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 2
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) {}
253 /********************************************************************
254 * Core slab cache functions
255 *******************************************************************/
257 int slab_is_available(void)
259 return slab_state >= UP;
262 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
265 return s->node[node];
267 return &s->local_node;
271 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
274 return s->cpu_slab[cpu];
280 static inline int check_valid_pointer(struct kmem_cache *s,
281 struct page *page, const void *object)
288 base = page_address(page);
289 if (object < base || object >= base + s->objects * s->size ||
290 (object - base) % s->size) {
298 * Slow version of get and set free pointer.
300 * This version requires touching the cache lines of kmem_cache which
301 * we avoid to do in the fast alloc free paths. There we obtain the offset
302 * from the page struct.
304 static inline void *get_freepointer(struct kmem_cache *s, void *object)
306 return *(void **)(object + s->offset);
309 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
311 *(void **)(object + s->offset) = fp;
314 /* Loop over all objects in a slab */
315 #define for_each_object(__p, __s, __addr) \
316 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
320 #define for_each_free_object(__p, __s, __free) \
321 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
323 /* Determine object index from a given position */
324 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
326 return (p - addr) / s->size;
329 #ifdef CONFIG_SLUB_DEBUG
333 #ifdef CONFIG_SLUB_DEBUG_ON
334 static int slub_debug = DEBUG_DEFAULT_FLAGS;
336 static int slub_debug;
339 static char *slub_debug_slabs;
344 static void print_section(char *text, u8 *addr, unsigned int length)
352 for (i = 0; i < length; i++) {
354 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
357 printk(" %02x", addr[i]);
359 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
361 printk(" %s\n",ascii);
372 printk(" %s\n", ascii);
376 static struct track *get_track(struct kmem_cache *s, void *object,
377 enum track_item alloc)
382 p = object + s->offset + sizeof(void *);
384 p = object + s->inuse;
389 static void set_track(struct kmem_cache *s, void *object,
390 enum track_item alloc, void *addr)
395 p = object + s->offset + sizeof(void *);
397 p = object + s->inuse;
402 p->cpu = smp_processor_id();
403 p->pid = current ? current->pid : -1;
406 memset(p, 0, sizeof(struct track));
409 static void init_tracking(struct kmem_cache *s, void *object)
411 if (!(s->flags & SLAB_STORE_USER))
414 set_track(s, object, TRACK_FREE, NULL);
415 set_track(s, object, TRACK_ALLOC, NULL);
418 static void print_track(const char *s, struct track *t)
423 printk(KERN_ERR "INFO: %s in ", s);
424 __print_symbol("%s", (unsigned long)t->addr);
425 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
428 static void print_tracking(struct kmem_cache *s, void *object)
430 if (!(s->flags & SLAB_STORE_USER))
433 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
434 print_track("Freed", get_track(s, object, TRACK_FREE));
437 static void print_page_info(struct page *page)
439 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
440 page, page->inuse, page->freelist, page->flags);
444 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
450 vsnprintf(buf, sizeof(buf), fmt, args);
452 printk(KERN_ERR "========================================"
453 "=====================================\n");
454 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
455 printk(KERN_ERR "----------------------------------------"
456 "-------------------------------------\n\n");
459 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
465 vsnprintf(buf, sizeof(buf), fmt, args);
467 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
470 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
472 unsigned int off; /* Offset of last byte */
473 u8 *addr = page_address(page);
475 print_tracking(s, p);
477 print_page_info(page);
479 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
480 p, p - addr, get_freepointer(s, p));
483 print_section("Bytes b4", p - 16, 16);
485 print_section("Object", p, min(s->objsize, 128));
487 if (s->flags & SLAB_RED_ZONE)
488 print_section("Redzone", p + s->objsize,
489 s->inuse - s->objsize);
492 off = s->offset + sizeof(void *);
496 if (s->flags & SLAB_STORE_USER)
497 off += 2 * sizeof(struct track);
500 /* Beginning of the filler is the free pointer */
501 print_section("Padding", p + off, s->size - off);
506 static void object_err(struct kmem_cache *s, struct page *page,
507 u8 *object, char *reason)
510 print_trailer(s, page, object);
513 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
519 vsnprintf(buf, sizeof(buf), fmt, args);
522 print_page_info(page);
526 static void init_object(struct kmem_cache *s, void *object, int active)
530 if (s->flags & __OBJECT_POISON) {
531 memset(p, POISON_FREE, s->objsize - 1);
532 p[s->objsize -1] = POISON_END;
535 if (s->flags & SLAB_RED_ZONE)
536 memset(p + s->objsize,
537 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
538 s->inuse - s->objsize);
541 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
544 if (*start != (u8)value)
552 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
553 void *from, void *to)
555 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
556 memset(from, data, to - from);
559 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
560 u8 *object, char *what,
561 u8* start, unsigned int value, unsigned int bytes)
566 fault = check_bytes(start, value, bytes);
571 while (end > fault && end[-1] == value)
574 slab_bug(s, "%s overwritten", what);
575 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
576 fault, end - 1, fault[0], value);
577 print_trailer(s, page, object);
579 restore_bytes(s, what, value, fault, end);
587 * Bytes of the object to be managed.
588 * If the freepointer may overlay the object then the free
589 * pointer is the first word of the object.
591 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
594 * object + s->objsize
595 * Padding to reach word boundary. This is also used for Redzoning.
596 * Padding is extended by another word if Redzoning is enabled and
599 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
600 * 0xcc (RED_ACTIVE) for objects in use.
603 * Meta data starts here.
605 * A. Free pointer (if we cannot overwrite object on free)
606 * B. Tracking data for SLAB_STORE_USER
607 * C. Padding to reach required alignment boundary or at mininum
608 * one word if debuggin is on to be able to detect writes
609 * before the word boundary.
611 * Padding is done using 0x5a (POISON_INUSE)
614 * Nothing is used beyond s->size.
616 * If slabcaches are merged then the objsize and inuse boundaries are mostly
617 * ignored. And therefore no slab options that rely on these boundaries
618 * may be used with merged slabcaches.
621 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
623 unsigned long off = s->inuse; /* The end of info */
626 /* Freepointer is placed after the object. */
627 off += sizeof(void *);
629 if (s->flags & SLAB_STORE_USER)
630 /* We also have user information there */
631 off += 2 * sizeof(struct track);
636 return check_bytes_and_report(s, page, p, "Object padding",
637 p + off, POISON_INUSE, s->size - off);
640 static int slab_pad_check(struct kmem_cache *s, struct page *page)
648 if (!(s->flags & SLAB_POISON))
651 start = page_address(page);
652 end = start + (PAGE_SIZE << s->order);
653 length = s->objects * s->size;
654 remainder = end - (start + length);
658 fault = check_bytes(start + length, POISON_INUSE, remainder);
661 while (end > fault && end[-1] == POISON_INUSE)
664 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
665 print_section("Padding", start, length);
667 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
671 static int check_object(struct kmem_cache *s, struct page *page,
672 void *object, int active)
675 u8 *endobject = object + s->objsize;
677 if (s->flags & SLAB_RED_ZONE) {
679 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
681 if (!check_bytes_and_report(s, page, object, "Redzone",
682 endobject, red, s->inuse - s->objsize))
685 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
686 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
687 POISON_INUSE, s->inuse - s->objsize);
690 if (s->flags & SLAB_POISON) {
691 if (!active && (s->flags & __OBJECT_POISON) &&
692 (!check_bytes_and_report(s, page, p, "Poison", p,
693 POISON_FREE, s->objsize - 1) ||
694 !check_bytes_and_report(s, page, p, "Poison",
695 p + s->objsize -1, POISON_END, 1)))
698 * check_pad_bytes cleans up on its own.
700 check_pad_bytes(s, page, p);
703 if (!s->offset && active)
705 * Object and freepointer overlap. Cannot check
706 * freepointer while object is allocated.
710 /* Check free pointer validity */
711 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
712 object_err(s, page, p, "Freepointer corrupt");
714 * No choice but to zap it and thus loose the remainder
715 * of the free objects in this slab. May cause
716 * another error because the object count is now wrong.
718 set_freepointer(s, p, NULL);
724 static int check_slab(struct kmem_cache *s, struct page *page)
726 VM_BUG_ON(!irqs_disabled());
728 if (!PageSlab(page)) {
729 slab_err(s, page, "Not a valid slab page");
732 if (page->inuse > s->objects) {
733 slab_err(s, page, "inuse %u > max %u",
734 s->name, page->inuse, s->objects);
737 /* Slab_pad_check fixes things up after itself */
738 slab_pad_check(s, page);
743 * Determine if a certain object on a page is on the freelist. Must hold the
744 * slab lock to guarantee that the chains are in a consistent state.
746 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
749 void *fp = page->freelist;
752 while (fp && nr <= s->objects) {
755 if (!check_valid_pointer(s, page, fp)) {
757 object_err(s, page, object,
758 "Freechain corrupt");
759 set_freepointer(s, object, NULL);
762 slab_err(s, page, "Freepointer corrupt");
763 page->freelist = NULL;
764 page->inuse = s->objects;
765 slab_fix(s, "Freelist cleared");
771 fp = get_freepointer(s, object);
775 if (page->inuse != s->objects - nr) {
776 slab_err(s, page, "Wrong object count. Counter is %d but "
777 "counted were %d", page->inuse, s->objects - nr);
778 page->inuse = s->objects - nr;
779 slab_fix(s, "Object count adjusted.");
781 return search == NULL;
784 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
786 if (s->flags & SLAB_TRACE) {
787 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
789 alloc ? "alloc" : "free",
794 print_section("Object", (void *)object, s->objsize);
801 * Tracking of fully allocated slabs for debugging purposes.
803 static void add_full(struct kmem_cache_node *n, struct page *page)
805 spin_lock(&n->list_lock);
806 list_add(&page->lru, &n->full);
807 spin_unlock(&n->list_lock);
810 static void remove_full(struct kmem_cache *s, struct page *page)
812 struct kmem_cache_node *n;
814 if (!(s->flags & SLAB_STORE_USER))
817 n = get_node(s, page_to_nid(page));
819 spin_lock(&n->list_lock);
820 list_del(&page->lru);
821 spin_unlock(&n->list_lock);
824 static void setup_object_debug(struct kmem_cache *s, struct page *page,
827 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
830 init_object(s, object, 0);
831 init_tracking(s, object);
834 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
835 void *object, void *addr)
837 if (!check_slab(s, page))
840 if (object && !on_freelist(s, page, object)) {
841 object_err(s, page, object, "Object already allocated");
845 if (!check_valid_pointer(s, page, object)) {
846 object_err(s, page, object, "Freelist Pointer check fails");
850 if (object && !check_object(s, page, object, 0))
853 /* Success perform special debug activities for allocs */
854 if (s->flags & SLAB_STORE_USER)
855 set_track(s, object, TRACK_ALLOC, addr);
856 trace(s, page, object, 1);
857 init_object(s, object, 1);
861 if (PageSlab(page)) {
863 * If this is a slab page then lets do the best we can
864 * to avoid issues in the future. Marking all objects
865 * as used avoids touching the remaining objects.
867 slab_fix(s, "Marking all objects used");
868 page->inuse = s->objects;
869 page->freelist = NULL;
874 static int free_debug_processing(struct kmem_cache *s, struct page *page,
875 void *object, void *addr)
877 if (!check_slab(s, page))
880 if (!check_valid_pointer(s, page, object)) {
881 slab_err(s, page, "Invalid object pointer 0x%p", object);
885 if (on_freelist(s, page, object)) {
886 object_err(s, page, object, "Object already free");
890 if (!check_object(s, page, object, 1))
893 if (unlikely(s != page->slab)) {
895 slab_err(s, page, "Attempt to free object(0x%p) "
896 "outside of slab", object);
900 "SLUB <none>: no slab for object 0x%p.\n",
905 object_err(s, page, object,
906 "page slab pointer corrupt.");
910 /* Special debug activities for freeing objects */
911 if (!SlabFrozen(page) && !page->freelist)
912 remove_full(s, page);
913 if (s->flags & SLAB_STORE_USER)
914 set_track(s, object, TRACK_FREE, addr);
915 trace(s, page, object, 0);
916 init_object(s, object, 0);
920 slab_fix(s, "Object at 0x%p not freed", object);
924 static int __init setup_slub_debug(char *str)
926 slub_debug = DEBUG_DEFAULT_FLAGS;
927 if (*str++ != '=' || !*str)
929 * No options specified. Switch on full debugging.
935 * No options but restriction on slabs. This means full
936 * debugging for slabs matching a pattern.
943 * Switch off all debugging measures.
948 * Determine which debug features should be switched on
950 for ( ;*str && *str != ','; str++) {
951 switch (tolower(*str)) {
953 slub_debug |= SLAB_DEBUG_FREE;
956 slub_debug |= SLAB_RED_ZONE;
959 slub_debug |= SLAB_POISON;
962 slub_debug |= SLAB_STORE_USER;
965 slub_debug |= SLAB_TRACE;
968 printk(KERN_ERR "slub_debug option '%c' "
969 "unknown. skipped\n",*str);
975 slub_debug_slabs = str + 1;
980 __setup("slub_debug", setup_slub_debug);
982 static unsigned long kmem_cache_flags(unsigned long objsize,
983 unsigned long flags, const char *name,
984 void (*ctor)(struct kmem_cache *, void *))
987 * The page->offset field is only 16 bit wide. This is an offset
988 * in units of words from the beginning of an object. If the slab
989 * size is bigger then we cannot move the free pointer behind the
992 * On 32 bit platforms the limit is 256k. On 64bit platforms
995 * Debugging or ctor may create a need to move the free
996 * pointer. Fail if this happens.
998 if (objsize >= 65535 * sizeof(void *)) {
999 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1000 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1004 * Enable debugging if selected on the kernel commandline.
1006 if (slub_debug && (!slub_debug_slabs ||
1007 strncmp(slub_debug_slabs, name,
1008 strlen(slub_debug_slabs)) == 0))
1009 flags |= slub_debug;
1015 static inline void setup_object_debug(struct kmem_cache *s,
1016 struct page *page, void *object) {}
1018 static inline int alloc_debug_processing(struct kmem_cache *s,
1019 struct page *page, void *object, void *addr) { return 0; }
1021 static inline int free_debug_processing(struct kmem_cache *s,
1022 struct page *page, void *object, void *addr) { return 0; }
1024 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1026 static inline int check_object(struct kmem_cache *s, struct page *page,
1027 void *object, int active) { return 1; }
1028 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1029 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1030 unsigned long flags, const char *name,
1031 void (*ctor)(struct kmem_cache *, void *))
1035 #define slub_debug 0
1038 * Slab allocation and freeing
1040 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1043 int pages = 1 << s->order;
1046 flags |= __GFP_COMP;
1048 if (s->flags & SLAB_CACHE_DMA)
1051 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1052 flags |= __GFP_RECLAIMABLE;
1055 page = alloc_pages(flags, s->order);
1057 page = alloc_pages_node(node, flags, s->order);
1062 mod_zone_page_state(page_zone(page),
1063 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1064 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1070 static void setup_object(struct kmem_cache *s, struct page *page,
1073 setup_object_debug(s, page, object);
1074 if (unlikely(s->ctor))
1078 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1081 struct kmem_cache_node *n;
1086 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1088 page = allocate_slab(s,
1089 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1093 n = get_node(s, page_to_nid(page));
1095 atomic_long_inc(&n->nr_slabs);
1097 page->flags |= 1 << PG_slab;
1098 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1099 SLAB_STORE_USER | SLAB_TRACE))
1102 start = page_address(page);
1104 if (unlikely(s->flags & SLAB_POISON))
1105 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1108 for_each_object(p, s, start) {
1109 setup_object(s, page, last);
1110 set_freepointer(s, last, p);
1113 setup_object(s, page, last);
1114 set_freepointer(s, last, NULL);
1116 page->freelist = start;
1122 static void __free_slab(struct kmem_cache *s, struct page *page)
1124 int pages = 1 << s->order;
1126 if (unlikely(SlabDebug(page))) {
1129 slab_pad_check(s, page);
1130 for_each_object(p, s, page_address(page))
1131 check_object(s, page, p, 0);
1132 ClearSlabDebug(page);
1135 mod_zone_page_state(page_zone(page),
1136 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1137 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1140 __free_pages(page, s->order);
1143 static void rcu_free_slab(struct rcu_head *h)
1147 page = container_of((struct list_head *)h, struct page, lru);
1148 __free_slab(page->slab, page);
1151 static void free_slab(struct kmem_cache *s, struct page *page)
1153 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1155 * RCU free overloads the RCU head over the LRU
1157 struct rcu_head *head = (void *)&page->lru;
1159 call_rcu(head, rcu_free_slab);
1161 __free_slab(s, page);
1164 static void discard_slab(struct kmem_cache *s, struct page *page)
1166 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1168 atomic_long_dec(&n->nr_slabs);
1169 reset_page_mapcount(page);
1170 __ClearPageSlab(page);
1175 * Per slab locking using the pagelock
1177 static __always_inline void slab_lock(struct page *page)
1179 bit_spin_lock(PG_locked, &page->flags);
1182 static __always_inline void slab_unlock(struct page *page)
1184 bit_spin_unlock(PG_locked, &page->flags);
1187 static __always_inline int slab_trylock(struct page *page)
1191 rc = bit_spin_trylock(PG_locked, &page->flags);
1196 * Management of partially allocated slabs
1198 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1200 spin_lock(&n->list_lock);
1202 list_add_tail(&page->lru, &n->partial);
1203 spin_unlock(&n->list_lock);
1206 static void add_partial(struct kmem_cache_node *n, struct page *page)
1208 spin_lock(&n->list_lock);
1210 list_add(&page->lru, &n->partial);
1211 spin_unlock(&n->list_lock);
1214 static void remove_partial(struct kmem_cache *s,
1217 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1219 spin_lock(&n->list_lock);
1220 list_del(&page->lru);
1222 spin_unlock(&n->list_lock);
1226 * Lock slab and remove from the partial list.
1228 * Must hold list_lock.
1230 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1232 if (slab_trylock(page)) {
1233 list_del(&page->lru);
1235 SetSlabFrozen(page);
1242 * Try to allocate a partial slab from a specific node.
1244 static struct page *get_partial_node(struct kmem_cache_node *n)
1249 * Racy check. If we mistakenly see no partial slabs then we
1250 * just allocate an empty slab. If we mistakenly try to get a
1251 * partial slab and there is none available then get_partials()
1254 if (!n || !n->nr_partial)
1257 spin_lock(&n->list_lock);
1258 list_for_each_entry(page, &n->partial, lru)
1259 if (lock_and_freeze_slab(n, page))
1263 spin_unlock(&n->list_lock);
1268 * Get a page from somewhere. Search in increasing NUMA distances.
1270 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1273 struct zonelist *zonelist;
1278 * The defrag ratio allows a configuration of the tradeoffs between
1279 * inter node defragmentation and node local allocations. A lower
1280 * defrag_ratio increases the tendency to do local allocations
1281 * instead of attempting to obtain partial slabs from other nodes.
1283 * If the defrag_ratio is set to 0 then kmalloc() always
1284 * returns node local objects. If the ratio is higher then kmalloc()
1285 * may return off node objects because partial slabs are obtained
1286 * from other nodes and filled up.
1288 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1289 * defrag_ratio = 1000) then every (well almost) allocation will
1290 * first attempt to defrag slab caches on other nodes. This means
1291 * scanning over all nodes to look for partial slabs which may be
1292 * expensive if we do it every time we are trying to find a slab
1293 * with available objects.
1295 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1298 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1299 ->node_zonelists[gfp_zone(flags)];
1300 for (z = zonelist->zones; *z; z++) {
1301 struct kmem_cache_node *n;
1303 n = get_node(s, zone_to_nid(*z));
1305 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1306 n->nr_partial > MIN_PARTIAL) {
1307 page = get_partial_node(n);
1317 * Get a partial page, lock it and return it.
1319 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1322 int searchnode = (node == -1) ? numa_node_id() : node;
1324 page = get_partial_node(get_node(s, searchnode));
1325 if (page || (flags & __GFP_THISNODE))
1328 return get_any_partial(s, flags);
1332 * Move a page back to the lists.
1334 * Must be called with the slab lock held.
1336 * On exit the slab lock will have been dropped.
1338 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1340 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1342 ClearSlabFrozen(page);
1346 add_partial(n, page);
1347 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1352 if (n->nr_partial < MIN_PARTIAL) {
1354 * Adding an empty slab to the partial slabs in order
1355 * to avoid page allocator overhead. This slab needs
1356 * to come after the other slabs with objects in
1357 * order to fill them up. That way the size of the
1358 * partial list stays small. kmem_cache_shrink can
1359 * reclaim empty slabs from the partial list.
1361 add_partial_tail(n, page);
1365 discard_slab(s, page);
1371 * Remove the cpu slab
1373 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1375 struct page *page = c->page;
1377 * Merge cpu freelist into freelist. Typically we get here
1378 * because both freelists are empty. So this is unlikely
1381 while (unlikely(c->freelist)) {
1384 /* Retrieve object from cpu_freelist */
1385 object = c->freelist;
1386 c->freelist = c->freelist[c->offset];
1388 /* And put onto the regular freelist */
1389 object[c->offset] = page->freelist;
1390 page->freelist = object;
1394 unfreeze_slab(s, page);
1397 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1400 deactivate_slab(s, c);
1405 * Called from IPI handler with interrupts disabled.
1407 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1409 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1411 if (likely(c && c->page))
1415 static void flush_cpu_slab(void *d)
1417 struct kmem_cache *s = d;
1419 __flush_cpu_slab(s, smp_processor_id());
1422 static void flush_all(struct kmem_cache *s)
1425 on_each_cpu(flush_cpu_slab, s, 1, 1);
1427 unsigned long flags;
1429 local_irq_save(flags);
1431 local_irq_restore(flags);
1436 * Check if the objects in a per cpu structure fit numa
1437 * locality expectations.
1439 static inline int node_match(struct kmem_cache_cpu *c, int node)
1442 if (node != -1 && c->node != node)
1449 * Slow path. The lockless freelist is empty or we need to perform
1452 * Interrupts are disabled.
1454 * Processing is still very fast if new objects have been freed to the
1455 * regular freelist. In that case we simply take over the regular freelist
1456 * as the lockless freelist and zap the regular freelist.
1458 * If that is not working then we fall back to the partial lists. We take the
1459 * first element of the freelist as the object to allocate now and move the
1460 * rest of the freelist to the lockless freelist.
1462 * And if we were unable to get a new slab from the partial slab lists then
1463 * we need to allocate a new slab. This is slowest path since we may sleep.
1465 static void *__slab_alloc(struct kmem_cache *s,
1466 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1471 /* We handle __GFP_ZERO in the caller */
1472 gfpflags &= ~__GFP_ZERO;
1478 if (unlikely(!node_match(c, node)))
1481 object = c->page->freelist;
1482 if (unlikely(!object))
1484 if (unlikely(SlabDebug(c->page)))
1487 object = c->page->freelist;
1488 c->freelist = object[c->offset];
1489 c->page->inuse = s->objects;
1490 c->page->freelist = NULL;
1491 c->node = page_to_nid(c->page);
1492 slab_unlock(c->page);
1496 deactivate_slab(s, c);
1499 new = get_partial(s, gfpflags, node);
1505 if (gfpflags & __GFP_WAIT)
1508 new = new_slab(s, gfpflags, node);
1510 if (gfpflags & __GFP_WAIT)
1511 local_irq_disable();
1514 c = get_cpu_slab(s, smp_processor_id());
1524 object = c->page->freelist;
1525 if (!alloc_debug_processing(s, c->page, object, addr))
1529 c->page->freelist = object[c->offset];
1531 slab_unlock(c->page);
1536 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1537 * have the fastpath folded into their functions. So no function call
1538 * overhead for requests that can be satisfied on the fastpath.
1540 * The fastpath works by first checking if the lockless freelist can be used.
1541 * If not then __slab_alloc is called for slow processing.
1543 * Otherwise we can simply pick the next object from the lockless free list.
1545 static void __always_inline *slab_alloc(struct kmem_cache *s,
1546 gfp_t gfpflags, int node, void *addr)
1549 unsigned long flags;
1550 struct kmem_cache_cpu *c;
1552 local_irq_save(flags);
1553 c = get_cpu_slab(s, smp_processor_id());
1554 if (unlikely(!c->freelist || !node_match(c, node)))
1556 object = __slab_alloc(s, gfpflags, node, addr, c);
1559 object = c->freelist;
1560 c->freelist = object[c->offset];
1562 local_irq_restore(flags);
1564 if (unlikely((gfpflags & __GFP_ZERO) && object))
1565 memset(object, 0, c->objsize);
1570 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1572 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1574 EXPORT_SYMBOL(kmem_cache_alloc);
1577 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1579 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1581 EXPORT_SYMBOL(kmem_cache_alloc_node);
1585 * Slow patch handling. This may still be called frequently since objects
1586 * have a longer lifetime than the cpu slabs in most processing loads.
1588 * So we still attempt to reduce cache line usage. Just take the slab
1589 * lock and free the item. If there is no additional partial page
1590 * handling required then we can return immediately.
1592 static void __slab_free(struct kmem_cache *s, struct page *page,
1593 void *x, void *addr, unsigned int offset)
1596 void **object = (void *)x;
1600 if (unlikely(SlabDebug(page)))
1603 prior = object[offset] = page->freelist;
1604 page->freelist = object;
1607 if (unlikely(SlabFrozen(page)))
1610 if (unlikely(!page->inuse))
1614 * Objects left in the slab. If it
1615 * was not on the partial list before
1618 if (unlikely(!prior))
1619 add_partial(get_node(s, page_to_nid(page)), page);
1628 * Slab still on the partial list.
1630 remove_partial(s, page);
1633 discard_slab(s, page);
1637 if (!free_debug_processing(s, page, x, addr))
1643 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1644 * can perform fastpath freeing without additional function calls.
1646 * The fastpath is only possible if we are freeing to the current cpu slab
1647 * of this processor. This typically the case if we have just allocated
1650 * If fastpath is not possible then fall back to __slab_free where we deal
1651 * with all sorts of special processing.
1653 static void __always_inline slab_free(struct kmem_cache *s,
1654 struct page *page, void *x, void *addr)
1656 void **object = (void *)x;
1657 unsigned long flags;
1658 struct kmem_cache_cpu *c;
1660 local_irq_save(flags);
1661 debug_check_no_locks_freed(object, s->objsize);
1662 c = get_cpu_slab(s, smp_processor_id());
1663 if (likely(page == c->page && c->node >= 0)) {
1664 object[c->offset] = c->freelist;
1665 c->freelist = object;
1667 __slab_free(s, page, x, addr, c->offset);
1669 local_irq_restore(flags);
1672 void kmem_cache_free(struct kmem_cache *s, void *x)
1676 page = virt_to_head_page(x);
1678 slab_free(s, page, x, __builtin_return_address(0));
1680 EXPORT_SYMBOL(kmem_cache_free);
1682 /* Figure out on which slab object the object resides */
1683 static struct page *get_object_page(const void *x)
1685 struct page *page = virt_to_head_page(x);
1687 if (!PageSlab(page))
1694 * Object placement in a slab is made very easy because we always start at
1695 * offset 0. If we tune the size of the object to the alignment then we can
1696 * get the required alignment by putting one properly sized object after
1699 * Notice that the allocation order determines the sizes of the per cpu
1700 * caches. Each processor has always one slab available for allocations.
1701 * Increasing the allocation order reduces the number of times that slabs
1702 * must be moved on and off the partial lists and is therefore a factor in
1707 * Mininum / Maximum order of slab pages. This influences locking overhead
1708 * and slab fragmentation. A higher order reduces the number of partial slabs
1709 * and increases the number of allocations possible without having to
1710 * take the list_lock.
1712 static int slub_min_order;
1713 static int slub_max_order = DEFAULT_MAX_ORDER;
1714 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1717 * Merge control. If this is set then no merging of slab caches will occur.
1718 * (Could be removed. This was introduced to pacify the merge skeptics.)
1720 static int slub_nomerge;
1723 * Calculate the order of allocation given an slab object size.
1725 * The order of allocation has significant impact on performance and other
1726 * system components. Generally order 0 allocations should be preferred since
1727 * order 0 does not cause fragmentation in the page allocator. Larger objects
1728 * be problematic to put into order 0 slabs because there may be too much
1729 * unused space left. We go to a higher order if more than 1/8th of the slab
1732 * In order to reach satisfactory performance we must ensure that a minimum
1733 * number of objects is in one slab. Otherwise we may generate too much
1734 * activity on the partial lists which requires taking the list_lock. This is
1735 * less a concern for large slabs though which are rarely used.
1737 * slub_max_order specifies the order where we begin to stop considering the
1738 * number of objects in a slab as critical. If we reach slub_max_order then
1739 * we try to keep the page order as low as possible. So we accept more waste
1740 * of space in favor of a small page order.
1742 * Higher order allocations also allow the placement of more objects in a
1743 * slab and thereby reduce object handling overhead. If the user has
1744 * requested a higher mininum order then we start with that one instead of
1745 * the smallest order which will fit the object.
1747 static inline int slab_order(int size, int min_objects,
1748 int max_order, int fract_leftover)
1752 int min_order = slub_min_order;
1754 for (order = max(min_order,
1755 fls(min_objects * size - 1) - PAGE_SHIFT);
1756 order <= max_order; order++) {
1758 unsigned long slab_size = PAGE_SIZE << order;
1760 if (slab_size < min_objects * size)
1763 rem = slab_size % size;
1765 if (rem <= slab_size / fract_leftover)
1773 static inline int calculate_order(int size)
1780 * Attempt to find best configuration for a slab. This
1781 * works by first attempting to generate a layout with
1782 * the best configuration and backing off gradually.
1784 * First we reduce the acceptable waste in a slab. Then
1785 * we reduce the minimum objects required in a slab.
1787 min_objects = slub_min_objects;
1788 while (min_objects > 1) {
1790 while (fraction >= 4) {
1791 order = slab_order(size, min_objects,
1792 slub_max_order, fraction);
1793 if (order <= slub_max_order)
1801 * We were unable to place multiple objects in a slab. Now
1802 * lets see if we can place a single object there.
1804 order = slab_order(size, 1, slub_max_order, 1);
1805 if (order <= slub_max_order)
1809 * Doh this slab cannot be placed using slub_max_order.
1811 order = slab_order(size, 1, MAX_ORDER, 1);
1812 if (order <= MAX_ORDER)
1818 * Figure out what the alignment of the objects will be.
1820 static unsigned long calculate_alignment(unsigned long flags,
1821 unsigned long align, unsigned long size)
1824 * If the user wants hardware cache aligned objects then
1825 * follow that suggestion if the object is sufficiently
1828 * The hardware cache alignment cannot override the
1829 * specified alignment though. If that is greater
1832 if ((flags & SLAB_HWCACHE_ALIGN) &&
1833 size > cache_line_size() / 2)
1834 return max_t(unsigned long, align, cache_line_size());
1836 if (align < ARCH_SLAB_MINALIGN)
1837 return ARCH_SLAB_MINALIGN;
1839 return ALIGN(align, sizeof(void *));
1842 static void init_kmem_cache_cpu(struct kmem_cache *s,
1843 struct kmem_cache_cpu *c)
1848 c->offset = s->offset / sizeof(void *);
1849 c->objsize = s->objsize;
1852 static void init_kmem_cache_node(struct kmem_cache_node *n)
1855 atomic_long_set(&n->nr_slabs, 0);
1856 spin_lock_init(&n->list_lock);
1857 INIT_LIST_HEAD(&n->partial);
1858 #ifdef CONFIG_SLUB_DEBUG
1859 INIT_LIST_HEAD(&n->full);
1865 * Per cpu array for per cpu structures.
1867 * The per cpu array places all kmem_cache_cpu structures from one processor
1868 * close together meaning that it becomes possible that multiple per cpu
1869 * structures are contained in one cacheline. This may be particularly
1870 * beneficial for the kmalloc caches.
1872 * A desktop system typically has around 60-80 slabs. With 100 here we are
1873 * likely able to get per cpu structures for all caches from the array defined
1874 * here. We must be able to cover all kmalloc caches during bootstrap.
1876 * If the per cpu array is exhausted then fall back to kmalloc
1877 * of individual cachelines. No sharing is possible then.
1879 #define NR_KMEM_CACHE_CPU 100
1881 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1882 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1884 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1885 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1887 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1888 int cpu, gfp_t flags)
1890 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1893 per_cpu(kmem_cache_cpu_free, cpu) =
1894 (void *)c->freelist;
1896 /* Table overflow: So allocate ourselves */
1898 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1899 flags, cpu_to_node(cpu));
1904 init_kmem_cache_cpu(s, c);
1908 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1910 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1911 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1915 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1916 per_cpu(kmem_cache_cpu_free, cpu) = c;
1919 static void free_kmem_cache_cpus(struct kmem_cache *s)
1923 for_each_online_cpu(cpu) {
1924 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1927 s->cpu_slab[cpu] = NULL;
1928 free_kmem_cache_cpu(c, cpu);
1933 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1937 for_each_online_cpu(cpu) {
1938 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1943 c = alloc_kmem_cache_cpu(s, cpu, flags);
1945 free_kmem_cache_cpus(s);
1948 s->cpu_slab[cpu] = c;
1954 * Initialize the per cpu array.
1956 static void init_alloc_cpu_cpu(int cpu)
1960 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1963 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1964 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1966 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1969 static void __init init_alloc_cpu(void)
1973 for_each_online_cpu(cpu)
1974 init_alloc_cpu_cpu(cpu);
1978 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
1979 static inline void init_alloc_cpu(void) {}
1981 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1983 init_kmem_cache_cpu(s, &s->cpu_slab);
1990 * No kmalloc_node yet so do it by hand. We know that this is the first
1991 * slab on the node for this slabcache. There are no concurrent accesses
1994 * Note that this function only works on the kmalloc_node_cache
1995 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
1996 * memory on a fresh node that has no slab structures yet.
1998 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2002 struct kmem_cache_node *n;
2004 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2006 page = new_slab(kmalloc_caches, gfpflags, node);
2009 if (page_to_nid(page) != node) {
2010 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2012 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2013 "in order to be able to continue\n");
2018 page->freelist = get_freepointer(kmalloc_caches, n);
2020 kmalloc_caches->node[node] = n;
2021 #ifdef CONFIG_SLUB_DEBUG
2022 init_object(kmalloc_caches, n, 1);
2023 init_tracking(kmalloc_caches, n);
2025 init_kmem_cache_node(n);
2026 atomic_long_inc(&n->nr_slabs);
2027 add_partial(n, page);
2031 static void free_kmem_cache_nodes(struct kmem_cache *s)
2035 for_each_node_state(node, N_NORMAL_MEMORY) {
2036 struct kmem_cache_node *n = s->node[node];
2037 if (n && n != &s->local_node)
2038 kmem_cache_free(kmalloc_caches, n);
2039 s->node[node] = NULL;
2043 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2048 if (slab_state >= UP)
2049 local_node = page_to_nid(virt_to_page(s));
2053 for_each_node_state(node, N_NORMAL_MEMORY) {
2054 struct kmem_cache_node *n;
2056 if (local_node == node)
2059 if (slab_state == DOWN) {
2060 n = early_kmem_cache_node_alloc(gfpflags,
2064 n = kmem_cache_alloc_node(kmalloc_caches,
2068 free_kmem_cache_nodes(s);
2074 init_kmem_cache_node(n);
2079 static void free_kmem_cache_nodes(struct kmem_cache *s)
2083 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2085 init_kmem_cache_node(&s->local_node);
2091 * calculate_sizes() determines the order and the distribution of data within
2094 static int calculate_sizes(struct kmem_cache *s)
2096 unsigned long flags = s->flags;
2097 unsigned long size = s->objsize;
2098 unsigned long align = s->align;
2101 * Determine if we can poison the object itself. If the user of
2102 * the slab may touch the object after free or before allocation
2103 * then we should never poison the object itself.
2105 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2107 s->flags |= __OBJECT_POISON;
2109 s->flags &= ~__OBJECT_POISON;
2112 * Round up object size to the next word boundary. We can only
2113 * place the free pointer at word boundaries and this determines
2114 * the possible location of the free pointer.
2116 size = ALIGN(size, sizeof(void *));
2118 #ifdef CONFIG_SLUB_DEBUG
2120 * If we are Redzoning then check if there is some space between the
2121 * end of the object and the free pointer. If not then add an
2122 * additional word to have some bytes to store Redzone information.
2124 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2125 size += sizeof(void *);
2129 * With that we have determined the number of bytes in actual use
2130 * by the object. This is the potential offset to the free pointer.
2134 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2137 * Relocate free pointer after the object if it is not
2138 * permitted to overwrite the first word of the object on
2141 * This is the case if we do RCU, have a constructor or
2142 * destructor or are poisoning the objects.
2145 size += sizeof(void *);
2148 #ifdef CONFIG_SLUB_DEBUG
2149 if (flags & SLAB_STORE_USER)
2151 * Need to store information about allocs and frees after
2154 size += 2 * sizeof(struct track);
2156 if (flags & SLAB_RED_ZONE)
2158 * Add some empty padding so that we can catch
2159 * overwrites from earlier objects rather than let
2160 * tracking information or the free pointer be
2161 * corrupted if an user writes before the start
2164 size += sizeof(void *);
2168 * Determine the alignment based on various parameters that the
2169 * user specified and the dynamic determination of cache line size
2172 align = calculate_alignment(flags, align, s->objsize);
2175 * SLUB stores one object immediately after another beginning from
2176 * offset 0. In order to align the objects we have to simply size
2177 * each object to conform to the alignment.
2179 size = ALIGN(size, align);
2182 s->order = calculate_order(size);
2187 * Determine the number of objects per slab
2189 s->objects = (PAGE_SIZE << s->order) / size;
2191 return !!s->objects;
2195 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2196 const char *name, size_t size,
2197 size_t align, unsigned long flags,
2198 void (*ctor)(struct kmem_cache *, void *))
2200 memset(s, 0, kmem_size);
2205 s->flags = kmem_cache_flags(size, flags, name, ctor);
2207 if (!calculate_sizes(s))
2212 s->defrag_ratio = 100;
2214 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2217 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2219 free_kmem_cache_nodes(s);
2221 if (flags & SLAB_PANIC)
2222 panic("Cannot create slab %s size=%lu realsize=%u "
2223 "order=%u offset=%u flags=%lx\n",
2224 s->name, (unsigned long)size, s->size, s->order,
2230 * Check if a given pointer is valid
2232 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2236 page = get_object_page(object);
2238 if (!page || s != page->slab)
2239 /* No slab or wrong slab */
2242 if (!check_valid_pointer(s, page, object))
2246 * We could also check if the object is on the slabs freelist.
2247 * But this would be too expensive and it seems that the main
2248 * purpose of kmem_ptr_valid is to check if the object belongs
2249 * to a certain slab.
2253 EXPORT_SYMBOL(kmem_ptr_validate);
2256 * Determine the size of a slab object
2258 unsigned int kmem_cache_size(struct kmem_cache *s)
2262 EXPORT_SYMBOL(kmem_cache_size);
2264 const char *kmem_cache_name(struct kmem_cache *s)
2268 EXPORT_SYMBOL(kmem_cache_name);
2271 * Attempt to free all slabs on a node. Return the number of slabs we
2272 * were unable to free.
2274 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2275 struct list_head *list)
2277 int slabs_inuse = 0;
2278 unsigned long flags;
2279 struct page *page, *h;
2281 spin_lock_irqsave(&n->list_lock, flags);
2282 list_for_each_entry_safe(page, h, list, lru)
2284 list_del(&page->lru);
2285 discard_slab(s, page);
2288 spin_unlock_irqrestore(&n->list_lock, flags);
2293 * Release all resources used by a slab cache.
2295 static inline int kmem_cache_close(struct kmem_cache *s)
2301 /* Attempt to free all objects */
2302 free_kmem_cache_cpus(s);
2303 for_each_node_state(node, N_NORMAL_MEMORY) {
2304 struct kmem_cache_node *n = get_node(s, node);
2306 n->nr_partial -= free_list(s, n, &n->partial);
2307 if (atomic_long_read(&n->nr_slabs))
2310 free_kmem_cache_nodes(s);
2315 * Close a cache and release the kmem_cache structure
2316 * (must be used for caches created using kmem_cache_create)
2318 void kmem_cache_destroy(struct kmem_cache *s)
2320 down_write(&slub_lock);
2324 up_write(&slub_lock);
2325 if (kmem_cache_close(s))
2327 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);
2612 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2613 * the remaining slabs by the number of items in use. The slabs with the
2614 * most items in use come first. New allocations will then fill those up
2615 * and thus they can be removed from the partial lists.
2617 * The slabs with the least items are placed last. This results in them
2618 * being allocated from last increasing the chance that the last objects
2619 * are freed in them.
2621 int kmem_cache_shrink(struct kmem_cache *s)
2625 struct kmem_cache_node *n;
2628 struct list_head *slabs_by_inuse =
2629 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2630 unsigned long flags;
2632 if (!slabs_by_inuse)
2636 for_each_node_state(node, N_NORMAL_MEMORY) {
2637 n = get_node(s, node);
2642 for (i = 0; i < s->objects; i++)
2643 INIT_LIST_HEAD(slabs_by_inuse + i);
2645 spin_lock_irqsave(&n->list_lock, flags);
2648 * Build lists indexed by the items in use in each slab.
2650 * Note that concurrent frees may occur while we hold the
2651 * list_lock. page->inuse here is the upper limit.
2653 list_for_each_entry_safe(page, t, &n->partial, lru) {
2654 if (!page->inuse && slab_trylock(page)) {
2656 * Must hold slab lock here because slab_free
2657 * may have freed the last object and be
2658 * waiting to release the slab.
2660 list_del(&page->lru);
2663 discard_slab(s, page);
2665 list_move(&page->lru,
2666 slabs_by_inuse + page->inuse);
2671 * Rebuild the partial list with the slabs filled up most
2672 * first and the least used slabs at the end.
2674 for (i = s->objects - 1; i >= 0; i--)
2675 list_splice(slabs_by_inuse + i, n->partial.prev);
2677 spin_unlock_irqrestore(&n->list_lock, flags);
2680 kfree(slabs_by_inuse);
2683 EXPORT_SYMBOL(kmem_cache_shrink);
2685 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2686 static int slab_mem_going_offline_callback(void *arg)
2688 struct kmem_cache *s;
2690 down_read(&slub_lock);
2691 list_for_each_entry(s, &slab_caches, list)
2692 kmem_cache_shrink(s);
2693 up_read(&slub_lock);
2698 static void slab_mem_offline_callback(void *arg)
2700 struct kmem_cache_node *n;
2701 struct kmem_cache *s;
2702 struct memory_notify *marg = arg;
2705 offline_node = marg->status_change_nid;
2708 * If the node still has available memory. we need kmem_cache_node
2711 if (offline_node < 0)
2714 down_read(&slub_lock);
2715 list_for_each_entry(s, &slab_caches, list) {
2716 n = get_node(s, offline_node);
2719 * if n->nr_slabs > 0, slabs still exist on the node
2720 * that is going down. We were unable to free them,
2721 * and offline_pages() function shoudn't call this
2722 * callback. So, we must fail.
2724 BUG_ON(atomic_long_read(&n->nr_slabs));
2726 s->node[offline_node] = NULL;
2727 kmem_cache_free(kmalloc_caches, n);
2730 up_read(&slub_lock);
2733 static int slab_mem_going_online_callback(void *arg)
2735 struct kmem_cache_node *n;
2736 struct kmem_cache *s;
2737 struct memory_notify *marg = arg;
2738 int nid = marg->status_change_nid;
2742 * If the node's memory is already available, then kmem_cache_node is
2743 * already created. Nothing to do.
2749 * We are bringing a node online. No memory is availabe yet. We must
2750 * allocate a kmem_cache_node structure in order to bring the node
2753 down_read(&slub_lock);
2754 list_for_each_entry(s, &slab_caches, list) {
2756 * XXX: kmem_cache_alloc_node will fallback to other nodes
2757 * since memory is not yet available from the node that
2760 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2765 init_kmem_cache_node(n);
2769 up_read(&slub_lock);
2773 static int slab_memory_callback(struct notifier_block *self,
2774 unsigned long action, void *arg)
2779 case MEM_GOING_ONLINE:
2780 ret = slab_mem_going_online_callback(arg);
2782 case MEM_GOING_OFFLINE:
2783 ret = slab_mem_going_offline_callback(arg);
2786 case MEM_CANCEL_ONLINE:
2787 slab_mem_offline_callback(arg);
2790 case MEM_CANCEL_OFFLINE:
2794 ret = notifier_from_errno(ret);
2798 #endif /* CONFIG_MEMORY_HOTPLUG */
2800 /********************************************************************
2801 * Basic setup of slabs
2802 *******************************************************************/
2804 void __init kmem_cache_init(void)
2813 * Must first have the slab cache available for the allocations of the
2814 * struct kmem_cache_node's. There is special bootstrap code in
2815 * kmem_cache_open for slab_state == DOWN.
2817 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2818 sizeof(struct kmem_cache_node), GFP_KERNEL);
2819 kmalloc_caches[0].refcount = -1;
2822 hotplug_memory_notifier(slab_memory_callback, 1);
2825 /* Able to allocate the per node structures */
2826 slab_state = PARTIAL;
2828 /* Caches that are not of the two-to-the-power-of size */
2829 if (KMALLOC_MIN_SIZE <= 64) {
2830 create_kmalloc_cache(&kmalloc_caches[1],
2831 "kmalloc-96", 96, GFP_KERNEL);
2834 if (KMALLOC_MIN_SIZE <= 128) {
2835 create_kmalloc_cache(&kmalloc_caches[2],
2836 "kmalloc-192", 192, GFP_KERNEL);
2840 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) {
2841 create_kmalloc_cache(&kmalloc_caches[i],
2842 "kmalloc", 1 << i, GFP_KERNEL);
2848 * Patch up the size_index table if we have strange large alignment
2849 * requirements for the kmalloc array. This is only the case for
2850 * mips it seems. The standard arches will not generate any code here.
2852 * Largest permitted alignment is 256 bytes due to the way we
2853 * handle the index determination for the smaller caches.
2855 * Make sure that nothing crazy happens if someone starts tinkering
2856 * around with ARCH_KMALLOC_MINALIGN
2858 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2859 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2861 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2862 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2866 /* Provide the correct kmalloc names now that the caches are up */
2867 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++)
2868 kmalloc_caches[i]. name =
2869 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2872 register_cpu_notifier(&slab_notifier);
2873 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2874 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2876 kmem_size = sizeof(struct kmem_cache);
2880 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2881 " CPUs=%d, Nodes=%d\n",
2882 caches, cache_line_size(),
2883 slub_min_order, slub_max_order, slub_min_objects,
2884 nr_cpu_ids, nr_node_ids);
2888 * Find a mergeable slab cache
2890 static int slab_unmergeable(struct kmem_cache *s)
2892 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2899 * We may have set a slab to be unmergeable during bootstrap.
2901 if (s->refcount < 0)
2907 static struct kmem_cache *find_mergeable(size_t size,
2908 size_t align, unsigned long flags, const char *name,
2909 void (*ctor)(struct kmem_cache *, void *))
2911 struct kmem_cache *s;
2913 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2919 size = ALIGN(size, sizeof(void *));
2920 align = calculate_alignment(flags, align, size);
2921 size = ALIGN(size, align);
2922 flags = kmem_cache_flags(size, flags, name, NULL);
2924 list_for_each_entry(s, &slab_caches, list) {
2925 if (slab_unmergeable(s))
2931 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
2934 * Check if alignment is compatible.
2935 * Courtesy of Adrian Drzewiecki
2937 if ((s->size & ~(align -1)) != s->size)
2940 if (s->size - size >= sizeof(void *))
2948 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2949 size_t align, unsigned long flags,
2950 void (*ctor)(struct kmem_cache *, void *))
2952 struct kmem_cache *s;
2954 down_write(&slub_lock);
2955 s = find_mergeable(size, align, flags, name, ctor);
2961 * Adjust the object sizes so that we clear
2962 * the complete object on kzalloc.
2964 s->objsize = max(s->objsize, (int)size);
2967 * And then we need to update the object size in the
2968 * per cpu structures
2970 for_each_online_cpu(cpu)
2971 get_cpu_slab(s, cpu)->objsize = s->objsize;
2972 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2973 up_write(&slub_lock);
2974 if (sysfs_slab_alias(s, name))
2978 s = kmalloc(kmem_size, GFP_KERNEL);
2980 if (kmem_cache_open(s, GFP_KERNEL, name,
2981 size, align, flags, ctor)) {
2982 list_add(&s->list, &slab_caches);
2983 up_write(&slub_lock);
2984 if (sysfs_slab_add(s))
2990 up_write(&slub_lock);
2993 if (flags & SLAB_PANIC)
2994 panic("Cannot create slabcache %s\n", name);
2999 EXPORT_SYMBOL(kmem_cache_create);
3003 * Use the cpu notifier to insure that the cpu slabs are flushed when
3006 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3007 unsigned long action, void *hcpu)
3009 long cpu = (long)hcpu;
3010 struct kmem_cache *s;
3011 unsigned long flags;
3014 case CPU_UP_PREPARE:
3015 case CPU_UP_PREPARE_FROZEN:
3016 init_alloc_cpu_cpu(cpu);
3017 down_read(&slub_lock);
3018 list_for_each_entry(s, &slab_caches, list)
3019 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3021 up_read(&slub_lock);
3024 case CPU_UP_CANCELED:
3025 case CPU_UP_CANCELED_FROZEN:
3027 case CPU_DEAD_FROZEN:
3028 down_read(&slub_lock);
3029 list_for_each_entry(s, &slab_caches, list) {
3030 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3032 local_irq_save(flags);
3033 __flush_cpu_slab(s, cpu);
3034 local_irq_restore(flags);
3035 free_kmem_cache_cpu(c, cpu);
3036 s->cpu_slab[cpu] = NULL;
3038 up_read(&slub_lock);
3046 static struct notifier_block __cpuinitdata slab_notifier =
3047 { &slab_cpuup_callback, NULL, 0 };
3051 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3053 struct kmem_cache *s;
3055 if (unlikely(size > PAGE_SIZE / 2))
3056 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3058 s = get_slab(size, gfpflags);
3060 if (unlikely(ZERO_OR_NULL_PTR(s)))
3063 return slab_alloc(s, gfpflags, -1, caller);
3066 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3067 int node, void *caller)
3069 struct kmem_cache *s;
3071 if (unlikely(size > PAGE_SIZE / 2))
3072 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
3074 s = get_slab(size, gfpflags);
3076 if (unlikely(ZERO_OR_NULL_PTR(s)))
3079 return slab_alloc(s, gfpflags, node, caller);
3082 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3083 static int validate_slab(struct kmem_cache *s, struct page *page,
3087 void *addr = page_address(page);
3089 if (!check_slab(s, page) ||
3090 !on_freelist(s, page, NULL))
3093 /* Now we know that a valid freelist exists */
3094 bitmap_zero(map, s->objects);
3096 for_each_free_object(p, s, page->freelist) {
3097 set_bit(slab_index(p, s, addr), map);
3098 if (!check_object(s, page, p, 0))
3102 for_each_object(p, s, addr)
3103 if (!test_bit(slab_index(p, s, addr), map))
3104 if (!check_object(s, page, p, 1))
3109 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3112 if (slab_trylock(page)) {
3113 validate_slab(s, page, map);
3116 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3119 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3120 if (!SlabDebug(page))
3121 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3122 "on slab 0x%p\n", s->name, page);
3124 if (SlabDebug(page))
3125 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3126 "slab 0x%p\n", s->name, page);
3130 static int validate_slab_node(struct kmem_cache *s,
3131 struct kmem_cache_node *n, unsigned long *map)
3133 unsigned long count = 0;
3135 unsigned long flags;
3137 spin_lock_irqsave(&n->list_lock, flags);
3139 list_for_each_entry(page, &n->partial, lru) {
3140 validate_slab_slab(s, page, map);
3143 if (count != n->nr_partial)
3144 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3145 "counter=%ld\n", s->name, count, n->nr_partial);
3147 if (!(s->flags & SLAB_STORE_USER))
3150 list_for_each_entry(page, &n->full, lru) {
3151 validate_slab_slab(s, page, map);
3154 if (count != atomic_long_read(&n->nr_slabs))
3155 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3156 "counter=%ld\n", s->name, count,
3157 atomic_long_read(&n->nr_slabs));
3160 spin_unlock_irqrestore(&n->list_lock, flags);
3164 static long validate_slab_cache(struct kmem_cache *s)
3167 unsigned long count = 0;
3168 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3169 sizeof(unsigned long), GFP_KERNEL);
3175 for_each_node_state(node, N_NORMAL_MEMORY) {
3176 struct kmem_cache_node *n = get_node(s, node);
3178 count += validate_slab_node(s, n, map);
3184 #ifdef SLUB_RESILIENCY_TEST
3185 static void resiliency_test(void)
3189 printk(KERN_ERR "SLUB resiliency testing\n");
3190 printk(KERN_ERR "-----------------------\n");
3191 printk(KERN_ERR "A. Corruption after allocation\n");
3193 p = kzalloc(16, GFP_KERNEL);
3195 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3196 " 0x12->0x%p\n\n", p + 16);
3198 validate_slab_cache(kmalloc_caches + 4);
3200 /* Hmmm... The next two are dangerous */
3201 p = kzalloc(32, GFP_KERNEL);
3202 p[32 + sizeof(void *)] = 0x34;
3203 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3204 " 0x34 -> -0x%p\n", p);
3205 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3207 validate_slab_cache(kmalloc_caches + 5);
3208 p = kzalloc(64, GFP_KERNEL);
3209 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3211 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3213 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
3214 validate_slab_cache(kmalloc_caches + 6);
3216 printk(KERN_ERR "\nB. Corruption after free\n");
3217 p = kzalloc(128, GFP_KERNEL);
3220 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3221 validate_slab_cache(kmalloc_caches + 7);
3223 p = kzalloc(256, GFP_KERNEL);
3226 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
3227 validate_slab_cache(kmalloc_caches + 8);
3229 p = kzalloc(512, GFP_KERNEL);
3232 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3233 validate_slab_cache(kmalloc_caches + 9);
3236 static void resiliency_test(void) {};
3240 * Generate lists of code addresses where slabcache objects are allocated
3245 unsigned long count;
3258 unsigned long count;
3259 struct location *loc;
3262 static void free_loc_track(struct loc_track *t)
3265 free_pages((unsigned long)t->loc,
3266 get_order(sizeof(struct location) * t->max));
3269 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3274 order = get_order(sizeof(struct location) * max);
3276 l = (void *)__get_free_pages(flags, order);
3281 memcpy(l, t->loc, sizeof(struct location) * t->count);
3289 static int add_location(struct loc_track *t, struct kmem_cache *s,
3290 const struct track *track)
3292 long start, end, pos;
3295 unsigned long age = jiffies - track->when;
3301 pos = start + (end - start + 1) / 2;
3304 * There is nothing at "end". If we end up there
3305 * we need to add something to before end.
3310 caddr = t->loc[pos].addr;
3311 if (track->addr == caddr) {
3317 if (age < l->min_time)
3319 if (age > l->max_time)
3322 if (track->pid < l->min_pid)
3323 l->min_pid = track->pid;
3324 if (track->pid > l->max_pid)
3325 l->max_pid = track->pid;
3327 cpu_set(track->cpu, l->cpus);
3329 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3333 if (track->addr < caddr)
3340 * Not found. Insert new tracking element.
3342 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3348 (t->count - pos) * sizeof(struct location));
3351 l->addr = track->addr;
3355 l->min_pid = track->pid;
3356 l->max_pid = track->pid;
3357 cpus_clear(l->cpus);
3358 cpu_set(track->cpu, l->cpus);
3359 nodes_clear(l->nodes);
3360 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3364 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3365 struct page *page, enum track_item alloc)
3367 void *addr = page_address(page);
3368 DECLARE_BITMAP(map, s->objects);
3371 bitmap_zero(map, s->objects);
3372 for_each_free_object(p, s, page->freelist)
3373 set_bit(slab_index(p, s, addr), map);
3375 for_each_object(p, s, addr)
3376 if (!test_bit(slab_index(p, s, addr), map))
3377 add_location(t, s, get_track(s, p, alloc));
3380 static int list_locations(struct kmem_cache *s, char *buf,
3381 enum track_item alloc)
3385 struct loc_track t = { 0, 0, NULL };
3388 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3390 return sprintf(buf, "Out of memory\n");
3392 /* Push back cpu slabs */
3395 for_each_node_state(node, N_NORMAL_MEMORY) {
3396 struct kmem_cache_node *n = get_node(s, node);
3397 unsigned long flags;
3400 if (!atomic_long_read(&n->nr_slabs))
3403 spin_lock_irqsave(&n->list_lock, flags);
3404 list_for_each_entry(page, &n->partial, lru)
3405 process_slab(&t, s, page, alloc);
3406 list_for_each_entry(page, &n->full, lru)
3407 process_slab(&t, s, page, alloc);
3408 spin_unlock_irqrestore(&n->list_lock, flags);
3411 for (i = 0; i < t.count; i++) {
3412 struct location *l = &t.loc[i];
3414 if (n > PAGE_SIZE - 100)
3416 n += sprintf(buf + n, "%7ld ", l->count);
3419 n += sprint_symbol(buf + n, (unsigned long)l->addr);
3421 n += sprintf(buf + n, "<not-available>");
3423 if (l->sum_time != l->min_time) {
3424 unsigned long remainder;
3426 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3428 div_long_long_rem(l->sum_time, l->count, &remainder),
3431 n += sprintf(buf + n, " age=%ld",
3434 if (l->min_pid != l->max_pid)
3435 n += sprintf(buf + n, " pid=%ld-%ld",
3436 l->min_pid, l->max_pid);
3438 n += sprintf(buf + n, " pid=%ld",
3441 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3442 n < PAGE_SIZE - 60) {
3443 n += sprintf(buf + n, " cpus=");
3444 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3448 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3449 n < PAGE_SIZE - 60) {
3450 n += sprintf(buf + n, " nodes=");
3451 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3455 n += sprintf(buf + n, "\n");
3460 n += sprintf(buf, "No data\n");
3464 static unsigned long count_partial(struct kmem_cache_node *n)
3466 unsigned long flags;
3467 unsigned long x = 0;
3470 spin_lock_irqsave(&n->list_lock, flags);
3471 list_for_each_entry(page, &n->partial, lru)
3473 spin_unlock_irqrestore(&n->list_lock, flags);
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) {
3505 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3515 if (flags & SO_CPU) {
3518 if (flags & SO_OBJECTS)
3529 for_each_node_state(node, N_NORMAL_MEMORY) {
3530 struct kmem_cache_node *n = get_node(s, node);
3532 if (flags & SO_PARTIAL) {
3533 if (flags & SO_OBJECTS)
3534 x = count_partial(n);
3541 if (flags & SO_FULL) {
3542 int full_slabs = atomic_long_read(&n->nr_slabs)
3546 if (flags & SO_OBJECTS)
3547 x = full_slabs * s->objects;
3555 x = sprintf(buf, "%lu", total);
3557 for_each_node_state(node, N_NORMAL_MEMORY)
3559 x += sprintf(buf + x, " N%d=%lu",
3563 return x + sprintf(buf + x, "\n");
3566 static int any_slab_objects(struct kmem_cache *s)
3571 for_each_possible_cpu(cpu) {
3572 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3578 for_each_online_node(node) {
3579 struct kmem_cache_node *n = get_node(s, node);
3584 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3590 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3591 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3593 struct slab_attribute {
3594 struct attribute attr;
3595 ssize_t (*show)(struct kmem_cache *s, char *buf);
3596 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3599 #define SLAB_ATTR_RO(_name) \
3600 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3602 #define SLAB_ATTR(_name) \
3603 static struct slab_attribute _name##_attr = \
3604 __ATTR(_name, 0644, _name##_show, _name##_store)
3606 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3608 return sprintf(buf, "%d\n", s->size);
3610 SLAB_ATTR_RO(slab_size);
3612 static ssize_t align_show(struct kmem_cache *s, char *buf)
3614 return sprintf(buf, "%d\n", s->align);
3616 SLAB_ATTR_RO(align);
3618 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3620 return sprintf(buf, "%d\n", s->objsize);
3622 SLAB_ATTR_RO(object_size);
3624 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3626 return sprintf(buf, "%d\n", s->objects);
3628 SLAB_ATTR_RO(objs_per_slab);
3630 static ssize_t order_show(struct kmem_cache *s, char *buf)
3632 return sprintf(buf, "%d\n", s->order);
3634 SLAB_ATTR_RO(order);
3636 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3639 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3641 return n + sprintf(buf + n, "\n");
3647 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3649 return sprintf(buf, "%d\n", s->refcount - 1);
3651 SLAB_ATTR_RO(aliases);
3653 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3655 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3657 SLAB_ATTR_RO(slabs);
3659 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3661 return slab_objects(s, buf, SO_PARTIAL);
3663 SLAB_ATTR_RO(partial);
3665 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3667 return slab_objects(s, buf, SO_CPU);
3669 SLAB_ATTR_RO(cpu_slabs);
3671 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3673 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3675 SLAB_ATTR_RO(objects);
3677 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3679 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3682 static ssize_t sanity_checks_store(struct kmem_cache *s,
3683 const char *buf, size_t length)
3685 s->flags &= ~SLAB_DEBUG_FREE;
3687 s->flags |= SLAB_DEBUG_FREE;
3690 SLAB_ATTR(sanity_checks);
3692 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3694 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3697 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3700 s->flags &= ~SLAB_TRACE;
3702 s->flags |= SLAB_TRACE;
3707 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3709 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3712 static ssize_t reclaim_account_store(struct kmem_cache *s,
3713 const char *buf, size_t length)
3715 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3717 s->flags |= SLAB_RECLAIM_ACCOUNT;
3720 SLAB_ATTR(reclaim_account);
3722 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3724 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3726 SLAB_ATTR_RO(hwcache_align);
3728 #ifdef CONFIG_ZONE_DMA
3729 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3731 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3733 SLAB_ATTR_RO(cache_dma);
3736 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3738 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3740 SLAB_ATTR_RO(destroy_by_rcu);
3742 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3744 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3747 static ssize_t red_zone_store(struct kmem_cache *s,
3748 const char *buf, size_t length)
3750 if (any_slab_objects(s))
3753 s->flags &= ~SLAB_RED_ZONE;
3755 s->flags |= SLAB_RED_ZONE;
3759 SLAB_ATTR(red_zone);
3761 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3763 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3766 static ssize_t poison_store(struct kmem_cache *s,
3767 const char *buf, size_t length)
3769 if (any_slab_objects(s))
3772 s->flags &= ~SLAB_POISON;
3774 s->flags |= SLAB_POISON;
3780 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3782 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3785 static ssize_t store_user_store(struct kmem_cache *s,
3786 const char *buf, size_t length)
3788 if (any_slab_objects(s))
3791 s->flags &= ~SLAB_STORE_USER;
3793 s->flags |= SLAB_STORE_USER;
3797 SLAB_ATTR(store_user);
3799 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3804 static ssize_t validate_store(struct kmem_cache *s,
3805 const char *buf, size_t length)
3809 if (buf[0] == '1') {
3810 ret = validate_slab_cache(s);
3816 SLAB_ATTR(validate);
3818 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3823 static ssize_t shrink_store(struct kmem_cache *s,
3824 const char *buf, size_t length)
3826 if (buf[0] == '1') {
3827 int rc = kmem_cache_shrink(s);
3837 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3839 if (!(s->flags & SLAB_STORE_USER))
3841 return list_locations(s, buf, TRACK_ALLOC);
3843 SLAB_ATTR_RO(alloc_calls);
3845 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3847 if (!(s->flags & SLAB_STORE_USER))
3849 return list_locations(s, buf, TRACK_FREE);
3851 SLAB_ATTR_RO(free_calls);
3854 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3856 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3859 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3860 const char *buf, size_t length)
3862 int n = simple_strtoul(buf, NULL, 10);
3865 s->defrag_ratio = n * 10;
3868 SLAB_ATTR(defrag_ratio);
3871 static struct attribute * slab_attrs[] = {
3872 &slab_size_attr.attr,
3873 &object_size_attr.attr,
3874 &objs_per_slab_attr.attr,
3879 &cpu_slabs_attr.attr,
3883 &sanity_checks_attr.attr,
3885 &hwcache_align_attr.attr,
3886 &reclaim_account_attr.attr,
3887 &destroy_by_rcu_attr.attr,
3888 &red_zone_attr.attr,
3890 &store_user_attr.attr,
3891 &validate_attr.attr,
3893 &alloc_calls_attr.attr,
3894 &free_calls_attr.attr,
3895 #ifdef CONFIG_ZONE_DMA
3896 &cache_dma_attr.attr,
3899 &defrag_ratio_attr.attr,
3904 static struct attribute_group slab_attr_group = {
3905 .attrs = slab_attrs,
3908 static ssize_t slab_attr_show(struct kobject *kobj,
3909 struct attribute *attr,
3912 struct slab_attribute *attribute;
3913 struct kmem_cache *s;
3916 attribute = to_slab_attr(attr);
3919 if (!attribute->show)
3922 err = attribute->show(s, buf);
3927 static ssize_t slab_attr_store(struct kobject *kobj,
3928 struct attribute *attr,
3929 const char *buf, size_t len)
3931 struct slab_attribute *attribute;
3932 struct kmem_cache *s;
3935 attribute = to_slab_attr(attr);
3938 if (!attribute->store)
3941 err = attribute->store(s, buf, len);
3946 static struct sysfs_ops slab_sysfs_ops = {
3947 .show = slab_attr_show,
3948 .store = slab_attr_store,
3951 static struct kobj_type slab_ktype = {
3952 .sysfs_ops = &slab_sysfs_ops,
3955 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3957 struct kobj_type *ktype = get_ktype(kobj);
3959 if (ktype == &slab_ktype)
3964 static struct kset_uevent_ops slab_uevent_ops = {
3965 .filter = uevent_filter,
3968 static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3970 #define ID_STR_LENGTH 64
3972 /* Create a unique string id for a slab cache:
3974 * :[flags-]size:[memory address of kmemcache]
3976 static char *create_unique_id(struct kmem_cache *s)
3978 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3985 * First flags affecting slabcache operations. We will only
3986 * get here for aliasable slabs so we do not need to support
3987 * too many flags. The flags here must cover all flags that
3988 * are matched during merging to guarantee that the id is
3991 if (s->flags & SLAB_CACHE_DMA)
3993 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3995 if (s->flags & SLAB_DEBUG_FREE)
3999 p += sprintf(p, "%07d", s->size);
4000 BUG_ON(p > name + ID_STR_LENGTH - 1);
4004 static int sysfs_slab_add(struct kmem_cache *s)
4010 if (slab_state < SYSFS)
4011 /* Defer until later */
4014 unmergeable = slab_unmergeable(s);
4017 * Slabcache can never be merged so we can use the name proper.
4018 * This is typically the case for debug situations. In that
4019 * case we can catch duplicate names easily.
4021 sysfs_remove_link(&slab_subsys.kobj, s->name);
4025 * Create a unique name for the slab as a target
4028 name = create_unique_id(s);
4031 kobj_set_kset_s(s, slab_subsys);
4032 kobject_set_name(&s->kobj, name);
4033 kobject_init(&s->kobj);
4034 err = kobject_add(&s->kobj);
4038 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4041 kobject_uevent(&s->kobj, KOBJ_ADD);
4043 /* Setup first alias */
4044 sysfs_slab_alias(s, s->name);
4050 static void sysfs_slab_remove(struct kmem_cache *s)
4052 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4053 kobject_del(&s->kobj);
4057 * Need to buffer aliases during bootup until sysfs becomes
4058 * available lest we loose that information.
4060 struct saved_alias {
4061 struct kmem_cache *s;
4063 struct saved_alias *next;
4066 static struct saved_alias *alias_list;
4068 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4070 struct saved_alias *al;
4072 if (slab_state == SYSFS) {
4074 * If we have a leftover link then remove it.
4076 sysfs_remove_link(&slab_subsys.kobj, name);
4077 return sysfs_create_link(&slab_subsys.kobj,
4081 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4087 al->next = alias_list;
4092 static int __init slab_sysfs_init(void)
4094 struct kmem_cache *s;
4097 err = subsystem_register(&slab_subsys);
4099 printk(KERN_ERR "Cannot register slab subsystem.\n");
4105 list_for_each_entry(s, &slab_caches, list) {
4106 err = sysfs_slab_add(s);
4108 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4109 " to sysfs\n", s->name);
4112 while (alias_list) {
4113 struct saved_alias *al = alias_list;
4115 alias_list = alias_list->next;
4116 err = sysfs_slab_alias(al->s, al->name);
4118 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4119 " %s to sysfs\n", s->name);
4127 __initcall(slab_sysfs_init);