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 __alignof__(unsigned long long)
151 #ifndef ARCH_SLAB_MINALIGN
152 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
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 static void init_tracking(struct kmem_cache *s, void *object)
302 if (s->flags & SLAB_STORE_USER) {
303 set_track(s, object, TRACK_FREE, NULL);
304 set_track(s, object, TRACK_ALLOC, NULL);
308 static void print_track(const char *s, struct track *t)
313 printk(KERN_ERR "%s: ", s);
314 __print_symbol("%s", (unsigned long)t->addr);
315 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
318 static void print_trailer(struct kmem_cache *s, u8 *p)
320 unsigned int off; /* Offset of last byte */
322 if (s->flags & SLAB_RED_ZONE)
323 print_section("Redzone", p + s->objsize,
324 s->inuse - s->objsize);
326 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
328 get_freepointer(s, p));
331 off = s->offset + sizeof(void *);
335 if (s->flags & SLAB_STORE_USER) {
336 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
337 print_track("Last free ", get_track(s, p, TRACK_FREE));
338 off += 2 * sizeof(struct track);
342 /* Beginning of the filler is the free pointer */
343 print_section("Filler", p + off, s->size - off);
346 static void object_err(struct kmem_cache *s, struct page *page,
347 u8 *object, char *reason)
349 u8 *addr = page_address(page);
351 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
352 s->name, reason, object, page);
353 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
354 object - addr, page->flags, page->inuse, page->freelist);
355 if (object > addr + 16)
356 print_section("Bytes b4", object - 16, 16);
357 print_section("Object", object, min(s->objsize, 128));
358 print_trailer(s, object);
362 static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
367 va_start(args, reason);
368 vsnprintf(buf, sizeof(buf), reason, args);
370 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
375 static void init_object(struct kmem_cache *s, void *object, int active)
379 if (s->flags & __OBJECT_POISON) {
380 memset(p, POISON_FREE, s->objsize - 1);
381 p[s->objsize -1] = POISON_END;
384 if (s->flags & SLAB_RED_ZONE)
385 memset(p + s->objsize,
386 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
387 s->inuse - s->objsize);
390 static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
393 if (*start != (u8)value)
402 static int check_valid_pointer(struct kmem_cache *s, struct page *page,
410 base = page_address(page);
411 if (object < base || object >= base + s->objects * s->size ||
412 (object - base) % s->size) {
423 * Bytes of the object to be managed.
424 * If the freepointer may overlay the object then the free
425 * pointer is the first word of the object.
426 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
429 * object + s->objsize
430 * Padding to reach word boundary. This is also used for Redzoning.
431 * Padding is extended to word size if Redzoning is enabled
432 * and objsize == inuse.
433 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
434 * 0xcc (RED_ACTIVE) for objects in use.
437 * A. Free pointer (if we cannot overwrite object on free)
438 * B. Tracking data for SLAB_STORE_USER
439 * C. Padding to reach required alignment boundary
440 * Padding is done using 0x5a (POISON_INUSE)
444 * If slabcaches are merged then the objsize and inuse boundaries are to
445 * be ignored. And therefore no slab options that rely on these boundaries
446 * may be used with merged slabcaches.
449 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
450 void *from, void *to)
452 printk(KERN_ERR "@@@ SLUB: %s Restoring %s (0x%x) from 0x%p-0x%p\n",
453 s->name, message, data, from, to - 1);
454 memset(from, data, to - from);
457 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
459 unsigned long off = s->inuse; /* The end of info */
462 /* Freepointer is placed after the object. */
463 off += sizeof(void *);
465 if (s->flags & SLAB_STORE_USER)
466 /* We also have user information there */
467 off += 2 * sizeof(struct track);
472 if (check_bytes(p + off, POISON_INUSE, s->size - off))
475 object_err(s, page, p, "Object padding check fails");
480 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
484 static int slab_pad_check(struct kmem_cache *s, struct page *page)
487 int length, remainder;
489 if (!(s->flags & SLAB_POISON))
492 p = page_address(page);
493 length = s->objects * s->size;
494 remainder = (PAGE_SIZE << s->order) - length;
498 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
499 printk(KERN_ERR "SLUB: %s slab 0x%p: Padding fails check\n",
502 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
503 p + length + remainder);
509 static int check_object(struct kmem_cache *s, struct page *page,
510 void *object, int active)
513 u8 *endobject = object + s->objsize;
515 if (s->flags & SLAB_RED_ZONE) {
517 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
519 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
520 object_err(s, page, object,
521 active ? "Redzone Active" : "Redzone Inactive");
522 restore_bytes(s, "redzone", red,
523 endobject, object + s->inuse);
527 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
528 !check_bytes(endobject, POISON_INUSE,
529 s->inuse - s->objsize)) {
530 object_err(s, page, p, "Alignment padding check fails");
532 * Fix it so that there will not be another report.
534 * Hmmm... We may be corrupting an object that now expects
535 * to be longer than allowed.
537 restore_bytes(s, "alignment padding", POISON_INUSE,
538 endobject, object + s->inuse);
542 if (s->flags & SLAB_POISON) {
543 if (!active && (s->flags & __OBJECT_POISON) &&
544 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
545 p[s->objsize - 1] != POISON_END)) {
547 object_err(s, page, p, "Poison check failed");
548 restore_bytes(s, "Poison", POISON_FREE,
549 p, p + s->objsize -1);
550 restore_bytes(s, "Poison", POISON_END,
551 p + s->objsize - 1, p + s->objsize);
555 * check_pad_bytes cleans up on its own.
557 check_pad_bytes(s, page, p);
560 if (!s->offset && active)
562 * Object and freepointer overlap. Cannot check
563 * freepointer while object is allocated.
567 /* Check free pointer validity */
568 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
569 object_err(s, page, p, "Freepointer corrupt");
571 * No choice but to zap it and thus loose the remainder
572 * of the free objects in this slab. May cause
573 * another error because the object count maybe
576 set_freepointer(s, p, NULL);
582 static int check_slab(struct kmem_cache *s, struct page *page)
584 VM_BUG_ON(!irqs_disabled());
586 if (!PageSlab(page)) {
587 printk(KERN_ERR "SLUB: %s Not a valid slab page @0x%p "
588 "flags=%lx mapping=0x%p count=%d \n",
589 s->name, page, page->flags, page->mapping,
593 if (page->offset * sizeof(void *) != s->offset) {
594 printk(KERN_ERR "SLUB: %s Corrupted offset %lu in slab @0x%p"
595 " flags=0x%lx mapping=0x%p count=%d\n",
597 (unsigned long)(page->offset * sizeof(void *)),
605 if (page->inuse > s->objects) {
606 printk(KERN_ERR "SLUB: %s Inuse %u > max %u in slab "
607 "page @0x%p flags=%lx mapping=0x%p count=%d\n",
608 s->name, page->inuse, s->objects, page, page->flags,
609 page->mapping, page_count(page));
613 /* Slab_pad_check fixes things up after itself */
614 slab_pad_check(s, page);
619 * Determine if a certain object on a page is on the freelist and
620 * therefore free. Must hold the slab lock for cpu slabs to
621 * guarantee that the chains are consistent.
623 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
626 void *fp = page->freelist;
629 while (fp && nr <= s->objects) {
632 if (!check_valid_pointer(s, page, fp)) {
634 object_err(s, page, object,
635 "Freechain corrupt");
636 set_freepointer(s, object, NULL);
639 printk(KERN_ERR "SLUB: %s slab 0x%p "
640 "freepointer 0x%p corrupted.\n",
643 page->freelist = NULL;
644 page->inuse = s->objects;
650 fp = get_freepointer(s, object);
654 if (page->inuse != s->objects - nr) {
655 printk(KERN_ERR "slab %s: page 0x%p wrong object count."
656 " counter is %d but counted were %d\n",
657 s->name, page, page->inuse,
659 page->inuse = s->objects - nr;
661 return search == NULL;
665 * Tracking of fully allocated slabs for debugging
667 static void add_full(struct kmem_cache *s, struct page *page)
669 struct kmem_cache_node *n;
671 VM_BUG_ON(!irqs_disabled());
673 if (!(s->flags & SLAB_STORE_USER))
676 n = get_node(s, page_to_nid(page));
677 spin_lock(&n->list_lock);
678 list_add(&page->lru, &n->full);
679 spin_unlock(&n->list_lock);
682 static void remove_full(struct kmem_cache *s, struct page *page)
684 struct kmem_cache_node *n;
686 if (!(s->flags & SLAB_STORE_USER))
689 n = get_node(s, page_to_nid(page));
691 spin_lock(&n->list_lock);
692 list_del(&page->lru);
693 spin_unlock(&n->list_lock);
696 static int alloc_object_checks(struct kmem_cache *s, struct page *page,
699 if (!check_slab(s, page))
702 if (object && !on_freelist(s, page, object)) {
703 printk(KERN_ERR "SLUB: %s Object 0x%p@0x%p "
704 "already allocated.\n",
705 s->name, object, page);
709 if (!check_valid_pointer(s, page, object)) {
710 object_err(s, page, object, "Freelist Pointer check fails");
717 if (!check_object(s, page, object, 0))
719 init_object(s, object, 1);
721 if (s->flags & SLAB_TRACE) {
722 printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
723 s->name, object, page->inuse,
731 if (PageSlab(page)) {
733 * If this is a slab page then lets do the best we can
734 * to avoid issues in the future. Marking all objects
735 * as used avoids touching the remainder.
737 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
739 page->inuse = s->objects;
740 page->freelist = NULL;
741 /* Fix up fields that may be corrupted */
742 page->offset = s->offset / sizeof(void *);
747 static int free_object_checks(struct kmem_cache *s, struct page *page,
750 if (!check_slab(s, page))
753 if (!check_valid_pointer(s, page, object)) {
754 printk(KERN_ERR "SLUB: %s slab 0x%p invalid "
755 "object pointer 0x%p\n",
756 s->name, page, object);
760 if (on_freelist(s, page, object)) {
761 printk(KERN_ERR "SLUB: %s slab 0x%p object "
762 "0x%p already free.\n", s->name, page, object);
766 if (!check_object(s, page, object, 1))
769 if (unlikely(s != page->slab)) {
771 printk(KERN_ERR "slab_free %s size %d: attempt to"
772 "free object(0x%p) outside of slab.\n",
773 s->name, s->size, object);
777 "slab_free : no slab(NULL) for object 0x%p.\n",
780 printk(KERN_ERR "slab_free %s(%d): object at 0x%p"
781 " belongs to slab %s(%d)\n",
782 s->name, s->size, object,
783 page->slab->name, page->slab->size);
786 if (s->flags & SLAB_TRACE) {
787 printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n",
788 s->name, object, page->inuse,
790 print_section("Object", object, s->objsize);
793 init_object(s, object, 0);
797 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
798 s->name, page, object);
803 * Slab allocation and freeing
805 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
808 int pages = 1 << s->order;
813 if (s->flags & SLAB_CACHE_DMA)
817 page = alloc_pages(flags, s->order);
819 page = alloc_pages_node(node, flags, s->order);
824 mod_zone_page_state(page_zone(page),
825 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
826 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
832 static void setup_object(struct kmem_cache *s, struct page *page,
835 if (PageError(page)) {
836 init_object(s, object, 0);
837 init_tracking(s, object);
840 if (unlikely(s->ctor)) {
841 int mode = SLAB_CTOR_CONSTRUCTOR;
843 if (!(s->flags & __GFP_WAIT))
844 mode |= SLAB_CTOR_ATOMIC;
846 s->ctor(object, s, mode);
850 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
853 struct kmem_cache_node *n;
859 if (flags & __GFP_NO_GROW)
862 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
864 if (flags & __GFP_WAIT)
867 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
871 n = get_node(s, page_to_nid(page));
873 atomic_long_inc(&n->nr_slabs);
874 page->offset = s->offset / sizeof(void *);
876 page->flags |= 1 << PG_slab;
877 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
878 SLAB_STORE_USER | SLAB_TRACE))
879 page->flags |= 1 << PG_error;
881 start = page_address(page);
882 end = start + s->objects * s->size;
884 if (unlikely(s->flags & SLAB_POISON))
885 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
888 for (p = start + s->size; p < end; p += s->size) {
889 setup_object(s, page, last);
890 set_freepointer(s, last, p);
893 setup_object(s, page, last);
894 set_freepointer(s, last, NULL);
896 page->freelist = start;
899 if (flags & __GFP_WAIT)
904 static void __free_slab(struct kmem_cache *s, struct page *page)
906 int pages = 1 << s->order;
908 if (unlikely(PageError(page) || s->dtor)) {
909 void *start = page_address(page);
910 void *end = start + (pages << PAGE_SHIFT);
913 slab_pad_check(s, page);
914 for (p = start; p <= end - s->size; p += s->size) {
917 check_object(s, page, p, 0);
921 mod_zone_page_state(page_zone(page),
922 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
923 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
926 page->mapping = NULL;
927 __free_pages(page, s->order);
930 static void rcu_free_slab(struct rcu_head *h)
934 page = container_of((struct list_head *)h, struct page, lru);
935 __free_slab(page->slab, page);
938 static void free_slab(struct kmem_cache *s, struct page *page)
940 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
942 * RCU free overloads the RCU head over the LRU
944 struct rcu_head *head = (void *)&page->lru;
946 call_rcu(head, rcu_free_slab);
948 __free_slab(s, page);
951 static void discard_slab(struct kmem_cache *s, struct page *page)
953 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
955 atomic_long_dec(&n->nr_slabs);
956 reset_page_mapcount(page);
957 page->flags &= ~(1 << PG_slab | 1 << PG_error);
962 * Per slab locking using the pagelock
964 static __always_inline void slab_lock(struct page *page)
966 bit_spin_lock(PG_locked, &page->flags);
969 static __always_inline void slab_unlock(struct page *page)
971 bit_spin_unlock(PG_locked, &page->flags);
974 static __always_inline int slab_trylock(struct page *page)
978 rc = bit_spin_trylock(PG_locked, &page->flags);
983 * Management of partially allocated slabs
985 static void add_partial(struct kmem_cache *s, struct page *page)
987 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
989 spin_lock(&n->list_lock);
991 list_add(&page->lru, &n->partial);
992 spin_unlock(&n->list_lock);
995 static void remove_partial(struct kmem_cache *s,
998 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1000 spin_lock(&n->list_lock);
1001 list_del(&page->lru);
1003 spin_unlock(&n->list_lock);
1007 * Lock page and remove it from the partial list
1009 * Must hold list_lock
1011 static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page)
1013 if (slab_trylock(page)) {
1014 list_del(&page->lru);
1022 * Try to get a partial slab from a specific node
1024 static struct page *get_partial_node(struct kmem_cache_node *n)
1029 * Racy check. If we mistakenly see no partial slabs then we
1030 * just allocate an empty slab. If we mistakenly try to get a
1031 * partial slab then get_partials() will return NULL.
1033 if (!n || !n->nr_partial)
1036 spin_lock(&n->list_lock);
1037 list_for_each_entry(page, &n->partial, lru)
1038 if (lock_and_del_slab(n, page))
1042 spin_unlock(&n->list_lock);
1047 * Get a page from somewhere. Search in increasing NUMA
1050 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1053 struct zonelist *zonelist;
1058 * The defrag ratio allows to configure the tradeoffs between
1059 * inter node defragmentation and node local allocations.
1060 * A lower defrag_ratio increases the tendency to do local
1061 * allocations instead of scanning throught the partial
1062 * lists on other nodes.
1064 * If defrag_ratio is set to 0 then kmalloc() always
1065 * returns node local objects. If its higher then kmalloc()
1066 * may return off node objects in order to avoid fragmentation.
1068 * A higher ratio means slabs may be taken from other nodes
1069 * thus reducing the number of partial slabs on those nodes.
1071 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1072 * defrag_ratio = 1000) then every (well almost) allocation
1073 * will first attempt to defrag slab caches on other nodes. This
1074 * means scanning over all nodes to look for partial slabs which
1075 * may be a bit expensive to do on every slab allocation.
1077 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1080 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1081 ->node_zonelists[gfp_zone(flags)];
1082 for (z = zonelist->zones; *z; z++) {
1083 struct kmem_cache_node *n;
1085 n = get_node(s, zone_to_nid(*z));
1087 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1088 n->nr_partial > 2) {
1089 page = get_partial_node(n);
1099 * Get a partial page, lock it and return it.
1101 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1104 int searchnode = (node == -1) ? numa_node_id() : node;
1106 page = get_partial_node(get_node(s, searchnode));
1107 if (page || (flags & __GFP_THISNODE))
1110 return get_any_partial(s, flags);
1114 * Move a page back to the lists.
1116 * Must be called with the slab lock held.
1118 * On exit the slab lock will have been dropped.
1120 static void putback_slab(struct kmem_cache *s, struct page *page)
1124 add_partial(s, page);
1125 else if (PageError(page))
1130 discard_slab(s, page);
1135 * Remove the cpu slab
1137 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1139 s->cpu_slab[cpu] = NULL;
1140 ClearPageActive(page);
1142 putback_slab(s, page);
1145 static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1148 deactivate_slab(s, page, cpu);
1153 * Called from IPI handler with interrupts disabled.
1155 static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1157 struct page *page = s->cpu_slab[cpu];
1160 flush_slab(s, page, cpu);
1163 static void flush_cpu_slab(void *d)
1165 struct kmem_cache *s = d;
1166 int cpu = smp_processor_id();
1168 __flush_cpu_slab(s, cpu);
1171 static void flush_all(struct kmem_cache *s)
1174 on_each_cpu(flush_cpu_slab, s, 1, 1);
1176 unsigned long flags;
1178 local_irq_save(flags);
1180 local_irq_restore(flags);
1185 * slab_alloc is optimized to only modify two cachelines on the fast path
1186 * (aside from the stack):
1188 * 1. The page struct
1189 * 2. The first cacheline of the object to be allocated.
1191 * The only cache lines that are read (apart from code) is the
1192 * per cpu array in the kmem_cache struct.
1194 * Fastpath is not possible if we need to get a new slab or have
1195 * debugging enabled (which means all slabs are marked with PageError)
1197 static void *slab_alloc(struct kmem_cache *s,
1198 gfp_t gfpflags, int node, void *addr)
1202 unsigned long flags;
1205 local_irq_save(flags);
1206 cpu = smp_processor_id();
1207 page = s->cpu_slab[cpu];
1212 if (unlikely(node != -1 && page_to_nid(page) != node))
1215 object = page->freelist;
1216 if (unlikely(!object))
1218 if (unlikely(PageError(page)))
1223 page->freelist = object[page->offset];
1225 local_irq_restore(flags);
1229 deactivate_slab(s, page, cpu);
1232 page = get_partial(s, gfpflags, node);
1235 s->cpu_slab[cpu] = page;
1236 SetPageActive(page);
1240 page = new_slab(s, gfpflags, node);
1242 cpu = smp_processor_id();
1243 if (s->cpu_slab[cpu]) {
1245 * Someone else populated the cpu_slab while we enabled
1246 * interrupts, or we have got scheduled on another cpu.
1247 * The page may not be on the requested node.
1250 page_to_nid(s->cpu_slab[cpu]) == node) {
1252 * Current cpuslab is acceptable and we
1253 * want the current one since its cache hot
1255 discard_slab(s, page);
1256 page = s->cpu_slab[cpu];
1260 /* Dump the current slab */
1261 flush_slab(s, s->cpu_slab[cpu], cpu);
1266 local_irq_restore(flags);
1269 if (!alloc_object_checks(s, page, object))
1271 if (s->flags & SLAB_STORE_USER)
1272 set_track(s, object, TRACK_ALLOC, addr);
1276 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1278 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1280 EXPORT_SYMBOL(kmem_cache_alloc);
1283 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1285 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1287 EXPORT_SYMBOL(kmem_cache_alloc_node);
1291 * The fastpath only writes the cacheline of the page struct and the first
1292 * cacheline of the object.
1294 * No special cachelines need to be read
1296 static void slab_free(struct kmem_cache *s, struct page *page,
1297 void *x, void *addr)
1300 void **object = (void *)x;
1301 unsigned long flags;
1303 local_irq_save(flags);
1306 if (unlikely(PageError(page)))
1309 prior = object[page->offset] = page->freelist;
1310 page->freelist = object;
1313 if (unlikely(PageActive(page)))
1315 * Cpu slabs are never on partial lists and are
1320 if (unlikely(!page->inuse))
1324 * Objects left in the slab. If it
1325 * was not on the partial list before
1328 if (unlikely(!prior))
1329 add_partial(s, page);
1333 local_irq_restore(flags);
1339 * Slab on the partial list.
1341 remove_partial(s, page);
1344 discard_slab(s, page);
1345 local_irq_restore(flags);
1349 if (!free_object_checks(s, page, x))
1351 if (!PageActive(page) && !page->freelist)
1352 remove_full(s, page);
1353 if (s->flags & SLAB_STORE_USER)
1354 set_track(s, x, TRACK_FREE, addr);
1358 void kmem_cache_free(struct kmem_cache *s, void *x)
1362 page = virt_to_head_page(x);
1364 slab_free(s, page, x, __builtin_return_address(0));
1366 EXPORT_SYMBOL(kmem_cache_free);
1368 /* Figure out on which slab object the object resides */
1369 static struct page *get_object_page(const void *x)
1371 struct page *page = virt_to_head_page(x);
1373 if (!PageSlab(page))
1380 * kmem_cache_open produces objects aligned at "size" and the first object
1381 * is placed at offset 0 in the slab (We have no metainformation on the
1382 * slab, all slabs are in essence "off slab").
1384 * In order to get the desired alignment one just needs to align the
1387 * Notice that the allocation order determines the sizes of the per cpu
1388 * caches. Each processor has always one slab available for allocations.
1389 * Increasing the allocation order reduces the number of times that slabs
1390 * must be moved on and off the partial lists and therefore may influence
1393 * The offset is used to relocate the free list link in each object. It is
1394 * therefore possible to move the free list link behind the object. This
1395 * is necessary for RCU to work properly and also useful for debugging.
1399 * Mininum / Maximum order of slab pages. This influences locking overhead
1400 * and slab fragmentation. A higher order reduces the number of partial slabs
1401 * and increases the number of allocations possible without having to
1402 * take the list_lock.
1404 static int slub_min_order;
1405 static int slub_max_order = DEFAULT_MAX_ORDER;
1408 * Minimum number of objects per slab. This is necessary in order to
1409 * reduce locking overhead. Similar to the queue size in SLAB.
1411 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1414 * Merge control. If this is set then no merging of slab caches will occur.
1416 static int slub_nomerge;
1421 static int slub_debug;
1423 static char *slub_debug_slabs;
1426 * Calculate the order of allocation given an slab object size.
1428 * The order of allocation has significant impact on other elements
1429 * of the system. Generally order 0 allocations should be preferred
1430 * since they do not cause fragmentation in the page allocator. Larger
1431 * objects may have problems with order 0 because there may be too much
1432 * space left unused in a slab. We go to a higher order if more than 1/8th
1433 * of the slab would be wasted.
1435 * In order to reach satisfactory performance we must ensure that
1436 * a minimum number of objects is in one slab. Otherwise we may
1437 * generate too much activity on the partial lists. This is less a
1438 * concern for large slabs though. slub_max_order specifies the order
1439 * where we begin to stop considering the number of objects in a slab.
1441 * Higher order allocations also allow the placement of more objects
1442 * in a slab and thereby reduce object handling overhead. If the user
1443 * has requested a higher mininum order then we start with that one
1446 static int calculate_order(int size)
1451 for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT);
1452 order < MAX_ORDER; order++) {
1453 unsigned long slab_size = PAGE_SIZE << order;
1455 if (slub_max_order > order &&
1456 slab_size < slub_min_objects * size)
1459 if (slab_size < size)
1462 rem = slab_size % size;
1464 if (rem <= (PAGE_SIZE << order) / 8)
1468 if (order >= MAX_ORDER)
1474 * Function to figure out which alignment to use from the
1475 * various ways of specifying it.
1477 static unsigned long calculate_alignment(unsigned long flags,
1478 unsigned long align, unsigned long size)
1481 * If the user wants hardware cache aligned objects then
1482 * follow that suggestion if the object is sufficiently
1485 * The hardware cache alignment cannot override the
1486 * specified alignment though. If that is greater
1489 if ((flags & (SLAB_MUST_HWCACHE_ALIGN | SLAB_HWCACHE_ALIGN)) &&
1490 size > L1_CACHE_BYTES / 2)
1491 return max_t(unsigned long, align, L1_CACHE_BYTES);
1493 if (align < ARCH_SLAB_MINALIGN)
1494 return ARCH_SLAB_MINALIGN;
1496 return ALIGN(align, sizeof(void *));
1499 static void init_kmem_cache_node(struct kmem_cache_node *n)
1502 atomic_long_set(&n->nr_slabs, 0);
1503 spin_lock_init(&n->list_lock);
1504 INIT_LIST_HEAD(&n->partial);
1505 INIT_LIST_HEAD(&n->full);
1510 * No kmalloc_node yet so do it by hand. We know that this is the first
1511 * slab on the node for this slabcache. There are no concurrent accesses
1514 * Note that this function only works on the kmalloc_node_cache
1515 * when allocating for the kmalloc_node_cache.
1517 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1521 struct kmem_cache_node *n;
1523 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1525 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1526 /* new_slab() disables interupts */
1532 page->freelist = get_freepointer(kmalloc_caches, n);
1534 kmalloc_caches->node[node] = n;
1535 init_object(kmalloc_caches, n, 1);
1536 init_kmem_cache_node(n);
1537 atomic_long_inc(&n->nr_slabs);
1538 add_partial(kmalloc_caches, page);
1542 static void free_kmem_cache_nodes(struct kmem_cache *s)
1546 for_each_online_node(node) {
1547 struct kmem_cache_node *n = s->node[node];
1548 if (n && n != &s->local_node)
1549 kmem_cache_free(kmalloc_caches, n);
1550 s->node[node] = NULL;
1554 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1559 if (slab_state >= UP)
1560 local_node = page_to_nid(virt_to_page(s));
1564 for_each_online_node(node) {
1565 struct kmem_cache_node *n;
1567 if (local_node == node)
1570 if (slab_state == DOWN) {
1571 n = early_kmem_cache_node_alloc(gfpflags,
1575 n = kmem_cache_alloc_node(kmalloc_caches,
1579 free_kmem_cache_nodes(s);
1585 init_kmem_cache_node(n);
1590 static void free_kmem_cache_nodes(struct kmem_cache *s)
1594 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1596 init_kmem_cache_node(&s->local_node);
1602 * calculate_sizes() determines the order and the distribution of data within
1605 static int calculate_sizes(struct kmem_cache *s)
1607 unsigned long flags = s->flags;
1608 unsigned long size = s->objsize;
1609 unsigned long align = s->align;
1612 * Determine if we can poison the object itself. If the user of
1613 * the slab may touch the object after free or before allocation
1614 * then we should never poison the object itself.
1616 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1617 !s->ctor && !s->dtor)
1618 s->flags |= __OBJECT_POISON;
1620 s->flags &= ~__OBJECT_POISON;
1623 * Round up object size to the next word boundary. We can only
1624 * place the free pointer at word boundaries and this determines
1625 * the possible location of the free pointer.
1627 size = ALIGN(size, sizeof(void *));
1630 * If we are redzoning then check if there is some space between the
1631 * end of the object and the free pointer. If not then add an
1632 * additional word, so that we can establish a redzone between
1633 * the object and the freepointer to be able to check for overwrites.
1635 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1636 size += sizeof(void *);
1639 * With that we have determined how much of the slab is in actual
1640 * use by the object. This is the potential offset to the free
1645 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1646 s->ctor || s->dtor)) {
1648 * Relocate free pointer after the object if it is not
1649 * permitted to overwrite the first word of the object on
1652 * This is the case if we do RCU, have a constructor or
1653 * destructor or are poisoning the objects.
1656 size += sizeof(void *);
1659 if (flags & SLAB_STORE_USER)
1661 * Need to store information about allocs and frees after
1664 size += 2 * sizeof(struct track);
1666 if (flags & DEBUG_DEFAULT_FLAGS)
1668 * Add some empty padding so that we can catch
1669 * overwrites from earlier objects rather than let
1670 * tracking information or the free pointer be
1671 * corrupted if an user writes before the start
1674 size += sizeof(void *);
1676 * Determine the alignment based on various parameters that the
1677 * user specified (this is unecessarily complex due to the attempt
1678 * to be compatible with SLAB. Should be cleaned up some day).
1680 align = calculate_alignment(flags, align, s->objsize);
1683 * SLUB stores one object immediately after another beginning from
1684 * offset 0. In order to align the objects we have to simply size
1685 * each object to conform to the alignment.
1687 size = ALIGN(size, align);
1690 s->order = calculate_order(size);
1695 * Determine the number of objects per slab
1697 s->objects = (PAGE_SIZE << s->order) / size;
1700 * Verify that the number of objects is within permitted limits.
1701 * The page->inuse field is only 16 bit wide! So we cannot have
1702 * more than 64k objects per slab.
1704 if (!s->objects || s->objects > 65535)
1710 static int __init finish_bootstrap(void)
1712 struct list_head *h;
1717 list_for_each(h, &slab_caches) {
1718 struct kmem_cache *s =
1719 container_of(h, struct kmem_cache, list);
1721 err = sysfs_slab_add(s);
1727 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1728 const char *name, size_t size,
1729 size_t align, unsigned long flags,
1730 void (*ctor)(void *, struct kmem_cache *, unsigned long),
1731 void (*dtor)(void *, struct kmem_cache *, unsigned long))
1733 memset(s, 0, kmem_size);
1741 BUG_ON(flags & SLUB_UNIMPLEMENTED);
1744 * The page->offset field is only 16 bit wide. This is an offset
1745 * in units of words from the beginning of an object. If the slab
1746 * size is bigger then we cannot move the free pointer behind the
1749 * On 32 bit platforms the limit is 256k. On 64bit platforms
1750 * the limit is 512k.
1752 * Debugging or ctor/dtors may create a need to move the free
1753 * pointer. Fail if this happens.
1755 if (s->size >= 65535 * sizeof(void *)) {
1756 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1757 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1758 BUG_ON(ctor || dtor);
1762 * Enable debugging if selected on the kernel commandline.
1764 if (slub_debug && (!slub_debug_slabs ||
1765 strncmp(slub_debug_slabs, name,
1766 strlen(slub_debug_slabs)) == 0))
1767 s->flags |= slub_debug;
1769 if (!calculate_sizes(s))
1774 s->defrag_ratio = 100;
1777 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
1780 if (flags & SLAB_PANIC)
1781 panic("Cannot create slab %s size=%lu realsize=%u "
1782 "order=%u offset=%u flags=%lx\n",
1783 s->name, (unsigned long)size, s->size, s->order,
1787 EXPORT_SYMBOL(kmem_cache_open);
1790 * Check if a given pointer is valid
1792 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
1797 page = get_object_page(object);
1799 if (!page || s != page->slab)
1800 /* No slab or wrong slab */
1803 addr = page_address(page);
1804 if (object < addr || object >= addr + s->objects * s->size)
1808 if ((object - addr) % s->size)
1809 /* Improperly aligned */
1813 * We could also check if the object is on the slabs freelist.
1814 * But this would be too expensive and it seems that the main
1815 * purpose of kmem_ptr_valid is to check if the object belongs
1816 * to a certain slab.
1820 EXPORT_SYMBOL(kmem_ptr_validate);
1823 * Determine the size of a slab object
1825 unsigned int kmem_cache_size(struct kmem_cache *s)
1829 EXPORT_SYMBOL(kmem_cache_size);
1831 const char *kmem_cache_name(struct kmem_cache *s)
1835 EXPORT_SYMBOL(kmem_cache_name);
1838 * Attempt to free all slabs on a node
1840 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
1841 struct list_head *list)
1843 int slabs_inuse = 0;
1844 unsigned long flags;
1845 struct page *page, *h;
1847 spin_lock_irqsave(&n->list_lock, flags);
1848 list_for_each_entry_safe(page, h, list, lru)
1850 list_del(&page->lru);
1851 discard_slab(s, page);
1854 spin_unlock_irqrestore(&n->list_lock, flags);
1859 * Release all resources used by slab cache
1861 static int kmem_cache_close(struct kmem_cache *s)
1867 /* Attempt to free all objects */
1868 for_each_online_node(node) {
1869 struct kmem_cache_node *n = get_node(s, node);
1871 free_list(s, n, &n->partial);
1872 if (atomic_long_read(&n->nr_slabs))
1875 free_kmem_cache_nodes(s);
1880 * Close a cache and release the kmem_cache structure
1881 * (must be used for caches created using kmem_cache_create)
1883 void kmem_cache_destroy(struct kmem_cache *s)
1885 down_write(&slub_lock);
1889 if (kmem_cache_close(s))
1891 sysfs_slab_remove(s);
1894 up_write(&slub_lock);
1896 EXPORT_SYMBOL(kmem_cache_destroy);
1898 /********************************************************************
1900 *******************************************************************/
1902 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
1903 EXPORT_SYMBOL(kmalloc_caches);
1905 #ifdef CONFIG_ZONE_DMA
1906 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
1909 static int __init setup_slub_min_order(char *str)
1911 get_option (&str, &slub_min_order);
1916 __setup("slub_min_order=", setup_slub_min_order);
1918 static int __init setup_slub_max_order(char *str)
1920 get_option (&str, &slub_max_order);
1925 __setup("slub_max_order=", setup_slub_max_order);
1927 static int __init setup_slub_min_objects(char *str)
1929 get_option (&str, &slub_min_objects);
1934 __setup("slub_min_objects=", setup_slub_min_objects);
1936 static int __init setup_slub_nomerge(char *str)
1942 __setup("slub_nomerge", setup_slub_nomerge);
1944 static int __init setup_slub_debug(char *str)
1946 if (!str || *str != '=')
1947 slub_debug = DEBUG_DEFAULT_FLAGS;
1950 if (*str == 0 || *str == ',')
1951 slub_debug = DEBUG_DEFAULT_FLAGS;
1953 for( ;*str && *str != ','; str++)
1955 case 'f' : case 'F' :
1956 slub_debug |= SLAB_DEBUG_FREE;
1958 case 'z' : case 'Z' :
1959 slub_debug |= SLAB_RED_ZONE;
1961 case 'p' : case 'P' :
1962 slub_debug |= SLAB_POISON;
1964 case 'u' : case 'U' :
1965 slub_debug |= SLAB_STORE_USER;
1967 case 't' : case 'T' :
1968 slub_debug |= SLAB_TRACE;
1971 printk(KERN_ERR "slub_debug option '%c' "
1972 "unknown. skipped\n",*str);
1977 slub_debug_slabs = str + 1;
1981 __setup("slub_debug", setup_slub_debug);
1983 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
1984 const char *name, int size, gfp_t gfp_flags)
1986 unsigned int flags = 0;
1988 if (gfp_flags & SLUB_DMA)
1989 flags = SLAB_CACHE_DMA;
1991 down_write(&slub_lock);
1992 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
1996 list_add(&s->list, &slab_caches);
1997 up_write(&slub_lock);
1998 if (sysfs_slab_add(s))
2003 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2006 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2008 int index = kmalloc_index(size);
2013 /* Allocation too large? */
2016 #ifdef CONFIG_ZONE_DMA
2017 if ((flags & SLUB_DMA)) {
2018 struct kmem_cache *s;
2019 struct kmem_cache *x;
2023 s = kmalloc_caches_dma[index];
2027 /* Dynamically create dma cache */
2028 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2030 panic("Unable to allocate memory for dma cache\n");
2032 if (index <= KMALLOC_SHIFT_HIGH)
2033 realsize = 1 << index;
2041 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2042 (unsigned int)realsize);
2043 s = create_kmalloc_cache(x, text, realsize, flags);
2044 kmalloc_caches_dma[index] = s;
2048 return &kmalloc_caches[index];
2051 void *__kmalloc(size_t size, gfp_t flags)
2053 struct kmem_cache *s = get_slab(size, flags);
2056 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2059 EXPORT_SYMBOL(__kmalloc);
2062 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2064 struct kmem_cache *s = get_slab(size, flags);
2067 return slab_alloc(s, flags, node, __builtin_return_address(0));
2070 EXPORT_SYMBOL(__kmalloc_node);
2073 size_t ksize(const void *object)
2075 struct page *page = get_object_page(object);
2076 struct kmem_cache *s;
2083 * Debugging requires use of the padding between object
2084 * and whatever may come after it.
2086 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2090 * If we have the need to store the freelist pointer
2091 * back there or track user information then we can
2092 * only use the space before that information.
2094 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2098 * Else we can use all the padding etc for the allocation
2102 EXPORT_SYMBOL(ksize);
2104 void kfree(const void *x)
2106 struct kmem_cache *s;
2112 page = virt_to_head_page(x);
2115 slab_free(s, page, (void *)x, __builtin_return_address(0));
2117 EXPORT_SYMBOL(kfree);
2120 * krealloc - reallocate memory. The contents will remain unchanged.
2122 * @p: object to reallocate memory for.
2123 * @new_size: how many bytes of memory are required.
2124 * @flags: the type of memory to allocate.
2126 * The contents of the object pointed to are preserved up to the
2127 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2128 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2129 * %NULL pointer, the object pointed to is freed.
2131 void *krealloc(const void *p, size_t new_size, gfp_t flags)
2133 struct kmem_cache *new_cache;
2138 return kmalloc(new_size, flags);
2140 if (unlikely(!new_size)) {
2145 page = virt_to_head_page(p);
2147 new_cache = get_slab(new_size, flags);
2150 * If new size fits in the current cache, bail out.
2152 if (likely(page->slab == new_cache))
2155 ret = kmalloc(new_size, flags);
2157 memcpy(ret, p, min(new_size, ksize(p)));
2162 EXPORT_SYMBOL(krealloc);
2164 /********************************************************************
2165 * Basic setup of slabs
2166 *******************************************************************/
2168 void __init kmem_cache_init(void)
2174 * Must first have the slab cache available for the allocations of the
2175 * struct kmalloc_cache_node's. There is special bootstrap code in
2176 * kmem_cache_open for slab_state == DOWN.
2178 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2179 sizeof(struct kmem_cache_node), GFP_KERNEL);
2182 /* Able to allocate the per node structures */
2183 slab_state = PARTIAL;
2185 /* Caches that are not of the two-to-the-power-of size */
2186 create_kmalloc_cache(&kmalloc_caches[1],
2187 "kmalloc-96", 96, GFP_KERNEL);
2188 create_kmalloc_cache(&kmalloc_caches[2],
2189 "kmalloc-192", 192, GFP_KERNEL);
2191 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2192 create_kmalloc_cache(&kmalloc_caches[i],
2193 "kmalloc", 1 << i, GFP_KERNEL);
2197 /* Provide the correct kmalloc names now that the caches are up */
2198 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2199 kmalloc_caches[i]. name =
2200 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2203 register_cpu_notifier(&slab_notifier);
2206 if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */
2207 kmem_size = offsetof(struct kmem_cache, cpu_slab)
2208 + nr_cpu_ids * sizeof(struct page *);
2210 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2211 " Processors=%d, Nodes=%d\n",
2212 KMALLOC_SHIFT_HIGH, L1_CACHE_BYTES,
2213 slub_min_order, slub_max_order, slub_min_objects,
2214 nr_cpu_ids, nr_node_ids);
2218 * Find a mergeable slab cache
2220 static int slab_unmergeable(struct kmem_cache *s)
2222 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2225 if (s->ctor || s->dtor)
2231 static struct kmem_cache *find_mergeable(size_t size,
2232 size_t align, unsigned long flags,
2233 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2234 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2236 struct list_head *h;
2238 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2244 size = ALIGN(size, sizeof(void *));
2245 align = calculate_alignment(flags, align, size);
2246 size = ALIGN(size, align);
2248 list_for_each(h, &slab_caches) {
2249 struct kmem_cache *s =
2250 container_of(h, struct kmem_cache, list);
2252 if (slab_unmergeable(s))
2258 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2259 (s->flags & SLUB_MERGE_SAME))
2262 * Check if alignment is compatible.
2263 * Courtesy of Adrian Drzewiecki
2265 if ((s->size & ~(align -1)) != s->size)
2268 if (s->size - size >= sizeof(void *))
2276 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2277 size_t align, unsigned long flags,
2278 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2279 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2281 struct kmem_cache *s;
2283 down_write(&slub_lock);
2284 s = find_mergeable(size, align, flags, dtor, ctor);
2288 * Adjust the object sizes so that we clear
2289 * the complete object on kzalloc.
2291 s->objsize = max(s->objsize, (int)size);
2292 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2293 if (sysfs_slab_alias(s, name))
2296 s = kmalloc(kmem_size, GFP_KERNEL);
2297 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2298 size, align, flags, ctor, dtor)) {
2299 if (sysfs_slab_add(s)) {
2303 list_add(&s->list, &slab_caches);
2307 up_write(&slub_lock);
2311 up_write(&slub_lock);
2312 if (flags & SLAB_PANIC)
2313 panic("Cannot create slabcache %s\n", name);
2318 EXPORT_SYMBOL(kmem_cache_create);
2320 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2324 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
2326 memset(x, 0, s->objsize);
2329 EXPORT_SYMBOL(kmem_cache_zalloc);
2332 static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2334 struct list_head *h;
2336 down_read(&slub_lock);
2337 list_for_each(h, &slab_caches) {
2338 struct kmem_cache *s =
2339 container_of(h, struct kmem_cache, list);
2343 up_read(&slub_lock);
2347 * Use the cpu notifier to insure that the slab are flushed
2350 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2351 unsigned long action, void *hcpu)
2353 long cpu = (long)hcpu;
2356 case CPU_UP_CANCELED:
2358 for_all_slabs(__flush_cpu_slab, cpu);
2366 static struct notifier_block __cpuinitdata slab_notifier =
2367 { &slab_cpuup_callback, NULL, 0 };
2371 /***************************************************************
2372 * Compatiblility definitions
2373 **************************************************************/
2375 int kmem_cache_shrink(struct kmem_cache *s)
2380 EXPORT_SYMBOL(kmem_cache_shrink);
2384 /*****************************************************************
2385 * Generic reaper used to support the page allocator
2386 * (the cpu slabs are reaped by a per slab workqueue).
2388 * Maybe move this to the page allocator?
2389 ****************************************************************/
2391 static DEFINE_PER_CPU(unsigned long, reap_node);
2393 static void init_reap_node(int cpu)
2397 node = next_node(cpu_to_node(cpu), node_online_map);
2398 if (node == MAX_NUMNODES)
2399 node = first_node(node_online_map);
2401 __get_cpu_var(reap_node) = node;
2404 static void next_reap_node(void)
2406 int node = __get_cpu_var(reap_node);
2409 * Also drain per cpu pages on remote zones
2411 if (node != numa_node_id())
2412 drain_node_pages(node);
2414 node = next_node(node, node_online_map);
2415 if (unlikely(node >= MAX_NUMNODES))
2416 node = first_node(node_online_map);
2417 __get_cpu_var(reap_node) = node;
2420 #define init_reap_node(cpu) do { } while (0)
2421 #define next_reap_node(void) do { } while (0)
2424 #define REAPTIMEOUT_CPUC (2*HZ)
2427 static DEFINE_PER_CPU(struct delayed_work, reap_work);
2429 static void cache_reap(struct work_struct *unused)
2432 refresh_cpu_vm_stats(smp_processor_id());
2433 schedule_delayed_work(&__get_cpu_var(reap_work),
2437 static void __devinit start_cpu_timer(int cpu)
2439 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
2442 * When this gets called from do_initcalls via cpucache_init(),
2443 * init_workqueues() has already run, so keventd will be setup
2446 if (keventd_up() && reap_work->work.func == NULL) {
2447 init_reap_node(cpu);
2448 INIT_DELAYED_WORK(reap_work, cache_reap);
2449 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
2453 static int __init cpucache_init(void)
2458 * Register the timers that drain pcp pages and update vm statistics
2460 for_each_online_cpu(cpu)
2461 start_cpu_timer(cpu);
2464 __initcall(cpucache_init);
2467 #ifdef SLUB_RESILIENCY_TEST
2468 static unsigned long validate_slab_cache(struct kmem_cache *s);
2470 static void resiliency_test(void)
2474 printk(KERN_ERR "SLUB resiliency testing\n");
2475 printk(KERN_ERR "-----------------------\n");
2476 printk(KERN_ERR "A. Corruption after allocation\n");
2478 p = kzalloc(16, GFP_KERNEL);
2480 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2481 " 0x12->0x%p\n\n", p + 16);
2483 validate_slab_cache(kmalloc_caches + 4);
2485 /* Hmmm... The next two are dangerous */
2486 p = kzalloc(32, GFP_KERNEL);
2487 p[32 + sizeof(void *)] = 0x34;
2488 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2489 " 0x34 -> -0x%p\n", p);
2490 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2492 validate_slab_cache(kmalloc_caches + 5);
2493 p = kzalloc(64, GFP_KERNEL);
2494 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2496 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2498 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2499 validate_slab_cache(kmalloc_caches + 6);
2501 printk(KERN_ERR "\nB. Corruption after free\n");
2502 p = kzalloc(128, GFP_KERNEL);
2505 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2506 validate_slab_cache(kmalloc_caches + 7);
2508 p = kzalloc(256, GFP_KERNEL);
2511 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2512 validate_slab_cache(kmalloc_caches + 8);
2514 p = kzalloc(512, GFP_KERNEL);
2517 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2518 validate_slab_cache(kmalloc_caches + 9);
2521 static void resiliency_test(void) {};
2525 * These are not as efficient as kmalloc for the non debug case.
2526 * We do not have the page struct available so we have to touch one
2527 * cacheline in struct kmem_cache to check slab flags.
2529 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2531 struct kmem_cache *s = get_slab(size, gfpflags);
2536 return slab_alloc(s, gfpflags, -1, caller);
2539 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2540 int node, void *caller)
2542 struct kmem_cache *s = get_slab(size, gfpflags);
2547 return slab_alloc(s, gfpflags, node, caller);
2552 static int validate_slab(struct kmem_cache *s, struct page *page)
2555 void *addr = page_address(page);
2556 unsigned long map[BITS_TO_LONGS(s->objects)];
2558 if (!check_slab(s, page) ||
2559 !on_freelist(s, page, NULL))
2562 /* Now we know that a valid freelist exists */
2563 bitmap_zero(map, s->objects);
2565 for(p = page->freelist; p; p = get_freepointer(s, p)) {
2566 set_bit((p - addr) / s->size, map);
2567 if (!check_object(s, page, p, 0))
2571 for(p = addr; p < addr + s->objects * s->size; p += s->size)
2572 if (!test_bit((p - addr) / s->size, map))
2573 if (!check_object(s, page, p, 1))
2578 static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2580 if (slab_trylock(page)) {
2581 validate_slab(s, page);
2584 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2587 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2588 if (!PageError(page))
2589 printk(KERN_ERR "SLUB %s: PageError not set "
2590 "on slab 0x%p\n", s->name, page);
2592 if (PageError(page))
2593 printk(KERN_ERR "SLUB %s: PageError set on "
2594 "slab 0x%p\n", s->name, page);
2598 static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2600 unsigned long count = 0;
2602 unsigned long flags;
2604 spin_lock_irqsave(&n->list_lock, flags);
2606 list_for_each_entry(page, &n->partial, lru) {
2607 validate_slab_slab(s, page);
2610 if (count != n->nr_partial)
2611 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2612 "counter=%ld\n", s->name, count, n->nr_partial);
2614 if (!(s->flags & SLAB_STORE_USER))
2617 list_for_each_entry(page, &n->full, lru) {
2618 validate_slab_slab(s, page);
2621 if (count != atomic_long_read(&n->nr_slabs))
2622 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2623 "counter=%ld\n", s->name, count,
2624 atomic_long_read(&n->nr_slabs));
2627 spin_unlock_irqrestore(&n->list_lock, flags);
2631 static unsigned long validate_slab_cache(struct kmem_cache *s)
2634 unsigned long count = 0;
2637 for_each_online_node(node) {
2638 struct kmem_cache_node *n = get_node(s, node);
2640 count += validate_slab_node(s, n);
2645 static unsigned long count_partial(struct kmem_cache_node *n)
2647 unsigned long flags;
2648 unsigned long x = 0;
2651 spin_lock_irqsave(&n->list_lock, flags);
2652 list_for_each_entry(page, &n->partial, lru)
2654 spin_unlock_irqrestore(&n->list_lock, flags);
2658 enum slab_stat_type {
2665 #define SO_FULL (1 << SL_FULL)
2666 #define SO_PARTIAL (1 << SL_PARTIAL)
2667 #define SO_CPU (1 << SL_CPU)
2668 #define SO_OBJECTS (1 << SL_OBJECTS)
2670 static unsigned long slab_objects(struct kmem_cache *s,
2671 char *buf, unsigned long flags)
2673 unsigned long total = 0;
2677 unsigned long *nodes;
2678 unsigned long *per_cpu;
2680 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
2681 per_cpu = nodes + nr_node_ids;
2683 for_each_possible_cpu(cpu) {
2684 struct page *page = s->cpu_slab[cpu];
2688 node = page_to_nid(page);
2689 if (flags & SO_CPU) {
2692 if (flags & SO_OBJECTS)
2703 for_each_online_node(node) {
2704 struct kmem_cache_node *n = get_node(s, node);
2706 if (flags & SO_PARTIAL) {
2707 if (flags & SO_OBJECTS)
2708 x = count_partial(n);
2715 if (flags & SO_FULL) {
2716 int full_slabs = atomic_read(&n->nr_slabs)
2720 if (flags & SO_OBJECTS)
2721 x = full_slabs * s->objects;
2729 x = sprintf(buf, "%lu", total);
2731 for_each_online_node(node)
2733 x += sprintf(buf + x, " N%d=%lu",
2737 return x + sprintf(buf + x, "\n");
2740 static int any_slab_objects(struct kmem_cache *s)
2745 for_each_possible_cpu(cpu)
2746 if (s->cpu_slab[cpu])
2749 for_each_node(node) {
2750 struct kmem_cache_node *n = get_node(s, node);
2752 if (n->nr_partial || atomic_read(&n->nr_slabs))
2758 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2759 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
2761 struct slab_attribute {
2762 struct attribute attr;
2763 ssize_t (*show)(struct kmem_cache *s, char *buf);
2764 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
2767 #define SLAB_ATTR_RO(_name) \
2768 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
2770 #define SLAB_ATTR(_name) \
2771 static struct slab_attribute _name##_attr = \
2772 __ATTR(_name, 0644, _name##_show, _name##_store)
2774 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
2776 return sprintf(buf, "%d\n", s->size);
2778 SLAB_ATTR_RO(slab_size);
2780 static ssize_t align_show(struct kmem_cache *s, char *buf)
2782 return sprintf(buf, "%d\n", s->align);
2784 SLAB_ATTR_RO(align);
2786 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
2788 return sprintf(buf, "%d\n", s->objsize);
2790 SLAB_ATTR_RO(object_size);
2792 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
2794 return sprintf(buf, "%d\n", s->objects);
2796 SLAB_ATTR_RO(objs_per_slab);
2798 static ssize_t order_show(struct kmem_cache *s, char *buf)
2800 return sprintf(buf, "%d\n", s->order);
2802 SLAB_ATTR_RO(order);
2804 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
2807 int n = sprint_symbol(buf, (unsigned long)s->ctor);
2809 return n + sprintf(buf + n, "\n");
2815 static ssize_t dtor_show(struct kmem_cache *s, char *buf)
2818 int n = sprint_symbol(buf, (unsigned long)s->dtor);
2820 return n + sprintf(buf + n, "\n");
2826 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
2828 return sprintf(buf, "%d\n", s->refcount - 1);
2830 SLAB_ATTR_RO(aliases);
2832 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
2834 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
2836 SLAB_ATTR_RO(slabs);
2838 static ssize_t partial_show(struct kmem_cache *s, char *buf)
2840 return slab_objects(s, buf, SO_PARTIAL);
2842 SLAB_ATTR_RO(partial);
2844 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
2846 return slab_objects(s, buf, SO_CPU);
2848 SLAB_ATTR_RO(cpu_slabs);
2850 static ssize_t objects_show(struct kmem_cache *s, char *buf)
2852 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
2854 SLAB_ATTR_RO(objects);
2856 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
2858 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
2861 static ssize_t sanity_checks_store(struct kmem_cache *s,
2862 const char *buf, size_t length)
2864 s->flags &= ~SLAB_DEBUG_FREE;
2866 s->flags |= SLAB_DEBUG_FREE;
2869 SLAB_ATTR(sanity_checks);
2871 static ssize_t trace_show(struct kmem_cache *s, char *buf)
2873 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
2876 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
2879 s->flags &= ~SLAB_TRACE;
2881 s->flags |= SLAB_TRACE;
2886 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
2888 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
2891 static ssize_t reclaim_account_store(struct kmem_cache *s,
2892 const char *buf, size_t length)
2894 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
2896 s->flags |= SLAB_RECLAIM_ACCOUNT;
2899 SLAB_ATTR(reclaim_account);
2901 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
2903 return sprintf(buf, "%d\n", !!(s->flags &
2904 (SLAB_HWCACHE_ALIGN|SLAB_MUST_HWCACHE_ALIGN)));
2906 SLAB_ATTR_RO(hwcache_align);
2908 #ifdef CONFIG_ZONE_DMA
2909 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
2911 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
2913 SLAB_ATTR_RO(cache_dma);
2916 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
2918 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
2920 SLAB_ATTR_RO(destroy_by_rcu);
2922 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
2924 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
2927 static ssize_t red_zone_store(struct kmem_cache *s,
2928 const char *buf, size_t length)
2930 if (any_slab_objects(s))
2933 s->flags &= ~SLAB_RED_ZONE;
2935 s->flags |= SLAB_RED_ZONE;
2939 SLAB_ATTR(red_zone);
2941 static ssize_t poison_show(struct kmem_cache *s, char *buf)
2943 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
2946 static ssize_t poison_store(struct kmem_cache *s,
2947 const char *buf, size_t length)
2949 if (any_slab_objects(s))
2952 s->flags &= ~SLAB_POISON;
2954 s->flags |= SLAB_POISON;
2960 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
2962 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
2965 static ssize_t store_user_store(struct kmem_cache *s,
2966 const char *buf, size_t length)
2968 if (any_slab_objects(s))
2971 s->flags &= ~SLAB_STORE_USER;
2973 s->flags |= SLAB_STORE_USER;
2977 SLAB_ATTR(store_user);
2979 static ssize_t validate_show(struct kmem_cache *s, char *buf)
2984 static ssize_t validate_store(struct kmem_cache *s,
2985 const char *buf, size_t length)
2988 validate_slab_cache(s);
2993 SLAB_ATTR(validate);
2996 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
2998 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3001 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3002 const char *buf, size_t length)
3004 int n = simple_strtoul(buf, NULL, 10);
3007 s->defrag_ratio = n * 10;
3010 SLAB_ATTR(defrag_ratio);
3013 static struct attribute * slab_attrs[] = {
3014 &slab_size_attr.attr,
3015 &object_size_attr.attr,
3016 &objs_per_slab_attr.attr,
3021 &cpu_slabs_attr.attr,
3026 &sanity_checks_attr.attr,
3028 &hwcache_align_attr.attr,
3029 &reclaim_account_attr.attr,
3030 &destroy_by_rcu_attr.attr,
3031 &red_zone_attr.attr,
3033 &store_user_attr.attr,
3034 &validate_attr.attr,
3035 #ifdef CONFIG_ZONE_DMA
3036 &cache_dma_attr.attr,
3039 &defrag_ratio_attr.attr,
3044 static struct attribute_group slab_attr_group = {
3045 .attrs = slab_attrs,
3048 static ssize_t slab_attr_show(struct kobject *kobj,
3049 struct attribute *attr,
3052 struct slab_attribute *attribute;
3053 struct kmem_cache *s;
3056 attribute = to_slab_attr(attr);
3059 if (!attribute->show)
3062 err = attribute->show(s, buf);
3067 static ssize_t slab_attr_store(struct kobject *kobj,
3068 struct attribute *attr,
3069 const char *buf, size_t len)
3071 struct slab_attribute *attribute;
3072 struct kmem_cache *s;
3075 attribute = to_slab_attr(attr);
3078 if (!attribute->store)
3081 err = attribute->store(s, buf, len);
3086 static struct sysfs_ops slab_sysfs_ops = {
3087 .show = slab_attr_show,
3088 .store = slab_attr_store,
3091 static struct kobj_type slab_ktype = {
3092 .sysfs_ops = &slab_sysfs_ops,
3095 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3097 struct kobj_type *ktype = get_ktype(kobj);
3099 if (ktype == &slab_ktype)
3104 static struct kset_uevent_ops slab_uevent_ops = {
3105 .filter = uevent_filter,
3108 decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3110 #define ID_STR_LENGTH 64
3112 /* Create a unique string id for a slab cache:
3114 * :[flags-]size:[memory address of kmemcache]
3116 static char *create_unique_id(struct kmem_cache *s)
3118 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3125 * First flags affecting slabcache operations. We will only
3126 * get here for aliasable slabs so we do not need to support
3127 * too many flags. The flags here must cover all flags that
3128 * are matched during merging to guarantee that the id is
3131 if (s->flags & SLAB_CACHE_DMA)
3133 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3135 if (s->flags & SLAB_DEBUG_FREE)
3139 p += sprintf(p, "%07d", s->size);
3140 BUG_ON(p > name + ID_STR_LENGTH - 1);
3144 static int sysfs_slab_add(struct kmem_cache *s)
3150 if (slab_state < SYSFS)
3151 /* Defer until later */
3154 unmergeable = slab_unmergeable(s);
3157 * Slabcache can never be merged so we can use the name proper.
3158 * This is typically the case for debug situations. In that
3159 * case we can catch duplicate names easily.
3161 sysfs_remove_link(&slab_subsys.kset.kobj, s->name);
3165 * Create a unique name for the slab as a target
3168 name = create_unique_id(s);
3171 kobj_set_kset_s(s, slab_subsys);
3172 kobject_set_name(&s->kobj, name);
3173 kobject_init(&s->kobj);
3174 err = kobject_add(&s->kobj);
3178 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3181 kobject_uevent(&s->kobj, KOBJ_ADD);
3183 /* Setup first alias */
3184 sysfs_slab_alias(s, s->name);
3190 static void sysfs_slab_remove(struct kmem_cache *s)
3192 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3193 kobject_del(&s->kobj);
3197 * Need to buffer aliases during bootup until sysfs becomes
3198 * available lest we loose that information.
3200 struct saved_alias {
3201 struct kmem_cache *s;
3203 struct saved_alias *next;
3206 struct saved_alias *alias_list;
3208 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3210 struct saved_alias *al;
3212 if (slab_state == SYSFS) {
3214 * If we have a leftover link then remove it.
3216 sysfs_remove_link(&slab_subsys.kset.kobj, name);
3217 return sysfs_create_link(&slab_subsys.kset.kobj,
3221 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3227 al->next = alias_list;
3232 static int __init slab_sysfs_init(void)
3236 err = subsystem_register(&slab_subsys);
3238 printk(KERN_ERR "Cannot register slab subsystem.\n");
3244 while (alias_list) {
3245 struct saved_alias *al = alias_list;
3247 alias_list = alias_list->next;
3248 err = sysfs_slab_alias(al->s, al->name);
3257 __initcall(slab_sysfs_init);
3259 __initcall(finish_bootstrap);