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 * - Variable sizing of the per node arrays
103 /* Enable to test recovery from slab corruption on boot */
104 #undef SLUB_RESILIENCY_TEST
109 * Small page size. Make sure that we do not fragment memory
111 #define DEFAULT_MAX_ORDER 1
112 #define DEFAULT_MIN_OBJECTS 4
117 * Large page machines are customarily able to handle larger
120 #define DEFAULT_MAX_ORDER 2
121 #define DEFAULT_MIN_OBJECTS 8
126 * Mininum number of partial slabs. These will be left on the partial
127 * lists even if they are empty. kmem_cache_shrink may reclaim them.
129 #define MIN_PARTIAL 2
132 * Maximum number of desirable partial slabs.
133 * The existence of more partial slabs makes kmem_cache_shrink
134 * sort the partial list by the number of objects in the.
136 #define MAX_PARTIAL 10
138 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
139 SLAB_POISON | SLAB_STORE_USER)
141 * Set of flags that will prevent slab merging
143 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
144 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
146 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
149 #ifndef ARCH_KMALLOC_MINALIGN
150 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
153 #ifndef ARCH_SLAB_MINALIGN
154 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
157 /* Internal SLUB flags */
158 #define __OBJECT_POISON 0x80000000 /* Poison object */
160 static int kmem_size = sizeof(struct kmem_cache);
163 static struct notifier_block slab_notifier;
167 DOWN, /* No slab functionality available */
168 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
169 UP, /* Everything works */
173 /* A list of all slab caches on the system */
174 static DECLARE_RWSEM(slub_lock);
175 LIST_HEAD(slab_caches);
178 static int sysfs_slab_add(struct kmem_cache *);
179 static int sysfs_slab_alias(struct kmem_cache *, const char *);
180 static void sysfs_slab_remove(struct kmem_cache *);
182 static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
183 static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
184 static void sysfs_slab_remove(struct kmem_cache *s) {}
187 /********************************************************************
188 * Core slab cache functions
189 *******************************************************************/
191 int slab_is_available(void)
193 return slab_state >= UP;
196 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
199 return s->node[node];
201 return &s->local_node;
208 static void print_section(char *text, u8 *addr, unsigned int length)
216 for (i = 0; i < length; i++) {
218 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
221 printk(" %02x", addr[i]);
223 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
225 printk(" %s\n",ascii);
236 printk(" %s\n", ascii);
241 * Slow version of get and set free pointer.
243 * This requires touching the cache lines of kmem_cache.
244 * The offset can also be obtained from the page. In that
245 * case it is in the cacheline that we already need to touch.
247 static void *get_freepointer(struct kmem_cache *s, void *object)
249 return *(void **)(object + s->offset);
252 static void set_freepointer(struct kmem_cache *s, void *object, void *fp)
254 *(void **)(object + s->offset) = fp;
258 * Tracking user of a slab.
261 void *addr; /* Called from address */
262 int cpu; /* Was running on cpu */
263 int pid; /* Pid context */
264 unsigned long when; /* When did the operation occur */
267 enum track_item { TRACK_ALLOC, TRACK_FREE };
269 static struct track *get_track(struct kmem_cache *s, void *object,
270 enum track_item alloc)
275 p = object + s->offset + sizeof(void *);
277 p = object + s->inuse;
282 static void set_track(struct kmem_cache *s, void *object,
283 enum track_item alloc, void *addr)
288 p = object + s->offset + sizeof(void *);
290 p = object + s->inuse;
295 p->cpu = smp_processor_id();
296 p->pid = current ? current->pid : -1;
299 memset(p, 0, sizeof(struct track));
302 static void init_tracking(struct kmem_cache *s, void *object)
304 if (s->flags & SLAB_STORE_USER) {
305 set_track(s, object, TRACK_FREE, NULL);
306 set_track(s, object, TRACK_ALLOC, NULL);
310 static void print_track(const char *s, struct track *t)
315 printk(KERN_ERR "%s: ", s);
316 __print_symbol("%s", (unsigned long)t->addr);
317 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
320 static void print_trailer(struct kmem_cache *s, u8 *p)
322 unsigned int off; /* Offset of last byte */
324 if (s->flags & SLAB_RED_ZONE)
325 print_section("Redzone", p + s->objsize,
326 s->inuse - s->objsize);
328 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
330 get_freepointer(s, p));
333 off = s->offset + sizeof(void *);
337 if (s->flags & SLAB_STORE_USER) {
338 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
339 print_track("Last free ", get_track(s, p, TRACK_FREE));
340 off += 2 * sizeof(struct track);
344 /* Beginning of the filler is the free pointer */
345 print_section("Filler", p + off, s->size - off);
348 static void object_err(struct kmem_cache *s, struct page *page,
349 u8 *object, char *reason)
351 u8 *addr = page_address(page);
353 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
354 s->name, reason, object, page);
355 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
356 object - addr, page->flags, page->inuse, page->freelist);
357 if (object > addr + 16)
358 print_section("Bytes b4", object - 16, 16);
359 print_section("Object", object, min(s->objsize, 128));
360 print_trailer(s, object);
364 static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
369 va_start(args, reason);
370 vsnprintf(buf, sizeof(buf), reason, args);
372 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
377 static void init_object(struct kmem_cache *s, void *object, int active)
381 if (s->flags & __OBJECT_POISON) {
382 memset(p, POISON_FREE, s->objsize - 1);
383 p[s->objsize -1] = POISON_END;
386 if (s->flags & SLAB_RED_ZONE)
387 memset(p + s->objsize,
388 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
389 s->inuse - s->objsize);
392 static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
395 if (*start != (u8)value)
404 static int check_valid_pointer(struct kmem_cache *s, struct page *page,
412 base = page_address(page);
413 if (object < base || object >= base + s->objects * s->size ||
414 (object - base) % s->size) {
425 * Bytes of the object to be managed.
426 * If the freepointer may overlay the object then the free
427 * pointer is the first word of the object.
428 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
431 * object + s->objsize
432 * Padding to reach word boundary. This is also used for Redzoning.
433 * Padding is extended to word size if Redzoning is enabled
434 * and objsize == inuse.
435 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
436 * 0xcc (RED_ACTIVE) for objects in use.
439 * A. Free pointer (if we cannot overwrite object on free)
440 * B. Tracking data for SLAB_STORE_USER
441 * C. Padding to reach required alignment boundary
442 * Padding is done using 0x5a (POISON_INUSE)
446 * If slabcaches are merged then the objsize and inuse boundaries are to
447 * be ignored. And therefore no slab options that rely on these boundaries
448 * may be used with merged slabcaches.
451 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
452 void *from, void *to)
454 printk(KERN_ERR "@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
455 s->name, message, data, from, to - 1);
456 memset(from, data, to - from);
459 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
461 unsigned long off = s->inuse; /* The end of info */
464 /* Freepointer is placed after the object. */
465 off += sizeof(void *);
467 if (s->flags & SLAB_STORE_USER)
468 /* We also have user information there */
469 off += 2 * sizeof(struct track);
474 if (check_bytes(p + off, POISON_INUSE, s->size - off))
477 object_err(s, page, p, "Object padding check fails");
482 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
486 static int slab_pad_check(struct kmem_cache *s, struct page *page)
489 int length, remainder;
491 if (!(s->flags & SLAB_POISON))
494 p = page_address(page);
495 length = s->objects * s->size;
496 remainder = (PAGE_SIZE << s->order) - length;
500 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
501 slab_err(s, page, "Padding check failed");
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 slab_err(s, page, "Not a valid slab page flags=%lx "
588 "mapping=0x%p count=%d", page->flags, page->mapping,
592 if (page->offset * sizeof(void *) != s->offset) {
593 slab_err(s, page, "Corrupted offset %lu flags=0x%lx "
594 "mapping=0x%p count=%d",
595 (unsigned long)(page->offset * sizeof(void *)),
601 if (page->inuse > s->objects) {
602 slab_err(s, page, "inuse %u > max %u @0x%p flags=%lx "
603 "mapping=0x%p count=%d",
604 s->name, page->inuse, s->objects, page->flags,
605 page->mapping, page_count(page));
608 /* Slab_pad_check fixes things up after itself */
609 slab_pad_check(s, page);
614 * Determine if a certain object on a page is on the freelist and
615 * therefore free. Must hold the slab lock for cpu slabs to
616 * guarantee that the chains are consistent.
618 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
621 void *fp = page->freelist;
624 while (fp && nr <= s->objects) {
627 if (!check_valid_pointer(s, page, fp)) {
629 object_err(s, page, object,
630 "Freechain corrupt");
631 set_freepointer(s, object, NULL);
634 slab_err(s, page, "Freepointer 0x%p corrupt",
636 page->freelist = NULL;
637 page->inuse = s->objects;
638 printk(KERN_ERR "@@@ SLUB %s: Freelist "
639 "cleared. Slab 0x%p\n",
646 fp = get_freepointer(s, object);
650 if (page->inuse != s->objects - nr) {
651 slab_err(s, page, "Wrong object count. Counter is %d but "
652 "counted were %d", s, page, page->inuse,
654 page->inuse = s->objects - nr;
655 printk(KERN_ERR "@@@ SLUB %s: Object count adjusted. "
656 "Slab @0x%p\n", s->name, page);
658 return search == NULL;
662 * Tracking of fully allocated slabs for debugging
664 static void add_full(struct kmem_cache_node *n, struct page *page)
666 spin_lock(&n->list_lock);
667 list_add(&page->lru, &n->full);
668 spin_unlock(&n->list_lock);
671 static void remove_full(struct kmem_cache *s, struct page *page)
673 struct kmem_cache_node *n;
675 if (!(s->flags & SLAB_STORE_USER))
678 n = get_node(s, page_to_nid(page));
680 spin_lock(&n->list_lock);
681 list_del(&page->lru);
682 spin_unlock(&n->list_lock);
685 static int alloc_object_checks(struct kmem_cache *s, struct page *page,
688 if (!check_slab(s, page))
691 if (object && !on_freelist(s, page, object)) {
692 slab_err(s, page, "Object 0x%p already allocated", object);
696 if (!check_valid_pointer(s, page, object)) {
697 object_err(s, page, object, "Freelist Pointer check fails");
704 if (!check_object(s, page, object, 0))
709 if (PageSlab(page)) {
711 * If this is a slab page then lets do the best we can
712 * to avoid issues in the future. Marking all objects
713 * as used avoids touching the remainder.
715 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
717 page->inuse = s->objects;
718 page->freelist = NULL;
719 /* Fix up fields that may be corrupted */
720 page->offset = s->offset / sizeof(void *);
725 static int free_object_checks(struct kmem_cache *s, struct page *page,
728 if (!check_slab(s, page))
731 if (!check_valid_pointer(s, page, object)) {
732 slab_err(s, page, "Invalid object pointer 0x%p", object);
736 if (on_freelist(s, page, object)) {
737 slab_err(s, page, "Object 0x%p already free", object);
741 if (!check_object(s, page, object, 1))
744 if (unlikely(s != page->slab)) {
746 slab_err(s, page, "Attempt to free object(0x%p) "
747 "outside of slab", object);
751 "SLUB <none>: no slab for object 0x%p.\n",
756 slab_err(s, page, "object at 0x%p belongs "
757 "to slab %s", object, page->slab->name);
762 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
763 s->name, page, object);
768 * Slab allocation and freeing
770 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
773 int pages = 1 << s->order;
778 if (s->flags & SLAB_CACHE_DMA)
782 page = alloc_pages(flags, s->order);
784 page = alloc_pages_node(node, flags, s->order);
789 mod_zone_page_state(page_zone(page),
790 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
791 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
797 static void setup_object(struct kmem_cache *s, struct page *page,
800 if (PageError(page)) {
801 init_object(s, object, 0);
802 init_tracking(s, object);
805 if (unlikely(s->ctor)) {
806 int mode = SLAB_CTOR_CONSTRUCTOR;
808 if (!(s->flags & __GFP_WAIT))
809 mode |= SLAB_CTOR_ATOMIC;
811 s->ctor(object, s, mode);
815 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
818 struct kmem_cache_node *n;
824 if (flags & __GFP_NO_GROW)
827 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
829 if (flags & __GFP_WAIT)
832 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
836 n = get_node(s, page_to_nid(page));
838 atomic_long_inc(&n->nr_slabs);
839 page->offset = s->offset / sizeof(void *);
841 page->flags |= 1 << PG_slab;
842 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
843 SLAB_STORE_USER | SLAB_TRACE))
844 page->flags |= 1 << PG_error;
846 start = page_address(page);
847 end = start + s->objects * s->size;
849 if (unlikely(s->flags & SLAB_POISON))
850 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
853 for (p = start + s->size; p < end; p += s->size) {
854 setup_object(s, page, last);
855 set_freepointer(s, last, p);
858 setup_object(s, page, last);
859 set_freepointer(s, last, NULL);
861 page->freelist = start;
864 if (flags & __GFP_WAIT)
869 static void __free_slab(struct kmem_cache *s, struct page *page)
871 int pages = 1 << s->order;
873 if (unlikely(PageError(page) || s->dtor)) {
874 void *start = page_address(page);
875 void *end = start + (pages << PAGE_SHIFT);
878 slab_pad_check(s, page);
879 for (p = start; p <= end - s->size; p += s->size) {
882 check_object(s, page, p, 0);
886 mod_zone_page_state(page_zone(page),
887 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
888 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
891 page->mapping = NULL;
892 __free_pages(page, s->order);
895 static void rcu_free_slab(struct rcu_head *h)
899 page = container_of((struct list_head *)h, struct page, lru);
900 __free_slab(page->slab, page);
903 static void free_slab(struct kmem_cache *s, struct page *page)
905 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
907 * RCU free overloads the RCU head over the LRU
909 struct rcu_head *head = (void *)&page->lru;
911 call_rcu(head, rcu_free_slab);
913 __free_slab(s, page);
916 static void discard_slab(struct kmem_cache *s, struct page *page)
918 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
920 atomic_long_dec(&n->nr_slabs);
921 reset_page_mapcount(page);
922 page->flags &= ~(1 << PG_slab | 1 << PG_error);
927 * Per slab locking using the pagelock
929 static __always_inline void slab_lock(struct page *page)
931 bit_spin_lock(PG_locked, &page->flags);
934 static __always_inline void slab_unlock(struct page *page)
936 bit_spin_unlock(PG_locked, &page->flags);
939 static __always_inline int slab_trylock(struct page *page)
943 rc = bit_spin_trylock(PG_locked, &page->flags);
948 * Management of partially allocated slabs
950 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
952 spin_lock(&n->list_lock);
954 list_add_tail(&page->lru, &n->partial);
955 spin_unlock(&n->list_lock);
958 static void add_partial(struct kmem_cache_node *n, struct page *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 > MIN_PARTIAL) {
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)
1093 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1098 add_partial(n, page);
1099 else if (PageError(page) && (s->flags & SLAB_STORE_USER))
1104 if (n->nr_partial < MIN_PARTIAL) {
1106 * Adding an empty page to the partial slabs in order
1107 * to avoid page allocator overhead. This page needs to
1108 * come after all the others that are not fully empty
1109 * in order to make sure that we do maximum
1112 add_partial_tail(n, page);
1116 discard_slab(s, page);
1122 * Remove the cpu slab
1124 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1126 s->cpu_slab[cpu] = NULL;
1127 ClearPageActive(page);
1129 putback_slab(s, page);
1132 static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1135 deactivate_slab(s, page, cpu);
1140 * Called from IPI handler with interrupts disabled.
1142 static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1144 struct page *page = s->cpu_slab[cpu];
1147 flush_slab(s, page, cpu);
1150 static void flush_cpu_slab(void *d)
1152 struct kmem_cache *s = d;
1153 int cpu = smp_processor_id();
1155 __flush_cpu_slab(s, cpu);
1158 static void flush_all(struct kmem_cache *s)
1161 on_each_cpu(flush_cpu_slab, s, 1, 1);
1163 unsigned long flags;
1165 local_irq_save(flags);
1167 local_irq_restore(flags);
1172 * slab_alloc is optimized to only modify two cachelines on the fast path
1173 * (aside from the stack):
1175 * 1. The page struct
1176 * 2. The first cacheline of the object to be allocated.
1178 * The only cache lines that are read (apart from code) is the
1179 * per cpu array in the kmem_cache struct.
1181 * Fastpath is not possible if we need to get a new slab or have
1182 * debugging enabled (which means all slabs are marked with PageError)
1184 static void *slab_alloc(struct kmem_cache *s,
1185 gfp_t gfpflags, int node, void *addr)
1189 unsigned long flags;
1192 local_irq_save(flags);
1193 cpu = smp_processor_id();
1194 page = s->cpu_slab[cpu];
1199 if (unlikely(node != -1 && page_to_nid(page) != node))
1202 object = page->freelist;
1203 if (unlikely(!object))
1205 if (unlikely(PageError(page)))
1210 page->freelist = object[page->offset];
1212 local_irq_restore(flags);
1216 deactivate_slab(s, page, cpu);
1219 page = get_partial(s, gfpflags, node);
1222 s->cpu_slab[cpu] = page;
1223 SetPageActive(page);
1227 page = new_slab(s, gfpflags, node);
1229 cpu = smp_processor_id();
1230 if (s->cpu_slab[cpu]) {
1232 * Someone else populated the cpu_slab while we enabled
1233 * interrupts, or we have got scheduled on another cpu.
1234 * The page may not be on the requested node.
1237 page_to_nid(s->cpu_slab[cpu]) == node) {
1239 * Current cpuslab is acceptable and we
1240 * want the current one since its cache hot
1242 discard_slab(s, page);
1243 page = s->cpu_slab[cpu];
1247 /* Dump the current slab */
1248 flush_slab(s, s->cpu_slab[cpu], cpu);
1253 local_irq_restore(flags);
1256 if (!alloc_object_checks(s, page, object))
1258 if (s->flags & SLAB_STORE_USER)
1259 set_track(s, object, TRACK_ALLOC, addr);
1260 if (s->flags & SLAB_TRACE) {
1261 printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
1262 s->name, object, page->inuse,
1266 init_object(s, object, 1);
1270 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1272 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1274 EXPORT_SYMBOL(kmem_cache_alloc);
1277 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1279 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1281 EXPORT_SYMBOL(kmem_cache_alloc_node);
1285 * The fastpath only writes the cacheline of the page struct and the first
1286 * cacheline of the object.
1288 * No special cachelines need to be read
1290 static void slab_free(struct kmem_cache *s, struct page *page,
1291 void *x, void *addr)
1294 void **object = (void *)x;
1295 unsigned long flags;
1297 local_irq_save(flags);
1300 if (unlikely(PageError(page)))
1303 prior = object[page->offset] = page->freelist;
1304 page->freelist = object;
1307 if (unlikely(PageActive(page)))
1309 * Cpu slabs are never on partial lists and are
1314 if (unlikely(!page->inuse))
1318 * Objects left in the slab. If it
1319 * was not on the partial list before
1322 if (unlikely(!prior))
1323 add_partial(get_node(s, page_to_nid(page)), page);
1327 local_irq_restore(flags);
1333 * Slab on the partial list.
1335 remove_partial(s, page);
1338 discard_slab(s, page);
1339 local_irq_restore(flags);
1343 if (!free_object_checks(s, page, x))
1345 if (!PageActive(page) && !page->freelist)
1346 remove_full(s, page);
1347 if (s->flags & SLAB_STORE_USER)
1348 set_track(s, x, TRACK_FREE, addr);
1349 if (s->flags & SLAB_TRACE) {
1350 printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n",
1351 s->name, object, page->inuse,
1353 print_section("Object", (void *)object, s->objsize);
1356 init_object(s, object, 0);
1360 void kmem_cache_free(struct kmem_cache *s, void *x)
1364 page = virt_to_head_page(x);
1366 slab_free(s, page, x, __builtin_return_address(0));
1368 EXPORT_SYMBOL(kmem_cache_free);
1370 /* Figure out on which slab object the object resides */
1371 static struct page *get_object_page(const void *x)
1373 struct page *page = virt_to_head_page(x);
1375 if (!PageSlab(page))
1382 * kmem_cache_open produces objects aligned at "size" and the first object
1383 * is placed at offset 0 in the slab (We have no metainformation on the
1384 * slab, all slabs are in essence "off slab").
1386 * In order to get the desired alignment one just needs to align the
1389 * Notice that the allocation order determines the sizes of the per cpu
1390 * caches. Each processor has always one slab available for allocations.
1391 * Increasing the allocation order reduces the number of times that slabs
1392 * must be moved on and off the partial lists and therefore may influence
1395 * The offset is used to relocate the free list link in each object. It is
1396 * therefore possible to move the free list link behind the object. This
1397 * is necessary for RCU to work properly and also useful for debugging.
1401 * Mininum / Maximum order of slab pages. This influences locking overhead
1402 * and slab fragmentation. A higher order reduces the number of partial slabs
1403 * and increases the number of allocations possible without having to
1404 * take the list_lock.
1406 static int slub_min_order;
1407 static int slub_max_order = DEFAULT_MAX_ORDER;
1410 * Minimum number of objects per slab. This is necessary in order to
1411 * reduce locking overhead. Similar to the queue size in SLAB.
1413 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1416 * Merge control. If this is set then no merging of slab caches will occur.
1418 static int slub_nomerge;
1423 static int slub_debug;
1425 static char *slub_debug_slabs;
1428 * Calculate the order of allocation given an slab object size.
1430 * The order of allocation has significant impact on other elements
1431 * of the system. Generally order 0 allocations should be preferred
1432 * since they do not cause fragmentation in the page allocator. Larger
1433 * objects may have problems with order 0 because there may be too much
1434 * space left unused in a slab. We go to a higher order if more than 1/8th
1435 * of the slab would be wasted.
1437 * In order to reach satisfactory performance we must ensure that
1438 * a minimum number of objects is in one slab. Otherwise we may
1439 * generate too much activity on the partial lists. This is less a
1440 * concern for large slabs though. slub_max_order specifies the order
1441 * where we begin to stop considering the number of objects in a slab.
1443 * Higher order allocations also allow the placement of more objects
1444 * in a slab and thereby reduce object handling overhead. If the user
1445 * has requested a higher mininum order then we start with that one
1448 static int calculate_order(int size)
1453 for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT);
1454 order < MAX_ORDER; order++) {
1455 unsigned long slab_size = PAGE_SIZE << order;
1457 if (slub_max_order > order &&
1458 slab_size < slub_min_objects * size)
1461 if (slab_size < size)
1464 rem = slab_size % size;
1466 if (rem <= (PAGE_SIZE << order) / 8)
1470 if (order >= MAX_ORDER)
1476 * Function to figure out which alignment to use from the
1477 * various ways of specifying it.
1479 static unsigned long calculate_alignment(unsigned long flags,
1480 unsigned long align, unsigned long size)
1483 * If the user wants hardware cache aligned objects then
1484 * follow that suggestion if the object is sufficiently
1487 * The hardware cache alignment cannot override the
1488 * specified alignment though. If that is greater
1491 if ((flags & SLAB_HWCACHE_ALIGN) &&
1492 size > L1_CACHE_BYTES / 2)
1493 return max_t(unsigned long, align, L1_CACHE_BYTES);
1495 if (align < ARCH_SLAB_MINALIGN)
1496 return ARCH_SLAB_MINALIGN;
1498 return ALIGN(align, sizeof(void *));
1501 static void init_kmem_cache_node(struct kmem_cache_node *n)
1504 atomic_long_set(&n->nr_slabs, 0);
1505 spin_lock_init(&n->list_lock);
1506 INIT_LIST_HEAD(&n->partial);
1507 INIT_LIST_HEAD(&n->full);
1512 * No kmalloc_node yet so do it by hand. We know that this is the first
1513 * slab on the node for this slabcache. There are no concurrent accesses
1516 * Note that this function only works on the kmalloc_node_cache
1517 * when allocating for the kmalloc_node_cache.
1519 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1523 struct kmem_cache_node *n;
1525 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1527 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1528 /* new_slab() disables interupts */
1534 page->freelist = get_freepointer(kmalloc_caches, n);
1536 kmalloc_caches->node[node] = n;
1537 init_object(kmalloc_caches, n, 1);
1538 init_kmem_cache_node(n);
1539 atomic_long_inc(&n->nr_slabs);
1540 add_partial(n, page);
1544 static void free_kmem_cache_nodes(struct kmem_cache *s)
1548 for_each_online_node(node) {
1549 struct kmem_cache_node *n = s->node[node];
1550 if (n && n != &s->local_node)
1551 kmem_cache_free(kmalloc_caches, n);
1552 s->node[node] = NULL;
1556 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1561 if (slab_state >= UP)
1562 local_node = page_to_nid(virt_to_page(s));
1566 for_each_online_node(node) {
1567 struct kmem_cache_node *n;
1569 if (local_node == node)
1572 if (slab_state == DOWN) {
1573 n = early_kmem_cache_node_alloc(gfpflags,
1577 n = kmem_cache_alloc_node(kmalloc_caches,
1581 free_kmem_cache_nodes(s);
1587 init_kmem_cache_node(n);
1592 static void free_kmem_cache_nodes(struct kmem_cache *s)
1596 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1598 init_kmem_cache_node(&s->local_node);
1604 * calculate_sizes() determines the order and the distribution of data within
1607 static int calculate_sizes(struct kmem_cache *s)
1609 unsigned long flags = s->flags;
1610 unsigned long size = s->objsize;
1611 unsigned long align = s->align;
1614 * Determine if we can poison the object itself. If the user of
1615 * the slab may touch the object after free or before allocation
1616 * then we should never poison the object itself.
1618 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1619 !s->ctor && !s->dtor)
1620 s->flags |= __OBJECT_POISON;
1622 s->flags &= ~__OBJECT_POISON;
1625 * Round up object size to the next word boundary. We can only
1626 * place the free pointer at word boundaries and this determines
1627 * the possible location of the free pointer.
1629 size = ALIGN(size, sizeof(void *));
1632 * If we are redzoning then check if there is some space between the
1633 * end of the object and the free pointer. If not then add an
1634 * additional word, so that we can establish a redzone between
1635 * the object and the freepointer to be able to check for overwrites.
1637 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1638 size += sizeof(void *);
1641 * With that we have determined how much of the slab is in actual
1642 * use by the object. This is the potential offset to the free
1647 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1648 s->ctor || s->dtor)) {
1650 * Relocate free pointer after the object if it is not
1651 * permitted to overwrite the first word of the object on
1654 * This is the case if we do RCU, have a constructor or
1655 * destructor or are poisoning the objects.
1658 size += sizeof(void *);
1661 if (flags & SLAB_STORE_USER)
1663 * Need to store information about allocs and frees after
1666 size += 2 * sizeof(struct track);
1668 if (flags & DEBUG_DEFAULT_FLAGS)
1670 * Add some empty padding so that we can catch
1671 * overwrites from earlier objects rather than let
1672 * tracking information or the free pointer be
1673 * corrupted if an user writes before the start
1676 size += sizeof(void *);
1678 * Determine the alignment based on various parameters that the
1679 * user specified (this is unecessarily complex due to the attempt
1680 * to be compatible with SLAB. Should be cleaned up some day).
1682 align = calculate_alignment(flags, align, s->objsize);
1685 * SLUB stores one object immediately after another beginning from
1686 * offset 0. In order to align the objects we have to simply size
1687 * each object to conform to the alignment.
1689 size = ALIGN(size, align);
1692 s->order = calculate_order(size);
1697 * Determine the number of objects per slab
1699 s->objects = (PAGE_SIZE << s->order) / size;
1702 * Verify that the number of objects is within permitted limits.
1703 * The page->inuse field is only 16 bit wide! So we cannot have
1704 * more than 64k objects per slab.
1706 if (!s->objects || s->objects > 65535)
1712 static int __init finish_bootstrap(void)
1714 struct list_head *h;
1719 list_for_each(h, &slab_caches) {
1720 struct kmem_cache *s =
1721 container_of(h, struct kmem_cache, list);
1723 err = sysfs_slab_add(s);
1729 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1730 const char *name, size_t size,
1731 size_t align, unsigned long flags,
1732 void (*ctor)(void *, struct kmem_cache *, unsigned long),
1733 void (*dtor)(void *, struct kmem_cache *, unsigned long))
1735 memset(s, 0, kmem_size);
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 n->nr_partial -= 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 * kmem_cache_shrink removes empty slabs from the partial lists
2121 * and then sorts the partially allocated slabs by the number
2122 * of items in use. The slabs with the most items in use
2123 * come first. New allocations will remove these from the
2124 * partial list because they are full. The slabs with the
2125 * least items are placed last. If it happens that the objects
2126 * are freed then the page can be returned to the page allocator.
2128 int kmem_cache_shrink(struct kmem_cache *s)
2132 struct kmem_cache_node *n;
2135 struct list_head *slabs_by_inuse =
2136 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2137 unsigned long flags;
2139 if (!slabs_by_inuse)
2143 for_each_online_node(node) {
2144 n = get_node(s, node);
2149 for (i = 0; i < s->objects; i++)
2150 INIT_LIST_HEAD(slabs_by_inuse + i);
2152 spin_lock_irqsave(&n->list_lock, flags);
2155 * Build lists indexed by the items in use in
2156 * each slab or free slabs if empty.
2158 * Note that concurrent frees may occur while
2159 * we hold the list_lock. page->inuse here is
2162 list_for_each_entry_safe(page, t, &n->partial, lru) {
2163 if (!page->inuse && slab_trylock(page)) {
2165 * Must hold slab lock here because slab_free
2166 * may have freed the last object and be
2167 * waiting to release the slab.
2169 list_del(&page->lru);
2172 discard_slab(s, page);
2174 if (n->nr_partial > MAX_PARTIAL)
2175 list_move(&page->lru,
2176 slabs_by_inuse + page->inuse);
2180 if (n->nr_partial <= MAX_PARTIAL)
2184 * Rebuild the partial list with the slabs filled up
2185 * most first and the least used slabs at the end.
2187 for (i = s->objects - 1; i >= 0; i--)
2188 list_splice(slabs_by_inuse + i, n->partial.prev);
2191 spin_unlock_irqrestore(&n->list_lock, flags);
2194 kfree(slabs_by_inuse);
2197 EXPORT_SYMBOL(kmem_cache_shrink);
2200 * krealloc - reallocate memory. The contents will remain unchanged.
2202 * @p: object to reallocate memory for.
2203 * @new_size: how many bytes of memory are required.
2204 * @flags: the type of memory to allocate.
2206 * The contents of the object pointed to are preserved up to the
2207 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2208 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2209 * %NULL pointer, the object pointed to is freed.
2211 void *krealloc(const void *p, size_t new_size, gfp_t flags)
2213 struct kmem_cache *new_cache;
2218 return kmalloc(new_size, flags);
2220 if (unlikely(!new_size)) {
2225 page = virt_to_head_page(p);
2227 new_cache = get_slab(new_size, flags);
2230 * If new size fits in the current cache, bail out.
2232 if (likely(page->slab == new_cache))
2235 ret = kmalloc(new_size, flags);
2237 memcpy(ret, p, min(new_size, ksize(p)));
2242 EXPORT_SYMBOL(krealloc);
2244 /********************************************************************
2245 * Basic setup of slabs
2246 *******************************************************************/
2248 void __init kmem_cache_init(void)
2254 * Must first have the slab cache available for the allocations of the
2255 * struct kmalloc_cache_node's. There is special bootstrap code in
2256 * kmem_cache_open for slab_state == DOWN.
2258 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2259 sizeof(struct kmem_cache_node), GFP_KERNEL);
2262 /* Able to allocate the per node structures */
2263 slab_state = PARTIAL;
2265 /* Caches that are not of the two-to-the-power-of size */
2266 create_kmalloc_cache(&kmalloc_caches[1],
2267 "kmalloc-96", 96, GFP_KERNEL);
2268 create_kmalloc_cache(&kmalloc_caches[2],
2269 "kmalloc-192", 192, GFP_KERNEL);
2271 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2272 create_kmalloc_cache(&kmalloc_caches[i],
2273 "kmalloc", 1 << i, GFP_KERNEL);
2277 /* Provide the correct kmalloc names now that the caches are up */
2278 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2279 kmalloc_caches[i]. name =
2280 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2283 register_cpu_notifier(&slab_notifier);
2286 if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */
2287 kmem_size = offsetof(struct kmem_cache, cpu_slab)
2288 + nr_cpu_ids * sizeof(struct page *);
2290 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2291 " Processors=%d, Nodes=%d\n",
2292 KMALLOC_SHIFT_HIGH, L1_CACHE_BYTES,
2293 slub_min_order, slub_max_order, slub_min_objects,
2294 nr_cpu_ids, nr_node_ids);
2298 * Find a mergeable slab cache
2300 static int slab_unmergeable(struct kmem_cache *s)
2302 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2305 if (s->ctor || s->dtor)
2311 static struct kmem_cache *find_mergeable(size_t size,
2312 size_t align, unsigned long flags,
2313 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2314 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2316 struct list_head *h;
2318 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2324 size = ALIGN(size, sizeof(void *));
2325 align = calculate_alignment(flags, align, size);
2326 size = ALIGN(size, align);
2328 list_for_each(h, &slab_caches) {
2329 struct kmem_cache *s =
2330 container_of(h, struct kmem_cache, list);
2332 if (slab_unmergeable(s))
2338 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2339 (s->flags & SLUB_MERGE_SAME))
2342 * Check if alignment is compatible.
2343 * Courtesy of Adrian Drzewiecki
2345 if ((s->size & ~(align -1)) != s->size)
2348 if (s->size - size >= sizeof(void *))
2356 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2357 size_t align, unsigned long flags,
2358 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2359 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2361 struct kmem_cache *s;
2363 down_write(&slub_lock);
2364 s = find_mergeable(size, align, flags, dtor, ctor);
2368 * Adjust the object sizes so that we clear
2369 * the complete object on kzalloc.
2371 s->objsize = max(s->objsize, (int)size);
2372 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2373 if (sysfs_slab_alias(s, name))
2376 s = kmalloc(kmem_size, GFP_KERNEL);
2377 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2378 size, align, flags, ctor, dtor)) {
2379 if (sysfs_slab_add(s)) {
2383 list_add(&s->list, &slab_caches);
2387 up_write(&slub_lock);
2391 up_write(&slub_lock);
2392 if (flags & SLAB_PANIC)
2393 panic("Cannot create slabcache %s\n", name);
2398 EXPORT_SYMBOL(kmem_cache_create);
2400 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2404 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
2406 memset(x, 0, s->objsize);
2409 EXPORT_SYMBOL(kmem_cache_zalloc);
2412 static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2414 struct list_head *h;
2416 down_read(&slub_lock);
2417 list_for_each(h, &slab_caches) {
2418 struct kmem_cache *s =
2419 container_of(h, struct kmem_cache, list);
2423 up_read(&slub_lock);
2427 * Use the cpu notifier to insure that the slab are flushed
2430 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2431 unsigned long action, void *hcpu)
2433 long cpu = (long)hcpu;
2436 case CPU_UP_CANCELED:
2438 for_all_slabs(__flush_cpu_slab, cpu);
2446 static struct notifier_block __cpuinitdata slab_notifier =
2447 { &slab_cpuup_callback, NULL, 0 };
2453 /*****************************************************************
2454 * Generic reaper used to support the page allocator
2455 * (the cpu slabs are reaped by a per slab workqueue).
2457 * Maybe move this to the page allocator?
2458 ****************************************************************/
2460 static DEFINE_PER_CPU(unsigned long, reap_node);
2462 static void init_reap_node(int cpu)
2466 node = next_node(cpu_to_node(cpu), node_online_map);
2467 if (node == MAX_NUMNODES)
2468 node = first_node(node_online_map);
2470 __get_cpu_var(reap_node) = node;
2473 static void next_reap_node(void)
2475 int node = __get_cpu_var(reap_node);
2478 * Also drain per cpu pages on remote zones
2480 if (node != numa_node_id())
2481 drain_node_pages(node);
2483 node = next_node(node, node_online_map);
2484 if (unlikely(node >= MAX_NUMNODES))
2485 node = first_node(node_online_map);
2486 __get_cpu_var(reap_node) = node;
2489 #define init_reap_node(cpu) do { } while (0)
2490 #define next_reap_node(void) do { } while (0)
2493 #define REAPTIMEOUT_CPUC (2*HZ)
2496 static DEFINE_PER_CPU(struct delayed_work, reap_work);
2498 static void cache_reap(struct work_struct *unused)
2501 refresh_cpu_vm_stats(smp_processor_id());
2502 schedule_delayed_work(&__get_cpu_var(reap_work),
2506 static void __devinit start_cpu_timer(int cpu)
2508 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
2511 * When this gets called from do_initcalls via cpucache_init(),
2512 * init_workqueues() has already run, so keventd will be setup
2515 if (keventd_up() && reap_work->work.func == NULL) {
2516 init_reap_node(cpu);
2517 INIT_DELAYED_WORK(reap_work, cache_reap);
2518 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
2522 static int __init cpucache_init(void)
2527 * Register the timers that drain pcp pages and update vm statistics
2529 for_each_online_cpu(cpu)
2530 start_cpu_timer(cpu);
2533 __initcall(cpucache_init);
2536 #ifdef SLUB_RESILIENCY_TEST
2537 static unsigned long validate_slab_cache(struct kmem_cache *s);
2539 static void resiliency_test(void)
2543 printk(KERN_ERR "SLUB resiliency testing\n");
2544 printk(KERN_ERR "-----------------------\n");
2545 printk(KERN_ERR "A. Corruption after allocation\n");
2547 p = kzalloc(16, GFP_KERNEL);
2549 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2550 " 0x12->0x%p\n\n", p + 16);
2552 validate_slab_cache(kmalloc_caches + 4);
2554 /* Hmmm... The next two are dangerous */
2555 p = kzalloc(32, GFP_KERNEL);
2556 p[32 + sizeof(void *)] = 0x34;
2557 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2558 " 0x34 -> -0x%p\n", p);
2559 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2561 validate_slab_cache(kmalloc_caches + 5);
2562 p = kzalloc(64, GFP_KERNEL);
2563 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2565 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2567 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2568 validate_slab_cache(kmalloc_caches + 6);
2570 printk(KERN_ERR "\nB. Corruption after free\n");
2571 p = kzalloc(128, GFP_KERNEL);
2574 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2575 validate_slab_cache(kmalloc_caches + 7);
2577 p = kzalloc(256, GFP_KERNEL);
2580 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2581 validate_slab_cache(kmalloc_caches + 8);
2583 p = kzalloc(512, GFP_KERNEL);
2586 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2587 validate_slab_cache(kmalloc_caches + 9);
2590 static void resiliency_test(void) {};
2594 * These are not as efficient as kmalloc for the non debug case.
2595 * We do not have the page struct available so we have to touch one
2596 * cacheline in struct kmem_cache to check slab flags.
2598 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2600 struct kmem_cache *s = get_slab(size, gfpflags);
2605 return slab_alloc(s, gfpflags, -1, caller);
2608 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2609 int node, void *caller)
2611 struct kmem_cache *s = get_slab(size, gfpflags);
2616 return slab_alloc(s, gfpflags, node, caller);
2621 static int validate_slab(struct kmem_cache *s, struct page *page)
2624 void *addr = page_address(page);
2625 unsigned long map[BITS_TO_LONGS(s->objects)];
2627 if (!check_slab(s, page) ||
2628 !on_freelist(s, page, NULL))
2631 /* Now we know that a valid freelist exists */
2632 bitmap_zero(map, s->objects);
2634 for(p = page->freelist; p; p = get_freepointer(s, p)) {
2635 set_bit((p - addr) / s->size, map);
2636 if (!check_object(s, page, p, 0))
2640 for(p = addr; p < addr + s->objects * s->size; p += s->size)
2641 if (!test_bit((p - addr) / s->size, map))
2642 if (!check_object(s, page, p, 1))
2647 static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2649 if (slab_trylock(page)) {
2650 validate_slab(s, page);
2653 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2656 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2657 if (!PageError(page))
2658 printk(KERN_ERR "SLUB %s: PageError not set "
2659 "on slab 0x%p\n", s->name, page);
2661 if (PageError(page))
2662 printk(KERN_ERR "SLUB %s: PageError set on "
2663 "slab 0x%p\n", s->name, page);
2667 static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2669 unsigned long count = 0;
2671 unsigned long flags;
2673 spin_lock_irqsave(&n->list_lock, flags);
2675 list_for_each_entry(page, &n->partial, lru) {
2676 validate_slab_slab(s, page);
2679 if (count != n->nr_partial)
2680 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2681 "counter=%ld\n", s->name, count, n->nr_partial);
2683 if (!(s->flags & SLAB_STORE_USER))
2686 list_for_each_entry(page, &n->full, lru) {
2687 validate_slab_slab(s, page);
2690 if (count != atomic_long_read(&n->nr_slabs))
2691 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2692 "counter=%ld\n", s->name, count,
2693 atomic_long_read(&n->nr_slabs));
2696 spin_unlock_irqrestore(&n->list_lock, flags);
2700 static unsigned long validate_slab_cache(struct kmem_cache *s)
2703 unsigned long count = 0;
2706 for_each_online_node(node) {
2707 struct kmem_cache_node *n = get_node(s, node);
2709 count += validate_slab_node(s, n);
2715 * Generate lists of locations where slabcache objects are allocated
2720 unsigned long count;
2726 unsigned long count;
2727 struct location *loc;
2730 static void free_loc_track(struct loc_track *t)
2733 free_pages((unsigned long)t->loc,
2734 get_order(sizeof(struct location) * t->max));
2737 static int alloc_loc_track(struct loc_track *t, unsigned long max)
2743 max = PAGE_SIZE / sizeof(struct location);
2745 order = get_order(sizeof(struct location) * max);
2747 l = (void *)__get_free_pages(GFP_KERNEL, order);
2753 memcpy(l, t->loc, sizeof(struct location) * t->count);
2761 static int add_location(struct loc_track *t, struct kmem_cache *s,
2764 long start, end, pos;
2772 pos = start + (end - start + 1) / 2;
2775 * There is nothing at "end". If we end up there
2776 * we need to add something to before end.
2781 caddr = t->loc[pos].addr;
2782 if (addr == caddr) {
2783 t->loc[pos].count++;
2794 * Not found. Insert new tracking element
2796 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max))
2802 (t->count - pos) * sizeof(struct location));
2809 static void process_slab(struct loc_track *t, struct kmem_cache *s,
2810 struct page *page, enum track_item alloc)
2812 void *addr = page_address(page);
2813 unsigned long map[BITS_TO_LONGS(s->objects)];
2816 bitmap_zero(map, s->objects);
2817 for (p = page->freelist; p; p = get_freepointer(s, p))
2818 set_bit((p - addr) / s->size, map);
2820 for (p = addr; p < addr + s->objects * s->size; p += s->size)
2821 if (!test_bit((p - addr) / s->size, map)) {
2822 void *addr = get_track(s, p, alloc)->addr;
2824 add_location(t, s, addr);
2828 static int list_locations(struct kmem_cache *s, char *buf,
2829 enum track_item alloc)
2839 /* Push back cpu slabs */
2842 for_each_online_node(node) {
2843 struct kmem_cache_node *n = get_node(s, node);
2844 unsigned long flags;
2847 if (!atomic_read(&n->nr_slabs))
2850 spin_lock_irqsave(&n->list_lock, flags);
2851 list_for_each_entry(page, &n->partial, lru)
2852 process_slab(&t, s, page, alloc);
2853 list_for_each_entry(page, &n->full, lru)
2854 process_slab(&t, s, page, alloc);
2855 spin_unlock_irqrestore(&n->list_lock, flags);
2858 for (i = 0; i < t.count; i++) {
2859 void *addr = t.loc[i].addr;
2861 if (n > PAGE_SIZE - 100)
2863 n += sprintf(buf + n, "%7ld ", t.loc[i].count);
2865 n += sprint_symbol(buf + n, (unsigned long)t.loc[i].addr);
2867 n += sprintf(buf + n, "<not-available>");
2868 n += sprintf(buf + n, "\n");
2873 n += sprintf(buf, "No data\n");
2877 static unsigned long count_partial(struct kmem_cache_node *n)
2879 unsigned long flags;
2880 unsigned long x = 0;
2883 spin_lock_irqsave(&n->list_lock, flags);
2884 list_for_each_entry(page, &n->partial, lru)
2886 spin_unlock_irqrestore(&n->list_lock, flags);
2890 enum slab_stat_type {
2897 #define SO_FULL (1 << SL_FULL)
2898 #define SO_PARTIAL (1 << SL_PARTIAL)
2899 #define SO_CPU (1 << SL_CPU)
2900 #define SO_OBJECTS (1 << SL_OBJECTS)
2902 static unsigned long slab_objects(struct kmem_cache *s,
2903 char *buf, unsigned long flags)
2905 unsigned long total = 0;
2909 unsigned long *nodes;
2910 unsigned long *per_cpu;
2912 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
2913 per_cpu = nodes + nr_node_ids;
2915 for_each_possible_cpu(cpu) {
2916 struct page *page = s->cpu_slab[cpu];
2920 node = page_to_nid(page);
2921 if (flags & SO_CPU) {
2924 if (flags & SO_OBJECTS)
2935 for_each_online_node(node) {
2936 struct kmem_cache_node *n = get_node(s, node);
2938 if (flags & SO_PARTIAL) {
2939 if (flags & SO_OBJECTS)
2940 x = count_partial(n);
2947 if (flags & SO_FULL) {
2948 int full_slabs = atomic_read(&n->nr_slabs)
2952 if (flags & SO_OBJECTS)
2953 x = full_slabs * s->objects;
2961 x = sprintf(buf, "%lu", total);
2963 for_each_online_node(node)
2965 x += sprintf(buf + x, " N%d=%lu",
2969 return x + sprintf(buf + x, "\n");
2972 static int any_slab_objects(struct kmem_cache *s)
2977 for_each_possible_cpu(cpu)
2978 if (s->cpu_slab[cpu])
2981 for_each_node(node) {
2982 struct kmem_cache_node *n = get_node(s, node);
2984 if (n->nr_partial || atomic_read(&n->nr_slabs))
2990 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2991 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
2993 struct slab_attribute {
2994 struct attribute attr;
2995 ssize_t (*show)(struct kmem_cache *s, char *buf);
2996 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
2999 #define SLAB_ATTR_RO(_name) \
3000 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3002 #define SLAB_ATTR(_name) \
3003 static struct slab_attribute _name##_attr = \
3004 __ATTR(_name, 0644, _name##_show, _name##_store)
3006 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3008 return sprintf(buf, "%d\n", s->size);
3010 SLAB_ATTR_RO(slab_size);
3012 static ssize_t align_show(struct kmem_cache *s, char *buf)
3014 return sprintf(buf, "%d\n", s->align);
3016 SLAB_ATTR_RO(align);
3018 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3020 return sprintf(buf, "%d\n", s->objsize);
3022 SLAB_ATTR_RO(object_size);
3024 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3026 return sprintf(buf, "%d\n", s->objects);
3028 SLAB_ATTR_RO(objs_per_slab);
3030 static ssize_t order_show(struct kmem_cache *s, char *buf)
3032 return sprintf(buf, "%d\n", s->order);
3034 SLAB_ATTR_RO(order);
3036 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3039 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3041 return n + sprintf(buf + n, "\n");
3047 static ssize_t dtor_show(struct kmem_cache *s, char *buf)
3050 int n = sprint_symbol(buf, (unsigned long)s->dtor);
3052 return n + sprintf(buf + n, "\n");
3058 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3060 return sprintf(buf, "%d\n", s->refcount - 1);
3062 SLAB_ATTR_RO(aliases);
3064 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3066 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3068 SLAB_ATTR_RO(slabs);
3070 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3072 return slab_objects(s, buf, SO_PARTIAL);
3074 SLAB_ATTR_RO(partial);
3076 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3078 return slab_objects(s, buf, SO_CPU);
3080 SLAB_ATTR_RO(cpu_slabs);
3082 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3084 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3086 SLAB_ATTR_RO(objects);
3088 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3090 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3093 static ssize_t sanity_checks_store(struct kmem_cache *s,
3094 const char *buf, size_t length)
3096 s->flags &= ~SLAB_DEBUG_FREE;
3098 s->flags |= SLAB_DEBUG_FREE;
3101 SLAB_ATTR(sanity_checks);
3103 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3105 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3108 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3111 s->flags &= ~SLAB_TRACE;
3113 s->flags |= SLAB_TRACE;
3118 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3123 static ssize_t reclaim_account_store(struct kmem_cache *s,
3124 const char *buf, size_t length)
3126 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3128 s->flags |= SLAB_RECLAIM_ACCOUNT;
3131 SLAB_ATTR(reclaim_account);
3133 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3135 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3137 SLAB_ATTR_RO(hwcache_align);
3139 #ifdef CONFIG_ZONE_DMA
3140 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3142 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3144 SLAB_ATTR_RO(cache_dma);
3147 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3149 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3151 SLAB_ATTR_RO(destroy_by_rcu);
3153 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3155 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3158 static ssize_t red_zone_store(struct kmem_cache *s,
3159 const char *buf, size_t length)
3161 if (any_slab_objects(s))
3164 s->flags &= ~SLAB_RED_ZONE;
3166 s->flags |= SLAB_RED_ZONE;
3170 SLAB_ATTR(red_zone);
3172 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3174 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3177 static ssize_t poison_store(struct kmem_cache *s,
3178 const char *buf, size_t length)
3180 if (any_slab_objects(s))
3183 s->flags &= ~SLAB_POISON;
3185 s->flags |= SLAB_POISON;
3191 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3193 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3196 static ssize_t store_user_store(struct kmem_cache *s,
3197 const char *buf, size_t length)
3199 if (any_slab_objects(s))
3202 s->flags &= ~SLAB_STORE_USER;
3204 s->flags |= SLAB_STORE_USER;
3208 SLAB_ATTR(store_user);
3210 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3215 static ssize_t validate_store(struct kmem_cache *s,
3216 const char *buf, size_t length)
3219 validate_slab_cache(s);
3224 SLAB_ATTR(validate);
3226 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3231 static ssize_t shrink_store(struct kmem_cache *s,
3232 const char *buf, size_t length)
3234 if (buf[0] == '1') {
3235 int rc = kmem_cache_shrink(s);
3245 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3247 if (!(s->flags & SLAB_STORE_USER))
3249 return list_locations(s, buf, TRACK_ALLOC);
3251 SLAB_ATTR_RO(alloc_calls);
3253 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3255 if (!(s->flags & SLAB_STORE_USER))
3257 return list_locations(s, buf, TRACK_FREE);
3259 SLAB_ATTR_RO(free_calls);
3262 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3264 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3267 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3268 const char *buf, size_t length)
3270 int n = simple_strtoul(buf, NULL, 10);
3273 s->defrag_ratio = n * 10;
3276 SLAB_ATTR(defrag_ratio);
3279 static struct attribute * slab_attrs[] = {
3280 &slab_size_attr.attr,
3281 &object_size_attr.attr,
3282 &objs_per_slab_attr.attr,
3287 &cpu_slabs_attr.attr,
3292 &sanity_checks_attr.attr,
3294 &hwcache_align_attr.attr,
3295 &reclaim_account_attr.attr,
3296 &destroy_by_rcu_attr.attr,
3297 &red_zone_attr.attr,
3299 &store_user_attr.attr,
3300 &validate_attr.attr,
3302 &alloc_calls_attr.attr,
3303 &free_calls_attr.attr,
3304 #ifdef CONFIG_ZONE_DMA
3305 &cache_dma_attr.attr,
3308 &defrag_ratio_attr.attr,
3313 static struct attribute_group slab_attr_group = {
3314 .attrs = slab_attrs,
3317 static ssize_t slab_attr_show(struct kobject *kobj,
3318 struct attribute *attr,
3321 struct slab_attribute *attribute;
3322 struct kmem_cache *s;
3325 attribute = to_slab_attr(attr);
3328 if (!attribute->show)
3331 err = attribute->show(s, buf);
3336 static ssize_t slab_attr_store(struct kobject *kobj,
3337 struct attribute *attr,
3338 const char *buf, size_t len)
3340 struct slab_attribute *attribute;
3341 struct kmem_cache *s;
3344 attribute = to_slab_attr(attr);
3347 if (!attribute->store)
3350 err = attribute->store(s, buf, len);
3355 static struct sysfs_ops slab_sysfs_ops = {
3356 .show = slab_attr_show,
3357 .store = slab_attr_store,
3360 static struct kobj_type slab_ktype = {
3361 .sysfs_ops = &slab_sysfs_ops,
3364 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3366 struct kobj_type *ktype = get_ktype(kobj);
3368 if (ktype == &slab_ktype)
3373 static struct kset_uevent_ops slab_uevent_ops = {
3374 .filter = uevent_filter,
3377 decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3379 #define ID_STR_LENGTH 64
3381 /* Create a unique string id for a slab cache:
3383 * :[flags-]size:[memory address of kmemcache]
3385 static char *create_unique_id(struct kmem_cache *s)
3387 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3394 * First flags affecting slabcache operations. We will only
3395 * get here for aliasable slabs so we do not need to support
3396 * too many flags. The flags here must cover all flags that
3397 * are matched during merging to guarantee that the id is
3400 if (s->flags & SLAB_CACHE_DMA)
3402 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3404 if (s->flags & SLAB_DEBUG_FREE)
3408 p += sprintf(p, "%07d", s->size);
3409 BUG_ON(p > name + ID_STR_LENGTH - 1);
3413 static int sysfs_slab_add(struct kmem_cache *s)
3419 if (slab_state < SYSFS)
3420 /* Defer until later */
3423 unmergeable = slab_unmergeable(s);
3426 * Slabcache can never be merged so we can use the name proper.
3427 * This is typically the case for debug situations. In that
3428 * case we can catch duplicate names easily.
3430 sysfs_remove_link(&slab_subsys.kset.kobj, s->name);
3434 * Create a unique name for the slab as a target
3437 name = create_unique_id(s);
3440 kobj_set_kset_s(s, slab_subsys);
3441 kobject_set_name(&s->kobj, name);
3442 kobject_init(&s->kobj);
3443 err = kobject_add(&s->kobj);
3447 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3450 kobject_uevent(&s->kobj, KOBJ_ADD);
3452 /* Setup first alias */
3453 sysfs_slab_alias(s, s->name);
3459 static void sysfs_slab_remove(struct kmem_cache *s)
3461 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3462 kobject_del(&s->kobj);
3466 * Need to buffer aliases during bootup until sysfs becomes
3467 * available lest we loose that information.
3469 struct saved_alias {
3470 struct kmem_cache *s;
3472 struct saved_alias *next;
3475 struct saved_alias *alias_list;
3477 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3479 struct saved_alias *al;
3481 if (slab_state == SYSFS) {
3483 * If we have a leftover link then remove it.
3485 sysfs_remove_link(&slab_subsys.kset.kobj, name);
3486 return sysfs_create_link(&slab_subsys.kset.kobj,
3490 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3496 al->next = alias_list;
3501 static int __init slab_sysfs_init(void)
3505 err = subsystem_register(&slab_subsys);
3507 printk(KERN_ERR "Cannot register slab subsystem.\n");
3513 while (alias_list) {
3514 struct saved_alias *al = alias_list;
3516 alias_list = alias_list->next;
3517 err = sysfs_slab_alias(al->s, al->name);
3526 __initcall(slab_sysfs_init);
3528 __initcall(finish_bootstrap);