUSE, s->size - off);
624 /* Check the pad bytes at the end of a slab page */
625 static int slab_pad_check(struct kmem_cache *s, struct page *page)
633 if (!(s->flags & SLAB_POISON))
636 start = page_address(page);
637 length = (PAGE_SIZE << compound_order(page));
638 end = start + length;
639 remainder = length % s->size;
643 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
646 while (end > fault && end[-1] == POISON_INUSE)
649 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
650 print_section("Padding", end - remainder, remainder);
652 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
656 static int check_object(struct kmem_cache *s, struct page *page,
657 void *object, int active)
660 u8 *endobject = object + s->objsize;
662 if (s->flags & SLAB_RED_ZONE) {
664 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
666 if (!check_bytes_and_report(s, page, object, "Redzone",
667 endobject, red, s->inuse - s->objsize))
670 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
671 check_bytes_and_report(s, page, p, "Alignment padding",
672 endobject, POISON_INUSE, s->inuse - s->objsize);
676 if (s->flags & SLAB_POISON) {
677 if (!active && (s->flags & __OBJECT_POISON) &&
678 (!check_bytes_and_report(s, page, p, "Poison", p,
679 POISON_FREE, s->objsize - 1) ||
680 !check_bytes_and_report(s, page, p, "Poison",
681 p + s->objsize - 1, POISON_END, 1)))
684 * check_pad_bytes cleans up on its own.
686 check_pad_bytes(s, page, p);
689 if (!s->offset && active)
691 * Object and freepointer overlap. Cannot check
692 * freepointer while object is allocated.
696 /* Check free pointer validity */
697 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
698 object_err(s, page, p, "Freepointer corrupt");
700 * No choice but to zap it and thus lose the remainder
701 * of the free objects in this slab. May cause
702 * another error because the object count is now wrong.
704 set_freepointer(s, p, NULL);
710 static int check_slab(struct kmem_cache *s, struct page *page)
714 VM_BUG_ON(!irqs_disabled());
716 if (!PageSlab(page)) {
717 slab_err(s, page, "Not a valid slab page");
721 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
722 if (page->objects > maxobj) {
723 slab_err(s, page, "objects %u > max %u",
724 s->name, page->objects, maxobj);
727 if (page->inuse > page->objects) {
728 slab_err(s, page, "inuse %u > max %u",
729 s->name, page->inuse, page->objects);
732 /* Slab_pad_check fixes things up after itself */
733 slab_pad_check(s, page);
738 * Determine if a certain object on a page is on the freelist. Must hold the
739 * slab lock to guarantee that the chains are in a consistent state.
741 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
744 void *fp = page->freelist;
746 unsigned long max_objects;
748 while (fp && nr <= page->objects) {
751 if (!check_valid_pointer(s, page, fp)) {
753 object_err(s, page, object,
754 "Freechain corrupt");
755 set_freepointer(s, object, NULL);
758 slab_err(s, page, "Freepointer corrupt");
759 page->freelist = NULL;
760 page->inuse = page->objects;
761 slab_fix(s, "Freelist cleared");
767 fp = get_freepointer(s, object);
771 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
772 if (max_objects > MAX_OBJS_PER_PAGE)
773 max_objects = MAX_OBJS_PER_PAGE;
775 if (page->objects != max_objects) {
776 slab_err(s, page, "Wrong number of objects. Found %d but "
777 "should be %d", page->objects, max_objects);
778 page->objects = max_objects;
779 slab_fix(s, "Number of objects adjusted.");
781 if (page->inuse != page->objects - nr) {
782 slab_err(s, page, "Wrong object count. Counter is %d but "
783 "counted were %d", page->inuse, page->objects - nr);
784 page->inuse = page->objects - nr;
785 slab_fix(s, "Object count adjusted.");
787 return search == NULL;
790 static void trace(struct kmem_cache *s, struct page *page, void *object,
793 if (s->flags & SLAB_TRACE) {
794 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
796 alloc ? "alloc" : "free",
801 print_section("Object", (void *)object, s->objsize);
808 * Tracking of fully allocated slabs for debugging purposes.
810 static void add_full(struct kmem_cache_node *n, struct page *page)
812 spin_lock(&n->list_lock);
813 list_add(&page->lru, &n->full);
814 spin_unlock(&n->list_lock);
817 static void remove_full(struct kmem_cache *s, struct page *page)
819 struct kmem_cache_node *n;
821 if (!(s->flags & SLAB_STORE_USER))
824 n = get_node(s, page_to_nid(page));
826 spin_lock(&n->list_lock);
827 list_del(&page->lru);
828 spin_unlock(&n->list_lock);
831 /* Tracking of the number of slabs for debugging purposes */
832 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
834 struct kmem_cache_node *n = get_node(s, node);
836 return atomic_long_read(&n->nr_slabs);
839 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
841 struct kmem_cache_node *n = get_node(s, node);
844 * May be called early in order to allocate a slab for the
845 * kmem_cache_node structure. Solve the chicken-egg
846 * dilemma by deferring the increment of the count during
847 * bootstrap (see early_kmem_cache_node_alloc).
849 if (!NUMA_BUILD || n) {
850 atomic_long_inc(&n->nr_slabs);
851 atomic_long_add(objects, &n->total_objects);
854 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
856 struct kmem_cache_node *n = get_node(s, node);
858 atomic_long_dec(&n->nr_slabs);
859 atomic_long_sub(objects, &n->total_objects);
862 /* Object debug checks for alloc/free paths */
863 static void setup_object_debug(struct kmem_cache *s, struct page *page,
866 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
869 init_object(s, object, 0);
870 init_tracking(s, object);
873 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
874 void *object, unsigned long addr)
876 if (!check_slab(s, page))
879 if (!on_freelist(s, page, object)) {
880 object_err(s, page, object, "Object already allocated");
884 if (!check_valid_pointer(s, page, object)) {
885 object_err(s, page, object, "Freelist Pointer check fails");
889 if (!check_object(s, page, object, 0))
892 /* Success perform special debug activities for allocs */
893 if (s->flags & SLAB_STORE_USER)
894 set_track(s, object, TRACK_ALLOC, addr);
895 trace(s, page, object, 1);
896 init_object(s, object, 1);
900 if (PageSlab(page)) {
902 * If this is a slab page then lets do the best we can
903 * to avoid issues in the future. Marking all objects
904 * as used avoids touching the remaining objects.
906 slab_fix(s, "Marking all objects used");
907 page->inuse = page->objects;
908 page->freelist = NULL;
913 static int free_debug_processing(struct kmem_cache *s, struct page *page,
914 void *object, unsigned long addr)
916 if (!check_slab(s, page))
919 if (!check_valid_pointer(s, page, object)) {
920 slab_err(s, page, "Invalid object pointer 0x%p", object);
924 if (on_freelist(s, page, object)) {
925 object_err(s, page, object, "Object already free");
929 if (!check_object(s, page, object, 1))
932 if (unlikely(s != page->slab)) {
933 if (!PageSlab(page)) {
934 slab_err(s, page, "Attempt to free object(0x%p) "
935 "outside of slab", object);
936 } else if (!page->slab) {
938 "SLUB <none>: no slab for object 0x%p.\n",
942 object_err(s, page, object,
943 "page slab pointer corrupt.");
947 /* Special debug activities for freeing objects */
948 if (!PageSlubFrozen(page) && !page->freelist)
949 remove_full(s, page);
950 if (s->flags & SLAB_STORE_USER)
951 set_track(s, object, TRACK_FREE, addr);
952 trace(s, page, object, 0);
953 init_object(s, object, 0);
957 slab_fix(s, "Object at 0x%p not freed", object);
961 static int __init setup_slub_debug(char *str)
963 slub_debug = DEBUG_DEFAULT_FLAGS;
964 if (*str++ != '=' || !*str)
966 * No options specified. Switch on full debugging.
972 * No options but restriction on slabs. This means full
973 * debugging for slabs matching a pattern.
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 printk(KERN_ERR "slub_debug option '%c' "
1006 "unknown. skipped\n", *str);
1012 slub_debug_slabs = str + 1;
1017 __setup("slub_debug", setup_slub_debug);
1019 static unsigned long kmem_cache_flags(unsigned long objsize,
1020 unsigned long flags, const char *name,
1021 void (*ctor)(void *))
1024 * Enable debugging if selected on the kernel commandline.
1026 if (slub_debug && (!slub_debug_slabs ||
1027 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1028 flags |= slub_debug;
1033 static inline void setup_object_debug(struct kmem_cache *s,
1034 struct page *page, void *object) {}
1036 static inline int alloc_debug_processing(struct kmem_cache *s,
1037 struct page *page, void *object, unsigned long addr) { return 0; }
1039 static inline int free_debug_processing(struct kmem_cache *s,
1040 struct page *page, void *object, unsigned long addr) { return 0; }
1042 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1044 static inline int check_object(struct kmem_cache *s, struct page *page,
1045 void *object, int active) { return 1; }
1046 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1047 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1048 unsigned long flags, const char *name,
1049 void (*ctor)(void *))
1053 #define slub_debug 0
1055 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1057 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1059 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1064 * Slab allocation and freeing
1066 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1067 struct kmem_cache_order_objects oo)
1069 int order = oo_order(oo);
1072 return alloc_pages(flags, order);
1074 return alloc_pages_node(node, flags, order);
1077 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1080 struct kmem_cache_order_objects oo = s->oo;
1082 flags |= s->allocflags;
1084 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1086 if (unlikely(!page)) {
1089 * Allocation may have failed due to fragmentation.
1090 * Try a lower order alloc if possible
1092 page = alloc_slab_page(flags, node, oo);
1096 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1098 page->objects = oo_objects(oo);
1099 mod_zone_page_state(page_zone(page),
1100 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1101 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1107 static void setup_object(struct kmem_cache *s, struct page *page,
1110 setup_object_debug(s, page, object);
1111 if (unlikely(s->ctor))
1115 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1122 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1124 page = allocate_slab(s,
1125 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1129 inc_slabs_node(s, page_to_nid(page), page->objects);
1131 page->flags |= 1 << PG_slab;
1132 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1133 SLAB_STORE_USER | SLAB_TRACE))
1134 __SetPageSlubDebug(page);
1136 start = page_address(page);
1138 if (unlikely(s->flags & SLAB_POISON))
1139 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1142 for_each_object(p, s, start, page->objects) {
1143 setup_object(s, page, last);
1144 set_freepointer(s, last, p);
1147 setup_object(s, page, last);
1148 set_freepointer(s, last, NULL);
1150 page->freelist = start;
1156 static void __free_slab(struct kmem_cache *s, struct page *page)
1158 int order = compound_order(page);
1159 int pages = 1 << order;
1161 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1164 slab_pad_check(s, page);
1165 for_each_object(p, s, page_address(page),
1167 check_object(s, page, p, 0);
1168 __ClearPageSlubDebug(page);
1171 mod_zone_page_state(page_zone(page),
1172 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1173 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1176 __ClearPageSlab(page);
1177 reset_page_mapcount(page);
1178 __free_pages(page, order);
1181 static void rcu_free_slab(struct rcu_head *h)
1185 page = container_of((struct list_head *)h, struct page, lru);
1186 __free_slab(page->slab, page);
1189 static void free_slab(struct kmem_cache *s, struct page *page)
1191 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1193 * RCU free overloads the RCU head over the LRU
1195 struct rcu_head *head = (void *)&page->lru;
1197 call_rcu(head, rcu_free_slab);
1199 __free_slab(s, page);
1202 static void discard_slab(struct kmem_cache *s, struct page *page)
1204 dec_slabs_node(s, page_to_nid(page), page->objects);
1209 * Per slab locking using the pagelock
1211 static __always_inline void slab_lock(struct page *page)
1213 bit_spin_lock(PG_locked, &page->flags);
1216 static __always_inline void slab_unlock(struct page *page)
1218 __bit_spin_unlock(PG_locked, &page->flags);
1221 static __always_inline int slab_trylock(struct page *page)
1225 rc = bit_spin_trylock(PG_locked, &page->flags);
1230 * Management of partially allocated slabs
1232 static void add_partial(struct kmem_cache_node *n,
1233 struct page *page, int tail)
1235 spin_lock(&n->list_lock);
1238 list_add_tail(&page->lru, &n->partial);
1240 list_add(&page->lru, &n->partial);
1241 spin_unlock(&n->list_lock);
1244 static void remove_partial(struct kmem_cache *s, struct page *page)
1246 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1248 spin_lock(&n->list_lock);
1249 list_del(&page->lru);
1251 spin_unlock(&n->list_lock);
1255 * Lock slab and remove from the partial list.
1257 * Must hold list_lock.
1259 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1262 if (slab_trylock(page)) {
1263 list_del(&page->lru);
1265 __SetPageSlubFrozen(page);
1272 * Try to allocate a partial slab from a specific node.
1274 static struct page *get_partial_node(struct kmem_cache_node *n)
1279 * Racy check. If we mistakenly see no partial slabs then we
1280 * just allocate an empty slab. If we mistakenly try to get a
1281 * partial slab and there is none available then get_partials()
1284 if (!n || !n->nr_partial)
1287 spin_lock(&n->list_lock);
1288 list_for_each_entry(page, &n->partial, lru)
1289 if (lock_and_freeze_slab(n, page))
1293 spin_unlock(&n->list_lock);
1298 * Get a page from somewhere. Search in increasing NUMA distances.
1300 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1303 struct zonelist *zonelist;
1306 enum zone_type high_zoneidx = gfp_zone(flags);
1310 * The defrag ratio allows a configuration of the tradeoffs between
1311 * inter node defragmentation and node local allocations. A lower
1312 * defrag_ratio increases the tendency to do local allocations
1313 * instead of attempting to obtain partial slabs from other nodes.
1315 * If the defrag_ratio is set to 0 then kmalloc() always
1316 * returns node local objects. If the ratio is higher then kmalloc()
1317 * may return off node objects because partial slabs are obtained
1318 * from other nodes and filled up.
1320 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1321 * defrag_ratio = 1000) then every (well almost) allocation will
1322 * first attempt to defrag slab caches on other nodes. This means
1323 * scanning over all nodes to look for partial slabs which may be
1324 * expensive if we do it every time we are trying to find a slab
1325 * with available objects.
1327 if (!s->remote_node_defrag_ratio ||
1328 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1331 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1332 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1333 struct kmem_cache_node *n;
1335 n = get_node(s, zone_to_nid(zone));
1337 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1338 n->nr_partial > s->min_partial) {
1339 page = get_partial_node(n);
1349 * Get a partial page, lock it and return it.
1351 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1354 int searchnode = (node == -1) ? numa_node_id() : node;
1356 page = get_partial_node(get_node(s, searchnode));
1357 if (page || (flags & __GFP_THISNODE))
1360 return get_any_partial(s, flags);
1364 * Move a page back to the lists.
1366 * Must be called with the slab lock held.
1368 * On exit the slab lock will have been dropped.
1370 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1372 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1373 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1375 __ClearPageSlubFrozen(page);
1378 if (page->freelist) {
1379 add_partial(n, page, tail);
1380 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1382 stat(c, DEACTIVATE_FULL);
1383 if (SLABDEBUG && PageSlubDebug(page) &&
1384 (s->flags & SLAB_STORE_USER))
1389 stat(c, DEACTIVATE_EMPTY);
1390 if (n->nr_partial < s->min_partial) {
1392 * Adding an empty slab to the partial slabs in order
1393 * to avoid page allocator overhead. This slab needs
1394 * to come after the other slabs with objects in
1395 * so that the others get filled first. That way the
1396 * size of the partial list stays small.
1398 * kmem_cache_shrink can reclaim any empty slabs from
1401 add_partial(n, page, 1);
1405 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1406 discard_slab(s, page);
1412 * Remove the cpu slab
1414 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1416 struct page *page = c->page;
1420 stat(c, DEACTIVATE_REMOTE_FREES);
1422 * Merge cpu freelist into slab freelist. Typically we get here
1423 * because both freelists are empty. So this is unlikely
1426 while (unlikely(c->freelist)) {
1429 tail = 0; /* Hot objects. Put the slab first */
1431 /* Retrieve object from cpu_freelist */
1432 object = c->freelist;
1433 c->freelist = c->freelist[c->offset];
1435 /* And put onto the regular freelist */
1436 object[c->offset] = page->freelist;
1437 page->freelist = object;
1441 unfreeze_slab(s, page, tail);
1444 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1446 stat(c, CPUSLAB_FLUSH);
1448 deactivate_slab(s, c);
1454 * Called from IPI handler with interrupts disabled.
1456 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1458 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1460 if (likely(c && c->page))
1464 static void flush_cpu_slab(void *d)
1466 struct kmem_cache *s = d;
1468 __flush_cpu_slab(s, smp_processor_id());
1471 static void flush_all(struct kmem_cache *s)
1473 on_each_cpu(flush_cpu_slab, s, 1);
1477 * Check if the objects in a per cpu structure fit numa
1478 * locality expectations.
1480 static inline int node_match(struct kmem_cache_cpu *c, int node)
1483 if (node != -1 && c->node != node)
1490 * Slow path. The lockless freelist is empty or we need to perform
1493 * Interrupts are disabled.
1495 * Processing is still very fast if new objects have been freed to the
1496 * regular freelist. In that case we simply take over the regular freelist
1497 * as the lockless freelist and zap the regular freelist.
1499 * If that is not working then we fall back to the partial lists. We take the
1500 * first element of the freelist as the object to allocate now and move the
1501 * rest of the freelist to the lockless freelist.
1503 * And if we were unable to get a new slab from the partial slab lists then
1504 * we need to allocate a new slab. This is the slowest path since it involves
1505 * a call to the page allocator and the setup of a new slab.
1507 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1508 unsigned long addr, struct kmem_cache_cpu *c)
1513 /* We handle __GFP_ZERO in the caller */
1514 gfpflags &= ~__GFP_ZERO;
1520 if (unlikely(!node_match(c, node)))
1523 stat(c, ALLOC_REFILL);
1526 object = c->page->freelist;
1527 if (unlikely(!object))
1529 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1532 c->freelist = object[c->offset];
1533 c->page->inuse = c->page->objects;
1534 c->page->freelist = NULL;
1535 c->node = page_to_nid(c->page);
1537 slab_unlock(c->page);
1538 stat(c, ALLOC_SLOWPATH);
1542 deactivate_slab(s, c);
1545 new = get_partial(s, gfpflags, node);
1548 stat(c, ALLOC_FROM_PARTIAL);
1552 if (gfpflags & __GFP_WAIT)
1555 new = new_slab(s, gfpflags, node);
1557 if (gfpflags & __GFP_WAIT)
1558 local_irq_disable();
1561 c = get_cpu_slab(s, smp_processor_id());
1562 stat(c, ALLOC_SLAB);
1566 __SetPageSlubFrozen(new);
1572 if (!alloc_debug_processing(s, c->page, object, addr))
1576 c->page->freelist = object[c->offset];
1582 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1583 * have the fastpath folded into their functions. So no function call
1584 * overhead for requests that can be satisfied on the fastpath.
1586 * The fastpath works by first checking if the lockless freelist can be used.
1587 * If not then __slab_alloc is called for slow processing.
1589 * Otherwise we can simply pick the next object from the lockless free list.
1591 static __always_inline void *slab_alloc(struct kmem_cache *s,
1592 gfp_t gfpflags, int node, unsigned long addr)
1595 struct kmem_cache_cpu *c;
1596 unsigned long flags;
1597 unsigned int objsize;
1599 might_sleep_if(gfpflags & __GFP_WAIT);
1601 if (should_failslab(s->objsize, gfpflags))
1604 local_irq_save(flags);
1605 c = get_cpu_slab(s, smp_processor_id());
1606 objsize = c->objsize;
1607 if (unlikely(!c->freelist || !node_match(c, node)))
1609 object = __slab_alloc(s, gfpflags, node, addr, c);
1612 object = c->freelist;
1613 c->freelist = object[c->offset];
1614 stat(c, ALLOC_FASTPATH);
1616 local_irq_restore(flags);
1618 if (unlikely((gfpflags & __GFP_ZERO) && object))
1619 memset(object, 0, objsize);
1624 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1626 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1628 EXPORT_SYMBOL(kmem_cache_alloc);
1631 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1633 return slab_alloc(s, gfpflags, node, _RET_IP_);
1635 EXPORT_SYMBOL(kmem_cache_alloc_node);
1639 * Slow patch handling. This may still be called frequently since objects
1640 * have a longer lifetime than the cpu slabs in most processing loads.
1642 * So we still attempt to reduce cache line usage. Just take the slab
1643 * lock and free the item. If there is no additional partial page
1644 * handling required then we can return immediately.
1646 static void __slab_free(struct kmem_cache *s, struct page *page,
1647 void *x, unsigned long addr, unsigned int offset)
1650 void **object = (void *)x;
1651 struct kmem_cache_cpu *c;
1653 c = get_cpu_slab(s, raw_smp_processor_id());
1654 stat(c, FREE_SLOWPATH);
1657 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1661 prior = object[offset] = page->freelist;
1662 page->freelist = object;
1665 if (unlikely(PageSlubFrozen(page))) {
1666 stat(c, FREE_FROZEN);
1670 if (unlikely(!page->inuse))
1674 * Objects left in the slab. If it was not on the partial list before
1677 if (unlikely(!prior)) {
1678 add_partial(get_node(s, page_to_nid(page)), page, 1);
1679 stat(c, FREE_ADD_PARTIAL);
1689 * Slab still on the partial list.
1691 remove_partial(s, page);
1692 stat(c, FREE_REMOVE_PARTIAL);
1696 discard_slab(s, page);
1700 if (!free_debug_processing(s, page, x, addr))
1706 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1707 * can perform fastpath freeing without additional function calls.
1709 * The fastpath is only possible if we are freeing to the current cpu slab
1710 * of this processor. This typically the case if we have just allocated
1713 * If fastpath is not possible then fall back to __slab_free where we deal
1714 * with all sorts of special processing.
1716 static __always_inline void slab_free(struct kmem_cache *s,
1717 struct page *page, void *x, unsigned long addr)
1719 void **object = (void *)x;
1720 struct kmem_cache_cpu *c;
1721 unsigned long flags;
1723 local_irq_save(flags);
1724 c = get_cpu_slab(s, smp_processor_id());
1725 debug_check_no_locks_freed(object, c->objsize);
1726 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1727 debug_check_no_obj_freed(object, s->objsize);
1728 if (likely(page == c->page && c->node >= 0)) {
1729 object[c->offset] = c->freelist;
1730 c->freelist = object;
1731 stat(c, FREE_FASTPATH);
1733 __slab_free(s, page, x, addr, c->offset);
1735 local_irq_restore(flags);
1738 void kmem_cache_free(struct kmem_cache *s, void *x)
1742 page = virt_to_head_page(x);
1744 slab_free(s, page, x, _RET_IP_);
1746 EXPORT_SYMBOL(kmem_cache_free);
1748 /* Figure out on which slab page the object resides */
1749 static struct page *get_object_page(const void *x)
1751 struct page *page = virt_to_head_page(x);
1753 if (!PageSlab(page))
1760 * Object placement in a slab is made very easy because we always start at
1761 * offset 0. If we tune the size of the object to the alignment then we can
1762 * get the required alignment by putting one properly sized object after
1765 * Notice that the allocation order determines the sizes of the per cpu
1766 * caches. Each processor has always one slab available for allocations.
1767 * Increasing the allocation order reduces the number of times that slabs
1768 * must be moved on and off the partial lists and is therefore a factor in
1773 * Mininum / Maximum order of slab pages. This influences locking overhead
1774 * and slab fragmentation. A higher order reduces the number of partial slabs
1775 * and increases the number of allocations possible without having to
1776 * take the list_lock.
1778 static int slub_min_order;
1779 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1780 static int slub_min_objects;
1783 * Merge control. If this is set then no merging of slab caches will occur.
1784 * (Could be removed. This was introduced to pacify the merge skeptics.)
1786 static int slub_nomerge;
1789 * Calculate the order of allocation given an slab object size.
1791 * The order of allocation has significant impact on performance and other
1792 * system components. Generally order 0 allocations should be preferred since
1793 * order 0 does not cause fragmentation in the page allocator. Larger objects
1794 * be problematic to put into order 0 slabs because there may be too much
1795 * unused space left. We go to a higher order if more than 1/16th of the slab
1798 * In order to reach satisfactory performance we must ensure that a minimum
1799 * number of objects is in one slab. Otherwise we may generate too much
1800 * activity on the partial lists which requires taking the list_lock. This is
1801 * less a concern for large slabs though which are rarely used.
1803 * slub_max_order specifies the order where we begin to stop considering the
1804 * number of objects in a slab as critical. If we reach slub_max_order then
1805 * we try to keep the page order as low as possible. So we accept more waste
1806 * of space in favor of a small page order.
1808 * Higher order allocations also allow the placement of more objects in a
1809 * slab and thereby reduce object handling overhead. If the user has
1810 * requested a higher mininum order then we start with that one instead of
1811 * the smallest order which will fit the object.
1813 static inline int slab_order(int size, int min_objects,
1814 int max_order, int fract_leftover)
1818 int min_order = slub_min_order;
1820 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1821 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1823 for (order = max(min_order,
1824 fls(min_objects * size - 1) - PAGE_SHIFT);
1825 order <= max_order; order++) {
1827 unsigned long slab_size = PAGE_SIZE << order;
1829 if (slab_size < min_objects * size)
1832 rem = slab_size % size;
1834 if (rem <= slab_size / fract_leftover)
1842 static inline int calculate_order(int size)
1849 * Attempt to find best configuration for a slab. This
1850 * works by first attempting to generate a layout with
1851 * the best configuration and backing off gradually.
1853 * First we reduce the acceptable waste in a slab. Then
1854 * we reduce the minimum objects required in a slab.
1856 min_objects = slub_min_objects;
1858 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1859 while (min_objects > 1) {
1861 while (fraction >= 4) {
1862 order = slab_order(size, min_objects,
1863 slub_max_order, fraction);
1864 if (order <= slub_max_order)
1872 * We were unable to place multiple objects in a slab. Now
1873 * lets see if we can place a single object there.
1875 order = slab_order(size, 1, slub_max_order, 1);
1876 if (order <= slub_max_order)
1880 * Doh this slab cannot be placed using slub_max_order.
1882 order = slab_order(size, 1, MAX_ORDER, 1);
1883 if (order <= MAX_ORDER)
1889 * Figure out what the alignment of the objects will be.
1891 static unsigned long calculate_alignment(unsigned long flags,
1892 unsigned long align, unsigned long size)
1895 * If the user wants hardware cache aligned objects then follow that
1896 * suggestion if the object is sufficiently large.
1898 * The hardware cache alignment cannot override the specified
1899 * alignment though. If that is greater then use it.
1901 if (flags & SLAB_HWCACHE_ALIGN) {
1902 unsigned long ralign = cache_line_size();
1903 while (size <= ralign / 2)
1905 align = max(align, ralign);
1908 if (align < ARCH_SLAB_MINALIGN)
1909 align = ARCH_SLAB_MINALIGN;
1911 return ALIGN(align, sizeof(void *));
1914 static void init_kmem_cache_cpu(struct kmem_cache *s,
1915 struct kmem_cache_cpu *c)
1920 c->offset = s->offset / sizeof(void *);
1921 c->objsize = s->objsize;
1922 #ifdef CONFIG_SLUB_STATS
1923 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1928 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1931 spin_lock_init(&n->list_lock);
1932 INIT_LIST_HEAD(&n->partial);
1933 #ifdef CONFIG_SLUB_DEBUG
1934 atomic_long_set(&n->nr_slabs, 0);
1935 atomic_long_set(&n->total_objects, 0);
1936 INIT_LIST_HEAD(&n->full);
1942 * Per cpu array for per cpu structures.
1944 * The per cpu array places all kmem_cache_cpu structures from one processor
1945 * close together meaning that it becomes possible that multiple per cpu
1946 * structures are contained in one cacheline. This may be particularly
1947 * beneficial for the kmalloc caches.
1949 * A desktop system typically has around 60-80 slabs. With 100 here we are
1950 * likely able to get per cpu structures for all caches from the array defined
1951 * here. We must be able to cover all kmalloc caches during bootstrap.
1953 * If the per cpu array is exhausted then fall back to kmalloc
1954 * of individual cachelines. No sharing is possible then.
1956 #define NR_KMEM_CACHE_CPU 100
1958 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1959 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1961 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1962 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
1964 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1965 int cpu, gfp_t flags)
1967 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1970 per_cpu(kmem_cache_cpu_free, cpu) =
1971 (void *)c->freelist;
1973 /* Table overflow: So allocate ourselves */
1975 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1976 flags, cpu_to_node(cpu));
1981 init_kmem_cache_cpu(s, c);
1985 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1987 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1988 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1992 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1993 per_cpu(kmem_cache_cpu_free, cpu) = c;
1996 static void free_kmem_cache_cpus(struct kmem_cache *s)
2000 for_each_online_cpu(cpu) {
2001 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2004 s->cpu_slab[cpu] = NULL;
2005 free_kmem_cache_cpu(c, cpu);
2010 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2014 for_each_online_cpu(cpu) {
2015 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2020 c = alloc_kmem_cache_cpu(s, cpu, flags);
2022 free_kmem_cache_cpus(s);
2025 s->cpu_slab[cpu] = c;
2031 * Initialize the per cpu array.
2033 static void init_alloc_cpu_cpu(int cpu)
2037 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2040 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2041 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2043 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2046 static void __init init_alloc_cpu(void)
2050 for_each_online_cpu(cpu)
2051 init_alloc_cpu_cpu(cpu);
2055 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2056 static inline void init_alloc_cpu(void) {}
2058 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2060 init_kmem_cache_cpu(s, &s->cpu_slab);
2067 * No kmalloc_node yet so do it by hand. We know that this is the first
2068 * slab on the node for this slabcache. There are no concurrent accesses
2071 * Note that this function only works on the kmalloc_node_cache
2072 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2073 * memory on a fresh node that has no slab structures yet.
2075 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2078 struct kmem_cache_node *n;
2079 unsigned long flags;
2081 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2083 page = new_slab(kmalloc_caches, gfpflags, node);
2086 if (page_to_nid(page) != node) {
2087 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2089 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2090 "in order to be able to continue\n");
2095 page->freelist = get_freepointer(kmalloc_caches, n);
2097 kmalloc_caches->node[node] = n;
2098 #ifdef CONFIG_SLUB_DEBUG
2099 init_object(kmalloc_caches, n, 1);
2100 init_tracking(kmalloc_caches, n);
2102 init_kmem_cache_node(n, kmalloc_caches);
2103 inc_slabs_node(kmalloc_caches, node, page->objects);
2106 * lockdep requires consistent irq usage for each lock
2107 * so even though there cannot be a race this early in
2108 * the boot sequence, we still disable irqs.
2110 local_irq_save(flags);
2111 add_partial(n, page, 0);
2112 local_irq_restore(flags);
2115 static void free_kmem_cache_nodes(struct kmem_cache *s)
2119 for_each_node_state(node, N_NORMAL_MEMORY) {
2120 struct kmem_cache_node *n = s->node[node];
2121 if (n && n != &s->local_node)
2122 kmem_cache_free(kmalloc_caches, n);
2123 s->node[node] = NULL;
2127 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2132 if (slab_state >= UP)
2133 local_node = page_to_nid(virt_to_page(s));
2137 for_each_node_state(node, N_NORMAL_MEMORY) {
2138 struct kmem_cache_node *n;
2140 if (local_node == node)
2143 if (slab_state == DOWN) {
2144 early_kmem_cache_node_alloc(gfpflags, node);
2147 n = kmem_cache_alloc_node(kmalloc_caches,
2151 free_kmem_cache_nodes(s);
2157 init_kmem_cache_node(n, s);
2162 static void free_kmem_cache_nodes(struct kmem_cache *s)
2166 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2168 init_kmem_cache_node(&s->local_node, s);
2173 static void calculate_min_partial(struct kmem_cache *s, unsigned long min)
2175 if (min < MIN_PARTIAL)
2177 else if (min > MAX_PARTIAL)
2179 s->min_partial = min;
2183 * calculate_sizes() determines the order and the distribution of data within
2186 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2188 unsigned long flags = s->flags;
2189 unsigned long size = s->objsize;
2190 unsigned long align = s->align;
2194 * Round up object size to the next word boundary. We can only
2195 * place the free pointer at word boundaries and this determines
2196 * the possible location of the free pointer.
2198 size = ALIGN(size, sizeof(void *));
2200 #ifdef CONFIG_SLUB_DEBUG
2202 * Determine if we can poison the object itself. If the user of
2203 * the slab may touch the object after free or before allocation
2204 * then we should never poison the object itself.
2206 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2208 s->flags |= __OBJECT_POISON;
2210 s->flags &= ~__OBJECT_POISON;
2214 * If we are Redzoning then check if there is some space between the
2215 * end of the object and the free pointer. If not then add an
2216 * additional word to have some bytes to store Redzone information.
2218 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2219 size += sizeof(void *);
2223 * With that we have determined the number of bytes in actual use
2224 * by the object. This is the potential offset to the free pointer.
2228 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2231 * Relocate free pointer after the object if it is not
2232 * permitted to overwrite the first word of the object on
2235 * This is the case if we do RCU, have a constructor or
2236 * destructor or are poisoning the objects.
2239 size += sizeof(void *);
2242 #ifdef CONFIG_SLUB_DEBUG
2243 if (flags & SLAB_STORE_USER)
2245 * Need to store information about allocs and frees after
2248 size += 2 * sizeof(struct track);
2250 if (flags & SLAB_RED_ZONE)
2252 * Add some empty padding so that we can catch
2253 * overwrites from earlier objects rather than let
2254 * tracking information or the free pointer be
2255 * corrupted if a user writes before the start
2258 size += sizeof(void *);
2262 * Determine the alignment based on various parameters that the
2263 * user specified and the dynamic determination of cache line size
2266 align = calculate_alignment(flags, align, s->objsize);
2269 * SLUB stores one object immediately after another beginning from
2270 * offset 0. In order to align the objects we have to simply size
2271 * each object to conform to the alignment.
2273 size = ALIGN(size, align);
2275 if (forced_order >= 0)
2276 order = forced_order;
2278 order = calculate_order(size);
2285 s->allocflags |= __GFP_COMP;
2287 if (s->flags & SLAB_CACHE_DMA)
2288 s->allocflags |= SLUB_DMA;
2290 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2291 s->allocflags |= __GFP_RECLAIMABLE;
2294 * Determine the number of objects per slab
2296 s->oo = oo_make(order, size);
2297 s->min = oo_make(get_order(size), size);
2298 if (oo_objects(s->oo) > oo_objects(s->max))
2301 return !!oo_objects(s->oo);
2305 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2306 const char *name, size_t size,
2307 size_t align, unsigned long flags,
2308 void (*ctor)(void *))
2310 memset(s, 0, kmem_size);
2315 s->flags = kmem_cache_flags(size, flags, name, ctor);
2317 if (!calculate_sizes(s, -1))
2321 * The larger the object size is, the more pages we want on the partial
2322 * list to avoid pounding the page allocator excessively.
2324 calculate_min_partial(s, ilog2(s->size));
2327 s->remote_node_defrag_ratio = 1000;
2329 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2332 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2334 free_kmem_cache_nodes(s);
2336 if (flags & SLAB_PANIC)
2337 panic("Cannot create slab %s size=%lu realsize=%u "
2338 "order=%u offset=%u flags=%lx\n",
2339 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2345 * Check if a given pointer is valid
2347 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2351 page = get_object_page(object);
2353 if (!page || s != page->slab)
2354 /* No slab or wrong slab */
2357 if (!check_valid_pointer(s, page, object))
2361 * We could also check if the object is on the slabs freelist.
2362 * But this would be too expensive and it seems that the main
2363 * purpose of kmem_ptr_valid() is to check if the object belongs
2364 * to a certain slab.
2368 EXPORT_SYMBOL(kmem_ptr_validate);
2371 * Determine the size of a slab object
2373 unsigned int kmem_cache_size(struct kmem_cache *s)
2377 EXPORT_SYMBOL(kmem_cache_size);
2379 const char *kmem_cache_name(struct kmem_cache *s)
2383 EXPORT_SYMBOL(kmem_cache_name);
2385 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2388 #ifdef CONFIG_SLUB_DEBUG
2389 void *addr = page_address(page);
2391 DECLARE_BITMAP(map, page->objects);
2393 bitmap_zero(map, page->objects);
2394 slab_err(s, page, "%s", text);
2396 for_each_free_object(p, s, page->freelist)
2397 set_bit(slab_index(p, s, addr), map);
2399 for_each_object(p, s, addr, page->objects) {
2401 if (!test_bit(slab_index(p, s, addr), map)) {
2402 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2404 print_tracking(s, p);
2412 * Attempt to free all partial slabs on a node.
2414 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2416 unsigned long flags;
2417 struct page *page, *h;
2419 spin_lock_irqsave(&n->list_lock, flags);
2420 list_for_each_entry_safe(page, h, &n->partial, lru) {
2422 list_del(&page->lru);
2423 discard_slab(s, page);
2426 list_slab_objects(s, page,
2427 "Objects remaining on kmem_cache_close()");
2430 spin_unlock_irqrestore(&n->list_lock, flags);
2434 * Release all resources used by a slab cache.
2436 static inline int kmem_cache_close(struct kmem_cache *s)
2442 /* Attempt to free all objects */
2443 free_kmem_cache_cpus(s);
2444 for_each_node_state(node, N_NORMAL_MEMORY) {
2445 struct kmem_cache_node *n = get_node(s, node);
2448 if (n->nr_partial || slabs_node(s, node))
2451 free_kmem_cache_nodes(s);
2456 * Close a cache and release the kmem_cache structure
2457 * (must be used for caches created using kmem_cache_create)
2459 void kmem_cache_destroy(struct kmem_cache *s)
2461 down_write(&slub_lock);
2465 up_write(&slub_lock);
2466 if (kmem_cache_close(s)) {
2467 printk(KERN_ERR "SLUB %s: %s called for cache that "
2468 "still has objects.\n", s->name, __func__);
2471 sysfs_slab_remove(s);
2473 up_write(&slub_lock);
2475 EXPORT_SYMBOL(kmem_cache_destroy);
2477 /********************************************************************
2479 *******************************************************************/
2481 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2482 EXPORT_SYMBOL(kmalloc_caches);
2484 static int __init setup_slub_min_order(char *str)
2486 get_option(&str, &slub_min_order);
2491 __setup("slub_min_order=", setup_slub_min_order);
2493 static int __init setup_slub_max_order(char *str)
2495 get_option(&str, &slub_max_order);
2500 __setup("slub_max_order=", setup_slub_max_order);
2502 static int __init setup_slub_min_objects(char *str)
2504 get_option(&str, &slub_min_objects);
2509 __setup("slub_min_objects=", setup_slub_min_objects);
2511 static int __init setup_slub_nomerge(char *str)
2517 __setup("slub_nomerge", setup_slub_nomerge);
2519 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2520 const char *name, int size, gfp_t gfp_flags)
2522 unsigned int flags = 0;
2524 if (gfp_flags & SLUB_DMA)
2525 flags = SLAB_CACHE_DMA;
2527 down_write(&slub_lock);
2528 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2532 list_add(&s->list, &slab_caches);
2533 up_write(&slub_lock);
2534 if (sysfs_slab_add(s))
2539 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2542 #ifdef CONFIG_ZONE_DMA
2543 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2545 static void sysfs_add_func(struct work_struct *w)
2547 struct kmem_cache *s;
2549 down_write(&slub_lock);
2550 list_for_each_entry(s, &slab_caches, list) {
2551 if (s->flags & __SYSFS_ADD_DEFERRED) {
2552 s->flags &= ~__SYSFS_ADD_DEFERRED;
2556 up_write(&slub_lock);
2559 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2561 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2563 struct kmem_cache *s;
2567 s = kmalloc_caches_dma[index];
2571 /* Dynamically create dma cache */
2572 if (flags & __GFP_WAIT)
2573 down_write(&slub_lock);
2575 if (!down_write_trylock(&slub_lock))
2579 if (kmalloc_caches_dma[index])
2582 realsize = kmalloc_caches[index].objsize;
2583 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2584 (unsigned int)realsize);
2585 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2587 if (!s || !text || !kmem_cache_open(s, flags, text,
2588 realsize, ARCH_KMALLOC_MINALIGN,
2589 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2595 list_add(&s->list, &slab_caches);
2596 kmalloc_caches_dma[index] = s;
2598 schedule_work(&sysfs_add_work);
2601 up_write(&slub_lock);
2603 return kmalloc_caches_dma[index];
2608 * Conversion table for small slabs sizes / 8 to the index in the
2609 * kmalloc array. This is necessary for slabs < 192 since we have non power
2610 * of two cache sizes there. The size of larger slabs can be determined using
2613 static s8 size_index[24] = {
2640 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2646 return ZERO_SIZE_PTR;
2648 index = size_index[(size - 1) / 8];
2650 index = fls(size - 1);
2652 #ifdef CONFIG_ZONE_DMA
2653 if (unlikely((flags & SLUB_DMA)))
2654 return dma_kmalloc_cache(index, flags);
2657 return &kmalloc_caches[index];
2660 void *__kmalloc(size_t size, gfp_t flags)
2662 struct kmem_cache *s;
2664 if (unlikely(size > PAGE_SIZE))
2665 return kmalloc_large(size, flags);
2667 s = get_slab(size, flags);
2669 if (unlikely(ZERO_OR_NULL_PTR(s)))
2672 return slab_alloc(s, flags, -1, _RET_IP_);
2674 EXPORT_SYMBOL(__kmalloc);
2676 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2678 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2682 return page_address(page);
2688 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2690 struct kmem_cache *s;
2692 if (unlikely(size > PAGE_SIZE))
2693 return kmalloc_large_node(size, flags, node);
2695 s = get_slab(size, flags);
2697 if (unlikely(ZERO_OR_NULL_PTR(s)))
2700 return slab_alloc(s, flags, node, _RET_IP_);
2702 EXPORT_SYMBOL(__kmalloc_node);
2705 size_t ksize(const void *object)
2708 struct kmem_cache *s;
2710 if (unlikely(object == ZERO_SIZE_PTR))
2713 page = virt_to_head_page(object);
2715 if (unlikely(!PageSlab(page))) {
2716 WARN_ON(!PageCompound(page));
2717 return PAGE_SIZE << compound_order(page);
2721 #ifdef CONFIG_SLUB_DEBUG
2723 * Debugging requires use of the padding between object
2724 * and whatever may come after it.
2726 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2731 * If we have the need to store the freelist pointer
2732 * back there or track user information then we can
2733 * only use the space before that information.
2735 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2738 * Else we can use all the padding etc for the allocation
2743 void kfree(const void *x)
2746 void *object = (void *)x;
2748 if (unlikely(ZERO_OR_NULL_PTR(x)))
2751 page = virt_to_head_page(x);
2752 if (unlikely(!PageSlab(page))) {
2753 BUG_ON(!PageCompound(page));
2757 slab_free(page->slab, page, object, _RET_IP_);
2759 EXPORT_SYMBOL(kfree);
2762 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2763 * the remaining slabs by the number of items in use. The slabs with the
2764 * most items in use come first. New allocations will then fill those up
2765 * and thus they can be removed from the partial lists.
2767 * The slabs with the least items are placed last. This results in them
2768 * being allocated from last increasing the chance that the last objects
2769 * are freed in them.
2771 int kmem_cache_shrink(struct kmem_cache *s)
2775 struct kmem_cache_node *n;
2778 int objects = oo_objects(s->max);
2779 struct list_head *slabs_by_inuse =
2780 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2781 unsigned long flags;
2783 if (!slabs_by_inuse)
2787 for_each_node_state(node, N_NORMAL_MEMORY) {
2788 n = get_node(s, node);
2793 for (i = 0; i < objects; i++)
2794 INIT_LIST_HEAD(slabs_by_inuse + i);
2796 spin_lock_irqsave(&n->list_lock, flags);
2799 * Build lists indexed by the items in use in each slab.
2801 * Note that concurrent frees may occur while we hold the
2802 * list_lock. page->inuse here is the upper limit.
2804 list_for_each_entry_safe(page, t, &n->partial, lru) {
2805 if (!page->inuse && slab_trylock(page)) {
2807 * Must hold slab lock here because slab_free
2808 * may have freed the last object and be
2809 * waiting to release the slab.
2811 list_del(&page->lru);
2814 discard_slab(s, page);
2816 list_move(&page->lru,
2817 slabs_by_inuse + page->inuse);
2822 * Rebuild the partial list with the slabs filled up most
2823 * first and the least used slabs at the end.
2825 for (i = objects - 1; i >= 0; i--)
2826 list_splice(slabs_by_inuse + i, n->partial.prev);
2828 spin_unlock_irqrestore(&n->list_lock, flags);
2831 kfree(slabs_by_inuse);
2834 EXPORT_SYMBOL(kmem_cache_shrink);
2836 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2837 static int slab_mem_going_offline_callback(void *arg)
2839 struct kmem_cache *s;
2841 down_read(&slub_lock);
2842 list_for_each_entry(s, &slab_caches, list)
2843 kmem_cache_shrink(s);
2844 up_read(&slub_lock);
2849 static void slab_mem_offline_callback(void *arg)
2851 struct kmem_cache_node *n;
2852 struct kmem_cache *s;
2853 struct memory_notify *marg = arg;
2856 offline_node = marg->status_change_nid;
2859 * If the node still has available memory. we need kmem_cache_node
2862 if (offline_node < 0)
2865 down_read(&slub_lock);
2866 list_for_each_entry(s, &slab_caches, list) {
2867 n = get_node(s, offline_node);
2870 * if n->nr_slabs > 0, slabs still exist on the node
2871 * that is going down. We were unable to free them,
2872 * and offline_pages() function shoudn't call this
2873 * callback. So, we must fail.
2875 BUG_ON(slabs_node(s, offline_node));
2877 s->node[offline_node] = NULL;
2878 kmem_cache_free(kmalloc_caches, n);
2881 up_read(&slub_lock);
2884 static int slab_mem_going_online_callback(void *arg)
2886 struct kmem_cache_node *n;
2887 struct kmem_cache *s;
2888 struct memory_notify *marg = arg;
2889 int nid = marg->status_change_nid;
2893 * If the node's memory is already available, then kmem_cache_node is
2894 * already created. Nothing to do.
2900 * We are bringing a node online. No memory is available yet. We must
2901 * allocate a kmem_cache_node structure in order to bring the node
2904 down_read(&slub_lock);
2905 list_for_each_entry(s, &slab_caches, list) {
2907 * XXX: kmem_cache_alloc_node will fallback to other nodes
2908 * since memory is not yet available from the node that
2911 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2916 init_kmem_cache_node(n, s);
2920 up_read(&slub_lock);
2924 static int slab_memory_callback(struct notifier_block *self,
2925 unsigned long action, void *arg)
2930 case MEM_GOING_ONLINE:
2931 ret = slab_mem_going_online_callback(arg);
2933 case MEM_GOING_OFFLINE:
2934 ret = slab_mem_going_offline_callback(arg);
2937 case MEM_CANCEL_ONLINE:
2938 slab_mem_offline_callback(arg);
2941 case MEM_CANCEL_OFFLINE:
2945 ret = notifier_from_errno(ret);
2951 #endif /* CONFIG_MEMORY_HOTPLUG */
2953 /********************************************************************
2954 * Basic setup of slabs
2955 *******************************************************************/
2957 void __init kmem_cache_init(void)
2966 * Must first have the slab cache available for the allocations of the
2967 * struct kmem_cache_node's. There is special bootstrap code in
2968 * kmem_cache_open for slab_state == DOWN.
2970 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2971 sizeof(struct kmem_cache_node), GFP_KERNEL);
2972 kmalloc_caches[0].refcount = -1;
2975 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
2978 /* Able to allocate the per node structures */
2979 slab_state = PARTIAL;
2981 /* Caches that are not of the two-to-the-power-of size */
2982 if (KMALLOC_MIN_SIZE <= 64) {
2983 create_kmalloc_cache(&kmalloc_caches[1],
2984 "kmalloc-96", 96, GFP_KERNEL);
2986 create_kmalloc_cache(&kmalloc_caches[2],
2987 "kmalloc-192", 192, GFP_KERNEL);
2991 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
2992 create_kmalloc_cache(&kmalloc_caches[i],
2993 "kmalloc", 1 << i, GFP_KERNEL);
2999 * Patch up the size_index table if we have strange large alignment
3000 * requirements for the kmalloc array. This is only the case for
3001 * MIPS it seems. The standard arches will not generate any code here.
3003 * Largest permitted alignment is 256 bytes due to the way we
3004 * handle the index determination for the smaller caches.
3006 * Make sure that nothing crazy happens if someone starts tinkering
3007 * around with ARCH_KMALLOC_MINALIGN
3009 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3010 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3012 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3013 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3015 if (KMALLOC_MIN_SIZE == 128) {
3017 * The 192 byte sized cache is not used if the alignment
3018 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3021 for (i = 128 + 8; i <= 192; i += 8)
3022 size_index[(i - 1) / 8] = 8;
3027 /* Provide the correct kmalloc names now that the caches are up */
3028 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3029 kmalloc_caches[i]. name =
3030 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3033 register_cpu_notifier(&slab_notifier);
3034 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3035 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3037 kmem_size = sizeof(struct kmem_cache);
3041 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3042 " CPUs=%d, Nodes=%d\n",
3043 caches, cache_line_size(),
3044 slub_min_order, slub_max_order, slub_min_objects,
3045 nr_cpu_ids, nr_node_ids);
3049 * Find a mergeable slab cache
3051 static int slab_unmergeable(struct kmem_cache *s)
3053 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3060 * We may have set a slab to be unmergeable during bootstrap.
3062 if (s->refcount < 0)
3068 static struct kmem_cache *find_mergeable(size_t size,
3069 size_t align, unsigned long flags, const char *name,
3070 void (*ctor)(void *))
3072 struct kmem_cache *s;
3074 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3080 size = ALIGN(size, sizeof(void *));
3081 align = calculate_alignment(flags, align, size);
3082 size = ALIGN(size, align);
3083 flags = kmem_cache_flags(size, flags, name, NULL);
3085 list_for_each_entry(s, &slab_caches, list) {
3086 if (slab_unmergeable(s))
3092 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3095 * Check if alignment is compatible.
3096 * Courtesy of Adrian Drzewiecki
3098 if ((s->size & ~(align - 1)) != s->size)
3101 if (s->size - size >= sizeof(void *))
3109 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3110 size_t align, unsigned long flags, void (*ctor)(void *))
3112 struct kmem_cache *s;
3114 down_write(&slub_lock);
3115 s = find_mergeable(size, align, flags, name, ctor);
3121 * Adjust the object sizes so that we clear
3122 * the complete object on kzalloc.
3124 s->objsize = max(s->objsize, (int)size);
3127 * And then we need to update the object size in the
3128 * per cpu structures
3130 for_each_online_cpu(cpu)
3131 get_cpu_slab(s, cpu)->objsize = s->objsize;
3133 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3134 up_write(&slub_lock);
3136 if (sysfs_slab_alias(s, name)) {
3137 down_write(&slub_lock);
3139 up_write(&slub_lock);
3145 s = kmalloc(kmem_size, GFP_KERNEL);
3147 if (kmem_cache_open(s, GFP_KERNEL, name,
3148 size, align, flags, ctor)) {
3149 list_add(&s->list, &slab_caches);
3150 up_write(&slub_lock);
3151 if (sysfs_slab_add(s)) {
3152 down_write(&slub_lock);
3154 up_write(&slub_lock);
3162 up_write(&slub_lock);
3165 if (flags & SLAB_PANIC)
3166 panic("Cannot create slabcache %s\n", name);
3171 EXPORT_SYMBOL(kmem_cache_create);
3175 * Use the cpu notifier to insure that the cpu slabs are flushed when
3178 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3179 unsigned long action, void *hcpu)
3181 long cpu = (long)hcpu;
3182 struct kmem_cache *s;
3183 unsigned long flags;
3186 case CPU_UP_PREPARE:
3187 case CPU_UP_PREPARE_FROZEN:
3188 init_alloc_cpu_cpu(cpu);
3189 down_read(&slub_lock);
3190 list_for_each_entry(s, &slab_caches, list)
3191 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3193 up_read(&slub_lock);
3196 case CPU_UP_CANCELED:
3197 case CPU_UP_CANCELED_FROZEN:
3199 case CPU_DEAD_FROZEN:
3200 down_read(&slub_lock);
3201 list_for_each_entry(s, &slab_caches, list) {
3202 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3204 local_irq_save(flags);
3205 __flush_cpu_slab(s, cpu);
3206 local_irq_restore(flags);
3207 free_kmem_cache_cpu(c, cpu);
3208 s->cpu_slab[cpu] = NULL;
3210 up_read(&slub_lock);
3218 static struct notifier_block __cpuinitdata slab_notifier = {
3219 .notifier_call = slab_cpuup_callback
3224 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3226 struct kmem_cache *s;
3228 if (unlikely(size > PAGE_SIZE))
3229 return kmalloc_large(size, gfpflags);
3231 s = get_slab(size, gfpflags);
3233 if (unlikely(ZERO_OR_NULL_PTR(s)))
3236 return slab_alloc(s, gfpflags, -1, caller);
3239 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3240 int node, unsigned long caller)
3242 struct kmem_cache *s;
3244 if (unlikely(size > PAGE_SIZE))
3245 return kmalloc_large_node(size, gfpflags, node);
3247 s = get_slab(size, gfpflags);
3249 if (unlikely(ZERO_OR_NULL_PTR(s)))
3252 return slab_alloc(s, gfpflags, node, caller);
3255 #ifdef CONFIG_SLUB_DEBUG
3256 static unsigned long count_partial(struct kmem_cache_node *n,
3257 int (*get_count)(struct page *))
3259 unsigned long flags;
3260 unsigned long x = 0;
3263 spin_lock_irqsave(&n->list_lock, flags);
3264 list_for_each_entry(page, &n->partial, lru)
3265 x += get_count(page);
3266 spin_unlock_irqrestore(&n->list_lock, flags);
3270 static int count_inuse(struct page *page)
3275 static int count_total(struct page *page)
3277 return page->objects;
3280 static int count_free(struct page *page)
3282 return page->objects - page->inuse;
3285 static int validate_slab(struct kmem_cache *s, struct page *page,
3289 void *addr = page_address(page);
3291 if (!check_slab(s, page) ||
3292 !on_freelist(s, page, NULL))
3295 /* Now we know that a valid freelist exists */
3296 bitmap_zero(map, page->objects);
3298 for_each_free_object(p, s, page->freelist) {
3299 set_bit(slab_index(p, s, addr), map);
3300 if (!check_object(s, page, p, 0))
3304 for_each_object(p, s, addr, page->objects)
3305 if (!test_bit(slab_index(p, s, addr), map))
3306 if (!check_object(s, page, p, 1))
3311 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3314 if (slab_trylock(page)) {
3315 validate_slab(s, page, map);
3318 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3321 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3322 if (!PageSlubDebug(page))
3323 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3324 "on slab 0x%p\n", s->name, page);
3326 if (PageSlubDebug(page))
3327 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3328 "slab 0x%p\n", s->name, page);
3332 static int validate_slab_node(struct kmem_cache *s,
3333 struct kmem_cache_node *n, unsigned long *map)
3335 unsigned long count = 0;
3337 unsigned long flags;
3339 spin_lock_irqsave(&n->list_lock, flags);
3341 list_for_each_entry(page, &n->partial, lru) {
3342 validate_slab_slab(s, page, map);
3345 if (count != n->nr_partial)
3346 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3347 "counter=%ld\n", s->name, count, n->nr_partial);
3349 if (!(s->flags & SLAB_STORE_USER))
3352 list_for_each_entry(page, &n->full, lru) {
3353 validate_slab_slab(s, page, map);
3356 if (count != atomic_long_read(&n->nr_slabs))
3357 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3358 "counter=%ld\n", s->name, count,
3359 atomic_long_read(&n->nr_slabs));
3362 spin_unlock_irqrestore(&n->list_lock, flags);
3366 static long validate_slab_cache(struct kmem_cache *s)
3369 unsigned long count = 0;
3370 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3371 sizeof(unsigned long), GFP_KERNEL);
3377 for_each_node_state(node, N_NORMAL_MEMORY) {
3378 struct kmem_cache_node *n = get_node(s, node);
3380 count += validate_slab_node(s, n, map);
3386 #ifdef SLUB_RESILIENCY_TEST
3387 static void resiliency_test(void)
3391 printk(KERN_ERR "SLUB resiliency testing\n");
3392 printk(KERN_ERR "-----------------------\n");
3393 printk(KERN_ERR "A. Corruption after allocation\n");
3395 p = kzalloc(16, GFP_KERNEL);
3397 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3398 " 0x12->0x%p\n\n", p + 16);
3400 validate_slab_cache(kmalloc_caches + 4);
3402 /* Hmmm... The next two are dangerous */
3403 p = kzalloc(32, GFP_KERNEL);
3404 p[32 + sizeof(void *)] = 0x34;
3405 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3406 " 0x34 -> -0x%p\n", p);
3408 "If allocated object is overwritten then not detectable\n\n");
3410 validate_slab_cache(kmalloc_caches + 5);
3411 p = kzalloc(64, GFP_KERNEL);
3412 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3414 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3417 "If allocated object is overwritten then not detectable\n\n");
3418 validate_slab_cache(kmalloc_caches + 6);
3420 printk(KERN_ERR "\nB. Corruption after free\n");
3421 p = kzalloc(128, GFP_KERNEL);
3424 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3425 validate_slab_cache(kmalloc_caches + 7);
3427 p = kzalloc(256, GFP_KERNEL);
3430 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3432 validate_slab_cache(kmalloc_caches + 8);
3434 p = kzalloc(512, GFP_KERNEL);
3437 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3438 validate_slab_cache(kmalloc_caches + 9);
3441 static void resiliency_test(void) {};
3445 * Generate lists of code addresses where slabcache objects are allocated
3450 unsigned long count;
3457 DECLARE_BITMAP(cpus, NR_CPUS);
3463 unsigned long count;
3464 struct location *loc;
3467 static void free_loc_track(struct loc_track *t)
3470 free_pages((unsigned long)t->loc,
3471 get_order(sizeof(struct location) * t->max));
3474 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3479 order = get_order(sizeof(struct location) * max);
3481 l = (void *)__get_free_pages(flags, order);
3486 memcpy(l, t->loc, sizeof(struct location) * t->count);
3494 static int add_location(struct loc_track *t, struct kmem_cache *s,
3495 const struct track *track)
3497 long start, end, pos;
3499 unsigned long caddr;
3500 unsigned long age = jiffies - track->when;
3506 pos = start + (end - start + 1) / 2;
3509 * There is nothing at "end". If we end up there
3510 * we need to add something to before end.
3515 caddr = t->loc[pos].addr;
3516 if (track->addr == caddr) {
3522 if (age < l->min_time)
3524 if (age > l->max_time)
3527 if (track->pid < l->min_pid)
3528 l->min_pid = track->pid;
3529 if (track->pid > l->max_pid)
3530 l->max_pid = track->pid;
3532 cpumask_set_cpu(track->cpu,
3533 to_cpumask(l->cpus));
3535 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3539 if (track->addr < caddr)
3546 * Not found. Insert new tracking element.
3548 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3554 (t->count - pos) * sizeof(struct location));
3557 l->addr = track->addr;
3561 l->min_pid = track->pid;
3562 l->max_pid = track->pid;
3563 cpumask_clear(to_cpumask(l->cpus));
3564 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3565 nodes_clear(l->nodes);
3566 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3570 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3571 struct page *page, enum track_item alloc)
3573 void *addr = page_address(page);
3574 DECLARE_BITMAP(map, page->objects);
3577 bitmap_zero(map, page->objects);
3578 for_each_free_object(p, s, page->freelist)
3579 set_bit(slab_index(p, s, addr), map);
3581 for_each_object(p, s, addr, page->objects)
3582 if (!test_bit(slab_index(p, s, addr), map))
3583 add_location(t, s, get_track(s, p, alloc));
3586 static int list_locations(struct kmem_cache *s, char *buf,
3587 enum track_item alloc)
3591 struct loc_track t = { 0, 0, NULL };
3594 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3596 return sprintf(buf, "Out of memory\n");
3598 /* Push back cpu slabs */
3601 for_each_node_state(node, N_NORMAL_MEMORY) {
3602 struct kmem_cache_node *n = get_node(s, node);
3603 unsigned long flags;
3606 if (!atomic_long_read(&n->nr_slabs))
3609 spin_lock_irqsave(&n->list_lock, flags);
3610 list_for_each_entry(page, &n->partial, lru)
3611 process_slab(&t, s, page, alloc);
3612 list_for_each_entry(page, &n->full, lru)
3613 process_slab(&t, s, page, alloc);
3614 spin_unlock_irqrestore(&n->list_lock, flags);
3617 for (i = 0; i < t.count; i++) {
3618 struct location *l = &t.loc[i];
3620 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3622 len += sprintf(buf + len, "%7ld ", l->count);
3625 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3627 len += sprintf(buf + len, "<not-available>");
3629 if (l->sum_time != l->min_time) {
3630 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3632 (long)div_u64(l->sum_time, l->count),
3635 len += sprintf(buf + len, " age=%ld",
3638 if (l->min_pid != l->max_pid)
3639 len += sprintf(buf + len, " pid=%ld-%ld",
3640 l->min_pid, l->max_pid);
3642 len += sprintf(buf + len, " pid=%ld",
3645 if (num_online_cpus() > 1 &&
3646 !cpumask_empty(to_cpumask(l->cpus)) &&
3647 len < PAGE_SIZE - 60) {
3648 len += sprintf(buf + len, " cpus=");
3649 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3650 to_cpumask(l->cpus));
3653 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3654 len < PAGE_SIZE - 60) {
3655 len += sprintf(buf + len, " nodes=");
3656 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3660 len += sprintf(buf + len, "\n");
3665 len += sprintf(buf, "No data\n");
3669 enum slab_stat_type {
3670 SL_ALL, /* All slabs */
3671 SL_PARTIAL, /* Only partially allocated slabs */
3672 SL_CPU, /* Only slabs used for cpu caches */
3673 SL_OBJECTS, /* Determine allocated objects not slabs */
3674 SL_TOTAL /* Determine object capacity not slabs */
3677 #define SO_ALL (1 << SL_ALL)
3678 #define SO_PARTIAL (1 << SL_PARTIAL)
3679 #define SO_CPU (1 << SL_CPU)
3680 #define SO_OBJECTS (1 << SL_OBJECTS)
3681 #define SO_TOTAL (1 << SL_TOTAL)
3683 static ssize_t show_slab_objects(struct kmem_cache *s,
3684 char *buf, unsigned long flags)
3686 unsigned long total = 0;
3689 unsigned long *nodes;
3690 unsigned long *per_cpu;
3692 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3695 per_cpu = nodes + nr_node_ids;
3697 if (flags & SO_CPU) {
3700 for_each_possible_cpu(cpu) {
3701 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3703 if (!c || c->node < 0)
3707 if (flags & SO_TOTAL)
3708 x = c->page->objects;
3709 else if (flags & SO_OBJECTS)
3715 nodes[c->node] += x;
3721 if (flags & SO_ALL) {
3722 for_each_node_state(node, N_NORMAL_MEMORY) {
3723 struct kmem_cache_node *n = get_node(s, node);
3725 if (flags & SO_TOTAL)
3726 x = atomic_long_read(&n->total_objects);
3727 else if (flags & SO_OBJECTS)
3728 x = atomic_long_read(&n->total_objects) -
3729 count_partial(n, count_free);
3732 x = atomic_long_read(&n->nr_slabs);
3737 } else if (flags & SO_PARTIAL) {
3738 for_each_node_state(node, N_NORMAL_MEMORY) {
3739 struct kmem_cache_node *n = get_node(s, node);
3741 if (flags & SO_TOTAL)
3742 x = count_partial(n, count_total);
3743 else if (flags & SO_OBJECTS)
3744 x = count_partial(n, count_inuse);
3751 x = sprintf(buf, "%lu", total);
3753 for_each_node_state(node, N_NORMAL_MEMORY)
3755 x += sprintf(buf + x, " N%d=%lu",
3759 return x + sprintf(buf + x, "\n");
3762 static int any_slab_objects(struct kmem_cache *s)
3766 for_each_online_node(node) {
3767 struct kmem_cache_node *n = get_node(s, node);
3772 if (atomic_long_read(&n->total_objects))
3778 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3779 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3781 struct slab_attribute {
3782 struct attribute attr;
3783 ssize_t (*show)(struct kmem_cache *s, char *buf);
3784 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3787 #define SLAB_ATTR_RO(_name) \
3788 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3790 #define SLAB_ATTR(_name) \
3791 static struct slab_attribute _name##_attr = \
3792 __ATTR(_name, 0644, _name##_show, _name##_store)
3794 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3796 return sprintf(buf, "%d\n", s->size);
3798 SLAB_ATTR_RO(slab_size);
3800 static ssize_t align_show(struct kmem_cache *s, char *buf)
3802 return sprintf(buf, "%d\n", s->align);
3804 SLAB_ATTR_RO(align);
3806 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3808 return sprintf(buf, "%d\n", s->objsize);
3810 SLAB_ATTR_RO(object_size);
3812 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3814 return sprintf(buf, "%d\n", oo_objects(s->oo));
3816 SLAB_ATTR_RO(objs_per_slab);
3818 static ssize_t order_store(struct kmem_cache *s,
3819 const char *buf, size_t length)
3821 unsigned long order;
3824 err = strict_strtoul(buf, 10, &order);
3828 if (order > slub_max_order || order < slub_min_order)
3831 calculate_sizes(s, order);
3835 static ssize_t order_show(struct kmem_cache *s, char *buf)
3837 return sprintf(buf, "%d\n", oo_order(s->oo));
3841 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3844 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3846 return n + sprintf(buf + n, "\n");
3852 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3854 return sprintf(buf, "%d\n", s->refcount - 1);
3856 SLAB_ATTR_RO(aliases);
3858 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3860 return show_slab_objects(s, buf, SO_ALL);
3862 SLAB_ATTR_RO(slabs);
3864 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3866 return show_slab_objects(s, buf, SO_PARTIAL);
3868 SLAB_ATTR_RO(partial);
3870 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3872 return show_slab_objects(s, buf, SO_CPU);
3874 SLAB_ATTR_RO(cpu_slabs);
3876 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3878 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3880 SLAB_ATTR_RO(objects);
3882 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3884 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3886 SLAB_ATTR_RO(objects_partial);
3888 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3890 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3892 SLAB_ATTR_RO(total_objects);
3894 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3896 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3899 static ssize_t sanity_checks_store(struct kmem_cache *s,
3900 const char *buf, size_t length)
3902 s->flags &= ~SLAB_DEBUG_FREE;
3904 s->flags |= SLAB_DEBUG_FREE;
3907 SLAB_ATTR(sanity_checks);
3909 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3911 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3914 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3917 s->flags &= ~SLAB_TRACE;
3919 s->flags |= SLAB_TRACE;
3924 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3926 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3929 static ssize_t reclaim_account_store(struct kmem_cache *s,
3930 const char *buf, size_t length)
3932 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3934 s->flags |= SLAB_RECLAIM_ACCOUNT;
3937 SLAB_ATTR(reclaim_account);
3939 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3941 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3943 SLAB_ATTR_RO(hwcache_align);
3945 #ifdef CONFIG_ZONE_DMA
3946 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3948 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3950 SLAB_ATTR_RO(cache_dma);
3953 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3955 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3957 SLAB_ATTR_RO(destroy_by_rcu);
3959 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3961 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3964 static ssize_t red_zone_store(struct kmem_cache *s,
3965 const char *buf, size_t length)
3967 if (any_slab_objects(s))
3970 s->flags &= ~SLAB_RED_ZONE;
3972 s->flags |= SLAB_RED_ZONE;
3973 calculate_sizes(s, -1);
3976 SLAB_ATTR(red_zone);
3978 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3980 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3983 static ssize_t poison_store(struct kmem_cache *s,
3984 const char *buf, size_t length)
3986 if (any_slab_objects(s))
3989 s->flags &= ~SLAB_POISON;
3991 s->flags |= SLAB_POISON;
3992 calculate_sizes(s, -1);
3997 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3999 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4002 static ssize_t store_user_store(struct kmem_cache *s,
4003 const char *buf, size_t length)
4005 if (any_slab_objects(s))
4008 s->flags &= ~SLAB_STORE_USER;
4010 s->flags |= SLAB_STORE_USER;
4011 calculate_sizes(s, -1);
4014 SLAB_ATTR(store_user);
4016 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4021 static ssize_t validate_store(struct kmem_cache *s,
4022 const char *buf, size_t length)
4026 if (buf[0] == '1') {
4027 ret = validate_slab_cache(s);
4033 SLAB_ATTR(validate);
4035 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4040 static ssize_t shrink_store(struct kmem_cache *s,
4041 const char *buf, size_t length)
4043 if (buf[0] == '1') {
4044 int rc = kmem_cache_shrink(s);
4054 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4056 if (!(s->flags & SLAB_STORE_USER))
4058 return list_locations(s, buf, TRACK_ALLOC);
4060 SLAB_ATTR_RO(alloc_calls);
4062 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4064 if (!(s->flags & SLAB_STORE_USER))
4066 return list_locations(s, buf, TRACK_FREE);
4068 SLAB_ATTR_RO(free_calls);
4071 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4073 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4076 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4077 const char *buf, size_t length)
4079 unsigned long ratio;
4082 err = strict_strtoul(buf, 10, &ratio);
4087 s->remote_node_defrag_ratio = ratio * 10;
4091 SLAB_ATTR(remote_node_defrag_ratio);
4094 #ifdef CONFIG_SLUB_STATS
4095 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4097 unsigned long sum = 0;
4100 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4105 for_each_online_cpu(cpu) {
4106 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4112 len = sprintf(buf, "%lu", sum);
4115 for_each_online_cpu(cpu) {
4116 if (data[cpu] && len < PAGE_SIZE - 20)
4117 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4121 return len + sprintf(buf + len, "\n");
4124 #define STAT_ATTR(si, text) \
4125 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4127 return show_stat(s, buf, si); \
4129 SLAB_ATTR_RO(text); \
4131 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4132 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4133 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4134 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4135 STAT_ATTR(FREE_FROZEN, free_frozen);
4136 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4137 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4138 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4139 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4140 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4141 STAT_ATTR(FREE_SLAB, free_slab);
4142 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4143 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4144 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4145 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4146 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4147 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4148 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4151 static struct attribute *slab_attrs[] = {
4152 &slab_size_attr.attr,
4153 &object_size_attr.attr,
4154 &objs_per_slab_attr.attr,
4157 &objects_partial_attr.attr,
4158 &total_objects_attr.attr,
4161 &cpu_slabs_attr.attr,
4165 &sanity_checks_attr.attr,
4167 &hwcache_align_attr.attr,
4168 &reclaim_account_attr.attr,
4169 &destroy_by_rcu_attr.attr,
4170 &red_zone_attr.attr,
4172 &store_user_attr.attr,
4173 &validate_attr.attr,
4175 &alloc_calls_attr.attr,
4176 &free_calls_attr.attr,
4177 #ifdef CONFIG_ZONE_DMA
4178 &cache_dma_attr.attr,
4181 &remote_node_defrag_ratio_attr.attr,
4183 #ifdef CONFIG_SLUB_STATS
4184 &alloc_fastpath_attr.attr,
4185 &alloc_slowpath_attr.attr,
4186 &free_fastpath_attr.attr,
4187 &free_slowpath_attr.attr,
4188 &free_frozen_attr.attr,
4189 &free_add_partial_attr.attr,
4190 &free_remove_partial_attr.attr,
4191 &alloc_from_partial_attr.attr,
4192 &alloc_slab_attr.attr,
4193 &alloc_refill_attr.attr,
4194 &free_slab_attr.attr,
4195 &cpuslab_flush_attr.attr,
4196 &deactivate_full_attr.attr,
4197 &deactivate_empty_attr.attr,
4198 &deactivate_to_head_attr.attr,
4199 &deactivate_to_tail_attr.attr,
4200 &deactivate_remote_frees_attr.attr,
4201 &order_fallback_attr.attr,
4206 static struct attribute_group slab_attr_group = {
4207 .attrs = slab_attrs,
4210 static ssize_t slab_attr_show(struct kobject *kobj,
4211 struct attribute *attr,
4214 struct slab_attribute *attribute;
4215 struct kmem_cache *s;
4218 attribute = to_slab_attr(attr);
4221 if (!attribute->show)
4224 err = attribute->show(s, buf);
4229 static ssize_t slab_attr_store(struct kobject *kobj,
4230 struct attribute *attr,
4231 const char *buf, size_t len)
4233 struct slab_attribute *attribute;
4234 struct kmem_cache *s;
4237 attribute = to_slab_attr(attr);
4240 if (!attribute->store)
4243 err = attribute->store(s, buf, len);
4248 static void kmem_cache_release(struct kobject *kobj)
4250 struct kmem_cache *s = to_slab(kobj);
4255 static struct sysfs_ops slab_sysfs_ops = {
4256 .show = slab_attr_show,
4257 .store = slab_attr_store,
4260 static struct kobj_type slab_ktype = {
4261 .sysfs_ops = &slab_sysfs_ops,
4262 .release = kmem_cache_release
4265 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4267 struct kobj_type *ktype = get_ktype(kobj);
4269 if (ktype == &slab_ktype)
4274 static struct kset_uevent_ops slab_uevent_ops = {
4275 .filter = uevent_filter,
4278 static struct kset *slab_kset;
4280 #define ID_STR_LENGTH 64
4282 /* Create a unique string id for a slab cache:
4284 * Format :[flags-]size
4286 static char *create_unique_id(struct kmem_cache *s)
4288 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4295 * First flags affecting slabcache operations. We will only
4296 * get here for aliasable slabs so we do not need to support
4297 * too many flags. The flags here must cover all flags that
4298 * are matched during merging to guarantee that the id is
4301 if (s->flags & SLAB_CACHE_DMA)
4303 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4305 if (s->flags & SLAB_DEBUG_FREE)
4309 p += sprintf(p, "%07d", s->size);
4310 BUG_ON(p > name + ID_STR_LENGTH - 1);
4314 static int sysfs_slab_add(struct kmem_cache *s)
4320 if (slab_state < SYSFS)
4321 /* Defer until later */
4324 unmergeable = slab_unmergeable(s);
4327 * Slabcache can never be merged so we can use the name proper.
4328 * This is typically the case for debug situations. In that
4329 * case we can catch duplicate names easily.
4331 sysfs_remove_link(&slab_kset->kobj, s->name);
4335 * Create a unique name for the slab as a target
4338 name = create_unique_id(s);
4341 s->kobj.kset = slab_kset;
4342 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4344 kobject_put(&s->kobj);
4348 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4351 kobject_uevent(&s->kobj, KOBJ_ADD);
4353 /* Setup first alias */
4354 sysfs_slab_alias(s, s->name);
4360 static void sysfs_slab_remove(struct kmem_cache *s)
4362 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4363 kobject_del(&s->kobj);
4364 kobject_put(&s->kobj);
4368 * Need to buffer aliases during bootup until sysfs becomes
4369 * available lest we lose that information.
4371 struct saved_alias {
4372 struct kmem_cache *s;
4374 struct saved_alias *next;
4377 static struct saved_alias *alias_list;
4379 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4381 struct saved_alias *al;
4383 if (slab_state == SYSFS) {
4385 * If we have a leftover link then remove it.
4387 sysfs_remove_link(&slab_kset->kobj, name);
4388 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4391 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4397 al->next = alias_list;
4402 static int __init slab_sysfs_init(void)
4404 struct kmem_cache *s;
4407 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4409 printk(KERN_ERR "Cannot register slab subsystem.\n");
4415 list_for_each_entry(s, &slab_caches, list) {
4416 err = sysfs_slab_add(s);
4418 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4419 " to sysfs\n", s->name);
4422 while (alias_list) {
4423 struct saved_alias *al = alias_list;
4425 alias_list = alias_list->next;
4426 err = sysfs_slab_alias(al->s, al->name);
4428 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4429 " %s to sysfs\n", s->name);
4437 __initcall(slab_sysfs_init);
4441 * The /proc/slabinfo ABI
4443 #ifdef CONFIG_SLABINFO
4444 static void print_slabinfo_header(struct seq_file *m)
4446 seq_puts(m, "slabinfo - version: 2.1\n");
4447 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4448 "<objperslab> <pagesperslab>");
4449 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4450 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4454 static void *s_start(struct seq_file *m, loff_t *pos)
4458 down_read(&slub_lock);
4460 print_slabinfo_header(m);
4462 return seq_list_start(&slab_caches, *pos);
4465 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4467 return seq_list_next(p, &slab_caches, pos);
4470 static void s_stop(struct seq_file *m, void *p)
4472 up_read(&slub_lock);
4475 static int s_show(struct seq_file *m, void *p)
4477 unsigned long nr_partials = 0;
4478 unsigned long nr_slabs = 0;
4479 unsigned long nr_inuse = 0;
4480 unsigned long nr_objs = 0;
4481 unsigned long nr_free = 0;
4482 struct kmem_cache *s;
4485 s = list_entry(p, struct kmem_cache, list);
4487 for_each_online_node(node) {
4488 struct kmem_cache_node *n = get_node(s, node);
4493 nr_partials += n->nr_partial;
4494 nr_slabs += atomic_long_read(&n->nr_slabs);
4495 nr_objs += atomic_long_read(&n->total_objects);
4496 nr_free += count_partial(n, count_free);
4499 nr_inuse = nr_objs - nr_free;
4501 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4502 nr_objs, s->size, oo_objects(s->oo),
4503 (1 << oo_order(s->oo)));
4504 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4505 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4511 static const struct seq_operations slabinfo_op = {
4518 static int slabinfo_open(struct inode *inode, struct file *file)
4520 return seq_open(file, &slabinfo_op);
4523 static const struct file_operations proc_slabinfo_operations = {
4524 .open = slabinfo_open,
4526 .llseek = seq_lseek,
4527 .release = seq_release,
4530 static int __init slab_proc_init(void)
4532 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4535 module_init(slab_proc_init);
4536 #endif /* CONFIG_SLABINFO */