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>
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
69 * Slabs with free elements are kept on a partial list.
70 * There is no list for full slabs. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * Otherwise there is no need to track full slabs unless we have to
73 * track full slabs for debugging purposes.
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
79 * Overloading of page flags that are otherwise used for LRU management.
81 * PageActive The slab is used as a cpu cache. Allocations
82 * may be performed from the slab. The slab is not
83 * on any slab list and cannot be moved onto one.
85 * PageError Slab requires special handling due to debug
86 * options set. This moves slab handling out of
91 * Issues still to be resolved:
93 * - The per cpu array is updated for each new slab and and is a remote
94 * cacheline for most nodes. This could become a bouncing cacheline given
95 * enough frequent updates. There are 16 pointers in a cacheline.so at
96 * max 16 cpus could compete. Likely okay.
98 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
100 * - Support DEBUG_SLAB_LEAK. Trouble is we do not know where the full
103 * - SLAB_DEBUG_INITIAL is not supported but I have never seen a use of
106 * - Variable sizing of the per node arrays
109 /* Enable to test recovery from slab corruption on boot */
110 #undef SLUB_RESILIENCY_TEST
115 * Small page size. Make sure that we do not fragment memory
117 #define DEFAULT_MAX_ORDER 1
118 #define DEFAULT_MIN_OBJECTS 4
123 * Large page machines are customarily able to handle larger
126 #define DEFAULT_MAX_ORDER 2
127 #define DEFAULT_MIN_OBJECTS 8
132 * Flags from the regular SLAB that SLUB does not support:
134 #define SLUB_UNIMPLEMENTED (SLAB_DEBUG_INITIAL)
136 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
137 SLAB_POISON | SLAB_STORE_USER)
139 * Set of flags that will prevent slab merging
141 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
142 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
144 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
147 #ifndef ARCH_KMALLOC_MINALIGN
148 #define ARCH_KMALLOC_MINALIGN sizeof(void *)
151 #ifndef ARCH_SLAB_MINALIGN
152 #define ARCH_SLAB_MINALIGN sizeof(void *)
155 /* Internal SLUB flags */
156 #define __OBJECT_POISON 0x80000000 /* Poison object */
158 static int kmem_size = sizeof(struct kmem_cache);
161 static struct notifier_block slab_notifier;
165 DOWN, /* No slab functionality available */
166 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
167 UP, /* Everything works */
171 /* A list of all slab caches on the system */
172 static DECLARE_RWSEM(slub_lock);
173 LIST_HEAD(slab_caches);
176 static int sysfs_slab_add(struct kmem_cache *);
177 static int sysfs_slab_alias(struct kmem_cache *, const char *);
178 static void sysfs_slab_remove(struct kmem_cache *);
180 static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
181 static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
182 static void sysfs_slab_remove(struct kmem_cache *s) {}
185 /********************************************************************
186 * Core slab cache functions
187 *******************************************************************/
189 int slab_is_available(void)
191 return slab_state >= UP;
194 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
197 return s->node[node];
199 return &s->local_node;
206 static void print_section(char *text, u8 *addr, unsigned int length)
214 for (i = 0; i < length; i++) {
216 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
219 printk(" %02x", addr[i]);
221 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
223 printk(" %s\n",ascii);
234 printk(" %s\n", ascii);
239 * Slow version of get and set free pointer.
241 * This requires touching the cache lines of kmem_cache.
242 * The offset can also be obtained from the page. In that
243 * case it is in the cacheline that we already need to touch.
245 static void *get_freepointer(struct kmem_cache *s, void *object)
247 return *(void **)(object + s->offset);
250 static void set_freepointer(struct kmem_cache *s, void *object, void *fp)
252 *(void **)(object + s->offset) = fp;
256 * Tracking user of a slab.
259 void *addr; /* Called from address */
260 int cpu; /* Was running on cpu */
261 int pid; /* Pid context */
262 unsigned long when; /* When did the operation occur */
265 enum track_item { TRACK_ALLOC, TRACK_FREE };
267 static struct track *get_track(struct kmem_cache *s, void *object,
268 enum track_item alloc)
273 p = object + s->offset + sizeof(void *);
275 p = object + s->inuse;
280 static void set_track(struct kmem_cache *s, void *object,
281 enum track_item alloc, void *addr)
286 p = object + s->offset + sizeof(void *);
288 p = object + s->inuse;
293 p->cpu = smp_processor_id();
294 p->pid = current ? current->pid : -1;
297 memset(p, 0, sizeof(struct track));
300 #define set_tracking(__s, __o, __a) set_track(__s, __o, __a, \
301 __builtin_return_address(0))
303 static void init_tracking(struct kmem_cache *s, void *object)
305 if (s->flags & SLAB_STORE_USER) {
306 set_track(s, object, TRACK_FREE, NULL);
307 set_track(s, object, TRACK_ALLOC, NULL);
311 static void print_track(const char *s, struct track *t)
316 printk(KERN_ERR "%s: ", s);
317 __print_symbol("%s", (unsigned long)t->addr);
318 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
321 static void print_trailer(struct kmem_cache *s, u8 *p)
323 unsigned int off; /* Offset of last byte */
325 if (s->flags & SLAB_RED_ZONE)
326 print_section("Redzone", p + s->objsize,
327 s->inuse - s->objsize);
329 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
331 get_freepointer(s, p));
334 off = s->offset + sizeof(void *);
338 if (s->flags & SLAB_STORE_USER) {
339 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
340 print_track("Last free ", get_track(s, p, TRACK_FREE));
341 off += 2 * sizeof(struct track);
345 /* Beginning of the filler is the free pointer */
346 print_section("Filler", p + off, s->size - off);
349 static void object_err(struct kmem_cache *s, struct page *page,
350 u8 *object, char *reason)
352 u8 *addr = page_address(page);
354 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
355 s->name, reason, object, page);
356 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
357 object - addr, page->flags, page->inuse, page->freelist);
358 if (object > addr + 16)
359 print_section("Bytes b4", object - 16, 16);
360 print_section("Object", object, min(s->objsize, 128));
361 print_trailer(s, object);
365 static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
370 va_start(args, reason);
371 vsnprintf(buf, sizeof(buf), reason, args);
373 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
378 static void init_object(struct kmem_cache *s, void *object, int active)
382 if (s->flags & __OBJECT_POISON) {
383 memset(p, POISON_FREE, s->objsize - 1);
384 p[s->objsize -1] = POISON_END;
387 if (s->flags & SLAB_RED_ZONE)
388 memset(p + s->objsize,
389 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
390 s->inuse - s->objsize);
393 static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
396 if (*start != (u8)value)
405 static int check_valid_pointer(struct kmem_cache *s, struct page *page,
413 base = page_address(page);
414 if (object < base || object >= base + s->objects * s->size ||
415 (object - base) % s->size) {
426 * Bytes of the object to be managed.
427 * If the freepointer may overlay the object then the free
428 * pointer is the first word of the object.
429 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
432 * object + s->objsize
433 * Padding to reach word boundary. This is also used for Redzoning.
434 * Padding is extended to word size if Redzoning is enabled
435 * and objsize == inuse.
436 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
437 * 0xcc (RED_ACTIVE) for objects in use.
440 * A. Free pointer (if we cannot overwrite object on free)
441 * B. Tracking data for SLAB_STORE_USER
442 * C. Padding to reach required alignment boundary
443 * Padding is done using 0x5a (POISON_INUSE)
447 * If slabcaches are merged then the objsize and inuse boundaries are to
448 * be ignored. And therefore no slab options that rely on these boundaries
449 * may be used with merged slabcaches.
452 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
453 void *from, void *to)
455 printk(KERN_ERR "@@@ SLUB: %s Restoring %s (0x%x) from 0x%p-0x%p\n",
456 s->name, message, data, from, to - 1);
457 memset(from, data, to - from);
460 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
462 unsigned long off = s->inuse; /* The end of info */
465 /* Freepointer is placed after the object. */
466 off += sizeof(void *);
468 if (s->flags & SLAB_STORE_USER)
469 /* We also have user information there */
470 off += 2 * sizeof(struct track);
475 if (check_bytes(p + off, POISON_INUSE, s->size - off))
478 object_err(s, page, p, "Object padding check fails");
483 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
487 static int slab_pad_check(struct kmem_cache *s, struct page *page)
490 int length, remainder;
492 if (!(s->flags & SLAB_POISON))
495 p = page_address(page);
496 length = s->objects * s->size;
497 remainder = (PAGE_SIZE << s->order) - length;
501 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
502 printk(KERN_ERR "SLUB: %s slab 0x%p: Padding fails check\n",
505 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
506 p + length + remainder);
512 static int check_object(struct kmem_cache *s, struct page *page,
513 void *object, int active)
516 u8 *endobject = object + s->objsize;
518 if (s->flags & SLAB_RED_ZONE) {
520 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
522 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
523 object_err(s, page, object,
524 active ? "Redzone Active" : "Redzone Inactive");
525 restore_bytes(s, "redzone", red,
526 endobject, object + s->inuse);
530 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
531 !check_bytes(endobject, POISON_INUSE,
532 s->inuse - s->objsize)) {
533 object_err(s, page, p, "Alignment padding check fails");
535 * Fix it so that there will not be another report.
537 * Hmmm... We may be corrupting an object that now expects
538 * to be longer than allowed.
540 restore_bytes(s, "alignment padding", POISON_INUSE,
541 endobject, object + s->inuse);
545 if (s->flags & SLAB_POISON) {
546 if (!active && (s->flags & __OBJECT_POISON) &&
547 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
548 p[s->objsize - 1] != POISON_END)) {
550 object_err(s, page, p, "Poison check failed");
551 restore_bytes(s, "Poison", POISON_FREE,
552 p, p + s->objsize -1);
553 restore_bytes(s, "Poison", POISON_END,
554 p + s->objsize - 1, p + s->objsize);
558 * check_pad_bytes cleans up on its own.
560 check_pad_bytes(s, page, p);
563 if (!s->offset && active)
565 * Object and freepointer overlap. Cannot check
566 * freepointer while object is allocated.
570 /* Check free pointer validity */
571 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
572 object_err(s, page, p, "Freepointer corrupt");
574 * No choice but to zap it and thus loose the remainder
575 * of the free objects in this slab. May cause
576 * another error because the object count maybe
579 set_freepointer(s, p, NULL);
585 static int check_slab(struct kmem_cache *s, struct page *page)
587 VM_BUG_ON(!irqs_disabled());
589 if (!PageSlab(page)) {
590 printk(KERN_ERR "SLUB: %s Not a valid slab page @0x%p "
591 "flags=%lx mapping=0x%p count=%d \n",
592 s->name, page, page->flags, page->mapping,
596 if (page->offset * sizeof(void *) != s->offset) {
597 printk(KERN_ERR "SLUB: %s Corrupted offset %lu in slab @0x%p"
598 " flags=0x%lx mapping=0x%p count=%d\n",
600 (unsigned long)(page->offset * sizeof(void *)),
608 if (page->inuse > s->objects) {
609 printk(KERN_ERR "SLUB: %s Inuse %u > max %u in slab "
610 "page @0x%p flags=%lx mapping=0x%p count=%d\n",
611 s->name, page->inuse, s->objects, page, page->flags,
612 page->mapping, page_count(page));
616 /* Slab_pad_check fixes things up after itself */
617 slab_pad_check(s, page);
622 * Determine if a certain object on a page is on the freelist and
623 * therefore free. Must hold the slab lock for cpu slabs to
624 * guarantee that the chains are consistent.
626 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
629 void *fp = page->freelist;
632 while (fp && nr <= s->objects) {
635 if (!check_valid_pointer(s, page, fp)) {
637 object_err(s, page, object,
638 "Freechain corrupt");
639 set_freepointer(s, object, NULL);
642 printk(KERN_ERR "SLUB: %s slab 0x%p "
643 "freepointer 0x%p corrupted.\n",
646 page->freelist = NULL;
647 page->inuse = s->objects;
653 fp = get_freepointer(s, object);
657 if (page->inuse != s->objects - nr) {
658 printk(KERN_ERR "slab %s: page 0x%p wrong object count."
659 " counter is %d but counted were %d\n",
660 s->name, page, page->inuse,
662 page->inuse = s->objects - nr;
664 return search == NULL;
667 static int alloc_object_checks(struct kmem_cache *s, struct page *page,
670 if (!check_slab(s, page))
673 if (object && !on_freelist(s, page, object)) {
674 printk(KERN_ERR "SLUB: %s Object 0x%p@0x%p "
675 "already allocated.\n",
676 s->name, object, page);
680 if (!check_valid_pointer(s, page, object)) {
681 object_err(s, page, object, "Freelist Pointer check fails");
688 if (!check_object(s, page, object, 0))
690 init_object(s, object, 1);
692 if (s->flags & SLAB_TRACE) {
693 printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
694 s->name, object, page->inuse,
702 if (PageSlab(page)) {
704 * If this is a slab page then lets do the best we can
705 * to avoid issues in the future. Marking all objects
706 * as used avoids touching the remainder.
708 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
710 page->inuse = s->objects;
711 page->freelist = NULL;
712 /* Fix up fields that may be corrupted */
713 page->offset = s->offset / sizeof(void *);
718 static int free_object_checks(struct kmem_cache *s, struct page *page,
721 if (!check_slab(s, page))
724 if (!check_valid_pointer(s, page, object)) {
725 printk(KERN_ERR "SLUB: %s slab 0x%p invalid "
726 "object pointer 0x%p\n",
727 s->name, page, object);
731 if (on_freelist(s, page, object)) {
732 printk(KERN_ERR "SLUB: %s slab 0x%p object "
733 "0x%p already free.\n", s->name, page, object);
737 if (!check_object(s, page, object, 1))
740 if (unlikely(s != page->slab)) {
742 printk(KERN_ERR "slab_free %s size %d: attempt to"
743 "free object(0x%p) outside of slab.\n",
744 s->name, s->size, object);
748 "slab_free : no slab(NULL) for object 0x%p.\n",
751 printk(KERN_ERR "slab_free %s(%d): object at 0x%p"
752 " belongs to slab %s(%d)\n",
753 s->name, s->size, object,
754 page->slab->name, page->slab->size);
757 if (s->flags & SLAB_TRACE) {
758 printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n",
759 s->name, object, page->inuse,
761 print_section("Object", object, s->objsize);
764 init_object(s, object, 0);
768 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
769 s->name, page, object);
774 * Slab allocation and freeing
776 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
779 int pages = 1 << s->order;
784 if (s->flags & SLAB_CACHE_DMA)
788 page = alloc_pages(flags, s->order);
790 page = alloc_pages_node(node, flags, s->order);
795 mod_zone_page_state(page_zone(page),
796 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
797 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
803 static void setup_object(struct kmem_cache *s, struct page *page,
806 if (PageError(page)) {
807 init_object(s, object, 0);
808 init_tracking(s, object);
811 if (unlikely(s->ctor)) {
812 int mode = SLAB_CTOR_CONSTRUCTOR;
814 if (!(s->flags & __GFP_WAIT))
815 mode |= SLAB_CTOR_ATOMIC;
817 s->ctor(object, s, mode);
821 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
824 struct kmem_cache_node *n;
830 if (flags & __GFP_NO_GROW)
833 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
835 if (flags & __GFP_WAIT)
838 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
842 n = get_node(s, page_to_nid(page));
844 atomic_long_inc(&n->nr_slabs);
845 page->offset = s->offset / sizeof(void *);
847 page->flags |= 1 << PG_slab;
848 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
849 SLAB_STORE_USER | SLAB_TRACE))
850 page->flags |= 1 << PG_error;
852 start = page_address(page);
853 end = start + s->objects * s->size;
855 if (unlikely(s->flags & SLAB_POISON))
856 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
859 for (p = start + s->size; p < end; p += s->size) {
860 setup_object(s, page, last);
861 set_freepointer(s, last, p);
864 setup_object(s, page, last);
865 set_freepointer(s, last, NULL);
867 page->freelist = start;
870 if (flags & __GFP_WAIT)
875 static void __free_slab(struct kmem_cache *s, struct page *page)
877 int pages = 1 << s->order;
879 if (unlikely(PageError(page) || s->dtor)) {
880 void *start = page_address(page);
881 void *end = start + (pages << PAGE_SHIFT);
884 slab_pad_check(s, page);
885 for (p = start; p <= end - s->size; p += s->size) {
888 check_object(s, page, p, 0);
892 mod_zone_page_state(page_zone(page),
893 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
894 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
897 page->mapping = NULL;
898 __free_pages(page, s->order);
901 static void rcu_free_slab(struct rcu_head *h)
905 page = container_of((struct list_head *)h, struct page, lru);
906 __free_slab(page->slab, page);
909 static void free_slab(struct kmem_cache *s, struct page *page)
911 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
913 * RCU free overloads the RCU head over the LRU
915 struct rcu_head *head = (void *)&page->lru;
917 call_rcu(head, rcu_free_slab);
919 __free_slab(s, page);
922 static void discard_slab(struct kmem_cache *s, struct page *page)
924 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
926 atomic_long_dec(&n->nr_slabs);
927 reset_page_mapcount(page);
928 page->flags &= ~(1 << PG_slab | 1 << PG_error);
933 * Per slab locking using the pagelock
935 static __always_inline void slab_lock(struct page *page)
937 bit_spin_lock(PG_locked, &page->flags);
940 static __always_inline void slab_unlock(struct page *page)
942 bit_spin_unlock(PG_locked, &page->flags);
945 static __always_inline int slab_trylock(struct page *page)
949 rc = bit_spin_trylock(PG_locked, &page->flags);
954 * Management of partially allocated slabs
956 static void add_partial(struct kmem_cache *s, struct page *page)
958 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
960 spin_lock(&n->list_lock);
962 list_add(&page->lru, &n->partial);
963 spin_unlock(&n->list_lock);
966 static void remove_partial(struct kmem_cache *s,
969 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
971 spin_lock(&n->list_lock);
972 list_del(&page->lru);
974 spin_unlock(&n->list_lock);
978 * Lock page and remove it from the partial list
980 * Must hold list_lock
982 static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page)
984 if (slab_trylock(page)) {
985 list_del(&page->lru);
993 * Try to get a partial slab from a specific node
995 static struct page *get_partial_node(struct kmem_cache_node *n)
1000 * Racy check. If we mistakenly see no partial slabs then we
1001 * just allocate an empty slab. If we mistakenly try to get a
1002 * partial slab then get_partials() will return NULL.
1004 if (!n || !n->nr_partial)
1007 spin_lock(&n->list_lock);
1008 list_for_each_entry(page, &n->partial, lru)
1009 if (lock_and_del_slab(n, page))
1013 spin_unlock(&n->list_lock);
1018 * Get a page from somewhere. Search in increasing NUMA
1021 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1024 struct zonelist *zonelist;
1029 * The defrag ratio allows to configure the tradeoffs between
1030 * inter node defragmentation and node local allocations.
1031 * A lower defrag_ratio increases the tendency to do local
1032 * allocations instead of scanning throught the partial
1033 * lists on other nodes.
1035 * If defrag_ratio is set to 0 then kmalloc() always
1036 * returns node local objects. If its higher then kmalloc()
1037 * may return off node objects in order to avoid fragmentation.
1039 * A higher ratio means slabs may be taken from other nodes
1040 * thus reducing the number of partial slabs on those nodes.
1042 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1043 * defrag_ratio = 1000) then every (well almost) allocation
1044 * will first attempt to defrag slab caches on other nodes. This
1045 * means scanning over all nodes to look for partial slabs which
1046 * may be a bit expensive to do on every slab allocation.
1048 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1051 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1052 ->node_zonelists[gfp_zone(flags)];
1053 for (z = zonelist->zones; *z; z++) {
1054 struct kmem_cache_node *n;
1056 n = get_node(s, zone_to_nid(*z));
1058 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1059 n->nr_partial > 2) {
1060 page = get_partial_node(n);
1070 * Get a partial page, lock it and return it.
1072 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1075 int searchnode = (node == -1) ? numa_node_id() : node;
1077 page = get_partial_node(get_node(s, searchnode));
1078 if (page || (flags & __GFP_THISNODE))
1081 return get_any_partial(s, flags);
1085 * Move a page back to the lists.
1087 * Must be called with the slab lock held.
1089 * On exit the slab lock will have been dropped.
1091 static void putback_slab(struct kmem_cache *s, struct page *page)
1095 add_partial(s, page);
1099 discard_slab(s, page);
1104 * Remove the cpu slab
1106 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1108 s->cpu_slab[cpu] = NULL;
1109 ClearPageActive(page);
1111 putback_slab(s, page);
1114 static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1117 deactivate_slab(s, page, cpu);
1122 * Called from IPI handler with interrupts disabled.
1124 static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1126 struct page *page = s->cpu_slab[cpu];
1129 flush_slab(s, page, cpu);
1132 static void flush_cpu_slab(void *d)
1134 struct kmem_cache *s = d;
1135 int cpu = smp_processor_id();
1137 __flush_cpu_slab(s, cpu);
1140 static void flush_all(struct kmem_cache *s)
1143 on_each_cpu(flush_cpu_slab, s, 1, 1);
1145 unsigned long flags;
1147 local_irq_save(flags);
1149 local_irq_restore(flags);
1154 * slab_alloc is optimized to only modify two cachelines on the fast path
1155 * (aside from the stack):
1157 * 1. The page struct
1158 * 2. The first cacheline of the object to be allocated.
1160 * The only cache lines that are read (apart from code) is the
1161 * per cpu array in the kmem_cache struct.
1163 * Fastpath is not possible if we need to get a new slab or have
1164 * debugging enabled (which means all slabs are marked with PageError)
1166 static __always_inline void *slab_alloc(struct kmem_cache *s,
1167 gfp_t gfpflags, int node)
1171 unsigned long flags;
1174 local_irq_save(flags);
1175 cpu = smp_processor_id();
1176 page = s->cpu_slab[cpu];
1181 if (unlikely(node != -1 && page_to_nid(page) != node))
1184 object = page->freelist;
1185 if (unlikely(!object))
1187 if (unlikely(PageError(page)))
1192 page->freelist = object[page->offset];
1194 local_irq_restore(flags);
1198 deactivate_slab(s, page, cpu);
1201 page = get_partial(s, gfpflags, node);
1204 s->cpu_slab[cpu] = page;
1205 SetPageActive(page);
1209 page = new_slab(s, gfpflags, node);
1211 cpu = smp_processor_id();
1212 if (s->cpu_slab[cpu]) {
1214 * Someone else populated the cpu_slab while we enabled
1215 * interrupts, or we have got scheduled on another cpu.
1216 * The page may not be on the requested node.
1219 page_to_nid(s->cpu_slab[cpu]) == node) {
1221 * Current cpuslab is acceptable and we
1222 * want the current one since its cache hot
1224 discard_slab(s, page);
1225 page = s->cpu_slab[cpu];
1229 /* Dump the current slab */
1230 flush_slab(s, s->cpu_slab[cpu], cpu);
1235 local_irq_restore(flags);
1238 if (!alloc_object_checks(s, page, object))
1240 if (s->flags & SLAB_STORE_USER)
1241 set_tracking(s, object, TRACK_ALLOC);
1245 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1247 return slab_alloc(s, gfpflags, -1);
1249 EXPORT_SYMBOL(kmem_cache_alloc);
1252 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1254 return slab_alloc(s, gfpflags, node);
1256 EXPORT_SYMBOL(kmem_cache_alloc_node);
1260 * The fastpath only writes the cacheline of the page struct and the first
1261 * cacheline of the object.
1263 * No special cachelines need to be read
1265 static void slab_free(struct kmem_cache *s, struct page *page, void *x)
1268 void **object = (void *)x;
1269 unsigned long flags;
1271 local_irq_save(flags);
1274 if (unlikely(PageError(page)))
1277 prior = object[page->offset] = page->freelist;
1278 page->freelist = object;
1281 if (unlikely(PageActive(page)))
1283 * Cpu slabs are never on partial lists and are
1288 if (unlikely(!page->inuse))
1292 * Objects left in the slab. If it
1293 * was not on the partial list before
1296 if (unlikely(!prior))
1297 add_partial(s, page);
1301 local_irq_restore(flags);
1307 * Partially used slab that is on the partial list.
1309 remove_partial(s, page);
1312 discard_slab(s, page);
1313 local_irq_restore(flags);
1317 if (free_object_checks(s, page, x))
1322 void kmem_cache_free(struct kmem_cache *s, void *x)
1326 page = virt_to_page(x);
1328 if (unlikely(PageCompound(page)))
1329 page = page->first_page;
1332 if (unlikely(PageError(page) && (s->flags & SLAB_STORE_USER)))
1333 set_tracking(s, x, TRACK_FREE);
1334 slab_free(s, page, x);
1336 EXPORT_SYMBOL(kmem_cache_free);
1338 /* Figure out on which slab object the object resides */
1339 static struct page *get_object_page(const void *x)
1341 struct page *page = virt_to_page(x);
1343 if (unlikely(PageCompound(page)))
1344 page = page->first_page;
1346 if (!PageSlab(page))
1353 * kmem_cache_open produces objects aligned at "size" and the first object
1354 * is placed at offset 0 in the slab (We have no metainformation on the
1355 * slab, all slabs are in essence "off slab").
1357 * In order to get the desired alignment one just needs to align the
1360 * Notice that the allocation order determines the sizes of the per cpu
1361 * caches. Each processor has always one slab available for allocations.
1362 * Increasing the allocation order reduces the number of times that slabs
1363 * must be moved on and off the partial lists and therefore may influence
1366 * The offset is used to relocate the free list link in each object. It is
1367 * therefore possible to move the free list link behind the object. This
1368 * is necessary for RCU to work properly and also useful for debugging.
1372 * Mininum / Maximum order of slab pages. This influences locking overhead
1373 * and slab fragmentation. A higher order reduces the number of partial slabs
1374 * and increases the number of allocations possible without having to
1375 * take the list_lock.
1377 static int slub_min_order;
1378 static int slub_max_order = DEFAULT_MAX_ORDER;
1381 * Minimum number of objects per slab. This is necessary in order to
1382 * reduce locking overhead. Similar to the queue size in SLAB.
1384 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1387 * Merge control. If this is set then no merging of slab caches will occur.
1389 static int slub_nomerge;
1394 static int slub_debug;
1396 static char *slub_debug_slabs;
1399 * Calculate the order of allocation given an slab object size.
1401 * The order of allocation has significant impact on other elements
1402 * of the system. Generally order 0 allocations should be preferred
1403 * since they do not cause fragmentation in the page allocator. Larger
1404 * objects may have problems with order 0 because there may be too much
1405 * space left unused in a slab. We go to a higher order if more than 1/8th
1406 * of the slab would be wasted.
1408 * In order to reach satisfactory performance we must ensure that
1409 * a minimum number of objects is in one slab. Otherwise we may
1410 * generate too much activity on the partial lists. This is less a
1411 * concern for large slabs though. slub_max_order specifies the order
1412 * where we begin to stop considering the number of objects in a slab.
1414 * Higher order allocations also allow the placement of more objects
1415 * in a slab and thereby reduce object handling overhead. If the user
1416 * has requested a higher mininum order then we start with that one
1419 static int calculate_order(int size)
1424 for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT);
1425 order < MAX_ORDER; order++) {
1426 unsigned long slab_size = PAGE_SIZE << order;
1428 if (slub_max_order > order &&
1429 slab_size < slub_min_objects * size)
1432 if (slab_size < size)
1435 rem = slab_size % size;
1437 if (rem <= (PAGE_SIZE << order) / 8)
1441 if (order >= MAX_ORDER)
1447 * Function to figure out which alignment to use from the
1448 * various ways of specifying it.
1450 static unsigned long calculate_alignment(unsigned long flags,
1451 unsigned long align, unsigned long size)
1454 * If the user wants hardware cache aligned objects then
1455 * follow that suggestion if the object is sufficiently
1458 * The hardware cache alignment cannot override the
1459 * specified alignment though. If that is greater
1462 if ((flags & (SLAB_MUST_HWCACHE_ALIGN | SLAB_HWCACHE_ALIGN)) &&
1463 size > L1_CACHE_BYTES / 2)
1464 return max_t(unsigned long, align, L1_CACHE_BYTES);
1466 if (align < ARCH_SLAB_MINALIGN)
1467 return ARCH_SLAB_MINALIGN;
1469 return ALIGN(align, sizeof(void *));
1472 static void init_kmem_cache_node(struct kmem_cache_node *n)
1475 atomic_long_set(&n->nr_slabs, 0);
1476 spin_lock_init(&n->list_lock);
1477 INIT_LIST_HEAD(&n->partial);
1482 * No kmalloc_node yet so do it by hand. We know that this is the first
1483 * slab on the node for this slabcache. There are no concurrent accesses
1486 * Note that this function only works on the kmalloc_node_cache
1487 * when allocating for the kmalloc_node_cache.
1489 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1493 struct kmem_cache_node *n;
1495 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1497 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1498 /* new_slab() disables interupts */
1504 page->freelist = get_freepointer(kmalloc_caches, n);
1506 kmalloc_caches->node[node] = n;
1507 init_object(kmalloc_caches, n, 1);
1508 init_kmem_cache_node(n);
1509 atomic_long_inc(&n->nr_slabs);
1510 add_partial(kmalloc_caches, page);
1514 static void free_kmem_cache_nodes(struct kmem_cache *s)
1518 for_each_online_node(node) {
1519 struct kmem_cache_node *n = s->node[node];
1520 if (n && n != &s->local_node)
1521 kmem_cache_free(kmalloc_caches, n);
1522 s->node[node] = NULL;
1526 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1531 if (slab_state >= UP)
1532 local_node = page_to_nid(virt_to_page(s));
1536 for_each_online_node(node) {
1537 struct kmem_cache_node *n;
1539 if (local_node == node)
1542 if (slab_state == DOWN) {
1543 n = early_kmem_cache_node_alloc(gfpflags,
1547 n = kmem_cache_alloc_node(kmalloc_caches,
1551 free_kmem_cache_nodes(s);
1557 init_kmem_cache_node(n);
1562 static void free_kmem_cache_nodes(struct kmem_cache *s)
1566 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1568 init_kmem_cache_node(&s->local_node);
1574 * calculate_sizes() determines the order and the distribution of data within
1577 static int calculate_sizes(struct kmem_cache *s)
1579 unsigned long flags = s->flags;
1580 unsigned long size = s->objsize;
1581 unsigned long align = s->align;
1584 * Determine if we can poison the object itself. If the user of
1585 * the slab may touch the object after free or before allocation
1586 * then we should never poison the object itself.
1588 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1589 !s->ctor && !s->dtor)
1590 s->flags |= __OBJECT_POISON;
1592 s->flags &= ~__OBJECT_POISON;
1595 * Round up object size to the next word boundary. We can only
1596 * place the free pointer at word boundaries and this determines
1597 * the possible location of the free pointer.
1599 size = ALIGN(size, sizeof(void *));
1602 * If we are redzoning then check if there is some space between the
1603 * end of the object and the free pointer. If not then add an
1604 * additional word, so that we can establish a redzone between
1605 * the object and the freepointer to be able to check for overwrites.
1607 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1608 size += sizeof(void *);
1611 * With that we have determined how much of the slab is in actual
1612 * use by the object. This is the potential offset to the free
1617 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1618 s->ctor || s->dtor)) {
1620 * Relocate free pointer after the object if it is not
1621 * permitted to overwrite the first word of the object on
1624 * This is the case if we do RCU, have a constructor or
1625 * destructor or are poisoning the objects.
1628 size += sizeof(void *);
1631 if (flags & SLAB_STORE_USER)
1633 * Need to store information about allocs and frees after
1636 size += 2 * sizeof(struct track);
1638 if (flags & DEBUG_DEFAULT_FLAGS)
1640 * Add some empty padding so that we can catch
1641 * overwrites from earlier objects rather than let
1642 * tracking information or the free pointer be
1643 * corrupted if an user writes before the start
1646 size += sizeof(void *);
1648 * Determine the alignment based on various parameters that the
1649 * user specified (this is unecessarily complex due to the attempt
1650 * to be compatible with SLAB. Should be cleaned up some day).
1652 align = calculate_alignment(flags, align, s->objsize);
1655 * SLUB stores one object immediately after another beginning from
1656 * offset 0. In order to align the objects we have to simply size
1657 * each object to conform to the alignment.
1659 size = ALIGN(size, align);
1662 s->order = calculate_order(size);
1667 * Determine the number of objects per slab
1669 s->objects = (PAGE_SIZE << s->order) / size;
1672 * Verify that the number of objects is within permitted limits.
1673 * The page->inuse field is only 16 bit wide! So we cannot have
1674 * more than 64k objects per slab.
1676 if (!s->objects || s->objects > 65535)
1682 static int __init finish_bootstrap(void)
1684 struct list_head *h;
1689 list_for_each(h, &slab_caches) {
1690 struct kmem_cache *s =
1691 container_of(h, struct kmem_cache, list);
1693 err = sysfs_slab_add(s);
1699 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1700 const char *name, size_t size,
1701 size_t align, unsigned long flags,
1702 void (*ctor)(void *, struct kmem_cache *, unsigned long),
1703 void (*dtor)(void *, struct kmem_cache *, unsigned long))
1705 memset(s, 0, kmem_size);
1713 BUG_ON(flags & SLUB_UNIMPLEMENTED);
1716 * The page->offset field is only 16 bit wide. This is an offset
1717 * in units of words from the beginning of an object. If the slab
1718 * size is bigger then we cannot move the free pointer behind the
1721 * On 32 bit platforms the limit is 256k. On 64bit platforms
1722 * the limit is 512k.
1724 * Debugging or ctor/dtors may create a need to move the free
1725 * pointer. Fail if this happens.
1727 if (s->size >= 65535 * sizeof(void *)) {
1728 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1729 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1730 BUG_ON(ctor || dtor);
1734 * Enable debugging if selected on the kernel commandline.
1736 if (slub_debug && (!slub_debug_slabs ||
1737 strncmp(slub_debug_slabs, name,
1738 strlen(slub_debug_slabs)) == 0))
1739 s->flags |= slub_debug;
1741 if (!calculate_sizes(s))
1746 s->defrag_ratio = 100;
1749 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
1752 if (flags & SLAB_PANIC)
1753 panic("Cannot create slab %s size=%lu realsize=%u "
1754 "order=%u offset=%u flags=%lx\n",
1755 s->name, (unsigned long)size, s->size, s->order,
1759 EXPORT_SYMBOL(kmem_cache_open);
1762 * Check if a given pointer is valid
1764 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
1769 page = get_object_page(object);
1771 if (!page || s != page->slab)
1772 /* No slab or wrong slab */
1775 addr = page_address(page);
1776 if (object < addr || object >= addr + s->objects * s->size)
1780 if ((object - addr) % s->size)
1781 /* Improperly aligned */
1785 * We could also check if the object is on the slabs freelist.
1786 * But this would be too expensive and it seems that the main
1787 * purpose of kmem_ptr_valid is to check if the object belongs
1788 * to a certain slab.
1792 EXPORT_SYMBOL(kmem_ptr_validate);
1795 * Determine the size of a slab object
1797 unsigned int kmem_cache_size(struct kmem_cache *s)
1801 EXPORT_SYMBOL(kmem_cache_size);
1803 const char *kmem_cache_name(struct kmem_cache *s)
1807 EXPORT_SYMBOL(kmem_cache_name);
1810 * Attempt to free all slabs on a node
1812 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
1813 struct list_head *list)
1815 int slabs_inuse = 0;
1816 unsigned long flags;
1817 struct page *page, *h;
1819 spin_lock_irqsave(&n->list_lock, flags);
1820 list_for_each_entry_safe(page, h, list, lru)
1822 list_del(&page->lru);
1823 discard_slab(s, page);
1826 spin_unlock_irqrestore(&n->list_lock, flags);
1831 * Release all resources used by slab cache
1833 static int kmem_cache_close(struct kmem_cache *s)
1839 /* Attempt to free all objects */
1840 for_each_online_node(node) {
1841 struct kmem_cache_node *n = get_node(s, node);
1843 free_list(s, n, &n->partial);
1844 if (atomic_long_read(&n->nr_slabs))
1847 free_kmem_cache_nodes(s);
1852 * Close a cache and release the kmem_cache structure
1853 * (must be used for caches created using kmem_cache_create)
1855 void kmem_cache_destroy(struct kmem_cache *s)
1857 down_write(&slub_lock);
1861 if (kmem_cache_close(s))
1863 sysfs_slab_remove(s);
1866 up_write(&slub_lock);
1868 EXPORT_SYMBOL(kmem_cache_destroy);
1870 /********************************************************************
1872 *******************************************************************/
1874 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
1875 EXPORT_SYMBOL(kmalloc_caches);
1877 #ifdef CONFIG_ZONE_DMA
1878 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
1881 static int __init setup_slub_min_order(char *str)
1883 get_option (&str, &slub_min_order);
1888 __setup("slub_min_order=", setup_slub_min_order);
1890 static int __init setup_slub_max_order(char *str)
1892 get_option (&str, &slub_max_order);
1897 __setup("slub_max_order=", setup_slub_max_order);
1899 static int __init setup_slub_min_objects(char *str)
1901 get_option (&str, &slub_min_objects);
1906 __setup("slub_min_objects=", setup_slub_min_objects);
1908 static int __init setup_slub_nomerge(char *str)
1914 __setup("slub_nomerge", setup_slub_nomerge);
1916 static int __init setup_slub_debug(char *str)
1918 if (!str || *str != '=')
1919 slub_debug = DEBUG_DEFAULT_FLAGS;
1922 if (*str == 0 || *str == ',')
1923 slub_debug = DEBUG_DEFAULT_FLAGS;
1925 for( ;*str && *str != ','; str++)
1927 case 'f' : case 'F' :
1928 slub_debug |= SLAB_DEBUG_FREE;
1930 case 'z' : case 'Z' :
1931 slub_debug |= SLAB_RED_ZONE;
1933 case 'p' : case 'P' :
1934 slub_debug |= SLAB_POISON;
1936 case 'u' : case 'U' :
1937 slub_debug |= SLAB_STORE_USER;
1939 case 't' : case 'T' :
1940 slub_debug |= SLAB_TRACE;
1943 printk(KERN_ERR "slub_debug option '%c' "
1944 "unknown. skipped\n",*str);
1949 slub_debug_slabs = str + 1;
1953 __setup("slub_debug", setup_slub_debug);
1955 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
1956 const char *name, int size, gfp_t gfp_flags)
1958 unsigned int flags = 0;
1960 if (gfp_flags & SLUB_DMA)
1961 flags = SLAB_CACHE_DMA;
1963 down_write(&slub_lock);
1964 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
1968 list_add(&s->list, &slab_caches);
1969 up_write(&slub_lock);
1970 if (sysfs_slab_add(s))
1975 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
1978 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
1980 int index = kmalloc_index(size);
1985 /* Allocation too large? */
1988 #ifdef CONFIG_ZONE_DMA
1989 if ((flags & SLUB_DMA)) {
1990 struct kmem_cache *s;
1991 struct kmem_cache *x;
1995 s = kmalloc_caches_dma[index];
1999 /* Dynamically create dma cache */
2000 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2002 panic("Unable to allocate memory for dma cache\n");
2004 if (index <= KMALLOC_SHIFT_HIGH)
2005 realsize = 1 << index;
2013 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2014 (unsigned int)realsize);
2015 s = create_kmalloc_cache(x, text, realsize, flags);
2016 kmalloc_caches_dma[index] = s;
2020 return &kmalloc_caches[index];
2023 void *__kmalloc(size_t size, gfp_t flags)
2025 struct kmem_cache *s = get_slab(size, flags);
2028 return kmem_cache_alloc(s, flags);
2031 EXPORT_SYMBOL(__kmalloc);
2034 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2036 struct kmem_cache *s = get_slab(size, flags);
2039 return kmem_cache_alloc_node(s, flags, node);
2042 EXPORT_SYMBOL(__kmalloc_node);
2045 size_t ksize(const void *object)
2047 struct page *page = get_object_page(object);
2048 struct kmem_cache *s;
2055 * Debugging requires use of the padding between object
2056 * and whatever may come after it.
2058 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2062 * If we have the need to store the freelist pointer
2063 * back there or track user information then we can
2064 * only use the space before that information.
2066 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2070 * Else we can use all the padding etc for the allocation
2074 EXPORT_SYMBOL(ksize);
2076 void kfree(const void *x)
2078 struct kmem_cache *s;
2084 page = virt_to_page(x);
2086 if (unlikely(PageCompound(page)))
2087 page = page->first_page;
2091 if (unlikely(PageError(page) && (s->flags & SLAB_STORE_USER)))
2092 set_tracking(s, (void *)x, TRACK_FREE);
2093 slab_free(s, page, (void *)x);
2095 EXPORT_SYMBOL(kfree);
2098 * krealloc - reallocate memory. The contents will remain unchanged.
2100 * @p: object to reallocate memory for.
2101 * @new_size: how many bytes of memory are required.
2102 * @flags: the type of memory to allocate.
2104 * The contents of the object pointed to are preserved up to the
2105 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2106 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2107 * %NULL pointer, the object pointed to is freed.
2109 void *krealloc(const void *p, size_t new_size, gfp_t flags)
2111 struct kmem_cache *new_cache;
2116 return kmalloc(new_size, flags);
2118 if (unlikely(!new_size)) {
2123 page = virt_to_page(p);
2125 if (unlikely(PageCompound(page)))
2126 page = page->first_page;
2128 new_cache = get_slab(new_size, flags);
2131 * If new size fits in the current cache, bail out.
2133 if (likely(page->slab == new_cache))
2136 ret = kmalloc(new_size, flags);
2138 memcpy(ret, p, min(new_size, ksize(p)));
2143 EXPORT_SYMBOL(krealloc);
2145 /********************************************************************
2146 * Basic setup of slabs
2147 *******************************************************************/
2149 void __init kmem_cache_init(void)
2155 * Must first have the slab cache available for the allocations of the
2156 * struct kmalloc_cache_node's. There is special bootstrap code in
2157 * kmem_cache_open for slab_state == DOWN.
2159 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2160 sizeof(struct kmem_cache_node), GFP_KERNEL);
2163 /* Able to allocate the per node structures */
2164 slab_state = PARTIAL;
2166 /* Caches that are not of the two-to-the-power-of size */
2167 create_kmalloc_cache(&kmalloc_caches[1],
2168 "kmalloc-96", 96, GFP_KERNEL);
2169 create_kmalloc_cache(&kmalloc_caches[2],
2170 "kmalloc-192", 192, GFP_KERNEL);
2172 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2173 create_kmalloc_cache(&kmalloc_caches[i],
2174 "kmalloc", 1 << i, GFP_KERNEL);
2178 /* Provide the correct kmalloc names now that the caches are up */
2179 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2180 kmalloc_caches[i]. name =
2181 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2184 register_cpu_notifier(&slab_notifier);
2187 if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */
2188 kmem_size = offsetof(struct kmem_cache, cpu_slab)
2189 + nr_cpu_ids * sizeof(struct page *);
2191 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2192 " Processors=%d, Nodes=%d\n",
2193 KMALLOC_SHIFT_HIGH, L1_CACHE_BYTES,
2194 slub_min_order, slub_max_order, slub_min_objects,
2195 nr_cpu_ids, nr_node_ids);
2199 * Find a mergeable slab cache
2201 static int slab_unmergeable(struct kmem_cache *s)
2203 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2206 if (s->ctor || s->dtor)
2212 static struct kmem_cache *find_mergeable(size_t size,
2213 size_t align, unsigned long flags,
2214 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2215 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2217 struct list_head *h;
2219 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2225 size = ALIGN(size, sizeof(void *));
2226 align = calculate_alignment(flags, align, size);
2227 size = ALIGN(size, align);
2229 list_for_each(h, &slab_caches) {
2230 struct kmem_cache *s =
2231 container_of(h, struct kmem_cache, list);
2233 if (slab_unmergeable(s))
2239 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2240 (s->flags & SLUB_MERGE_SAME))
2243 * Check if alignment is compatible.
2244 * Courtesy of Adrian Drzewiecki
2246 if ((s->size & ~(align -1)) != s->size)
2249 if (s->size - size >= sizeof(void *))
2257 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2258 size_t align, unsigned long flags,
2259 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2260 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2262 struct kmem_cache *s;
2264 down_write(&slub_lock);
2265 s = find_mergeable(size, align, flags, dtor, ctor);
2269 * Adjust the object sizes so that we clear
2270 * the complete object on kzalloc.
2272 s->objsize = max(s->objsize, (int)size);
2273 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2274 if (sysfs_slab_alias(s, name))
2277 s = kmalloc(kmem_size, GFP_KERNEL);
2278 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2279 size, align, flags, ctor, dtor)) {
2280 if (sysfs_slab_add(s)) {
2284 list_add(&s->list, &slab_caches);
2288 up_write(&slub_lock);
2292 up_write(&slub_lock);
2293 if (flags & SLAB_PANIC)
2294 panic("Cannot create slabcache %s\n", name);
2299 EXPORT_SYMBOL(kmem_cache_create);
2301 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2305 x = kmem_cache_alloc(s, flags);
2307 memset(x, 0, s->objsize);
2310 EXPORT_SYMBOL(kmem_cache_zalloc);
2313 static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2315 struct list_head *h;
2317 down_read(&slub_lock);
2318 list_for_each(h, &slab_caches) {
2319 struct kmem_cache *s =
2320 container_of(h, struct kmem_cache, list);
2324 up_read(&slub_lock);
2328 * Use the cpu notifier to insure that the slab are flushed
2331 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2332 unsigned long action, void *hcpu)
2334 long cpu = (long)hcpu;
2337 case CPU_UP_CANCELED:
2339 for_all_slabs(__flush_cpu_slab, cpu);
2347 static struct notifier_block __cpuinitdata slab_notifier =
2348 { &slab_cpuup_callback, NULL, 0 };
2352 /***************************************************************
2353 * Compatiblility definitions
2354 **************************************************************/
2356 int kmem_cache_shrink(struct kmem_cache *s)
2361 EXPORT_SYMBOL(kmem_cache_shrink);
2365 /*****************************************************************
2366 * Generic reaper used to support the page allocator
2367 * (the cpu slabs are reaped by a per slab workqueue).
2369 * Maybe move this to the page allocator?
2370 ****************************************************************/
2372 static DEFINE_PER_CPU(unsigned long, reap_node);
2374 static void init_reap_node(int cpu)
2378 node = next_node(cpu_to_node(cpu), node_online_map);
2379 if (node == MAX_NUMNODES)
2380 node = first_node(node_online_map);
2382 __get_cpu_var(reap_node) = node;
2385 static void next_reap_node(void)
2387 int node = __get_cpu_var(reap_node);
2390 * Also drain per cpu pages on remote zones
2392 if (node != numa_node_id())
2393 drain_node_pages(node);
2395 node = next_node(node, node_online_map);
2396 if (unlikely(node >= MAX_NUMNODES))
2397 node = first_node(node_online_map);
2398 __get_cpu_var(reap_node) = node;
2401 #define init_reap_node(cpu) do { } while (0)
2402 #define next_reap_node(void) do { } while (0)
2405 #define REAPTIMEOUT_CPUC (2*HZ)
2408 static DEFINE_PER_CPU(struct delayed_work, reap_work);
2410 static void cache_reap(struct work_struct *unused)
2413 refresh_cpu_vm_stats(smp_processor_id());
2414 schedule_delayed_work(&__get_cpu_var(reap_work),
2418 static void __devinit start_cpu_timer(int cpu)
2420 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
2423 * When this gets called from do_initcalls via cpucache_init(),
2424 * init_workqueues() has already run, so keventd will be setup
2427 if (keventd_up() && reap_work->work.func == NULL) {
2428 init_reap_node(cpu);
2429 INIT_DELAYED_WORK(reap_work, cache_reap);
2430 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
2434 static int __init cpucache_init(void)
2439 * Register the timers that drain pcp pages and update vm statistics
2441 for_each_online_cpu(cpu)
2442 start_cpu_timer(cpu);
2445 __initcall(cpucache_init);
2448 #ifdef SLUB_RESILIENCY_TEST
2449 static unsigned long validate_slab_cache(struct kmem_cache *s);
2451 static void resiliency_test(void)
2455 printk(KERN_ERR "SLUB resiliency testing\n");
2456 printk(KERN_ERR "-----------------------\n");
2457 printk(KERN_ERR "A. Corruption after allocation\n");
2459 p = kzalloc(16, GFP_KERNEL);
2461 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2462 " 0x12->0x%p\n\n", p + 16);
2464 validate_slab_cache(kmalloc_caches + 4);
2466 /* Hmmm... The next two are dangerous */
2467 p = kzalloc(32, GFP_KERNEL);
2468 p[32 + sizeof(void *)] = 0x34;
2469 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2470 " 0x34 -> -0x%p\n", p);
2471 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2473 validate_slab_cache(kmalloc_caches + 5);
2474 p = kzalloc(64, GFP_KERNEL);
2475 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2477 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2479 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2480 validate_slab_cache(kmalloc_caches + 6);
2482 printk(KERN_ERR "\nB. Corruption after free\n");
2483 p = kzalloc(128, GFP_KERNEL);
2486 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2487 validate_slab_cache(kmalloc_caches + 7);
2489 p = kzalloc(256, GFP_KERNEL);
2492 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2493 validate_slab_cache(kmalloc_caches + 8);
2495 p = kzalloc(512, GFP_KERNEL);
2498 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2499 validate_slab_cache(kmalloc_caches + 9);
2502 static void resiliency_test(void) {};
2506 * These are not as efficient as kmalloc for the non debug case.
2507 * We do not have the page struct available so we have to touch one
2508 * cacheline in struct kmem_cache to check slab flags.
2510 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2512 struct kmem_cache *s = get_slab(size, gfpflags);
2518 object = kmem_cache_alloc(s, gfpflags);
2520 if (object && (s->flags & SLAB_STORE_USER))
2521 set_track(s, object, TRACK_ALLOC, caller);
2526 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2527 int node, void *caller)
2529 struct kmem_cache *s = get_slab(size, gfpflags);
2535 object = kmem_cache_alloc_node(s, gfpflags, node);
2537 if (object && (s->flags & SLAB_STORE_USER))
2538 set_track(s, object, TRACK_ALLOC, caller);
2545 static unsigned long count_partial(struct kmem_cache_node *n)
2547 unsigned long flags;
2548 unsigned long x = 0;
2551 spin_lock_irqsave(&n->list_lock, flags);
2552 list_for_each_entry(page, &n->partial, lru)
2554 spin_unlock_irqrestore(&n->list_lock, flags);
2558 enum slab_stat_type {
2565 #define SO_FULL (1 << SL_FULL)
2566 #define SO_PARTIAL (1 << SL_PARTIAL)
2567 #define SO_CPU (1 << SL_CPU)
2568 #define SO_OBJECTS (1 << SL_OBJECTS)
2570 static unsigned long slab_objects(struct kmem_cache *s,
2571 char *buf, unsigned long flags)
2573 unsigned long total = 0;
2577 unsigned long *nodes;
2578 unsigned long *per_cpu;
2580 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
2581 per_cpu = nodes + nr_node_ids;
2583 for_each_possible_cpu(cpu) {
2584 struct page *page = s->cpu_slab[cpu];
2588 node = page_to_nid(page);
2589 if (flags & SO_CPU) {
2592 if (flags & SO_OBJECTS)
2603 for_each_online_node(node) {
2604 struct kmem_cache_node *n = get_node(s, node);
2606 if (flags & SO_PARTIAL) {
2607 if (flags & SO_OBJECTS)
2608 x = count_partial(n);
2615 if (flags & SO_FULL) {
2616 int full_slabs = atomic_read(&n->nr_slabs)
2620 if (flags & SO_OBJECTS)
2621 x = full_slabs * s->objects;
2629 x = sprintf(buf, "%lu", total);
2631 for_each_online_node(node)
2633 x += sprintf(buf + x, " N%d=%lu",
2637 return x + sprintf(buf + x, "\n");
2640 static int any_slab_objects(struct kmem_cache *s)
2645 for_each_possible_cpu(cpu)
2646 if (s->cpu_slab[cpu])
2649 for_each_node(node) {
2650 struct kmem_cache_node *n = get_node(s, node);
2652 if (n->nr_partial || atomic_read(&n->nr_slabs))
2658 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2659 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
2661 struct slab_attribute {
2662 struct attribute attr;
2663 ssize_t (*show)(struct kmem_cache *s, char *buf);
2664 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
2667 #define SLAB_ATTR_RO(_name) \
2668 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
2670 #define SLAB_ATTR(_name) \
2671 static struct slab_attribute _name##_attr = \
2672 __ATTR(_name, 0644, _name##_show, _name##_store)
2675 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
2677 return sprintf(buf, "%d\n", s->size);
2679 SLAB_ATTR_RO(slab_size);
2681 static ssize_t align_show(struct kmem_cache *s, char *buf)
2683 return sprintf(buf, "%d\n", s->align);
2685 SLAB_ATTR_RO(align);
2687 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
2689 return sprintf(buf, "%d\n", s->objsize);
2691 SLAB_ATTR_RO(object_size);
2693 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
2695 return sprintf(buf, "%d\n", s->objects);
2697 SLAB_ATTR_RO(objs_per_slab);
2699 static ssize_t order_show(struct kmem_cache *s, char *buf)
2701 return sprintf(buf, "%d\n", s->order);
2703 SLAB_ATTR_RO(order);
2705 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
2708 int n = sprint_symbol(buf, (unsigned long)s->ctor);
2710 return n + sprintf(buf + n, "\n");
2716 static ssize_t dtor_show(struct kmem_cache *s, char *buf)
2719 int n = sprint_symbol(buf, (unsigned long)s->dtor);
2721 return n + sprintf(buf + n, "\n");
2727 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
2729 return sprintf(buf, "%d\n", s->refcount - 1);
2731 SLAB_ATTR_RO(aliases);
2733 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
2735 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
2737 SLAB_ATTR_RO(slabs);
2739 static ssize_t partial_show(struct kmem_cache *s, char *buf)
2741 return slab_objects(s, buf, SO_PARTIAL);
2743 SLAB_ATTR_RO(partial);
2745 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
2747 return slab_objects(s, buf, SO_CPU);
2749 SLAB_ATTR_RO(cpu_slabs);
2751 static ssize_t objects_show(struct kmem_cache *s, char *buf)
2753 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
2755 SLAB_ATTR_RO(objects);
2757 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
2759 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
2762 static ssize_t sanity_checks_store(struct kmem_cache *s,
2763 const char *buf, size_t length)
2765 s->flags &= ~SLAB_DEBUG_FREE;
2767 s->flags |= SLAB_DEBUG_FREE;
2770 SLAB_ATTR(sanity_checks);
2772 static ssize_t trace_show(struct kmem_cache *s, char *buf)
2774 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
2777 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
2780 s->flags &= ~SLAB_TRACE;
2782 s->flags |= SLAB_TRACE;
2787 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
2789 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
2792 static ssize_t reclaim_account_store(struct kmem_cache *s,
2793 const char *buf, size_t length)
2795 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
2797 s->flags |= SLAB_RECLAIM_ACCOUNT;
2800 SLAB_ATTR(reclaim_account);
2802 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
2804 return sprintf(buf, "%d\n", !!(s->flags &
2805 (SLAB_HWCACHE_ALIGN|SLAB_MUST_HWCACHE_ALIGN)));
2807 SLAB_ATTR_RO(hwcache_align);
2809 #ifdef CONFIG_ZONE_DMA
2810 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
2812 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
2814 SLAB_ATTR_RO(cache_dma);
2817 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
2819 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
2821 SLAB_ATTR_RO(destroy_by_rcu);
2823 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
2825 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
2828 static ssize_t red_zone_store(struct kmem_cache *s,
2829 const char *buf, size_t length)
2831 if (any_slab_objects(s))
2834 s->flags &= ~SLAB_RED_ZONE;
2836 s->flags |= SLAB_RED_ZONE;
2840 SLAB_ATTR(red_zone);
2842 static ssize_t poison_show(struct kmem_cache *s, char *buf)
2844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
2847 static ssize_t poison_store(struct kmem_cache *s,
2848 const char *buf, size_t length)
2850 if (any_slab_objects(s))
2853 s->flags &= ~SLAB_POISON;
2855 s->flags |= SLAB_POISON;
2861 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
2863 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
2866 static ssize_t store_user_store(struct kmem_cache *s,
2867 const char *buf, size_t length)
2869 if (any_slab_objects(s))
2872 s->flags &= ~SLAB_STORE_USER;
2874 s->flags |= SLAB_STORE_USER;
2878 SLAB_ATTR(store_user);
2881 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
2883 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
2886 static ssize_t defrag_ratio_store(struct kmem_cache *s,
2887 const char *buf, size_t length)
2889 int n = simple_strtoul(buf, NULL, 10);
2892 s->defrag_ratio = n * 10;
2895 SLAB_ATTR(defrag_ratio);
2898 static struct attribute * slab_attrs[] = {
2899 &slab_size_attr.attr,
2900 &object_size_attr.attr,
2901 &objs_per_slab_attr.attr,
2906 &cpu_slabs_attr.attr,
2911 &sanity_checks_attr.attr,
2913 &hwcache_align_attr.attr,
2914 &reclaim_account_attr.attr,
2915 &destroy_by_rcu_attr.attr,
2916 &red_zone_attr.attr,
2918 &store_user_attr.attr,
2919 #ifdef CONFIG_ZONE_DMA
2920 &cache_dma_attr.attr,
2923 &defrag_ratio_attr.attr,
2928 static struct attribute_group slab_attr_group = {
2929 .attrs = slab_attrs,
2932 static ssize_t slab_attr_show(struct kobject *kobj,
2933 struct attribute *attr,
2936 struct slab_attribute *attribute;
2937 struct kmem_cache *s;
2940 attribute = to_slab_attr(attr);
2943 if (!attribute->show)
2946 err = attribute->show(s, buf);
2951 static ssize_t slab_attr_store(struct kobject *kobj,
2952 struct attribute *attr,
2953 const char *buf, size_t len)
2955 struct slab_attribute *attribute;
2956 struct kmem_cache *s;
2959 attribute = to_slab_attr(attr);
2962 if (!attribute->store)
2965 err = attribute->store(s, buf, len);
2970 static struct sysfs_ops slab_sysfs_ops = {
2971 .show = slab_attr_show,
2972 .store = slab_attr_store,
2975 static struct kobj_type slab_ktype = {
2976 .sysfs_ops = &slab_sysfs_ops,
2979 static int uevent_filter(struct kset *kset, struct kobject *kobj)
2981 struct kobj_type *ktype = get_ktype(kobj);
2983 if (ktype == &slab_ktype)
2988 static struct kset_uevent_ops slab_uevent_ops = {
2989 .filter = uevent_filter,
2992 decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
2994 #define ID_STR_LENGTH 64
2996 /* Create a unique string id for a slab cache:
2998 * :[flags-]size:[memory address of kmemcache]
3000 static char *create_unique_id(struct kmem_cache *s)
3002 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3009 * First flags affecting slabcache operations. We will only
3010 * get here for aliasable slabs so we do not need to support
3011 * too many flags. The flags here must cover all flags that
3012 * are matched during merging to guarantee that the id is
3015 if (s->flags & SLAB_CACHE_DMA)
3017 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3019 if (s->flags & SLAB_DEBUG_FREE)
3023 p += sprintf(p, "%07d", s->size);
3024 BUG_ON(p > name + ID_STR_LENGTH - 1);
3028 static int sysfs_slab_add(struct kmem_cache *s)
3034 if (slab_state < SYSFS)
3035 /* Defer until later */
3038 unmergeable = slab_unmergeable(s);
3041 * Slabcache can never be merged so we can use the name proper.
3042 * This is typically the case for debug situations. In that
3043 * case we can catch duplicate names easily.
3045 sysfs_remove_link(&slab_subsys.kset.kobj, s->name);
3049 * Create a unique name for the slab as a target
3052 name = create_unique_id(s);
3055 kobj_set_kset_s(s, slab_subsys);
3056 kobject_set_name(&s->kobj, name);
3057 kobject_init(&s->kobj);
3058 err = kobject_add(&s->kobj);
3062 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3065 kobject_uevent(&s->kobj, KOBJ_ADD);
3067 /* Setup first alias */
3068 sysfs_slab_alias(s, s->name);
3074 static void sysfs_slab_remove(struct kmem_cache *s)
3076 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3077 kobject_del(&s->kobj);
3081 * Need to buffer aliases during bootup until sysfs becomes
3082 * available lest we loose that information.
3084 struct saved_alias {
3085 struct kmem_cache *s;
3087 struct saved_alias *next;
3090 struct saved_alias *alias_list;
3092 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3094 struct saved_alias *al;
3096 if (slab_state == SYSFS) {
3098 * If we have a leftover link then remove it.
3100 sysfs_remove_link(&slab_subsys.kset.kobj, name);
3101 return sysfs_create_link(&slab_subsys.kset.kobj,
3105 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3111 al->next = alias_list;
3116 static int __init slab_sysfs_init(void)
3120 err = subsystem_register(&slab_subsys);
3122 printk(KERN_ERR "Cannot register slab subsystem.\n");
3128 while (alias_list) {
3129 struct saved_alias *al = alias_list;
3131 alias_list = alias_list->next;
3132 err = sysfs_slab_alias(al->s, al->name);
3141 __initcall(slab_sysfs_init);
3143 __initcall(finish_bootstrap);