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 <clameter@sgi.com>
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/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
23 #include <linux/memory.h>
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
72 * freed then the slab will show up again on the partial lists.
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
80 * Overloading of page flags that are otherwise used for LRU management.
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
94 * freelist that allows lockless access to
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
100 * the fast path and disables lockless freelists.
103 #define FROZEN (1 << PG_active)
105 #ifdef CONFIG_SLUB_DEBUG
106 #define SLABDEBUG (1 << PG_error)
111 static inline int SlabFrozen(struct page *page)
113 return page->flags & FROZEN;
116 static inline void SetSlabFrozen(struct page *page)
118 page->flags |= FROZEN;
121 static inline void ClearSlabFrozen(struct page *page)
123 page->flags &= ~FROZEN;
126 static inline int SlabDebug(struct page *page)
128 return page->flags & SLABDEBUG;
131 static inline void SetSlabDebug(struct page *page)
133 page->flags |= SLABDEBUG;
136 static inline void ClearSlabDebug(struct page *page)
138 page->flags &= ~SLABDEBUG;
142 * Issues still to be resolved:
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
146 * - Variable sizing of the per node arrays
149 /* Enable to test recovery from slab corruption on boot */
150 #undef SLUB_RESILIENCY_TEST
155 * Small page size. Make sure that we do not fragment memory
157 #define DEFAULT_MAX_ORDER 1
158 #define DEFAULT_MIN_OBJECTS 4
163 * Large page machines are customarily able to handle larger
166 #define DEFAULT_MAX_ORDER 2
167 #define DEFAULT_MIN_OBJECTS 8
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
175 #define MIN_PARTIAL 5
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
182 #define MAX_PARTIAL 10
184 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
188 * Set of flags that will prevent slab merging
190 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
193 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
196 #ifndef ARCH_KMALLOC_MINALIGN
197 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
200 #ifndef ARCH_SLAB_MINALIGN
201 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 /* Internal SLUB flags */
205 #define __OBJECT_POISON 0x80000000 /* Poison object */
206 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
208 /* Not all arches define cache_line_size */
209 #ifndef cache_line_size
210 #define cache_line_size() L1_CACHE_BYTES
213 static int kmem_size = sizeof(struct kmem_cache);
216 static struct notifier_block slab_notifier;
220 DOWN, /* No slab functionality available */
221 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
222 UP, /* Everything works but does not show up in sysfs */
226 /* A list of all slab caches on the system */
227 static DECLARE_RWSEM(slub_lock);
228 static LIST_HEAD(slab_caches);
231 * Tracking user of a slab.
234 void *addr; /* Called from address */
235 int cpu; /* Was running on cpu */
236 int pid; /* Pid context */
237 unsigned long when; /* When did the operation occur */
240 enum track_item { TRACK_ALLOC, TRACK_FREE };
242 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
243 static int sysfs_slab_add(struct kmem_cache *);
244 static int sysfs_slab_alias(struct kmem_cache *, const char *);
245 static void sysfs_slab_remove(struct kmem_cache *);
247 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
248 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
250 static inline void sysfs_slab_remove(struct kmem_cache *s)
256 /********************************************************************
257 * Core slab cache functions
258 *******************************************************************/
260 int slab_is_available(void)
262 return slab_state >= UP;
265 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
268 return s->node[node];
270 return &s->local_node;
274 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
277 return s->cpu_slab[cpu];
283 static inline int check_valid_pointer(struct kmem_cache *s,
284 struct page *page, const void *object)
291 base = page_address(page);
292 if (object < base || object >= base + s->objects * s->size ||
293 (object - base) % s->size) {
301 * Slow version of get and set free pointer.
303 * This version requires touching the cache lines of kmem_cache which
304 * we avoid to do in the fast alloc free paths. There we obtain the offset
305 * from the page struct.
307 static inline void *get_freepointer(struct kmem_cache *s, void *object)
309 return *(void **)(object + s->offset);
312 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
314 *(void **)(object + s->offset) = fp;
317 /* Loop over all objects in a slab */
318 #define for_each_object(__p, __s, __addr) \
319 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
323 #define for_each_free_object(__p, __s, __free) \
324 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
326 /* Determine object index from a given position */
327 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
329 return (p - addr) / s->size;
332 #ifdef CONFIG_SLUB_DEBUG
336 #ifdef CONFIG_SLUB_DEBUG_ON
337 static int slub_debug = DEBUG_DEFAULT_FLAGS;
339 static int slub_debug;
342 static char *slub_debug_slabs;
347 static void print_section(char *text, u8 *addr, unsigned int length)
355 for (i = 0; i < length; i++) {
357 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
360 printk(" %02x", addr[i]);
362 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
364 printk(" %s\n",ascii);
375 printk(" %s\n", ascii);
379 static struct track *get_track(struct kmem_cache *s, void *object,
380 enum track_item alloc)
385 p = object + s->offset + sizeof(void *);
387 p = object + s->inuse;
392 static void set_track(struct kmem_cache *s, void *object,
393 enum track_item alloc, void *addr)
398 p = object + s->offset + sizeof(void *);
400 p = object + s->inuse;
405 p->cpu = smp_processor_id();
406 p->pid = current ? current->pid : -1;
409 memset(p, 0, sizeof(struct track));
412 static void init_tracking(struct kmem_cache *s, void *object)
414 if (!(s->flags & SLAB_STORE_USER))
417 set_track(s, object, TRACK_FREE, NULL);
418 set_track(s, object, TRACK_ALLOC, NULL);
421 static void print_track(const char *s, struct track *t)
426 printk(KERN_ERR "INFO: %s in ", s);
427 __print_symbol("%s", (unsigned long)t->addr);
428 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
431 static void print_tracking(struct kmem_cache *s, void *object)
433 if (!(s->flags & SLAB_STORE_USER))
436 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
437 print_track("Freed", get_track(s, object, TRACK_FREE));
440 static void print_page_info(struct page *page)
442 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
443 page, page->inuse, page->freelist, page->flags);
447 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
453 vsnprintf(buf, sizeof(buf), fmt, args);
455 printk(KERN_ERR "========================================"
456 "=====================================\n");
457 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
458 printk(KERN_ERR "----------------------------------------"
459 "-------------------------------------\n\n");
462 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
468 vsnprintf(buf, sizeof(buf), fmt, args);
470 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
473 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
475 unsigned int off; /* Offset of last byte */
476 u8 *addr = page_address(page);
478 print_tracking(s, p);
480 print_page_info(page);
482 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
483 p, p - addr, get_freepointer(s, p));
486 print_section("Bytes b4", p - 16, 16);
488 print_section("Object", p, min(s->objsize, 128));
490 if (s->flags & SLAB_RED_ZONE)
491 print_section("Redzone", p + s->objsize,
492 s->inuse - s->objsize);
495 off = s->offset + sizeof(void *);
499 if (s->flags & SLAB_STORE_USER)
500 off += 2 * sizeof(struct track);
503 /* Beginning of the filler is the free pointer */
504 print_section("Padding", p + off, s->size - off);
509 static void object_err(struct kmem_cache *s, struct page *page,
510 u8 *object, char *reason)
513 print_trailer(s, page, object);
516 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
522 vsnprintf(buf, sizeof(buf), fmt, args);
525 print_page_info(page);
529 static void init_object(struct kmem_cache *s, void *object, int active)
533 if (s->flags & __OBJECT_POISON) {
534 memset(p, POISON_FREE, s->objsize - 1);
535 p[s->objsize -1] = POISON_END;
538 if (s->flags & SLAB_RED_ZONE)
539 memset(p + s->objsize,
540 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
541 s->inuse - s->objsize);
544 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
547 if (*start != (u8)value)
555 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
556 void *from, void *to)
558 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
559 memset(from, data, to - from);
562 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
563 u8 *object, char *what,
564 u8* start, unsigned int value, unsigned int bytes)
569 fault = check_bytes(start, value, bytes);
574 while (end > fault && end[-1] == value)
577 slab_bug(s, "%s overwritten", what);
578 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
579 fault, end - 1, fault[0], value);
580 print_trailer(s, page, object);
582 restore_bytes(s, what, value, fault, end);
590 * Bytes of the object to be managed.
591 * If the freepointer may overlay the object then the free
592 * pointer is the first word of the object.
594 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
597 * object + s->objsize
598 * Padding to reach word boundary. This is also used for Redzoning.
599 * Padding is extended by another word if Redzoning is enabled and
602 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
603 * 0xcc (RED_ACTIVE) for objects in use.
606 * Meta data starts here.
608 * A. Free pointer (if we cannot overwrite object on free)
609 * B. Tracking data for SLAB_STORE_USER
610 * C. Padding to reach required alignment boundary or at mininum
611 * one word if debuggin is on to be able to detect writes
612 * before the word boundary.
614 * Padding is done using 0x5a (POISON_INUSE)
617 * Nothing is used beyond s->size.
619 * If slabcaches are merged then the objsize and inuse boundaries are mostly
620 * ignored. And therefore no slab options that rely on these boundaries
621 * may be used with merged slabcaches.
624 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
626 unsigned long off = s->inuse; /* The end of info */
629 /* Freepointer is placed after the object. */
630 off += sizeof(void *);
632 if (s->flags & SLAB_STORE_USER)
633 /* We also have user information there */
634 off += 2 * sizeof(struct track);
639 return check_bytes_and_report(s, page, p, "Object padding",
640 p + off, POISON_INUSE, s->size - off);
643 static int slab_pad_check(struct kmem_cache *s, struct page *page)
651 if (!(s->flags & SLAB_POISON))
654 start = page_address(page);
655 end = start + (PAGE_SIZE << s->order);
656 length = s->objects * s->size;
657 remainder = end - (start + length);
661 fault = check_bytes(start + length, POISON_INUSE, remainder);
664 while (end > fault && end[-1] == POISON_INUSE)
667 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
668 print_section("Padding", start, length);
670 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
674 static int check_object(struct kmem_cache *s, struct page *page,
675 void *object, int active)
678 u8 *endobject = object + s->objsize;
680 if (s->flags & SLAB_RED_ZONE) {
682 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
684 if (!check_bytes_and_report(s, page, object, "Redzone",
685 endobject, red, s->inuse - s->objsize))
688 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
689 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
690 POISON_INUSE, s->inuse - s->objsize);
693 if (s->flags & SLAB_POISON) {
694 if (!active && (s->flags & __OBJECT_POISON) &&
695 (!check_bytes_and_report(s, page, p, "Poison", p,
696 POISON_FREE, s->objsize - 1) ||
697 !check_bytes_and_report(s, page, p, "Poison",
698 p + s->objsize -1, POISON_END, 1)))
701 * check_pad_bytes cleans up on its own.
703 check_pad_bytes(s, page, p);
706 if (!s->offset && active)
708 * Object and freepointer overlap. Cannot check
709 * freepointer while object is allocated.
713 /* Check free pointer validity */
714 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
715 object_err(s, page, p, "Freepointer corrupt");
717 * No choice but to zap it and thus loose the remainder
718 * of the free objects in this slab. May cause
719 * another error because the object count is now wrong.
721 set_freepointer(s, p, NULL);
727 static int check_slab(struct kmem_cache *s, struct page *page)
729 VM_BUG_ON(!irqs_disabled());
731 if (!PageSlab(page)) {
732 slab_err(s, page, "Not a valid slab page");
735 if (page->inuse > s->objects) {
736 slab_err(s, page, "inuse %u > max %u",
737 s->name, page->inuse, s->objects);
740 /* Slab_pad_check fixes things up after itself */
741 slab_pad_check(s, page);
746 * Determine if a certain object on a page is on the freelist. Must hold the
747 * slab lock to guarantee that the chains are in a consistent state.
749 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
752 void *fp = page->freelist;
755 while (fp && nr <= s->objects) {
758 if (!check_valid_pointer(s, page, fp)) {
760 object_err(s, page, object,
761 "Freechain corrupt");
762 set_freepointer(s, object, NULL);
765 slab_err(s, page, "Freepointer corrupt");
766 page->freelist = NULL;
767 page->inuse = s->objects;
768 slab_fix(s, "Freelist cleared");
774 fp = get_freepointer(s, object);
778 if (page->inuse != s->objects - nr) {
779 slab_err(s, page, "Wrong object count. Counter is %d but "
780 "counted were %d", page->inuse, s->objects - nr);
781 page->inuse = s->objects - nr;
782 slab_fix(s, "Object count adjusted.");
784 return search == NULL;
787 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
789 if (s->flags & SLAB_TRACE) {
790 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
792 alloc ? "alloc" : "free",
797 print_section("Object", (void *)object, s->objsize);
804 * Tracking of fully allocated slabs for debugging purposes.
806 static void add_full(struct kmem_cache_node *n, struct page *page)
808 spin_lock(&n->list_lock);
809 list_add(&page->lru, &n->full);
810 spin_unlock(&n->list_lock);
813 static void remove_full(struct kmem_cache *s, struct page *page)
815 struct kmem_cache_node *n;
817 if (!(s->flags & SLAB_STORE_USER))
820 n = get_node(s, page_to_nid(page));
822 spin_lock(&n->list_lock);
823 list_del(&page->lru);
824 spin_unlock(&n->list_lock);
827 static void setup_object_debug(struct kmem_cache *s, struct page *page,
830 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
833 init_object(s, object, 0);
834 init_tracking(s, object);
837 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
838 void *object, void *addr)
840 if (!check_slab(s, page))
843 if (object && !on_freelist(s, page, object)) {
844 object_err(s, page, object, "Object already allocated");
848 if (!check_valid_pointer(s, page, object)) {
849 object_err(s, page, object, "Freelist Pointer check fails");
853 if (object && !check_object(s, page, object, 0))
856 /* Success perform special debug activities for allocs */
857 if (s->flags & SLAB_STORE_USER)
858 set_track(s, object, TRACK_ALLOC, addr);
859 trace(s, page, object, 1);
860 init_object(s, object, 1);
864 if (PageSlab(page)) {
866 * If this is a slab page then lets do the best we can
867 * to avoid issues in the future. Marking all objects
868 * as used avoids touching the remaining objects.
870 slab_fix(s, "Marking all objects used");
871 page->inuse = s->objects;
872 page->freelist = NULL;
877 static int free_debug_processing(struct kmem_cache *s, struct page *page,
878 void *object, void *addr)
880 if (!check_slab(s, page))
883 if (!check_valid_pointer(s, page, object)) {
884 slab_err(s, page, "Invalid object pointer 0x%p", object);
888 if (on_freelist(s, page, object)) {
889 object_err(s, page, object, "Object already free");
893 if (!check_object(s, page, object, 1))
896 if (unlikely(s != page->slab)) {
898 slab_err(s, page, "Attempt to free object(0x%p) "
899 "outside of slab", object);
903 "SLUB <none>: no slab for object 0x%p.\n",
908 object_err(s, page, object,
909 "page slab pointer corrupt.");
913 /* Special debug activities for freeing objects */
914 if (!SlabFrozen(page) && !page->freelist)
915 remove_full(s, page);
916 if (s->flags & SLAB_STORE_USER)
917 set_track(s, object, TRACK_FREE, addr);
918 trace(s, page, object, 0);
919 init_object(s, object, 0);
923 slab_fix(s, "Object at 0x%p not freed", object);
927 static int __init setup_slub_debug(char *str)
929 slub_debug = DEBUG_DEFAULT_FLAGS;
930 if (*str++ != '=' || !*str)
932 * No options specified. Switch on full debugging.
938 * No options but restriction on slabs. This means full
939 * debugging for slabs matching a pattern.
946 * Switch off all debugging measures.
951 * Determine which debug features should be switched on
953 for ( ;*str && *str != ','; str++) {
954 switch (tolower(*str)) {
956 slub_debug |= SLAB_DEBUG_FREE;
959 slub_debug |= SLAB_RED_ZONE;
962 slub_debug |= SLAB_POISON;
965 slub_debug |= SLAB_STORE_USER;
968 slub_debug |= SLAB_TRACE;
971 printk(KERN_ERR "slub_debug option '%c' "
972 "unknown. skipped\n",*str);
978 slub_debug_slabs = str + 1;
983 __setup("slub_debug", setup_slub_debug);
985 static unsigned long kmem_cache_flags(unsigned long objsize,
986 unsigned long flags, const char *name,
987 void (*ctor)(struct kmem_cache *, void *))
990 * The page->offset field is only 16 bit wide. This is an offset
991 * in units of words from the beginning of an object. If the slab
992 * size is bigger then we cannot move the free pointer behind the
995 * On 32 bit platforms the limit is 256k. On 64bit platforms
998 * Debugging or ctor may create a need to move the free
999 * pointer. Fail if this happens.
1001 if (objsize >= 65535 * sizeof(void *)) {
1002 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1003 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1007 * Enable debugging if selected on the kernel commandline.
1009 if (slub_debug && (!slub_debug_slabs ||
1010 strncmp(slub_debug_slabs, name,
1011 strlen(slub_debug_slabs)) == 0))
1012 flags |= slub_debug;
1018 static inline void setup_object_debug(struct kmem_cache *s,
1019 struct page *page, void *object) {}
1021 static inline int alloc_debug_processing(struct kmem_cache *s,
1022 struct page *page, void *object, void *addr) { return 0; }
1024 static inline int free_debug_processing(struct kmem_cache *s,
1025 struct page *page, void *object, void *addr) { return 0; }
1027 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1029 static inline int check_object(struct kmem_cache *s, struct page *page,
1030 void *object, int active) { return 1; }
1031 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1032 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1033 unsigned long flags, const char *name,
1034 void (*ctor)(struct kmem_cache *, void *))
1038 #define slub_debug 0
1041 * Slab allocation and freeing
1043 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1046 int pages = 1 << s->order;
1049 flags |= __GFP_COMP;
1051 if (s->flags & SLAB_CACHE_DMA)
1054 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1055 flags |= __GFP_RECLAIMABLE;
1058 page = alloc_pages(flags, s->order);
1060 page = alloc_pages_node(node, flags, s->order);
1065 mod_zone_page_state(page_zone(page),
1066 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1067 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1073 static void setup_object(struct kmem_cache *s, struct page *page,
1076 setup_object_debug(s, page, object);
1077 if (unlikely(s->ctor))
1081 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1084 struct kmem_cache_node *n;
1089 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1091 page = allocate_slab(s,
1092 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1096 n = get_node(s, page_to_nid(page));
1098 atomic_long_inc(&n->nr_slabs);
1100 page->flags |= 1 << PG_slab;
1101 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1102 SLAB_STORE_USER | SLAB_TRACE))
1105 start = page_address(page);
1107 if (unlikely(s->flags & SLAB_POISON))
1108 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1111 for_each_object(p, s, start) {
1112 setup_object(s, page, last);
1113 set_freepointer(s, last, p);
1116 setup_object(s, page, last);
1117 set_freepointer(s, last, NULL);
1119 page->freelist = start;
1125 static void __free_slab(struct kmem_cache *s, struct page *page)
1127 int pages = 1 << s->order;
1129 if (unlikely(SlabDebug(page))) {
1132 slab_pad_check(s, page);
1133 for_each_object(p, s, page_address(page))
1134 check_object(s, page, p, 0);
1135 ClearSlabDebug(page);
1138 mod_zone_page_state(page_zone(page),
1139 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1140 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1143 __free_pages(page, s->order);
1146 static void rcu_free_slab(struct rcu_head *h)
1150 page = container_of((struct list_head *)h, struct page, lru);
1151 __free_slab(page->slab, page);
1154 static void free_slab(struct kmem_cache *s, struct page *page)
1156 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1158 * RCU free overloads the RCU head over the LRU
1160 struct rcu_head *head = (void *)&page->lru;
1162 call_rcu(head, rcu_free_slab);
1164 __free_slab(s, page);
1167 static void discard_slab(struct kmem_cache *s, struct page *page)
1169 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1171 atomic_long_dec(&n->nr_slabs);
1172 reset_page_mapcount(page);
1173 __ClearPageSlab(page);
1178 * Per slab locking using the pagelock
1180 static __always_inline void slab_lock(struct page *page)
1182 bit_spin_lock(PG_locked, &page->flags);
1185 static __always_inline void slab_unlock(struct page *page)
1187 bit_spin_unlock(PG_locked, &page->flags);
1190 static __always_inline int slab_trylock(struct page *page)
1194 rc = bit_spin_trylock(PG_locked, &page->flags);
1199 * Management of partially allocated slabs
1201 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1203 spin_lock(&n->list_lock);
1205 list_add_tail(&page->lru, &n->partial);
1206 spin_unlock(&n->list_lock);
1209 static void add_partial(struct kmem_cache_node *n, struct page *page)
1211 spin_lock(&n->list_lock);
1213 list_add(&page->lru, &n->partial);
1214 spin_unlock(&n->list_lock);
1217 static void remove_partial(struct kmem_cache *s,
1220 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1222 spin_lock(&n->list_lock);
1223 list_del(&page->lru);
1225 spin_unlock(&n->list_lock);
1229 * Lock slab and remove from the partial list.
1231 * Must hold list_lock.
1233 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1235 if (slab_trylock(page)) {
1236 list_del(&page->lru);
1238 SetSlabFrozen(page);
1245 * Try to allocate a partial slab from a specific node.
1247 static struct page *get_partial_node(struct kmem_cache_node *n)
1252 * Racy check. If we mistakenly see no partial slabs then we
1253 * just allocate an empty slab. If we mistakenly try to get a
1254 * partial slab and there is none available then get_partials()
1257 if (!n || !n->nr_partial)
1260 spin_lock(&n->list_lock);
1261 list_for_each_entry(page, &n->partial, lru)
1262 if (lock_and_freeze_slab(n, page))
1266 spin_unlock(&n->list_lock);
1271 * Get a page from somewhere. Search in increasing NUMA distances.
1273 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1276 struct zonelist *zonelist;
1281 * The defrag ratio allows a configuration of the tradeoffs between
1282 * inter node defragmentation and node local allocations. A lower
1283 * defrag_ratio increases the tendency to do local allocations
1284 * instead of attempting to obtain partial slabs from other nodes.
1286 * If the defrag_ratio is set to 0 then kmalloc() always
1287 * returns node local objects. If the ratio is higher then kmalloc()
1288 * may return off node objects because partial slabs are obtained
1289 * from other nodes and filled up.
1291 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1292 * defrag_ratio = 1000) then every (well almost) allocation will
1293 * first attempt to defrag slab caches on other nodes. This means
1294 * scanning over all nodes to look for partial slabs which may be
1295 * expensive if we do it every time we are trying to find a slab
1296 * with available objects.
1298 if (!s->remote_node_defrag_ratio ||
1299 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1302 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1303 ->node_zonelists[gfp_zone(flags)];
1304 for (z = zonelist->zones; *z; z++) {
1305 struct kmem_cache_node *n;
1307 n = get_node(s, zone_to_nid(*z));
1309 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1310 n->nr_partial > MIN_PARTIAL) {
1311 page = get_partial_node(n);
1321 * Get a partial page, lock it and return it.
1323 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1326 int searchnode = (node == -1) ? numa_node_id() : node;
1328 page = get_partial_node(get_node(s, searchnode));
1329 if (page || (flags & __GFP_THISNODE))
1332 return get_any_partial(s, flags);
1336 * Move a page back to the lists.
1338 * Must be called with the slab lock held.
1340 * On exit the slab lock will have been dropped.
1342 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1344 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1346 ClearSlabFrozen(page);
1350 add_partial(n, page);
1351 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1356 if (n->nr_partial < MIN_PARTIAL) {
1358 * Adding an empty slab to the partial slabs in order
1359 * to avoid page allocator overhead. This slab needs
1360 * to come after the other slabs with objects in
1361 * order to fill them up. That way the size of the
1362 * partial list stays small. kmem_cache_shrink can
1363 * reclaim empty slabs from the partial list.
1365 add_partial_tail(n, page);
1369 discard_slab(s, page);
1375 * Remove the cpu slab
1377 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1379 struct page *page = c->page;
1381 * Merge cpu freelist into freelist. Typically we get here
1382 * because both freelists are empty. So this is unlikely
1385 while (unlikely(c->freelist)) {
1388 /* Retrieve object from cpu_freelist */
1389 object = c->freelist;
1390 c->freelist = c->freelist[c->offset];
1392 /* And put onto the regular freelist */
1393 object[c->offset] = page->freelist;
1394 page->freelist = object;
1398 unfreeze_slab(s, page);
1401 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1404 deactivate_slab(s, c);
1409 * Called from IPI handler with interrupts disabled.
1411 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1413 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1415 if (likely(c && c->page))
1419 static void flush_cpu_slab(void *d)
1421 struct kmem_cache *s = d;
1423 __flush_cpu_slab(s, smp_processor_id());
1426 static void flush_all(struct kmem_cache *s)
1429 on_each_cpu(flush_cpu_slab, s, 1, 1);
1431 unsigned long flags;
1433 local_irq_save(flags);
1435 local_irq_restore(flags);
1440 * Check if the objects in a per cpu structure fit numa
1441 * locality expectations.
1443 static inline int node_match(struct kmem_cache_cpu *c, int node)
1446 if (node != -1 && c->node != node)
1453 * Slow path. The lockless freelist is empty or we need to perform
1456 * Interrupts are disabled.
1458 * Processing is still very fast if new objects have been freed to the
1459 * regular freelist. In that case we simply take over the regular freelist
1460 * as the lockless freelist and zap the regular freelist.
1462 * If that is not working then we fall back to the partial lists. We take the
1463 * first element of the freelist as the object to allocate now and move the
1464 * rest of the freelist to the lockless freelist.
1466 * And if we were unable to get a new slab from the partial slab lists then
1467 * we need to allocate a new slab. This is slowest path since we may sleep.
1469 static void *__slab_alloc(struct kmem_cache *s,
1470 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1479 if (unlikely(!node_match(c, node)))
1482 object = c->page->freelist;
1483 if (unlikely(!object))
1485 if (unlikely(SlabDebug(c->page)))
1488 object = c->page->freelist;
1489 c->freelist = object[c->offset];
1490 c->page->inuse = s->objects;
1491 c->page->freelist = NULL;
1492 c->node = page_to_nid(c->page);
1493 slab_unlock(c->page);
1497 deactivate_slab(s, c);
1500 new = get_partial(s, gfpflags, node);
1506 if (gfpflags & __GFP_WAIT)
1509 new = new_slab(s, gfpflags, node);
1511 if (gfpflags & __GFP_WAIT)
1512 local_irq_disable();
1515 c = get_cpu_slab(s, smp_processor_id());
1525 object = c->page->freelist;
1526 if (!alloc_debug_processing(s, c->page, object, addr))
1530 c->page->freelist = object[c->offset];
1532 slab_unlock(c->page);
1537 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1538 * have the fastpath folded into their functions. So no function call
1539 * overhead for requests that can be satisfied on the fastpath.
1541 * The fastpath works by first checking if the lockless freelist can be used.
1542 * If not then __slab_alloc is called for slow processing.
1544 * Otherwise we can simply pick the next object from the lockless free list.
1546 static void __always_inline *slab_alloc(struct kmem_cache *s,
1547 gfp_t gfpflags, int node, void *addr)
1550 unsigned long flags;
1551 struct kmem_cache_cpu *c;
1553 local_irq_save(flags);
1554 c = get_cpu_slab(s, smp_processor_id());
1555 if (unlikely(!c->freelist || !node_match(c, node)))
1557 object = __slab_alloc(s, gfpflags, node, addr, c);
1560 object = c->freelist;
1561 c->freelist = object[c->offset];
1563 local_irq_restore(flags);
1565 if (unlikely((gfpflags & __GFP_ZERO) && object))
1566 memset(object, 0, c->objsize);
1571 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1573 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1575 EXPORT_SYMBOL(kmem_cache_alloc);
1578 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1580 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1582 EXPORT_SYMBOL(kmem_cache_alloc_node);
1586 * Slow patch handling. This may still be called frequently since objects
1587 * have a longer lifetime than the cpu slabs in most processing loads.
1589 * So we still attempt to reduce cache line usage. Just take the slab
1590 * lock and free the item. If there is no additional partial page
1591 * handling required then we can return immediately.
1593 static void __slab_free(struct kmem_cache *s, struct page *page,
1594 void *x, void *addr, unsigned int offset)
1597 void **object = (void *)x;
1601 if (unlikely(SlabDebug(page)))
1604 prior = object[offset] = page->freelist;
1605 page->freelist = object;
1608 if (unlikely(SlabFrozen(page)))
1611 if (unlikely(!page->inuse))
1615 * Objects left in the slab. If it
1616 * was not on the partial list before
1619 if (unlikely(!prior))
1620 add_partial_tail(get_node(s, page_to_nid(page)), page);
1629 * Slab still on the partial list.
1631 remove_partial(s, page);
1634 discard_slab(s, page);
1638 if (!free_debug_processing(s, page, x, addr))
1644 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1645 * can perform fastpath freeing without additional function calls.
1647 * The fastpath is only possible if we are freeing to the current cpu slab
1648 * of this processor. This typically the case if we have just allocated
1651 * If fastpath is not possible then fall back to __slab_free where we deal
1652 * with all sorts of special processing.
1654 static void __always_inline slab_free(struct kmem_cache *s,
1655 struct page *page, void *x, void *addr)
1657 void **object = (void *)x;
1658 unsigned long flags;
1659 struct kmem_cache_cpu *c;
1661 local_irq_save(flags);
1662 debug_check_no_locks_freed(object, s->objsize);
1663 c = get_cpu_slab(s, smp_processor_id());
1664 if (likely(page == c->page && c->node >= 0)) {
1665 object[c->offset] = c->freelist;
1666 c->freelist = object;
1668 __slab_free(s, page, x, addr, c->offset);
1670 local_irq_restore(flags);
1673 void kmem_cache_free(struct kmem_cache *s, void *x)
1677 page = virt_to_head_page(x);
1679 slab_free(s, page, x, __builtin_return_address(0));
1681 EXPORT_SYMBOL(kmem_cache_free);
1683 /* Figure out on which slab object the object resides */
1684 static struct page *get_object_page(const void *x)
1686 struct page *page = virt_to_head_page(x);
1688 if (!PageSlab(page))
1695 * Object placement in a slab is made very easy because we always start at
1696 * offset 0. If we tune the size of the object to the alignment then we can
1697 * get the required alignment by putting one properly sized object after
1700 * Notice that the allocation order determines the sizes of the per cpu
1701 * caches. Each processor has always one slab available for allocations.
1702 * Increasing the allocation order reduces the number of times that slabs
1703 * must be moved on and off the partial lists and is therefore a factor in
1708 * Mininum / Maximum order of slab pages. This influences locking overhead
1709 * and slab fragmentation. A higher order reduces the number of partial slabs
1710 * and increases the number of allocations possible without having to
1711 * take the list_lock.
1713 static int slub_min_order;
1714 static int slub_max_order = DEFAULT_MAX_ORDER;
1715 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1718 * Merge control. If this is set then no merging of slab caches will occur.
1719 * (Could be removed. This was introduced to pacify the merge skeptics.)
1721 static int slub_nomerge;
1724 * Calculate the order of allocation given an slab object size.
1726 * The order of allocation has significant impact on performance and other
1727 * system components. Generally order 0 allocations should be preferred since
1728 * order 0 does not cause fragmentation in the page allocator. Larger objects
1729 * be problematic to put into order 0 slabs because there may be too much
1730 * unused space left. We go to a higher order if more than 1/8th of the slab
1733 * In order to reach satisfactory performance we must ensure that a minimum
1734 * number of objects is in one slab. Otherwise we may generate too much
1735 * activity on the partial lists which requires taking the list_lock. This is
1736 * less a concern for large slabs though which are rarely used.
1738 * slub_max_order specifies the order where we begin to stop considering the
1739 * number of objects in a slab as critical. If we reach slub_max_order then
1740 * we try to keep the page order as low as possible. So we accept more waste
1741 * of space in favor of a small page order.
1743 * Higher order allocations also allow the placement of more objects in a
1744 * slab and thereby reduce object handling overhead. If the user has
1745 * requested a higher mininum order then we start with that one instead of
1746 * the smallest order which will fit the object.
1748 static inline int slab_order(int size, int min_objects,
1749 int max_order, int fract_leftover)
1753 int min_order = slub_min_order;
1755 for (order = max(min_order,
1756 fls(min_objects * size - 1) - PAGE_SHIFT);
1757 order <= max_order; order++) {
1759 unsigned long slab_size = PAGE_SIZE << order;
1761 if (slab_size < min_objects * size)
1764 rem = slab_size % size;
1766 if (rem <= slab_size / fract_leftover)
1774 static inline int calculate_order(int size)
1781 * Attempt to find best configuration for a slab. This
1782 * works by first attempting to generate a layout with
1783 * the best configuration and backing off gradually.
1785 * First we reduce the acceptable waste in a slab. Then
1786 * we reduce the minimum objects required in a slab.
1788 min_objects = slub_min_objects;
1789 while (min_objects > 1) {
1791 while (fraction >= 4) {
1792 order = slab_order(size, min_objects,
1793 slub_max_order, fraction);
1794 if (order <= slub_max_order)
1802 * We were unable to place multiple objects in a slab. Now
1803 * lets see if we can place a single object there.
1805 order = slab_order(size, 1, slub_max_order, 1);
1806 if (order <= slub_max_order)
1810 * Doh this slab cannot be placed using slub_max_order.
1812 order = slab_order(size, 1, MAX_ORDER, 1);
1813 if (order <= MAX_ORDER)
1819 * Figure out what the alignment of the objects will be.
1821 static unsigned long calculate_alignment(unsigned long flags,
1822 unsigned long align, unsigned long size)
1825 * If the user wants hardware cache aligned objects then
1826 * follow that suggestion if the object is sufficiently
1829 * The hardware cache alignment cannot override the
1830 * specified alignment though. If that is greater
1833 if ((flags & SLAB_HWCACHE_ALIGN) &&
1834 size > cache_line_size() / 2)
1835 return max_t(unsigned long, align, cache_line_size());
1837 if (align < ARCH_SLAB_MINALIGN)
1838 return ARCH_SLAB_MINALIGN;
1840 return ALIGN(align, sizeof(void *));
1843 static void init_kmem_cache_cpu(struct kmem_cache *s,
1844 struct kmem_cache_cpu *c)
1849 c->offset = s->offset / sizeof(void *);
1850 c->objsize = s->objsize;
1853 static void init_kmem_cache_node(struct kmem_cache_node *n)
1856 atomic_long_set(&n->nr_slabs, 0);
1857 spin_lock_init(&n->list_lock);
1858 INIT_LIST_HEAD(&n->partial);
1859 #ifdef CONFIG_SLUB_DEBUG
1860 INIT_LIST_HEAD(&n->full);
1866 * Per cpu array for per cpu structures.
1868 * The per cpu array places all kmem_cache_cpu structures from one processor
1869 * close together meaning that it becomes possible that multiple per cpu
1870 * structures are contained in one cacheline. This may be particularly
1871 * beneficial for the kmalloc caches.
1873 * A desktop system typically has around 60-80 slabs. With 100 here we are
1874 * likely able to get per cpu structures for all caches from the array defined
1875 * here. We must be able to cover all kmalloc caches during bootstrap.
1877 * If the per cpu array is exhausted then fall back to kmalloc
1878 * of individual cachelines. No sharing is possible then.
1880 #define NR_KMEM_CACHE_CPU 100
1882 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1883 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1885 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1886 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1888 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1889 int cpu, gfp_t flags)
1891 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1894 per_cpu(kmem_cache_cpu_free, cpu) =
1895 (void *)c->freelist;
1897 /* Table overflow: So allocate ourselves */
1899 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1900 flags, cpu_to_node(cpu));
1905 init_kmem_cache_cpu(s, c);
1909 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1911 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1912 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1916 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1917 per_cpu(kmem_cache_cpu_free, cpu) = c;
1920 static void free_kmem_cache_cpus(struct kmem_cache *s)
1924 for_each_online_cpu(cpu) {
1925 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1928 s->cpu_slab[cpu] = NULL;
1929 free_kmem_cache_cpu(c, cpu);
1934 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1938 for_each_online_cpu(cpu) {
1939 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1944 c = alloc_kmem_cache_cpu(s, cpu, flags);
1946 free_kmem_cache_cpus(s);
1949 s->cpu_slab[cpu] = c;
1955 * Initialize the per cpu array.
1957 static void init_alloc_cpu_cpu(int cpu)
1961 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1964 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1965 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1967 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1970 static void __init init_alloc_cpu(void)
1974 for_each_online_cpu(cpu)
1975 init_alloc_cpu_cpu(cpu);
1979 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
1980 static inline void init_alloc_cpu(void) {}
1982 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1984 init_kmem_cache_cpu(s, &s->cpu_slab);
1991 * No kmalloc_node yet so do it by hand. We know that this is the first
1992 * slab on the node for this slabcache. There are no concurrent accesses
1995 * Note that this function only works on the kmalloc_node_cache
1996 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
1997 * memory on a fresh node that has no slab structures yet.
1999 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2003 struct kmem_cache_node *n;
2005 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2007 page = new_slab(kmalloc_caches, gfpflags, node);
2010 if (page_to_nid(page) != node) {
2011 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2013 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2014 "in order to be able to continue\n");
2019 page->freelist = get_freepointer(kmalloc_caches, n);
2021 kmalloc_caches->node[node] = n;
2022 #ifdef CONFIG_SLUB_DEBUG
2023 init_object(kmalloc_caches, n, 1);
2024 init_tracking(kmalloc_caches, n);
2026 init_kmem_cache_node(n);
2027 atomic_long_inc(&n->nr_slabs);
2028 add_partial(n, page);
2032 static void free_kmem_cache_nodes(struct kmem_cache *s)
2036 for_each_node_state(node, N_NORMAL_MEMORY) {
2037 struct kmem_cache_node *n = s->node[node];
2038 if (n && n != &s->local_node)
2039 kmem_cache_free(kmalloc_caches, n);
2040 s->node[node] = NULL;
2044 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2049 if (slab_state >= UP)
2050 local_node = page_to_nid(virt_to_page(s));
2054 for_each_node_state(node, N_NORMAL_MEMORY) {
2055 struct kmem_cache_node *n;
2057 if (local_node == node)
2060 if (slab_state == DOWN) {
2061 n = early_kmem_cache_node_alloc(gfpflags,
2065 n = kmem_cache_alloc_node(kmalloc_caches,
2069 free_kmem_cache_nodes(s);
2075 init_kmem_cache_node(n);
2080 static void free_kmem_cache_nodes(struct kmem_cache *s)
2084 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2086 init_kmem_cache_node(&s->local_node);
2092 * calculate_sizes() determines the order and the distribution of data within
2095 static int calculate_sizes(struct kmem_cache *s)
2097 unsigned long flags = s->flags;
2098 unsigned long size = s->objsize;
2099 unsigned long align = s->align;
2102 * Determine if we can poison the object itself. If the user of
2103 * the slab may touch the object after free or before allocation
2104 * then we should never poison the object itself.
2106 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2108 s->flags |= __OBJECT_POISON;
2110 s->flags &= ~__OBJECT_POISON;
2113 * Round up object size to the next word boundary. We can only
2114 * place the free pointer at word boundaries and this determines
2115 * the possible location of the free pointer.
2117 size = ALIGN(size, sizeof(void *));
2119 #ifdef CONFIG_SLUB_DEBUG
2121 * If we are Redzoning then check if there is some space between the
2122 * end of the object and the free pointer. If not then add an
2123 * additional word to have some bytes to store Redzone information.
2125 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2126 size += sizeof(void *);
2130 * With that we have determined the number of bytes in actual use
2131 * by the object. This is the potential offset to the free pointer.
2135 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2138 * Relocate free pointer after the object if it is not
2139 * permitted to overwrite the first word of the object on
2142 * This is the case if we do RCU, have a constructor or
2143 * destructor or are poisoning the objects.
2146 size += sizeof(void *);
2149 #ifdef CONFIG_SLUB_DEBUG
2150 if (flags & SLAB_STORE_USER)
2152 * Need to store information about allocs and frees after
2155 size += 2 * sizeof(struct track);
2157 if (flags & SLAB_RED_ZONE)
2159 * Add some empty padding so that we can catch
2160 * overwrites from earlier objects rather than let
2161 * tracking information or the free pointer be
2162 * corrupted if an user writes before the start
2165 size += sizeof(void *);
2169 * Determine the alignment based on various parameters that the
2170 * user specified and the dynamic determination of cache line size
2173 align = calculate_alignment(flags, align, s->objsize);
2176 * SLUB stores one object immediately after another beginning from
2177 * offset 0. In order to align the objects we have to simply size
2178 * each object to conform to the alignment.
2180 size = ALIGN(size, align);
2183 s->order = calculate_order(size);
2188 * Determine the number of objects per slab
2190 s->objects = (PAGE_SIZE << s->order) / size;
2192 return !!s->objects;
2196 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2197 const char *name, size_t size,
2198 size_t align, unsigned long flags,
2199 void (*ctor)(struct kmem_cache *, void *))
2201 memset(s, 0, kmem_size);
2206 s->flags = kmem_cache_flags(size, flags, name, ctor);
2208 if (!calculate_sizes(s))
2213 s->remote_node_defrag_ratio = 100;
2215 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2218 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2220 free_kmem_cache_nodes(s);
2222 if (flags & SLAB_PANIC)
2223 panic("Cannot create slab %s size=%lu realsize=%u "
2224 "order=%u offset=%u flags=%lx\n",
2225 s->name, (unsigned long)size, s->size, s->order,
2231 * Check if a given pointer is valid
2233 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2237 page = get_object_page(object);
2239 if (!page || s != page->slab)
2240 /* No slab or wrong slab */
2243 if (!check_valid_pointer(s, page, object))
2247 * We could also check if the object is on the slabs freelist.
2248 * But this would be too expensive and it seems that the main
2249 * purpose of kmem_ptr_valid is to check if the object belongs
2250 * to a certain slab.
2254 EXPORT_SYMBOL(kmem_ptr_validate);
2257 * Determine the size of a slab object
2259 unsigned int kmem_cache_size(struct kmem_cache *s)
2263 EXPORT_SYMBOL(kmem_cache_size);
2265 const char *kmem_cache_name(struct kmem_cache *s)
2269 EXPORT_SYMBOL(kmem_cache_name);
2272 * Attempt to free all slabs on a node. Return the number of slabs we
2273 * were unable to free.
2275 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2276 struct list_head *list)
2278 int slabs_inuse = 0;
2279 unsigned long flags;
2280 struct page *page, *h;
2282 spin_lock_irqsave(&n->list_lock, flags);
2283 list_for_each_entry_safe(page, h, list, lru)
2285 list_del(&page->lru);
2286 discard_slab(s, page);
2289 spin_unlock_irqrestore(&n->list_lock, flags);
2294 * Release all resources used by a slab cache.
2296 static inline int kmem_cache_close(struct kmem_cache *s)
2302 /* Attempt to free all objects */
2303 free_kmem_cache_cpus(s);
2304 for_each_node_state(node, N_NORMAL_MEMORY) {
2305 struct kmem_cache_node *n = get_node(s, node);
2307 n->nr_partial -= free_list(s, n, &n->partial);
2308 if (atomic_long_read(&n->nr_slabs))
2311 free_kmem_cache_nodes(s);
2316 * Close a cache and release the kmem_cache structure
2317 * (must be used for caches created using kmem_cache_create)
2319 void kmem_cache_destroy(struct kmem_cache *s)
2321 down_write(&slub_lock);
2325 up_write(&slub_lock);
2326 if (kmem_cache_close(s))
2328 sysfs_slab_remove(s);
2330 up_write(&slub_lock);
2332 EXPORT_SYMBOL(kmem_cache_destroy);
2334 /********************************************************************
2336 *******************************************************************/
2338 struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned;
2339 EXPORT_SYMBOL(kmalloc_caches);
2341 #ifdef CONFIG_ZONE_DMA
2342 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT];
2345 static int __init setup_slub_min_order(char *str)
2347 get_option (&str, &slub_min_order);
2352 __setup("slub_min_order=", setup_slub_min_order);
2354 static int __init setup_slub_max_order(char *str)
2356 get_option (&str, &slub_max_order);
2361 __setup("slub_max_order=", setup_slub_max_order);
2363 static int __init setup_slub_min_objects(char *str)
2365 get_option (&str, &slub_min_objects);
2370 __setup("slub_min_objects=", setup_slub_min_objects);
2372 static int __init setup_slub_nomerge(char *str)
2378 __setup("slub_nomerge", setup_slub_nomerge);
2380 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2381 const char *name, int size, gfp_t gfp_flags)
2383 unsigned int flags = 0;
2385 if (gfp_flags & SLUB_DMA)
2386 flags = SLAB_CACHE_DMA;
2388 down_write(&slub_lock);
2389 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2393 list_add(&s->list, &slab_caches);
2394 up_write(&slub_lock);
2395 if (sysfs_slab_add(s))
2400 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2403 #ifdef CONFIG_ZONE_DMA
2405 static void sysfs_add_func(struct work_struct *w)
2407 struct kmem_cache *s;
2409 down_write(&slub_lock);
2410 list_for_each_entry(s, &slab_caches, list) {
2411 if (s->flags & __SYSFS_ADD_DEFERRED) {
2412 s->flags &= ~__SYSFS_ADD_DEFERRED;
2416 up_write(&slub_lock);
2419 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2421 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2423 struct kmem_cache *s;
2427 s = kmalloc_caches_dma[index];
2431 /* Dynamically create dma cache */
2432 if (flags & __GFP_WAIT)
2433 down_write(&slub_lock);
2435 if (!down_write_trylock(&slub_lock))
2439 if (kmalloc_caches_dma[index])
2442 realsize = kmalloc_caches[index].objsize;
2443 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize),
2444 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2446 if (!s || !text || !kmem_cache_open(s, flags, text,
2447 realsize, ARCH_KMALLOC_MINALIGN,
2448 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2454 list_add(&s->list, &slab_caches);
2455 kmalloc_caches_dma[index] = s;
2457 schedule_work(&sysfs_add_work);
2460 up_write(&slub_lock);
2462 return kmalloc_caches_dma[index];
2467 * Conversion table for small slabs sizes / 8 to the index in the
2468 * kmalloc array. This is necessary for slabs < 192 since we have non power
2469 * of two cache sizes there. The size of larger slabs can be determined using
2472 static s8 size_index[24] = {
2499 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2505 return ZERO_SIZE_PTR;
2507 index = size_index[(size - 1) / 8];
2509 index = fls(size - 1);
2511 #ifdef CONFIG_ZONE_DMA
2512 if (unlikely((flags & SLUB_DMA)))
2513 return dma_kmalloc_cache(index, flags);
2516 return &kmalloc_caches[index];
2519 void *__kmalloc(size_t size, gfp_t flags)
2521 struct kmem_cache *s;
2523 if (unlikely(size > PAGE_SIZE / 2))
2524 return (void *)__get_free_pages(flags | __GFP_COMP,
2527 s = get_slab(size, flags);
2529 if (unlikely(ZERO_OR_NULL_PTR(s)))
2532 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2534 EXPORT_SYMBOL(__kmalloc);
2537 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2539 struct kmem_cache *s;
2541 if (unlikely(size > PAGE_SIZE / 2))
2542 return (void *)__get_free_pages(flags | __GFP_COMP,
2545 s = get_slab(size, flags);
2547 if (unlikely(ZERO_OR_NULL_PTR(s)))
2550 return slab_alloc(s, flags, node, __builtin_return_address(0));
2552 EXPORT_SYMBOL(__kmalloc_node);
2555 size_t ksize(const void *object)
2558 struct kmem_cache *s;
2561 if (unlikely(object == ZERO_SIZE_PTR))
2564 page = virt_to_head_page(object);
2567 if (unlikely(!PageSlab(page)))
2568 return PAGE_SIZE << compound_order(page);
2574 * Debugging requires use of the padding between object
2575 * and whatever may come after it.
2577 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2581 * If we have the need to store the freelist pointer
2582 * back there or track user information then we can
2583 * only use the space before that information.
2585 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2589 * Else we can use all the padding etc for the allocation
2593 EXPORT_SYMBOL(ksize);
2595 void kfree(const void *x)
2599 if (unlikely(ZERO_OR_NULL_PTR(x)))
2602 page = virt_to_head_page(x);
2603 if (unlikely(!PageSlab(page))) {
2607 slab_free(page->slab, page, (void *)x, __builtin_return_address(0));
2609 EXPORT_SYMBOL(kfree);
2611 static unsigned long count_partial(struct kmem_cache_node *n)
2613 unsigned long flags;
2614 unsigned long x = 0;
2617 spin_lock_irqsave(&n->list_lock, flags);
2618 list_for_each_entry(page, &n->partial, lru)
2620 spin_unlock_irqrestore(&n->list_lock, flags);
2625 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2626 * the remaining slabs by the number of items in use. The slabs with the
2627 * most items in use come first. New allocations will then fill those up
2628 * and thus they can be removed from the partial lists.
2630 * The slabs with the least items are placed last. This results in them
2631 * being allocated from last increasing the chance that the last objects
2632 * are freed in them.
2634 int kmem_cache_shrink(struct kmem_cache *s)
2638 struct kmem_cache_node *n;
2641 struct list_head *slabs_by_inuse =
2642 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2643 unsigned long flags;
2645 if (!slabs_by_inuse)
2649 for_each_node_state(node, N_NORMAL_MEMORY) {
2650 n = get_node(s, node);
2655 for (i = 0; i < s->objects; i++)
2656 INIT_LIST_HEAD(slabs_by_inuse + i);
2658 spin_lock_irqsave(&n->list_lock, flags);
2661 * Build lists indexed by the items in use in each slab.
2663 * Note that concurrent frees may occur while we hold the
2664 * list_lock. page->inuse here is the upper limit.
2666 list_for_each_entry_safe(page, t, &n->partial, lru) {
2667 if (!page->inuse && slab_trylock(page)) {
2669 * Must hold slab lock here because slab_free
2670 * may have freed the last object and be
2671 * waiting to release the slab.
2673 list_del(&page->lru);
2676 discard_slab(s, page);
2678 list_move(&page->lru,
2679 slabs_by_inuse + page->inuse);
2684 * Rebuild the partial list with the slabs filled up most
2685 * first and the least used slabs at the end.
2687 for (i = s->objects - 1; i >= 0; i--)
2688 list_splice(slabs_by_inuse + i, n->partial.prev);
2690 spin_unlock_irqrestore(&n->list_lock, flags);
2693 kfree(slabs_by_inuse);
2696 EXPORT_SYMBOL(kmem_cache_shrink);
2698 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2699 static int slab_mem_going_offline_callback(void *arg)
2701 struct kmem_cache *s;
2703 down_read(&slub_lock);
2704 list_for_each_entry(s, &slab_caches, list)
2705 kmem_cache_shrink(s);
2706 up_read(&slub_lock);
2711 static void slab_mem_offline_callback(void *arg)
2713 struct kmem_cache_node *n;
2714 struct kmem_cache *s;
2715 struct memory_notify *marg = arg;
2718 offline_node = marg->status_change_nid;
2721 * If the node still has available memory. we need kmem_cache_node
2724 if (offline_node < 0)
2727 down_read(&slub_lock);
2728 list_for_each_entry(s, &slab_caches, list) {
2729 n = get_node(s, offline_node);
2732 * if n->nr_slabs > 0, slabs still exist on the node
2733 * that is going down. We were unable to free them,
2734 * and offline_pages() function shoudn't call this
2735 * callback. So, we must fail.
2737 BUG_ON(atomic_long_read(&n->nr_slabs));
2739 s->node[offline_node] = NULL;
2740 kmem_cache_free(kmalloc_caches, n);
2743 up_read(&slub_lock);
2746 static int slab_mem_going_online_callback(void *arg)
2748 struct kmem_cache_node *n;
2749 struct kmem_cache *s;
2750 struct memory_notify *marg = arg;
2751 int nid = marg->status_change_nid;
2755 * If the node's memory is already available, then kmem_cache_node is
2756 * already created. Nothing to do.
2762 * We are bringing a node online. No memory is availabe yet. We must
2763 * allocate a kmem_cache_node structure in order to bring the node
2766 down_read(&slub_lock);
2767 list_for_each_entry(s, &slab_caches, list) {
2769 * XXX: kmem_cache_alloc_node will fallback to other nodes
2770 * since memory is not yet available from the node that
2773 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2778 init_kmem_cache_node(n);
2782 up_read(&slub_lock);
2786 static int slab_memory_callback(struct notifier_block *self,
2787 unsigned long action, void *arg)
2792 case MEM_GOING_ONLINE:
2793 ret = slab_mem_going_online_callback(arg);
2795 case MEM_GOING_OFFLINE:
2796 ret = slab_mem_going_offline_callback(arg);
2799 case MEM_CANCEL_ONLINE:
2800 slab_mem_offline_callback(arg);
2803 case MEM_CANCEL_OFFLINE:
2807 ret = notifier_from_errno(ret);
2811 #endif /* CONFIG_MEMORY_HOTPLUG */
2813 /********************************************************************
2814 * Basic setup of slabs
2815 *******************************************************************/
2817 void __init kmem_cache_init(void)
2826 * Must first have the slab cache available for the allocations of the
2827 * struct kmem_cache_node's. There is special bootstrap code in
2828 * kmem_cache_open for slab_state == DOWN.
2830 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2831 sizeof(struct kmem_cache_node), GFP_KERNEL);
2832 kmalloc_caches[0].refcount = -1;
2835 hotplug_memory_notifier(slab_memory_callback, 1);
2838 /* Able to allocate the per node structures */
2839 slab_state = PARTIAL;
2841 /* Caches that are not of the two-to-the-power-of size */
2842 if (KMALLOC_MIN_SIZE <= 64) {
2843 create_kmalloc_cache(&kmalloc_caches[1],
2844 "kmalloc-96", 96, GFP_KERNEL);
2847 if (KMALLOC_MIN_SIZE <= 128) {
2848 create_kmalloc_cache(&kmalloc_caches[2],
2849 "kmalloc-192", 192, GFP_KERNEL);
2853 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) {
2854 create_kmalloc_cache(&kmalloc_caches[i],
2855 "kmalloc", 1 << i, GFP_KERNEL);
2861 * Patch up the size_index table if we have strange large alignment
2862 * requirements for the kmalloc array. This is only the case for
2863 * mips it seems. The standard arches will not generate any code here.
2865 * Largest permitted alignment is 256 bytes due to the way we
2866 * handle the index determination for the smaller caches.
2868 * Make sure that nothing crazy happens if someone starts tinkering
2869 * around with ARCH_KMALLOC_MINALIGN
2871 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2872 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2874 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2875 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2879 /* Provide the correct kmalloc names now that the caches are up */
2880 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++)
2881 kmalloc_caches[i]. name =
2882 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2885 register_cpu_notifier(&slab_notifier);
2886 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2887 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2889 kmem_size = sizeof(struct kmem_cache);
2893 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2894 " CPUs=%d, Nodes=%d\n",
2895 caches, cache_line_size(),
2896 slub_min_order, slub_max_order, slub_min_objects,
2897 nr_cpu_ids, nr_node_ids);
2901 * Find a mergeable slab cache
2903 static int slab_unmergeable(struct kmem_cache *s)
2905 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2912 * We may have set a slab to be unmergeable during bootstrap.
2914 if (s->refcount < 0)
2920 static struct kmem_cache *find_mergeable(size_t size,
2921 size_t align, unsigned long flags, const char *name,
2922 void (*ctor)(struct kmem_cache *, void *))
2924 struct kmem_cache *s;
2926 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2932 size = ALIGN(size, sizeof(void *));
2933 align = calculate_alignment(flags, align, size);
2934 size = ALIGN(size, align);
2935 flags = kmem_cache_flags(size, flags, name, NULL);
2937 list_for_each_entry(s, &slab_caches, list) {
2938 if (slab_unmergeable(s))
2944 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
2947 * Check if alignment is compatible.
2948 * Courtesy of Adrian Drzewiecki
2950 if ((s->size & ~(align -1)) != s->size)
2953 if (s->size - size >= sizeof(void *))
2961 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2962 size_t align, unsigned long flags,
2963 void (*ctor)(struct kmem_cache *, void *))
2965 struct kmem_cache *s;
2967 down_write(&slub_lock);
2968 s = find_mergeable(size, align, flags, name, ctor);
2974 * Adjust the object sizes so that we clear
2975 * the complete object on kzalloc.
2977 s->objsize = max(s->objsize, (int)size);
2980 * And then we need to update the object size in the
2981 * per cpu structures
2983 for_each_online_cpu(cpu)
2984 get_cpu_slab(s, cpu)->objsize = s->objsize;
2985 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2986 up_write(&slub_lock);
2987 if (sysfs_slab_alias(s, name))
2991 s = kmalloc(kmem_size, GFP_KERNEL);
2993 if (kmem_cache_open(s, GFP_KERNEL, name,
2994 size, align, flags, ctor)) {
2995 list_add(&s->list, &slab_caches);
2996 up_write(&slub_lock);
2997 if (sysfs_slab_add(s))
3003 up_write(&slub_lock);
3006 if (flags & SLAB_PANIC)
3007 panic("Cannot create slabcache %s\n", name);
3012 EXPORT_SYMBOL(kmem_cache_create);
3016 * Use the cpu notifier to insure that the cpu slabs are flushed when
3019 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3020 unsigned long action, void *hcpu)
3022 long cpu = (long)hcpu;
3023 struct kmem_cache *s;
3024 unsigned long flags;
3027 case CPU_UP_PREPARE:
3028 case CPU_UP_PREPARE_FROZEN:
3029 init_alloc_cpu_cpu(cpu);
3030 down_read(&slub_lock);
3031 list_for_each_entry(s, &slab_caches, list)
3032 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3034 up_read(&slub_lock);
3037 case CPU_UP_CANCELED:
3038 case CPU_UP_CANCELED_FROZEN:
3040 case CPU_DEAD_FROZEN:
3041 down_read(&slub_lock);
3042 list_for_each_entry(s, &slab_caches, list) {
3043 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3045 local_irq_save(flags);
3046 __flush_cpu_slab(s, cpu);
3047 local_irq_restore(flags);
3048 free_kmem_cache_cpu(c, cpu);
3049 s->cpu_slab[cpu] = NULL;
3051 up_read(&slub_lock);
3059 static struct notifier_block __cpuinitdata slab_notifier =
3060 { &slab_cpuup_callback, NULL, 0 };
3064 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3066 struct kmem_cache *s;
3068 if (unlikely(size > PAGE_SIZE / 2))
3069 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3071 s = get_slab(size, gfpflags);
3073 if (unlikely(ZERO_OR_NULL_PTR(s)))
3076 return slab_alloc(s, gfpflags, -1, caller);
3079 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3080 int node, void *caller)
3082 struct kmem_cache *s;
3084 if (unlikely(size > PAGE_SIZE / 2))
3085 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3087 s = get_slab(size, gfpflags);
3089 if (unlikely(ZERO_OR_NULL_PTR(s)))
3092 return slab_alloc(s, gfpflags, node, caller);
3095 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3096 static int validate_slab(struct kmem_cache *s, struct page *page,
3100 void *addr = page_address(page);
3102 if (!check_slab(s, page) ||
3103 !on_freelist(s, page, NULL))
3106 /* Now we know that a valid freelist exists */
3107 bitmap_zero(map, s->objects);
3109 for_each_free_object(p, s, page->freelist) {
3110 set_bit(slab_index(p, s, addr), map);
3111 if (!check_object(s, page, p, 0))
3115 for_each_object(p, s, addr)
3116 if (!test_bit(slab_index(p, s, addr), map))
3117 if (!check_object(s, page, p, 1))
3122 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3125 if (slab_trylock(page)) {
3126 validate_slab(s, page, map);
3129 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3132 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3133 if (!SlabDebug(page))
3134 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3135 "on slab 0x%p\n", s->name, page);
3137 if (SlabDebug(page))
3138 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3139 "slab 0x%p\n", s->name, page);
3143 static int validate_slab_node(struct kmem_cache *s,
3144 struct kmem_cache_node *n, unsigned long *map)
3146 unsigned long count = 0;
3148 unsigned long flags;
3150 spin_lock_irqsave(&n->list_lock, flags);
3152 list_for_each_entry(page, &n->partial, lru) {
3153 validate_slab_slab(s, page, map);
3156 if (count != n->nr_partial)
3157 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3158 "counter=%ld\n", s->name, count, n->nr_partial);
3160 if (!(s->flags & SLAB_STORE_USER))
3163 list_for_each_entry(page, &n->full, lru) {
3164 validate_slab_slab(s, page, map);
3167 if (count != atomic_long_read(&n->nr_slabs))
3168 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3169 "counter=%ld\n", s->name, count,
3170 atomic_long_read(&n->nr_slabs));
3173 spin_unlock_irqrestore(&n->list_lock, flags);
3177 static long validate_slab_cache(struct kmem_cache *s)
3180 unsigned long count = 0;
3181 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3182 sizeof(unsigned long), GFP_KERNEL);
3188 for_each_node_state(node, N_NORMAL_MEMORY) {
3189 struct kmem_cache_node *n = get_node(s, node);
3191 count += validate_slab_node(s, n, map);
3197 #ifdef SLUB_RESILIENCY_TEST
3198 static void resiliency_test(void)
3202 printk(KERN_ERR "SLUB resiliency testing\n");
3203 printk(KERN_ERR "-----------------------\n");
3204 printk(KERN_ERR "A. Corruption after allocation\n");
3206 p = kzalloc(16, GFP_KERNEL);
3208 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3209 " 0x12->0x%p\n\n", p + 16);
3211 validate_slab_cache(kmalloc_caches + 4);
3213 /* Hmmm... The next two are dangerous */
3214 p = kzalloc(32, GFP_KERNEL);
3215 p[32 + sizeof(void *)] = 0x34;
3216 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3217 " 0x34 -> -0x%p\n", p);
3218 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3220 validate_slab_cache(kmalloc_caches + 5);
3221 p = kzalloc(64, GFP_KERNEL);
3222 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3224 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3226 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3227 validate_slab_cache(kmalloc_caches + 6);
3229 printk(KERN_ERR "\nB. Corruption after free\n");
3230 p = kzalloc(128, GFP_KERNEL);
3233 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3234 validate_slab_cache(kmalloc_caches + 7);
3236 p = kzalloc(256, GFP_KERNEL);
3239 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
3240 validate_slab_cache(kmalloc_caches + 8);
3242 p = kzalloc(512, GFP_KERNEL);
3245 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3246 validate_slab_cache(kmalloc_caches + 9);
3249 static void resiliency_test(void) {};
3253 * Generate lists of code addresses where slabcache objects are allocated
3258 unsigned long count;
3271 unsigned long count;
3272 struct location *loc;
3275 static void free_loc_track(struct loc_track *t)
3278 free_pages((unsigned long)t->loc,
3279 get_order(sizeof(struct location) * t->max));
3282 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3287 order = get_order(sizeof(struct location) * max);
3289 l = (void *)__get_free_pages(flags, order);
3294 memcpy(l, t->loc, sizeof(struct location) * t->count);
3302 static int add_location(struct loc_track *t, struct kmem_cache *s,
3303 const struct track *track)
3305 long start, end, pos;
3308 unsigned long age = jiffies - track->when;
3314 pos = start + (end - start + 1) / 2;
3317 * There is nothing at "end". If we end up there
3318 * we need to add something to before end.
3323 caddr = t->loc[pos].addr;
3324 if (track->addr == caddr) {
3330 if (age < l->min_time)
3332 if (age > l->max_time)
3335 if (track->pid < l->min_pid)
3336 l->min_pid = track->pid;
3337 if (track->pid > l->max_pid)
3338 l->max_pid = track->pid;
3340 cpu_set(track->cpu, l->cpus);
3342 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3346 if (track->addr < caddr)
3353 * Not found. Insert new tracking element.
3355 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3361 (t->count - pos) * sizeof(struct location));
3364 l->addr = track->addr;
3368 l->min_pid = track->pid;
3369 l->max_pid = track->pid;
3370 cpus_clear(l->cpus);
3371 cpu_set(track->cpu, l->cpus);
3372 nodes_clear(l->nodes);
3373 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3377 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3378 struct page *page, enum track_item alloc)
3380 void *addr = page_address(page);
3381 DECLARE_BITMAP(map, s->objects);
3384 bitmap_zero(map, s->objects);
3385 for_each_free_object(p, s, page->freelist)
3386 set_bit(slab_index(p, s, addr), map);
3388 for_each_object(p, s, addr)
3389 if (!test_bit(slab_index(p, s, addr), map))
3390 add_location(t, s, get_track(s, p, alloc));
3393 static int list_locations(struct kmem_cache *s, char *buf,
3394 enum track_item alloc)
3398 struct loc_track t = { 0, 0, NULL };
3401 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3403 return sprintf(buf, "Out of memory\n");
3405 /* Push back cpu slabs */
3408 for_each_node_state(node, N_NORMAL_MEMORY) {
3409 struct kmem_cache_node *n = get_node(s, node);
3410 unsigned long flags;
3413 if (!atomic_long_read(&n->nr_slabs))
3416 spin_lock_irqsave(&n->list_lock, flags);
3417 list_for_each_entry(page, &n->partial, lru)
3418 process_slab(&t, s, page, alloc);
3419 list_for_each_entry(page, &n->full, lru)
3420 process_slab(&t, s, page, alloc);
3421 spin_unlock_irqrestore(&n->list_lock, flags);
3424 for (i = 0; i < t.count; i++) {
3425 struct location *l = &t.loc[i];
3427 if (len > PAGE_SIZE - 100)
3429 len += sprintf(buf + len, "%7ld ", l->count);
3432 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3434 len += sprintf(buf + len, "<not-available>");
3436 if (l->sum_time != l->min_time) {
3437 unsigned long remainder;
3439 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3441 div_long_long_rem(l->sum_time, l->count, &remainder),
3444 len += sprintf(buf + len, " age=%ld",
3447 if (l->min_pid != l->max_pid)
3448 len += sprintf(buf + len, " pid=%ld-%ld",
3449 l->min_pid, l->max_pid);
3451 len += sprintf(buf + len, " pid=%ld",
3454 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3455 len < PAGE_SIZE - 60) {
3456 len += sprintf(buf + len, " cpus=");
3457 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3461 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3462 len < PAGE_SIZE - 60) {
3463 len += sprintf(buf + len, " nodes=");
3464 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3468 len += sprintf(buf + len, "\n");
3473 len += sprintf(buf, "No data\n");
3477 enum slab_stat_type {
3484 #define SO_FULL (1 << SL_FULL)
3485 #define SO_PARTIAL (1 << SL_PARTIAL)
3486 #define SO_CPU (1 << SL_CPU)
3487 #define SO_OBJECTS (1 << SL_OBJECTS)
3489 static unsigned long slab_objects(struct kmem_cache *s,
3490 char *buf, unsigned long flags)
3492 unsigned long total = 0;
3496 unsigned long *nodes;
3497 unsigned long *per_cpu;
3499 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3500 per_cpu = nodes + nr_node_ids;
3502 for_each_possible_cpu(cpu) {
3504 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3514 if (flags & SO_CPU) {
3515 if (flags & SO_OBJECTS)
3526 for_each_node_state(node, N_NORMAL_MEMORY) {
3527 struct kmem_cache_node *n = get_node(s, node);
3529 if (flags & SO_PARTIAL) {
3530 if (flags & SO_OBJECTS)
3531 x = count_partial(n);
3538 if (flags & SO_FULL) {
3539 int full_slabs = atomic_long_read(&n->nr_slabs)
3543 if (flags & SO_OBJECTS)
3544 x = full_slabs * s->objects;
3552 x = sprintf(buf, "%lu", total);
3554 for_each_node_state(node, N_NORMAL_MEMORY)
3556 x += sprintf(buf + x, " N%d=%lu",
3560 return x + sprintf(buf + x, "\n");
3563 static int any_slab_objects(struct kmem_cache *s)
3568 for_each_possible_cpu(cpu) {
3569 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3575 for_each_online_node(node) {
3576 struct kmem_cache_node *n = get_node(s, node);
3581 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3587 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3588 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3590 struct slab_attribute {
3591 struct attribute attr;
3592 ssize_t (*show)(struct kmem_cache *s, char *buf);
3593 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3596 #define SLAB_ATTR_RO(_name) \
3597 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3599 #define SLAB_ATTR(_name) \
3600 static struct slab_attribute _name##_attr = \
3601 __ATTR(_name, 0644, _name##_show, _name##_store)
3603 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3605 return sprintf(buf, "%d\n", s->size);
3607 SLAB_ATTR_RO(slab_size);
3609 static ssize_t align_show(struct kmem_cache *s, char *buf)
3611 return sprintf(buf, "%d\n", s->align);
3613 SLAB_ATTR_RO(align);
3615 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3617 return sprintf(buf, "%d\n", s->objsize);
3619 SLAB_ATTR_RO(object_size);
3621 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3623 return sprintf(buf, "%d\n", s->objects);
3625 SLAB_ATTR_RO(objs_per_slab);
3627 static ssize_t order_show(struct kmem_cache *s, char *buf)
3629 return sprintf(buf, "%d\n", s->order);
3631 SLAB_ATTR_RO(order);
3633 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3636 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3638 return n + sprintf(buf + n, "\n");
3644 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3646 return sprintf(buf, "%d\n", s->refcount - 1);
3648 SLAB_ATTR_RO(aliases);
3650 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3652 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3654 SLAB_ATTR_RO(slabs);
3656 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3658 return slab_objects(s, buf, SO_PARTIAL);
3660 SLAB_ATTR_RO(partial);
3662 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3664 return slab_objects(s, buf, SO_CPU);
3666 SLAB_ATTR_RO(cpu_slabs);
3668 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3670 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3672 SLAB_ATTR_RO(objects);
3674 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3676 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3679 static ssize_t sanity_checks_store(struct kmem_cache *s,
3680 const char *buf, size_t length)
3682 s->flags &= ~SLAB_DEBUG_FREE;
3684 s->flags |= SLAB_DEBUG_FREE;
3687 SLAB_ATTR(sanity_checks);
3689 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3691 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3694 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3697 s->flags &= ~SLAB_TRACE;
3699 s->flags |= SLAB_TRACE;
3704 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3709 static ssize_t reclaim_account_store(struct kmem_cache *s,
3710 const char *buf, size_t length)
3712 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3714 s->flags |= SLAB_RECLAIM_ACCOUNT;
3717 SLAB_ATTR(reclaim_account);
3719 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3721 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3723 SLAB_ATTR_RO(hwcache_align);
3725 #ifdef CONFIG_ZONE_DMA
3726 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3728 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3730 SLAB_ATTR_RO(cache_dma);
3733 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3735 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3737 SLAB_ATTR_RO(destroy_by_rcu);
3739 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3741 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3744 static ssize_t red_zone_store(struct kmem_cache *s,
3745 const char *buf, size_t length)
3747 if (any_slab_objects(s))
3750 s->flags &= ~SLAB_RED_ZONE;
3752 s->flags |= SLAB_RED_ZONE;
3756 SLAB_ATTR(red_zone);
3758 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3760 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3763 static ssize_t poison_store(struct kmem_cache *s,
3764 const char *buf, size_t length)
3766 if (any_slab_objects(s))
3769 s->flags &= ~SLAB_POISON;
3771 s->flags |= SLAB_POISON;
3777 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3782 static ssize_t store_user_store(struct kmem_cache *s,
3783 const char *buf, size_t length)
3785 if (any_slab_objects(s))
3788 s->flags &= ~SLAB_STORE_USER;
3790 s->flags |= SLAB_STORE_USER;
3794 SLAB_ATTR(store_user);
3796 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3801 static ssize_t validate_store(struct kmem_cache *s,
3802 const char *buf, size_t length)
3806 if (buf[0] == '1') {
3807 ret = validate_slab_cache(s);
3813 SLAB_ATTR(validate);
3815 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3820 static ssize_t shrink_store(struct kmem_cache *s,
3821 const char *buf, size_t length)
3823 if (buf[0] == '1') {
3824 int rc = kmem_cache_shrink(s);
3834 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3836 if (!(s->flags & SLAB_STORE_USER))
3838 return list_locations(s, buf, TRACK_ALLOC);
3840 SLAB_ATTR_RO(alloc_calls);
3842 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3844 if (!(s->flags & SLAB_STORE_USER))
3846 return list_locations(s, buf, TRACK_FREE);
3848 SLAB_ATTR_RO(free_calls);
3851 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
3853 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
3856 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
3857 const char *buf, size_t length)
3859 int n = simple_strtoul(buf, NULL, 10);
3862 s->remote_node_defrag_ratio = n * 10;
3865 SLAB_ATTR(remote_node_defrag_ratio);
3868 static struct attribute * slab_attrs[] = {
3869 &slab_size_attr.attr,
3870 &object_size_attr.attr,
3871 &objs_per_slab_attr.attr,
3876 &cpu_slabs_attr.attr,
3880 &sanity_checks_attr.attr,
3882 &hwcache_align_attr.attr,
3883 &reclaim_account_attr.attr,
3884 &destroy_by_rcu_attr.attr,
3885 &red_zone_attr.attr,
3887 &store_user_attr.attr,
3888 &validate_attr.attr,
3890 &alloc_calls_attr.attr,
3891 &free_calls_attr.attr,
3892 #ifdef CONFIG_ZONE_DMA
3893 &cache_dma_attr.attr,
3896 &remote_node_defrag_ratio_attr.attr,
3901 static struct attribute_group slab_attr_group = {
3902 .attrs = slab_attrs,
3905 static ssize_t slab_attr_show(struct kobject *kobj,
3906 struct attribute *attr,
3909 struct slab_attribute *attribute;
3910 struct kmem_cache *s;
3913 attribute = to_slab_attr(attr);
3916 if (!attribute->show)
3919 err = attribute->show(s, buf);
3924 static ssize_t slab_attr_store(struct kobject *kobj,
3925 struct attribute *attr,
3926 const char *buf, size_t len)
3928 struct slab_attribute *attribute;
3929 struct kmem_cache *s;
3932 attribute = to_slab_attr(attr);
3935 if (!attribute->store)
3938 err = attribute->store(s, buf, len);
3943 static void kmem_cache_release(struct kobject *kobj)
3945 struct kmem_cache *s = to_slab(kobj);
3950 static struct sysfs_ops slab_sysfs_ops = {
3951 .show = slab_attr_show,
3952 .store = slab_attr_store,
3955 static struct kobj_type slab_ktype = {
3956 .sysfs_ops = &slab_sysfs_ops,
3957 .release = kmem_cache_release
3960 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3962 struct kobj_type *ktype = get_ktype(kobj);
3964 if (ktype == &slab_ktype)
3969 static struct kset_uevent_ops slab_uevent_ops = {
3970 .filter = uevent_filter,
3973 static struct kset *slab_kset;
3975 #define ID_STR_LENGTH 64
3977 /* Create a unique string id for a slab cache:
3979 * :[flags-]size:[memory address of kmemcache]
3981 static char *create_unique_id(struct kmem_cache *s)
3983 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3990 * First flags affecting slabcache operations. We will only
3991 * get here for aliasable slabs so we do not need to support
3992 * too many flags. The flags here must cover all flags that
3993 * are matched during merging to guarantee that the id is
3996 if (s->flags & SLAB_CACHE_DMA)
3998 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4000 if (s->flags & SLAB_DEBUG_FREE)
4004 p += sprintf(p, "%07d", s->size);
4005 BUG_ON(p > name + ID_STR_LENGTH - 1);
4009 static int sysfs_slab_add(struct kmem_cache *s)
4015 if (slab_state < SYSFS)
4016 /* Defer until later */
4019 unmergeable = slab_unmergeable(s);
4022 * Slabcache can never be merged so we can use the name proper.
4023 * This is typically the case for debug situations. In that
4024 * case we can catch duplicate names easily.
4026 sysfs_remove_link(&slab_kset->kobj, s->name);
4030 * Create a unique name for the slab as a target
4033 name = create_unique_id(s);
4036 s->kobj.kset = slab_kset;
4037 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4039 kobject_put(&s->kobj);
4043 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4046 kobject_uevent(&s->kobj, KOBJ_ADD);
4048 /* Setup first alias */
4049 sysfs_slab_alias(s, s->name);
4055 static void sysfs_slab_remove(struct kmem_cache *s)
4057 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4058 kobject_del(&s->kobj);
4059 kobject_put(&s->kobj);
4063 * Need to buffer aliases during bootup until sysfs becomes
4064 * available lest we loose that information.
4066 struct saved_alias {
4067 struct kmem_cache *s;
4069 struct saved_alias *next;
4072 static struct saved_alias *alias_list;
4074 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4076 struct saved_alias *al;
4078 if (slab_state == SYSFS) {
4080 * If we have a leftover link then remove it.
4082 sysfs_remove_link(&slab_kset->kobj, name);
4083 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4086 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4092 al->next = alias_list;
4097 static int __init slab_sysfs_init(void)
4099 struct kmem_cache *s;
4102 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4104 printk(KERN_ERR "Cannot register slab subsystem.\n");
4110 list_for_each_entry(s, &slab_caches, list) {
4111 err = sysfs_slab_add(s);
4113 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4114 " to sysfs\n", s->name);
4117 while (alias_list) {
4118 struct saved_alias *al = alias_list;
4120 alias_list = alias_list->next;
4121 err = sysfs_slab_alias(al->s, al->name);
4123 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4124 " %s to sysfs\n", s->name);
4132 __initcall(slab_sysfs_init);
4136 * The /proc/slabinfo ABI
4138 #ifdef CONFIG_SLABINFO
4140 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4141 size_t count, loff_t *ppos)
4147 static void print_slabinfo_header(struct seq_file *m)
4149 seq_puts(m, "slabinfo - version: 2.1\n");
4150 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4151 "<objperslab> <pagesperslab>");
4152 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4153 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4157 static void *s_start(struct seq_file *m, loff_t *pos)
4161 down_read(&slub_lock);
4163 print_slabinfo_header(m);
4165 return seq_list_start(&slab_caches, *pos);
4168 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4170 return seq_list_next(p, &slab_caches, pos);
4173 static void s_stop(struct seq_file *m, void *p)
4175 up_read(&slub_lock);
4178 static int s_show(struct seq_file *m, void *p)
4180 unsigned long nr_partials = 0;
4181 unsigned long nr_slabs = 0;
4182 unsigned long nr_inuse = 0;
4183 unsigned long nr_objs;
4184 struct kmem_cache *s;
4187 s = list_entry(p, struct kmem_cache, list);
4189 for_each_online_node(node) {
4190 struct kmem_cache_node *n = get_node(s, node);
4195 nr_partials += n->nr_partial;
4196 nr_slabs += atomic_long_read(&n->nr_slabs);
4197 nr_inuse += count_partial(n);
4200 nr_objs = nr_slabs * s->objects;
4201 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4203 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4204 nr_objs, s->size, s->objects, (1 << s->order));
4205 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4206 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4212 const struct seq_operations slabinfo_op = {
4219 #endif /* CONFIG_SLABINFO */