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 and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
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 frozen and exempt from list processing.
82 * This means that the slab is dedicated to a purpose
83 * such as satisfying allocations for a specific
84 * processor. Objects may be freed in the slab while
85 * it is frozen but slab_free will then skip the usual
86 * list operations. It is up to the processor holding
87 * the slab to integrate the slab into the slab lists
88 * when the slab is no longer needed.
90 * One use of this flag is to mark slabs that are
91 * used for allocations. Then such a slab becomes a cpu
92 * slab. The cpu slab may be equipped with an additional
93 * lockless_freelist that allows lockless access to
94 * free objects in addition to the regular freelist
95 * that requires the slab lock.
97 * PageError Slab requires special handling due to debug
98 * options set. This moves slab handling out of
99 * the fast path and disables lockless freelists.
102 #define FROZEN (1 << PG_active)
104 #ifdef CONFIG_SLUB_DEBUG
105 #define SLABDEBUG (1 << PG_error)
110 static inline int SlabFrozen(struct page *page)
112 return page->flags & FROZEN;
115 static inline void SetSlabFrozen(struct page *page)
117 page->flags |= FROZEN;
120 static inline void ClearSlabFrozen(struct page *page)
122 page->flags &= ~FROZEN;
125 static inline int SlabDebug(struct page *page)
127 return page->flags & SLABDEBUG;
130 static inline void SetSlabDebug(struct page *page)
132 page->flags |= SLABDEBUG;
135 static inline void ClearSlabDebug(struct page *page)
137 page->flags &= ~SLABDEBUG;
141 * Issues still to be resolved:
143 * - The per cpu array is updated for each new slab and and is a remote
144 * cacheline for most nodes. This could become a bouncing cacheline given
145 * enough frequent updates. There are 16 pointers in a cacheline, so at
146 * max 16 cpus could compete for the cacheline which may be okay.
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
159 * Small page size. Make sure that we do not fragment memory
161 #define DEFAULT_MAX_ORDER 1
162 #define DEFAULT_MIN_OBJECTS 4
167 * Large page machines are customarily able to handle larger
170 #define DEFAULT_MAX_ORDER 2
171 #define DEFAULT_MIN_OBJECTS 8
176 * Mininum number of partial slabs. These will be left on the partial
177 * lists even if they are empty. kmem_cache_shrink may reclaim them.
179 #define MIN_PARTIAL 2
182 * Maximum number of desirable partial slabs.
183 * The existence of more partial slabs makes kmem_cache_shrink
184 * sort the partial list by the number of objects in the.
186 #define MAX_PARTIAL 10
188 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
189 SLAB_POISON | SLAB_STORE_USER)
192 * Set of flags that will prevent slab merging
194 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
195 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
197 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
200 #ifndef ARCH_KMALLOC_MINALIGN
201 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
204 #ifndef ARCH_SLAB_MINALIGN
205 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
208 /* Internal SLUB flags */
209 #define __OBJECT_POISON 0x80000000 /* Poison object */
211 /* Not all arches define cache_line_size */
212 #ifndef cache_line_size
213 #define cache_line_size() L1_CACHE_BYTES
216 static int kmem_size = sizeof(struct kmem_cache);
219 static struct notifier_block slab_notifier;
223 DOWN, /* No slab functionality available */
224 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
225 UP, /* Everything works but does not show up in sysfs */
229 /* A list of all slab caches on the system */
230 static DECLARE_RWSEM(slub_lock);
231 LIST_HEAD(slab_caches);
234 * Tracking user of a slab.
237 void *addr; /* Called from address */
238 int cpu; /* Was running on cpu */
239 int pid; /* Pid context */
240 unsigned long when; /* When did the operation occur */
243 enum track_item { TRACK_ALLOC, TRACK_FREE };
245 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
246 static int sysfs_slab_add(struct kmem_cache *);
247 static int sysfs_slab_alias(struct kmem_cache *, const char *);
248 static void sysfs_slab_remove(struct kmem_cache *);
250 static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251 static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
252 static void sysfs_slab_remove(struct kmem_cache *s) {}
255 /********************************************************************
256 * Core slab cache functions
257 *******************************************************************/
259 int slab_is_available(void)
261 return slab_state >= UP;
264 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
267 return s->node[node];
269 return &s->local_node;
273 static inline int check_valid_pointer(struct kmem_cache *s,
274 struct page *page, const void *object)
281 base = page_address(page);
282 if (object < base || object >= base + s->objects * s->size ||
283 (object - base) % s->size) {
291 * Slow version of get and set free pointer.
293 * This version requires touching the cache lines of kmem_cache which
294 * we avoid to do in the fast alloc free paths. There we obtain the offset
295 * from the page struct.
297 static inline void *get_freepointer(struct kmem_cache *s, void *object)
299 return *(void **)(object + s->offset);
302 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
304 *(void **)(object + s->offset) = fp;
307 /* Loop over all objects in a slab */
308 #define for_each_object(__p, __s, __addr) \
309 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
313 #define for_each_free_object(__p, __s, __free) \
314 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
316 /* Determine object index from a given position */
317 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
319 return (p - addr) / s->size;
322 #ifdef CONFIG_SLUB_DEBUG
326 static int slub_debug;
328 static char *slub_debug_slabs;
333 static void print_section(char *text, u8 *addr, unsigned int length)
341 for (i = 0; i < length; i++) {
343 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
346 printk(" %02x", addr[i]);
348 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
350 printk(" %s\n",ascii);
361 printk(" %s\n", ascii);
365 static struct track *get_track(struct kmem_cache *s, void *object,
366 enum track_item alloc)
371 p = object + s->offset + sizeof(void *);
373 p = object + s->inuse;
378 static void set_track(struct kmem_cache *s, void *object,
379 enum track_item alloc, void *addr)
384 p = object + s->offset + sizeof(void *);
386 p = object + s->inuse;
391 p->cpu = smp_processor_id();
392 p->pid = current ? current->pid : -1;
395 memset(p, 0, sizeof(struct track));
398 static void init_tracking(struct kmem_cache *s, void *object)
400 if (s->flags & SLAB_STORE_USER) {
401 set_track(s, object, TRACK_FREE, NULL);
402 set_track(s, object, TRACK_ALLOC, NULL);
406 static void print_track(const char *s, struct track *t)
411 printk(KERN_ERR "%s: ", s);
412 __print_symbol("%s", (unsigned long)t->addr);
413 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
416 static void print_trailer(struct kmem_cache *s, u8 *p)
418 unsigned int off; /* Offset of last byte */
420 if (s->flags & SLAB_RED_ZONE)
421 print_section("Redzone", p + s->objsize,
422 s->inuse - s->objsize);
424 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
426 get_freepointer(s, p));
429 off = s->offset + sizeof(void *);
433 if (s->flags & SLAB_STORE_USER) {
434 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
435 print_track("Last free ", get_track(s, p, TRACK_FREE));
436 off += 2 * sizeof(struct track);
440 /* Beginning of the filler is the free pointer */
441 print_section("Filler", p + off, s->size - off);
444 static void object_err(struct kmem_cache *s, struct page *page,
445 u8 *object, char *reason)
447 u8 *addr = page_address(page);
449 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
450 s->name, reason, object, page);
451 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
452 object - addr, page->flags, page->inuse, page->freelist);
453 if (object > addr + 16)
454 print_section("Bytes b4", object - 16, 16);
455 print_section("Object", object, min(s->objsize, 128));
456 print_trailer(s, object);
460 static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
465 va_start(args, reason);
466 vsnprintf(buf, sizeof(buf), reason, args);
468 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
473 static void init_object(struct kmem_cache *s, void *object, int active)
477 if (s->flags & __OBJECT_POISON) {
478 memset(p, POISON_FREE, s->objsize - 1);
479 p[s->objsize -1] = POISON_END;
482 if (s->flags & SLAB_RED_ZONE)
483 memset(p + s->objsize,
484 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
485 s->inuse - s->objsize);
488 static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
491 if (*start != (u8)value)
503 * Bytes of the object to be managed.
504 * If the freepointer may overlay the object then the free
505 * pointer is the first word of the object.
507 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
510 * object + s->objsize
511 * Padding to reach word boundary. This is also used for Redzoning.
512 * Padding is extended by another word if Redzoning is enabled and
515 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
516 * 0xcc (RED_ACTIVE) for objects in use.
519 * Meta data starts here.
521 * A. Free pointer (if we cannot overwrite object on free)
522 * B. Tracking data for SLAB_STORE_USER
523 * C. Padding to reach required alignment boundary or at mininum
524 * one word if debuggin is on to be able to detect writes
525 * before the word boundary.
527 * Padding is done using 0x5a (POISON_INUSE)
530 * Nothing is used beyond s->size.
532 * If slabcaches are merged then the objsize and inuse boundaries are mostly
533 * ignored. And therefore no slab options that rely on these boundaries
534 * may be used with merged slabcaches.
537 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
538 void *from, void *to)
540 printk(KERN_ERR "@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
541 s->name, message, data, from, to - 1);
542 memset(from, data, to - from);
545 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
547 unsigned long off = s->inuse; /* The end of info */
550 /* Freepointer is placed after the object. */
551 off += sizeof(void *);
553 if (s->flags & SLAB_STORE_USER)
554 /* We also have user information there */
555 off += 2 * sizeof(struct track);
560 if (check_bytes(p + off, POISON_INUSE, s->size - off))
563 object_err(s, page, p, "Object padding check fails");
568 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
572 static int slab_pad_check(struct kmem_cache *s, struct page *page)
575 int length, remainder;
577 if (!(s->flags & SLAB_POISON))
580 p = page_address(page);
581 length = s->objects * s->size;
582 remainder = (PAGE_SIZE << s->order) - length;
586 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
587 slab_err(s, page, "Padding check failed");
588 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
589 p + length + remainder);
595 static int check_object(struct kmem_cache *s, struct page *page,
596 void *object, int active)
599 u8 *endobject = object + s->objsize;
601 if (s->flags & SLAB_RED_ZONE) {
603 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
605 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
606 object_err(s, page, object,
607 active ? "Redzone Active" : "Redzone Inactive");
608 restore_bytes(s, "redzone", red,
609 endobject, object + s->inuse);
613 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
614 !check_bytes(endobject, POISON_INUSE,
615 s->inuse - s->objsize)) {
616 object_err(s, page, p, "Alignment padding check fails");
618 * Fix it so that there will not be another report.
620 * Hmmm... We may be corrupting an object that now expects
621 * to be longer than allowed.
623 restore_bytes(s, "alignment padding", POISON_INUSE,
624 endobject, object + s->inuse);
628 if (s->flags & SLAB_POISON) {
629 if (!active && (s->flags & __OBJECT_POISON) &&
630 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
631 p[s->objsize - 1] != POISON_END)) {
633 object_err(s, page, p, "Poison check failed");
634 restore_bytes(s, "Poison", POISON_FREE,
635 p, p + s->objsize -1);
636 restore_bytes(s, "Poison", POISON_END,
637 p + s->objsize - 1, p + s->objsize);
641 * check_pad_bytes cleans up on its own.
643 check_pad_bytes(s, page, p);
646 if (!s->offset && active)
648 * Object and freepointer overlap. Cannot check
649 * freepointer while object is allocated.
653 /* Check free pointer validity */
654 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
655 object_err(s, page, p, "Freepointer corrupt");
657 * No choice but to zap it and thus loose the remainder
658 * of the free objects in this slab. May cause
659 * another error because the object count is now wrong.
661 set_freepointer(s, p, NULL);
667 static int check_slab(struct kmem_cache *s, struct page *page)
669 VM_BUG_ON(!irqs_disabled());
671 if (!PageSlab(page)) {
672 slab_err(s, page, "Not a valid slab page flags=%lx "
673 "mapping=0x%p count=%d", page->flags, page->mapping,
677 if (page->offset * sizeof(void *) != s->offset) {
678 slab_err(s, page, "Corrupted offset %lu flags=0x%lx "
679 "mapping=0x%p count=%d",
680 (unsigned long)(page->offset * sizeof(void *)),
686 if (page->inuse > s->objects) {
687 slab_err(s, page, "inuse %u > max %u @0x%p flags=%lx "
688 "mapping=0x%p count=%d",
689 s->name, page->inuse, s->objects, page->flags,
690 page->mapping, page_count(page));
693 /* Slab_pad_check fixes things up after itself */
694 slab_pad_check(s, page);
699 * Determine if a certain object on a page is on the freelist. Must hold the
700 * slab lock to guarantee that the chains are in a consistent state.
702 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
705 void *fp = page->freelist;
708 while (fp && nr <= s->objects) {
711 if (!check_valid_pointer(s, page, fp)) {
713 object_err(s, page, object,
714 "Freechain corrupt");
715 set_freepointer(s, object, NULL);
718 slab_err(s, page, "Freepointer 0x%p corrupt",
720 page->freelist = NULL;
721 page->inuse = s->objects;
722 printk(KERN_ERR "@@@ SLUB %s: Freelist "
723 "cleared. Slab 0x%p\n",
730 fp = get_freepointer(s, object);
734 if (page->inuse != s->objects - nr) {
735 slab_err(s, page, "Wrong object count. Counter is %d but "
736 "counted were %d", s, page, page->inuse,
738 page->inuse = s->objects - nr;
739 printk(KERN_ERR "@@@ SLUB %s: Object count adjusted. "
740 "Slab @0x%p\n", s->name, page);
742 return search == NULL;
745 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
747 if (s->flags & SLAB_TRACE) {
748 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
750 alloc ? "alloc" : "free",
755 print_section("Object", (void *)object, s->objsize);
762 * Tracking of fully allocated slabs for debugging purposes.
764 static void add_full(struct kmem_cache_node *n, struct page *page)
766 spin_lock(&n->list_lock);
767 list_add(&page->lru, &n->full);
768 spin_unlock(&n->list_lock);
771 static void remove_full(struct kmem_cache *s, struct page *page)
773 struct kmem_cache_node *n;
775 if (!(s->flags & SLAB_STORE_USER))
778 n = get_node(s, page_to_nid(page));
780 spin_lock(&n->list_lock);
781 list_del(&page->lru);
782 spin_unlock(&n->list_lock);
785 static void setup_object_debug(struct kmem_cache *s, struct page *page,
788 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
791 init_object(s, object, 0);
792 init_tracking(s, object);
795 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
796 void *object, void *addr)
798 if (!check_slab(s, page))
801 if (object && !on_freelist(s, page, object)) {
802 slab_err(s, page, "Object 0x%p already allocated", object);
806 if (!check_valid_pointer(s, page, object)) {
807 object_err(s, page, object, "Freelist Pointer check fails");
811 if (object && !check_object(s, page, object, 0))
814 /* Success perform special debug activities for allocs */
815 if (s->flags & SLAB_STORE_USER)
816 set_track(s, object, TRACK_ALLOC, addr);
817 trace(s, page, object, 1);
818 init_object(s, object, 1);
822 if (PageSlab(page)) {
824 * If this is a slab page then lets do the best we can
825 * to avoid issues in the future. Marking all objects
826 * as used avoids touching the remaining objects.
828 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
830 page->inuse = s->objects;
831 page->freelist = NULL;
832 /* Fix up fields that may be corrupted */
833 page->offset = s->offset / sizeof(void *);
838 static int free_debug_processing(struct kmem_cache *s, struct page *page,
839 void *object, void *addr)
841 if (!check_slab(s, page))
844 if (!check_valid_pointer(s, page, object)) {
845 slab_err(s, page, "Invalid object pointer 0x%p", object);
849 if (on_freelist(s, page, object)) {
850 slab_err(s, page, "Object 0x%p already free", object);
854 if (!check_object(s, page, object, 1))
857 if (unlikely(s != page->slab)) {
859 slab_err(s, page, "Attempt to free object(0x%p) "
860 "outside of slab", object);
864 "SLUB <none>: no slab for object 0x%p.\n",
869 slab_err(s, page, "object at 0x%p belongs "
870 "to slab %s", object, page->slab->name);
874 /* Special debug activities for freeing objects */
875 if (!SlabFrozen(page) && !page->freelist)
876 remove_full(s, page);
877 if (s->flags & SLAB_STORE_USER)
878 set_track(s, object, TRACK_FREE, addr);
879 trace(s, page, object, 0);
880 init_object(s, object, 0);
884 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
885 s->name, page, object);
889 static int __init setup_slub_debug(char *str)
891 if (!str || *str != '=')
892 slub_debug = DEBUG_DEFAULT_FLAGS;
895 if (*str == 0 || *str == ',')
896 slub_debug = DEBUG_DEFAULT_FLAGS;
898 for( ;*str && *str != ','; str++)
900 case 'f' : case 'F' :
901 slub_debug |= SLAB_DEBUG_FREE;
903 case 'z' : case 'Z' :
904 slub_debug |= SLAB_RED_ZONE;
906 case 'p' : case 'P' :
907 slub_debug |= SLAB_POISON;
909 case 'u' : case 'U' :
910 slub_debug |= SLAB_STORE_USER;
912 case 't' : case 'T' :
913 slub_debug |= SLAB_TRACE;
916 printk(KERN_ERR "slub_debug option '%c' "
917 "unknown. skipped\n",*str);
922 slub_debug_slabs = str + 1;
926 __setup("slub_debug", setup_slub_debug);
928 static void kmem_cache_open_debug_check(struct kmem_cache *s)
931 * The page->offset field is only 16 bit wide. This is an offset
932 * in units of words from the beginning of an object. If the slab
933 * size is bigger then we cannot move the free pointer behind the
936 * On 32 bit platforms the limit is 256k. On 64bit platforms
939 * Debugging or ctor may create a need to move the free
940 * pointer. Fail if this happens.
942 if (s->objsize >= 65535 * sizeof(void *)) {
943 BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON |
944 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
949 * Enable debugging if selected on the kernel commandline.
951 if (slub_debug && (!slub_debug_slabs ||
952 strncmp(slub_debug_slabs, s->name,
953 strlen(slub_debug_slabs)) == 0))
954 s->flags |= slub_debug;
957 static inline void setup_object_debug(struct kmem_cache *s,
958 struct page *page, void *object) {}
960 static inline int alloc_debug_processing(struct kmem_cache *s,
961 struct page *page, void *object, void *addr) { return 0; }
963 static inline int free_debug_processing(struct kmem_cache *s,
964 struct page *page, void *object, void *addr) { return 0; }
966 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
968 static inline int check_object(struct kmem_cache *s, struct page *page,
969 void *object, int active) { return 1; }
970 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
971 static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {}
975 * Slab allocation and freeing
977 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
980 int pages = 1 << s->order;
985 if (s->flags & SLAB_CACHE_DMA)
989 page = alloc_pages(flags, s->order);
991 page = alloc_pages_node(node, flags, s->order);
996 mod_zone_page_state(page_zone(page),
997 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
998 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1004 static void setup_object(struct kmem_cache *s, struct page *page,
1007 setup_object_debug(s, page, object);
1008 if (unlikely(s->ctor))
1009 s->ctor(object, s, 0);
1012 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1015 struct kmem_cache_node *n;
1021 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
1023 if (flags & __GFP_WAIT)
1026 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
1030 n = get_node(s, page_to_nid(page));
1032 atomic_long_inc(&n->nr_slabs);
1033 page->offset = s->offset / sizeof(void *);
1035 page->flags |= 1 << PG_slab;
1036 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1037 SLAB_STORE_USER | SLAB_TRACE))
1040 start = page_address(page);
1041 end = start + s->objects * s->size;
1043 if (unlikely(s->flags & SLAB_POISON))
1044 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1047 for_each_object(p, s, start) {
1048 setup_object(s, page, last);
1049 set_freepointer(s, last, p);
1052 setup_object(s, page, last);
1053 set_freepointer(s, last, NULL);
1055 page->freelist = start;
1056 page->lockless_freelist = NULL;
1059 if (flags & __GFP_WAIT)
1060 local_irq_disable();
1064 static void __free_slab(struct kmem_cache *s, struct page *page)
1066 int pages = 1 << s->order;
1068 if (unlikely(SlabDebug(page))) {
1071 slab_pad_check(s, page);
1072 for_each_object(p, s, page_address(page))
1073 check_object(s, page, p, 0);
1076 mod_zone_page_state(page_zone(page),
1077 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1078 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1081 page->mapping = NULL;
1082 __free_pages(page, s->order);
1085 static void rcu_free_slab(struct rcu_head *h)
1089 page = container_of((struct list_head *)h, struct page, lru);
1090 __free_slab(page->slab, page);
1093 static void free_slab(struct kmem_cache *s, struct page *page)
1095 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1097 * RCU free overloads the RCU head over the LRU
1099 struct rcu_head *head = (void *)&page->lru;
1101 call_rcu(head, rcu_free_slab);
1103 __free_slab(s, page);
1106 static void discard_slab(struct kmem_cache *s, struct page *page)
1108 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1110 atomic_long_dec(&n->nr_slabs);
1111 reset_page_mapcount(page);
1112 ClearSlabDebug(page);
1113 __ClearPageSlab(page);
1118 * Per slab locking using the pagelock
1120 static __always_inline void slab_lock(struct page *page)
1122 bit_spin_lock(PG_locked, &page->flags);
1125 static __always_inline void slab_unlock(struct page *page)
1127 bit_spin_unlock(PG_locked, &page->flags);
1130 static __always_inline int slab_trylock(struct page *page)
1134 rc = bit_spin_trylock(PG_locked, &page->flags);
1139 * Management of partially allocated slabs
1141 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1143 spin_lock(&n->list_lock);
1145 list_add_tail(&page->lru, &n->partial);
1146 spin_unlock(&n->list_lock);
1149 static void add_partial(struct kmem_cache_node *n, struct page *page)
1151 spin_lock(&n->list_lock);
1153 list_add(&page->lru, &n->partial);
1154 spin_unlock(&n->list_lock);
1157 static void remove_partial(struct kmem_cache *s,
1160 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1162 spin_lock(&n->list_lock);
1163 list_del(&page->lru);
1165 spin_unlock(&n->list_lock);
1169 * Lock slab and remove from the partial list.
1171 * Must hold list_lock.
1173 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1175 if (slab_trylock(page)) {
1176 list_del(&page->lru);
1178 SetSlabFrozen(page);
1185 * Try to allocate a partial slab from a specific node.
1187 static struct page *get_partial_node(struct kmem_cache_node *n)
1192 * Racy check. If we mistakenly see no partial slabs then we
1193 * just allocate an empty slab. If we mistakenly try to get a
1194 * partial slab and there is none available then get_partials()
1197 if (!n || !n->nr_partial)
1200 spin_lock(&n->list_lock);
1201 list_for_each_entry(page, &n->partial, lru)
1202 if (lock_and_freeze_slab(n, page))
1206 spin_unlock(&n->list_lock);
1211 * Get a page from somewhere. Search in increasing NUMA distances.
1213 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1216 struct zonelist *zonelist;
1221 * The defrag ratio allows a configuration of the tradeoffs between
1222 * inter node defragmentation and node local allocations. A lower
1223 * defrag_ratio increases the tendency to do local allocations
1224 * instead of attempting to obtain partial slabs from other nodes.
1226 * If the defrag_ratio is set to 0 then kmalloc() always
1227 * returns node local objects. If the ratio is higher then kmalloc()
1228 * may return off node objects because partial slabs are obtained
1229 * from other nodes and filled up.
1231 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1232 * defrag_ratio = 1000) then every (well almost) allocation will
1233 * first attempt to defrag slab caches on other nodes. This means
1234 * scanning over all nodes to look for partial slabs which may be
1235 * expensive if we do it every time we are trying to find a slab
1236 * with available objects.
1238 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1241 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1242 ->node_zonelists[gfp_zone(flags)];
1243 for (z = zonelist->zones; *z; z++) {
1244 struct kmem_cache_node *n;
1246 n = get_node(s, zone_to_nid(*z));
1248 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1249 n->nr_partial > MIN_PARTIAL) {
1250 page = get_partial_node(n);
1260 * Get a partial page, lock it and return it.
1262 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1265 int searchnode = (node == -1) ? numa_node_id() : node;
1267 page = get_partial_node(get_node(s, searchnode));
1268 if (page || (flags & __GFP_THISNODE))
1271 return get_any_partial(s, flags);
1275 * Move a page back to the lists.
1277 * Must be called with the slab lock held.
1279 * On exit the slab lock will have been dropped.
1281 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1283 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1285 ClearSlabFrozen(page);
1289 add_partial(n, page);
1290 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1295 if (n->nr_partial < MIN_PARTIAL) {
1297 * Adding an empty slab to the partial slabs in order
1298 * to avoid page allocator overhead. This slab needs
1299 * to come after the other slabs with objects in
1300 * order to fill them up. That way the size of the
1301 * partial list stays small. kmem_cache_shrink can
1302 * reclaim empty slabs from the partial list.
1304 add_partial_tail(n, page);
1308 discard_slab(s, page);
1314 * Remove the cpu slab
1316 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1319 * Merge cpu freelist into freelist. Typically we get here
1320 * because both freelists are empty. So this is unlikely
1323 while (unlikely(page->lockless_freelist)) {
1326 /* Retrieve object from cpu_freelist */
1327 object = page->lockless_freelist;
1328 page->lockless_freelist = page->lockless_freelist[page->offset];
1330 /* And put onto the regular freelist */
1331 object[page->offset] = page->freelist;
1332 page->freelist = object;
1335 s->cpu_slab[cpu] = NULL;
1336 unfreeze_slab(s, page);
1339 static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1342 deactivate_slab(s, page, cpu);
1347 * Called from IPI handler with interrupts disabled.
1349 static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1351 struct page *page = s->cpu_slab[cpu];
1354 flush_slab(s, page, cpu);
1357 static void flush_cpu_slab(void *d)
1359 struct kmem_cache *s = d;
1360 int cpu = smp_processor_id();
1362 __flush_cpu_slab(s, cpu);
1365 static void flush_all(struct kmem_cache *s)
1368 on_each_cpu(flush_cpu_slab, s, 1, 1);
1370 unsigned long flags;
1372 local_irq_save(flags);
1374 local_irq_restore(flags);
1379 * Slow path. The lockless freelist is empty or we need to perform
1382 * Interrupts are disabled.
1384 * Processing is still very fast if new objects have been freed to the
1385 * regular freelist. In that case we simply take over the regular freelist
1386 * as the lockless freelist and zap the regular freelist.
1388 * If that is not working then we fall back to the partial lists. We take the
1389 * first element of the freelist as the object to allocate now and move the
1390 * rest of the freelist to the lockless freelist.
1392 * And if we were unable to get a new slab from the partial slab lists then
1393 * we need to allocate a new slab. This is slowest path since we may sleep.
1395 static void *__slab_alloc(struct kmem_cache *s,
1396 gfp_t gfpflags, int node, void *addr, struct page *page)
1399 int cpu = smp_processor_id();
1405 if (unlikely(node != -1 && page_to_nid(page) != node))
1408 object = page->freelist;
1409 if (unlikely(!object))
1411 if (unlikely(SlabDebug(page)))
1414 object = page->freelist;
1415 page->lockless_freelist = object[page->offset];
1416 page->inuse = s->objects;
1417 page->freelist = NULL;
1422 deactivate_slab(s, page, cpu);
1425 page = get_partial(s, gfpflags, node);
1427 s->cpu_slab[cpu] = page;
1431 page = new_slab(s, gfpflags, node);
1433 cpu = smp_processor_id();
1434 if (s->cpu_slab[cpu]) {
1436 * Someone else populated the cpu_slab while we
1437 * enabled interrupts, or we have gotten scheduled
1438 * on another cpu. The page may not be on the
1439 * requested node even if __GFP_THISNODE was
1440 * specified. So we need to recheck.
1443 page_to_nid(s->cpu_slab[cpu]) == node) {
1445 * Current cpuslab is acceptable and we
1446 * want the current one since its cache hot
1448 discard_slab(s, page);
1449 page = s->cpu_slab[cpu];
1453 /* New slab does not fit our expectations */
1454 flush_slab(s, s->cpu_slab[cpu], cpu);
1457 SetSlabFrozen(page);
1458 s->cpu_slab[cpu] = page;
1463 object = page->freelist;
1464 if (!alloc_debug_processing(s, page, object, addr))
1468 page->freelist = object[page->offset];
1474 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1475 * have the fastpath folded into their functions. So no function call
1476 * overhead for requests that can be satisfied on the fastpath.
1478 * The fastpath works by first checking if the lockless freelist can be used.
1479 * If not then __slab_alloc is called for slow processing.
1481 * Otherwise we can simply pick the next object from the lockless free list.
1483 static void __always_inline *slab_alloc(struct kmem_cache *s,
1484 gfp_t gfpflags, int node, void *addr)
1488 unsigned long flags;
1490 local_irq_save(flags);
1491 page = s->cpu_slab[smp_processor_id()];
1492 if (unlikely(!page || !page->lockless_freelist ||
1493 (node != -1 && page_to_nid(page) != node)))
1495 object = __slab_alloc(s, gfpflags, node, addr, page);
1498 object = page->lockless_freelist;
1499 page->lockless_freelist = object[page->offset];
1501 local_irq_restore(flags);
1505 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1507 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1509 EXPORT_SYMBOL(kmem_cache_alloc);
1512 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1514 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1516 EXPORT_SYMBOL(kmem_cache_alloc_node);
1520 * Slow patch handling. This may still be called frequently since objects
1521 * have a longer lifetime than the cpu slabs in most processing loads.
1523 * So we still attempt to reduce cache line usage. Just take the slab
1524 * lock and free the item. If there is no additional partial page
1525 * handling required then we can return immediately.
1527 static void __slab_free(struct kmem_cache *s, struct page *page,
1528 void *x, void *addr)
1531 void **object = (void *)x;
1535 if (unlikely(SlabDebug(page)))
1538 prior = object[page->offset] = page->freelist;
1539 page->freelist = object;
1542 if (unlikely(SlabFrozen(page)))
1545 if (unlikely(!page->inuse))
1549 * Objects left in the slab. If it
1550 * was not on the partial list before
1553 if (unlikely(!prior))
1554 add_partial(get_node(s, page_to_nid(page)), page);
1563 * Slab still on the partial list.
1565 remove_partial(s, page);
1568 discard_slab(s, page);
1572 if (!free_debug_processing(s, page, x, addr))
1578 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1579 * can perform fastpath freeing without additional function calls.
1581 * The fastpath is only possible if we are freeing to the current cpu slab
1582 * of this processor. This typically the case if we have just allocated
1585 * If fastpath is not possible then fall back to __slab_free where we deal
1586 * with all sorts of special processing.
1588 static void __always_inline slab_free(struct kmem_cache *s,
1589 struct page *page, void *x, void *addr)
1591 void **object = (void *)x;
1592 unsigned long flags;
1594 local_irq_save(flags);
1595 if (likely(page == s->cpu_slab[smp_processor_id()] &&
1596 !SlabDebug(page))) {
1597 object[page->offset] = page->lockless_freelist;
1598 page->lockless_freelist = object;
1600 __slab_free(s, page, x, addr);
1602 local_irq_restore(flags);
1605 void kmem_cache_free(struct kmem_cache *s, void *x)
1609 page = virt_to_head_page(x);
1611 slab_free(s, page, x, __builtin_return_address(0));
1613 EXPORT_SYMBOL(kmem_cache_free);
1615 /* Figure out on which slab object the object resides */
1616 static struct page *get_object_page(const void *x)
1618 struct page *page = virt_to_head_page(x);
1620 if (!PageSlab(page))
1627 * Object placement in a slab is made very easy because we always start at
1628 * offset 0. If we tune the size of the object to the alignment then we can
1629 * get the required alignment by putting one properly sized object after
1632 * Notice that the allocation order determines the sizes of the per cpu
1633 * caches. Each processor has always one slab available for allocations.
1634 * Increasing the allocation order reduces the number of times that slabs
1635 * must be moved on and off the partial lists and is therefore a factor in
1640 * Mininum / Maximum order of slab pages. This influences locking overhead
1641 * and slab fragmentation. A higher order reduces the number of partial slabs
1642 * and increases the number of allocations possible without having to
1643 * take the list_lock.
1645 static int slub_min_order;
1646 static int slub_max_order = DEFAULT_MAX_ORDER;
1647 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1650 * Merge control. If this is set then no merging of slab caches will occur.
1651 * (Could be removed. This was introduced to pacify the merge skeptics.)
1653 static int slub_nomerge;
1656 * Calculate the order of allocation given an slab object size.
1658 * The order of allocation has significant impact on performance and other
1659 * system components. Generally order 0 allocations should be preferred since
1660 * order 0 does not cause fragmentation in the page allocator. Larger objects
1661 * be problematic to put into order 0 slabs because there may be too much
1662 * unused space left. We go to a higher order if more than 1/8th of the slab
1665 * In order to reach satisfactory performance we must ensure that a minimum
1666 * number of objects is in one slab. Otherwise we may generate too much
1667 * activity on the partial lists which requires taking the list_lock. This is
1668 * less a concern for large slabs though which are rarely used.
1670 * slub_max_order specifies the order where we begin to stop considering the
1671 * number of objects in a slab as critical. If we reach slub_max_order then
1672 * we try to keep the page order as low as possible. So we accept more waste
1673 * of space in favor of a small page order.
1675 * Higher order allocations also allow the placement of more objects in a
1676 * slab and thereby reduce object handling overhead. If the user has
1677 * requested a higher mininum order then we start with that one instead of
1678 * the smallest order which will fit the object.
1680 static inline int slab_order(int size, int min_objects,
1681 int max_order, int fract_leftover)
1686 for (order = max(slub_min_order,
1687 fls(min_objects * size - 1) - PAGE_SHIFT);
1688 order <= max_order; order++) {
1690 unsigned long slab_size = PAGE_SIZE << order;
1692 if (slab_size < min_objects * size)
1695 rem = slab_size % size;
1697 if (rem <= slab_size / fract_leftover)
1705 static inline int calculate_order(int size)
1712 * Attempt to find best configuration for a slab. This
1713 * works by first attempting to generate a layout with
1714 * the best configuration and backing off gradually.
1716 * First we reduce the acceptable waste in a slab. Then
1717 * we reduce the minimum objects required in a slab.
1719 min_objects = slub_min_objects;
1720 while (min_objects > 1) {
1722 while (fraction >= 4) {
1723 order = slab_order(size, min_objects,
1724 slub_max_order, fraction);
1725 if (order <= slub_max_order)
1733 * We were unable to place multiple objects in a slab. Now
1734 * lets see if we can place a single object there.
1736 order = slab_order(size, 1, slub_max_order, 1);
1737 if (order <= slub_max_order)
1741 * Doh this slab cannot be placed using slub_max_order.
1743 order = slab_order(size, 1, MAX_ORDER, 1);
1744 if (order <= MAX_ORDER)
1750 * Figure out what the alignment of the objects will be.
1752 static unsigned long calculate_alignment(unsigned long flags,
1753 unsigned long align, unsigned long size)
1756 * If the user wants hardware cache aligned objects then
1757 * follow that suggestion if the object is sufficiently
1760 * The hardware cache alignment cannot override the
1761 * specified alignment though. If that is greater
1764 if ((flags & SLAB_HWCACHE_ALIGN) &&
1765 size > cache_line_size() / 2)
1766 return max_t(unsigned long, align, cache_line_size());
1768 if (align < ARCH_SLAB_MINALIGN)
1769 return ARCH_SLAB_MINALIGN;
1771 return ALIGN(align, sizeof(void *));
1774 static void init_kmem_cache_node(struct kmem_cache_node *n)
1777 atomic_long_set(&n->nr_slabs, 0);
1778 spin_lock_init(&n->list_lock);
1779 INIT_LIST_HEAD(&n->partial);
1780 INIT_LIST_HEAD(&n->full);
1785 * No kmalloc_node yet so do it by hand. We know that this is the first
1786 * slab on the node for this slabcache. There are no concurrent accesses
1789 * Note that this function only works on the kmalloc_node_cache
1790 * when allocating for the kmalloc_node_cache.
1792 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1796 struct kmem_cache_node *n;
1798 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1800 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1801 /* new_slab() disables interupts */
1807 page->freelist = get_freepointer(kmalloc_caches, n);
1809 kmalloc_caches->node[node] = n;
1810 setup_object_debug(kmalloc_caches, page, n);
1811 init_kmem_cache_node(n);
1812 atomic_long_inc(&n->nr_slabs);
1813 add_partial(n, page);
1817 static void free_kmem_cache_nodes(struct kmem_cache *s)
1821 for_each_online_node(node) {
1822 struct kmem_cache_node *n = s->node[node];
1823 if (n && n != &s->local_node)
1824 kmem_cache_free(kmalloc_caches, n);
1825 s->node[node] = NULL;
1829 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1834 if (slab_state >= UP)
1835 local_node = page_to_nid(virt_to_page(s));
1839 for_each_online_node(node) {
1840 struct kmem_cache_node *n;
1842 if (local_node == node)
1845 if (slab_state == DOWN) {
1846 n = early_kmem_cache_node_alloc(gfpflags,
1850 n = kmem_cache_alloc_node(kmalloc_caches,
1854 free_kmem_cache_nodes(s);
1860 init_kmem_cache_node(n);
1865 static void free_kmem_cache_nodes(struct kmem_cache *s)
1869 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1871 init_kmem_cache_node(&s->local_node);
1877 * calculate_sizes() determines the order and the distribution of data within
1880 static int calculate_sizes(struct kmem_cache *s)
1882 unsigned long flags = s->flags;
1883 unsigned long size = s->objsize;
1884 unsigned long align = s->align;
1887 * Determine if we can poison the object itself. If the user of
1888 * the slab may touch the object after free or before allocation
1889 * then we should never poison the object itself.
1891 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1893 s->flags |= __OBJECT_POISON;
1895 s->flags &= ~__OBJECT_POISON;
1898 * Round up object size to the next word boundary. We can only
1899 * place the free pointer at word boundaries and this determines
1900 * the possible location of the free pointer.
1902 size = ALIGN(size, sizeof(void *));
1904 #ifdef CONFIG_SLUB_DEBUG
1906 * If we are Redzoning then check if there is some space between the
1907 * end of the object and the free pointer. If not then add an
1908 * additional word to have some bytes to store Redzone information.
1910 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1911 size += sizeof(void *);
1915 * With that we have determined the number of bytes in actual use
1916 * by the object. This is the potential offset to the free pointer.
1920 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1923 * Relocate free pointer after the object if it is not
1924 * permitted to overwrite the first word of the object on
1927 * This is the case if we do RCU, have a constructor or
1928 * destructor or are poisoning the objects.
1931 size += sizeof(void *);
1934 #ifdef CONFIG_SLUB_DEBUG
1935 if (flags & SLAB_STORE_USER)
1937 * Need to store information about allocs and frees after
1940 size += 2 * sizeof(struct track);
1942 if (flags & SLAB_RED_ZONE)
1944 * Add some empty padding so that we can catch
1945 * overwrites from earlier objects rather than let
1946 * tracking information or the free pointer be
1947 * corrupted if an user writes before the start
1950 size += sizeof(void *);
1954 * Determine the alignment based on various parameters that the
1955 * user specified and the dynamic determination of cache line size
1958 align = calculate_alignment(flags, align, s->objsize);
1961 * SLUB stores one object immediately after another beginning from
1962 * offset 0. In order to align the objects we have to simply size
1963 * each object to conform to the alignment.
1965 size = ALIGN(size, align);
1968 s->order = calculate_order(size);
1973 * Determine the number of objects per slab
1975 s->objects = (PAGE_SIZE << s->order) / size;
1978 * Verify that the number of objects is within permitted limits.
1979 * The page->inuse field is only 16 bit wide! So we cannot have
1980 * more than 64k objects per slab.
1982 if (!s->objects || s->objects > 65535)
1988 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1989 const char *name, size_t size,
1990 size_t align, unsigned long flags,
1991 void (*ctor)(void *, struct kmem_cache *, unsigned long))
1993 memset(s, 0, kmem_size);
1999 kmem_cache_open_debug_check(s);
2001 if (!calculate_sizes(s))
2006 s->defrag_ratio = 100;
2009 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2012 if (flags & SLAB_PANIC)
2013 panic("Cannot create slab %s size=%lu realsize=%u "
2014 "order=%u offset=%u flags=%lx\n",
2015 s->name, (unsigned long)size, s->size, s->order,
2019 EXPORT_SYMBOL(kmem_cache_open);
2022 * Check if a given pointer is valid
2024 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2028 page = get_object_page(object);
2030 if (!page || s != page->slab)
2031 /* No slab or wrong slab */
2034 if (!check_valid_pointer(s, page, object))
2038 * We could also check if the object is on the slabs freelist.
2039 * But this would be too expensive and it seems that the main
2040 * purpose of kmem_ptr_valid is to check if the object belongs
2041 * to a certain slab.
2045 EXPORT_SYMBOL(kmem_ptr_validate);
2048 * Determine the size of a slab object
2050 unsigned int kmem_cache_size(struct kmem_cache *s)
2054 EXPORT_SYMBOL(kmem_cache_size);
2056 const char *kmem_cache_name(struct kmem_cache *s)
2060 EXPORT_SYMBOL(kmem_cache_name);
2063 * Attempt to free all slabs on a node. Return the number of slabs we
2064 * were unable to free.
2066 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2067 struct list_head *list)
2069 int slabs_inuse = 0;
2070 unsigned long flags;
2071 struct page *page, *h;
2073 spin_lock_irqsave(&n->list_lock, flags);
2074 list_for_each_entry_safe(page, h, list, lru)
2076 list_del(&page->lru);
2077 discard_slab(s, page);
2080 spin_unlock_irqrestore(&n->list_lock, flags);
2085 * Release all resources used by a slab cache.
2087 static int kmem_cache_close(struct kmem_cache *s)
2093 /* Attempt to free all objects */
2094 for_each_online_node(node) {
2095 struct kmem_cache_node *n = get_node(s, node);
2097 n->nr_partial -= free_list(s, n, &n->partial);
2098 if (atomic_long_read(&n->nr_slabs))
2101 free_kmem_cache_nodes(s);
2106 * Close a cache and release the kmem_cache structure
2107 * (must be used for caches created using kmem_cache_create)
2109 void kmem_cache_destroy(struct kmem_cache *s)
2111 down_write(&slub_lock);
2115 if (kmem_cache_close(s))
2117 sysfs_slab_remove(s);
2120 up_write(&slub_lock);
2122 EXPORT_SYMBOL(kmem_cache_destroy);
2124 /********************************************************************
2126 *******************************************************************/
2128 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
2129 EXPORT_SYMBOL(kmalloc_caches);
2131 #ifdef CONFIG_ZONE_DMA
2132 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
2135 static int __init setup_slub_min_order(char *str)
2137 get_option (&str, &slub_min_order);
2142 __setup("slub_min_order=", setup_slub_min_order);
2144 static int __init setup_slub_max_order(char *str)
2146 get_option (&str, &slub_max_order);
2151 __setup("slub_max_order=", setup_slub_max_order);
2153 static int __init setup_slub_min_objects(char *str)
2155 get_option (&str, &slub_min_objects);
2160 __setup("slub_min_objects=", setup_slub_min_objects);
2162 static int __init setup_slub_nomerge(char *str)
2168 __setup("slub_nomerge", setup_slub_nomerge);
2170 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2171 const char *name, int size, gfp_t gfp_flags)
2173 unsigned int flags = 0;
2175 if (gfp_flags & SLUB_DMA)
2176 flags = SLAB_CACHE_DMA;
2178 down_write(&slub_lock);
2179 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2183 list_add(&s->list, &slab_caches);
2184 up_write(&slub_lock);
2185 if (sysfs_slab_add(s))
2190 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2193 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2195 int index = kmalloc_index(size);
2200 /* Allocation too large? */
2203 #ifdef CONFIG_ZONE_DMA
2204 if ((flags & SLUB_DMA)) {
2205 struct kmem_cache *s;
2206 struct kmem_cache *x;
2210 s = kmalloc_caches_dma[index];
2214 /* Dynamically create dma cache */
2215 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2217 panic("Unable to allocate memory for dma cache\n");
2219 if (index <= KMALLOC_SHIFT_HIGH)
2220 realsize = 1 << index;
2228 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2229 (unsigned int)realsize);
2230 s = create_kmalloc_cache(x, text, realsize, flags);
2231 kmalloc_caches_dma[index] = s;
2235 return &kmalloc_caches[index];
2238 void *__kmalloc(size_t size, gfp_t flags)
2240 struct kmem_cache *s = get_slab(size, flags);
2243 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2244 return ZERO_SIZE_PTR;
2246 EXPORT_SYMBOL(__kmalloc);
2249 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2251 struct kmem_cache *s = get_slab(size, flags);
2254 return slab_alloc(s, flags, node, __builtin_return_address(0));
2255 return ZERO_SIZE_PTR;
2257 EXPORT_SYMBOL(__kmalloc_node);
2260 size_t ksize(const void *object)
2263 struct kmem_cache *s;
2265 if (object == ZERO_SIZE_PTR)
2268 page = get_object_page(object);
2274 * Debugging requires use of the padding between object
2275 * and whatever may come after it.
2277 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2281 * If we have the need to store the freelist pointer
2282 * back there or track user information then we can
2283 * only use the space before that information.
2285 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2289 * Else we can use all the padding etc for the allocation
2293 EXPORT_SYMBOL(ksize);
2295 void kfree(const void *x)
2297 struct kmem_cache *s;
2301 * This has to be an unsigned comparison. According to Linus
2302 * some gcc version treat a pointer as a signed entity. Then
2303 * this comparison would be true for all "negative" pointers
2304 * (which would cover the whole upper half of the address space).
2306 if ((unsigned long)x <= (unsigned long)ZERO_SIZE_PTR)
2309 page = virt_to_head_page(x);
2312 slab_free(s, page, (void *)x, __builtin_return_address(0));
2314 EXPORT_SYMBOL(kfree);
2317 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2318 * the remaining slabs by the number of items in use. The slabs with the
2319 * most items in use come first. New allocations will then fill those up
2320 * and thus they can be removed from the partial lists.
2322 * The slabs with the least items are placed last. This results in them
2323 * being allocated from last increasing the chance that the last objects
2324 * are freed in them.
2326 int kmem_cache_shrink(struct kmem_cache *s)
2330 struct kmem_cache_node *n;
2333 struct list_head *slabs_by_inuse =
2334 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2335 unsigned long flags;
2337 if (!slabs_by_inuse)
2341 for_each_online_node(node) {
2342 n = get_node(s, node);
2347 for (i = 0; i < s->objects; i++)
2348 INIT_LIST_HEAD(slabs_by_inuse + i);
2350 spin_lock_irqsave(&n->list_lock, flags);
2353 * Build lists indexed by the items in use in each slab.
2355 * Note that concurrent frees may occur while we hold the
2356 * list_lock. page->inuse here is the upper limit.
2358 list_for_each_entry_safe(page, t, &n->partial, lru) {
2359 if (!page->inuse && slab_trylock(page)) {
2361 * Must hold slab lock here because slab_free
2362 * may have freed the last object and be
2363 * waiting to release the slab.
2365 list_del(&page->lru);
2368 discard_slab(s, page);
2370 if (n->nr_partial > MAX_PARTIAL)
2371 list_move(&page->lru,
2372 slabs_by_inuse + page->inuse);
2376 if (n->nr_partial <= MAX_PARTIAL)
2380 * Rebuild the partial list with the slabs filled up most
2381 * first and the least used slabs at the end.
2383 for (i = s->objects - 1; i >= 0; i--)
2384 list_splice(slabs_by_inuse + i, n->partial.prev);
2387 spin_unlock_irqrestore(&n->list_lock, flags);
2390 kfree(slabs_by_inuse);
2393 EXPORT_SYMBOL(kmem_cache_shrink);
2396 * krealloc - reallocate memory. The contents will remain unchanged.
2397 * @p: object to reallocate memory for.
2398 * @new_size: how many bytes of memory are required.
2399 * @flags: the type of memory to allocate.
2401 * The contents of the object pointed to are preserved up to the
2402 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2403 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2404 * %NULL pointer, the object pointed to is freed.
2406 void *krealloc(const void *p, size_t new_size, gfp_t flags)
2411 if (unlikely(!p || p == ZERO_SIZE_PTR))
2412 return kmalloc(new_size, flags);
2414 if (unlikely(!new_size)) {
2416 return ZERO_SIZE_PTR;
2423 ret = kmalloc(new_size, flags);
2425 memcpy(ret, p, min(new_size, ks));
2430 EXPORT_SYMBOL(krealloc);
2432 /********************************************************************
2433 * Basic setup of slabs
2434 *******************************************************************/
2436 void __init kmem_cache_init(void)
2443 * Must first have the slab cache available for the allocations of the
2444 * struct kmem_cache_node's. There is special bootstrap code in
2445 * kmem_cache_open for slab_state == DOWN.
2447 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2448 sizeof(struct kmem_cache_node), GFP_KERNEL);
2449 kmalloc_caches[0].refcount = -1;
2453 /* Able to allocate the per node structures */
2454 slab_state = PARTIAL;
2456 /* Caches that are not of the two-to-the-power-of size */
2457 if (KMALLOC_MIN_SIZE <= 64) {
2458 create_kmalloc_cache(&kmalloc_caches[1],
2459 "kmalloc-96", 96, GFP_KERNEL);
2462 if (KMALLOC_MIN_SIZE <= 128) {
2463 create_kmalloc_cache(&kmalloc_caches[2],
2464 "kmalloc-192", 192, GFP_KERNEL);
2468 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
2469 create_kmalloc_cache(&kmalloc_caches[i],
2470 "kmalloc", 1 << i, GFP_KERNEL);
2476 /* Provide the correct kmalloc names now that the caches are up */
2477 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2478 kmalloc_caches[i]. name =
2479 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2482 register_cpu_notifier(&slab_notifier);
2485 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2486 nr_cpu_ids * sizeof(struct page *);
2488 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2489 " CPUs=%d, Nodes=%d\n",
2490 caches, cache_line_size(),
2491 slub_min_order, slub_max_order, slub_min_objects,
2492 nr_cpu_ids, nr_node_ids);
2496 * Find a mergeable slab cache
2498 static int slab_unmergeable(struct kmem_cache *s)
2500 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2507 * We may have set a slab to be unmergeable during bootstrap.
2509 if (s->refcount < 0)
2515 static struct kmem_cache *find_mergeable(size_t size,
2516 size_t align, unsigned long flags,
2517 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2519 struct list_head *h;
2521 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2527 size = ALIGN(size, sizeof(void *));
2528 align = calculate_alignment(flags, align, size);
2529 size = ALIGN(size, align);
2531 list_for_each(h, &slab_caches) {
2532 struct kmem_cache *s =
2533 container_of(h, struct kmem_cache, list);
2535 if (slab_unmergeable(s))
2541 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2542 (s->flags & SLUB_MERGE_SAME))
2545 * Check if alignment is compatible.
2546 * Courtesy of Adrian Drzewiecki
2548 if ((s->size & ~(align -1)) != s->size)
2551 if (s->size - size >= sizeof(void *))
2559 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2560 size_t align, unsigned long flags,
2561 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2562 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2564 struct kmem_cache *s;
2567 down_write(&slub_lock);
2568 s = find_mergeable(size, align, flags, ctor);
2572 * Adjust the object sizes so that we clear
2573 * the complete object on kzalloc.
2575 s->objsize = max(s->objsize, (int)size);
2576 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2577 if (sysfs_slab_alias(s, name))
2580 s = kmalloc(kmem_size, GFP_KERNEL);
2581 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2582 size, align, flags, ctor)) {
2583 if (sysfs_slab_add(s)) {
2587 list_add(&s->list, &slab_caches);
2591 up_write(&slub_lock);
2595 up_write(&slub_lock);
2596 if (flags & SLAB_PANIC)
2597 panic("Cannot create slabcache %s\n", name);
2602 EXPORT_SYMBOL(kmem_cache_create);
2604 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2608 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
2610 memset(x, 0, s->objsize);
2613 EXPORT_SYMBOL(kmem_cache_zalloc);
2616 static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2618 struct list_head *h;
2620 down_read(&slub_lock);
2621 list_for_each(h, &slab_caches) {
2622 struct kmem_cache *s =
2623 container_of(h, struct kmem_cache, list);
2627 up_read(&slub_lock);
2631 * Version of __flush_cpu_slab for the case that interrupts
2634 static void cpu_slab_flush(struct kmem_cache *s, int cpu)
2636 unsigned long flags;
2638 local_irq_save(flags);
2639 __flush_cpu_slab(s, cpu);
2640 local_irq_restore(flags);
2644 * Use the cpu notifier to insure that the cpu slabs are flushed when
2647 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2648 unsigned long action, void *hcpu)
2650 long cpu = (long)hcpu;
2653 case CPU_UP_CANCELED:
2654 case CPU_UP_CANCELED_FROZEN:
2656 case CPU_DEAD_FROZEN:
2657 for_all_slabs(cpu_slab_flush, cpu);
2665 static struct notifier_block __cpuinitdata slab_notifier =
2666 { &slab_cpuup_callback, NULL, 0 };
2670 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2672 struct kmem_cache *s = get_slab(size, gfpflags);
2675 return ZERO_SIZE_PTR;
2677 return slab_alloc(s, gfpflags, -1, caller);
2680 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2681 int node, void *caller)
2683 struct kmem_cache *s = get_slab(size, gfpflags);
2686 return ZERO_SIZE_PTR;
2688 return slab_alloc(s, gfpflags, node, caller);
2691 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2692 static int validate_slab(struct kmem_cache *s, struct page *page)
2695 void *addr = page_address(page);
2696 DECLARE_BITMAP(map, s->objects);
2698 if (!check_slab(s, page) ||
2699 !on_freelist(s, page, NULL))
2702 /* Now we know that a valid freelist exists */
2703 bitmap_zero(map, s->objects);
2705 for_each_free_object(p, s, page->freelist) {
2706 set_bit(slab_index(p, s, addr), map);
2707 if (!check_object(s, page, p, 0))
2711 for_each_object(p, s, addr)
2712 if (!test_bit(slab_index(p, s, addr), map))
2713 if (!check_object(s, page, p, 1))
2718 static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2720 if (slab_trylock(page)) {
2721 validate_slab(s, page);
2724 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2727 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2728 if (!SlabDebug(page))
2729 printk(KERN_ERR "SLUB %s: SlabDebug not set "
2730 "on slab 0x%p\n", s->name, page);
2732 if (SlabDebug(page))
2733 printk(KERN_ERR "SLUB %s: SlabDebug set on "
2734 "slab 0x%p\n", s->name, page);
2738 static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2740 unsigned long count = 0;
2742 unsigned long flags;
2744 spin_lock_irqsave(&n->list_lock, flags);
2746 list_for_each_entry(page, &n->partial, lru) {
2747 validate_slab_slab(s, page);
2750 if (count != n->nr_partial)
2751 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2752 "counter=%ld\n", s->name, count, n->nr_partial);
2754 if (!(s->flags & SLAB_STORE_USER))
2757 list_for_each_entry(page, &n->full, lru) {
2758 validate_slab_slab(s, page);
2761 if (count != atomic_long_read(&n->nr_slabs))
2762 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2763 "counter=%ld\n", s->name, count,
2764 atomic_long_read(&n->nr_slabs));
2767 spin_unlock_irqrestore(&n->list_lock, flags);
2771 static unsigned long validate_slab_cache(struct kmem_cache *s)
2774 unsigned long count = 0;
2777 for_each_online_node(node) {
2778 struct kmem_cache_node *n = get_node(s, node);
2780 count += validate_slab_node(s, n);
2785 #ifdef SLUB_RESILIENCY_TEST
2786 static void resiliency_test(void)
2790 printk(KERN_ERR "SLUB resiliency testing\n");
2791 printk(KERN_ERR "-----------------------\n");
2792 printk(KERN_ERR "A. Corruption after allocation\n");
2794 p = kzalloc(16, GFP_KERNEL);
2796 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2797 " 0x12->0x%p\n\n", p + 16);
2799 validate_slab_cache(kmalloc_caches + 4);
2801 /* Hmmm... The next two are dangerous */
2802 p = kzalloc(32, GFP_KERNEL);
2803 p[32 + sizeof(void *)] = 0x34;
2804 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2805 " 0x34 -> -0x%p\n", p);
2806 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2808 validate_slab_cache(kmalloc_caches + 5);
2809 p = kzalloc(64, GFP_KERNEL);
2810 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2812 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2814 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2815 validate_slab_cache(kmalloc_caches + 6);
2817 printk(KERN_ERR "\nB. Corruption after free\n");
2818 p = kzalloc(128, GFP_KERNEL);
2821 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2822 validate_slab_cache(kmalloc_caches + 7);
2824 p = kzalloc(256, GFP_KERNEL);
2827 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2828 validate_slab_cache(kmalloc_caches + 8);
2830 p = kzalloc(512, GFP_KERNEL);
2833 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2834 validate_slab_cache(kmalloc_caches + 9);
2837 static void resiliency_test(void) {};
2841 * Generate lists of code addresses where slabcache objects are allocated
2846 unsigned long count;
2859 unsigned long count;
2860 struct location *loc;
2863 static void free_loc_track(struct loc_track *t)
2866 free_pages((unsigned long)t->loc,
2867 get_order(sizeof(struct location) * t->max));
2870 static int alloc_loc_track(struct loc_track *t, unsigned long max)
2876 max = PAGE_SIZE / sizeof(struct location);
2878 order = get_order(sizeof(struct location) * max);
2880 l = (void *)__get_free_pages(GFP_ATOMIC, order);
2886 memcpy(l, t->loc, sizeof(struct location) * t->count);
2894 static int add_location(struct loc_track *t, struct kmem_cache *s,
2895 const struct track *track)
2897 long start, end, pos;
2900 unsigned long age = jiffies - track->when;
2906 pos = start + (end - start + 1) / 2;
2909 * There is nothing at "end". If we end up there
2910 * we need to add something to before end.
2915 caddr = t->loc[pos].addr;
2916 if (track->addr == caddr) {
2922 if (age < l->min_time)
2924 if (age > l->max_time)
2927 if (track->pid < l->min_pid)
2928 l->min_pid = track->pid;
2929 if (track->pid > l->max_pid)
2930 l->max_pid = track->pid;
2932 cpu_set(track->cpu, l->cpus);
2934 node_set(page_to_nid(virt_to_page(track)), l->nodes);
2938 if (track->addr < caddr)
2945 * Not found. Insert new tracking element.
2947 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max))
2953 (t->count - pos) * sizeof(struct location));
2956 l->addr = track->addr;
2960 l->min_pid = track->pid;
2961 l->max_pid = track->pid;
2962 cpus_clear(l->cpus);
2963 cpu_set(track->cpu, l->cpus);
2964 nodes_clear(l->nodes);
2965 node_set(page_to_nid(virt_to_page(track)), l->nodes);
2969 static void process_slab(struct loc_track *t, struct kmem_cache *s,
2970 struct page *page, enum track_item alloc)
2972 void *addr = page_address(page);
2973 DECLARE_BITMAP(map, s->objects);
2976 bitmap_zero(map, s->objects);
2977 for_each_free_object(p, s, page->freelist)
2978 set_bit(slab_index(p, s, addr), map);
2980 for_each_object(p, s, addr)
2981 if (!test_bit(slab_index(p, s, addr), map))
2982 add_location(t, s, get_track(s, p, alloc));
2985 static int list_locations(struct kmem_cache *s, char *buf,
2986 enum track_item alloc)
2996 /* Push back cpu slabs */
2999 for_each_online_node(node) {
3000 struct kmem_cache_node *n = get_node(s, node);
3001 unsigned long flags;
3004 if (!atomic_read(&n->nr_slabs))
3007 spin_lock_irqsave(&n->list_lock, flags);
3008 list_for_each_entry(page, &n->partial, lru)
3009 process_slab(&t, s, page, alloc);
3010 list_for_each_entry(page, &n->full, lru)
3011 process_slab(&t, s, page, alloc);
3012 spin_unlock_irqrestore(&n->list_lock, flags);
3015 for (i = 0; i < t.count; i++) {
3016 struct location *l = &t.loc[i];
3018 if (n > PAGE_SIZE - 100)
3020 n += sprintf(buf + n, "%7ld ", l->count);
3023 n += sprint_symbol(buf + n, (unsigned long)l->addr);
3025 n += sprintf(buf + n, "<not-available>");
3027 if (l->sum_time != l->min_time) {
3028 unsigned long remainder;
3030 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3032 div_long_long_rem(l->sum_time, l->count, &remainder),
3035 n += sprintf(buf + n, " age=%ld",
3038 if (l->min_pid != l->max_pid)
3039 n += sprintf(buf + n, " pid=%ld-%ld",
3040 l->min_pid, l->max_pid);
3042 n += sprintf(buf + n, " pid=%ld",
3045 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3046 n < PAGE_SIZE - 60) {
3047 n += sprintf(buf + n, " cpus=");
3048 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3052 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3053 n < PAGE_SIZE - 60) {
3054 n += sprintf(buf + n, " nodes=");
3055 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3059 n += sprintf(buf + n, "\n");
3064 n += sprintf(buf, "No data\n");
3068 static unsigned long count_partial(struct kmem_cache_node *n)
3070 unsigned long flags;
3071 unsigned long x = 0;
3074 spin_lock_irqsave(&n->list_lock, flags);
3075 list_for_each_entry(page, &n->partial, lru)
3077 spin_unlock_irqrestore(&n->list_lock, flags);
3081 enum slab_stat_type {
3088 #define SO_FULL (1 << SL_FULL)
3089 #define SO_PARTIAL (1 << SL_PARTIAL)
3090 #define SO_CPU (1 << SL_CPU)
3091 #define SO_OBJECTS (1 << SL_OBJECTS)
3093 static unsigned long slab_objects(struct kmem_cache *s,
3094 char *buf, unsigned long flags)
3096 unsigned long total = 0;
3100 unsigned long *nodes;
3101 unsigned long *per_cpu;
3103 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3104 per_cpu = nodes + nr_node_ids;
3106 for_each_possible_cpu(cpu) {
3107 struct page *page = s->cpu_slab[cpu];
3111 node = page_to_nid(page);
3112 if (flags & SO_CPU) {
3115 if (flags & SO_OBJECTS)
3126 for_each_online_node(node) {
3127 struct kmem_cache_node *n = get_node(s, node);
3129 if (flags & SO_PARTIAL) {
3130 if (flags & SO_OBJECTS)
3131 x = count_partial(n);
3138 if (flags & SO_FULL) {
3139 int full_slabs = atomic_read(&n->nr_slabs)
3143 if (flags & SO_OBJECTS)
3144 x = full_slabs * s->objects;
3152 x = sprintf(buf, "%lu", total);
3154 for_each_online_node(node)
3156 x += sprintf(buf + x, " N%d=%lu",
3160 return x + sprintf(buf + x, "\n");
3163 static int any_slab_objects(struct kmem_cache *s)
3168 for_each_possible_cpu(cpu)
3169 if (s->cpu_slab[cpu])
3172 for_each_node(node) {
3173 struct kmem_cache_node *n = get_node(s, node);
3175 if (n->nr_partial || atomic_read(&n->nr_slabs))
3181 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3182 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3184 struct slab_attribute {
3185 struct attribute attr;
3186 ssize_t (*show)(struct kmem_cache *s, char *buf);
3187 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3190 #define SLAB_ATTR_RO(_name) \
3191 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3193 #define SLAB_ATTR(_name) \
3194 static struct slab_attribute _name##_attr = \
3195 __ATTR(_name, 0644, _name##_show, _name##_store)
3197 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3199 return sprintf(buf, "%d\n", s->size);
3201 SLAB_ATTR_RO(slab_size);
3203 static ssize_t align_show(struct kmem_cache *s, char *buf)
3205 return sprintf(buf, "%d\n", s->align);
3207 SLAB_ATTR_RO(align);
3209 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3211 return sprintf(buf, "%d\n", s->objsize);
3213 SLAB_ATTR_RO(object_size);
3215 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3217 return sprintf(buf, "%d\n", s->objects);
3219 SLAB_ATTR_RO(objs_per_slab);
3221 static ssize_t order_show(struct kmem_cache *s, char *buf)
3223 return sprintf(buf, "%d\n", s->order);
3225 SLAB_ATTR_RO(order);
3227 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3230 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3232 return n + sprintf(buf + n, "\n");
3238 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3240 return sprintf(buf, "%d\n", s->refcount - 1);
3242 SLAB_ATTR_RO(aliases);
3244 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3246 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3248 SLAB_ATTR_RO(slabs);
3250 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3252 return slab_objects(s, buf, SO_PARTIAL);
3254 SLAB_ATTR_RO(partial);
3256 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3258 return slab_objects(s, buf, SO_CPU);
3260 SLAB_ATTR_RO(cpu_slabs);
3262 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3264 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3266 SLAB_ATTR_RO(objects);
3268 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3270 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3273 static ssize_t sanity_checks_store(struct kmem_cache *s,
3274 const char *buf, size_t length)
3276 s->flags &= ~SLAB_DEBUG_FREE;
3278 s->flags |= SLAB_DEBUG_FREE;
3281 SLAB_ATTR(sanity_checks);
3283 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3285 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3288 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3291 s->flags &= ~SLAB_TRACE;
3293 s->flags |= SLAB_TRACE;
3298 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3300 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3303 static ssize_t reclaim_account_store(struct kmem_cache *s,
3304 const char *buf, size_t length)
3306 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3308 s->flags |= SLAB_RECLAIM_ACCOUNT;
3311 SLAB_ATTR(reclaim_account);
3313 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3315 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3317 SLAB_ATTR_RO(hwcache_align);
3319 #ifdef CONFIG_ZONE_DMA
3320 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3322 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3324 SLAB_ATTR_RO(cache_dma);
3327 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3329 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3331 SLAB_ATTR_RO(destroy_by_rcu);
3333 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3335 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3338 static ssize_t red_zone_store(struct kmem_cache *s,
3339 const char *buf, size_t length)
3341 if (any_slab_objects(s))
3344 s->flags &= ~SLAB_RED_ZONE;
3346 s->flags |= SLAB_RED_ZONE;
3350 SLAB_ATTR(red_zone);
3352 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3354 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3357 static ssize_t poison_store(struct kmem_cache *s,
3358 const char *buf, size_t length)
3360 if (any_slab_objects(s))
3363 s->flags &= ~SLAB_POISON;
3365 s->flags |= SLAB_POISON;
3371 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3373 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3376 static ssize_t store_user_store(struct kmem_cache *s,
3377 const char *buf, size_t length)
3379 if (any_slab_objects(s))
3382 s->flags &= ~SLAB_STORE_USER;
3384 s->flags |= SLAB_STORE_USER;
3388 SLAB_ATTR(store_user);
3390 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3395 static ssize_t validate_store(struct kmem_cache *s,
3396 const char *buf, size_t length)
3399 validate_slab_cache(s);
3404 SLAB_ATTR(validate);
3406 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3411 static ssize_t shrink_store(struct kmem_cache *s,
3412 const char *buf, size_t length)
3414 if (buf[0] == '1') {
3415 int rc = kmem_cache_shrink(s);
3425 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3427 if (!(s->flags & SLAB_STORE_USER))
3429 return list_locations(s, buf, TRACK_ALLOC);
3431 SLAB_ATTR_RO(alloc_calls);
3433 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3435 if (!(s->flags & SLAB_STORE_USER))
3437 return list_locations(s, buf, TRACK_FREE);
3439 SLAB_ATTR_RO(free_calls);
3442 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3444 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3447 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3448 const char *buf, size_t length)
3450 int n = simple_strtoul(buf, NULL, 10);
3453 s->defrag_ratio = n * 10;
3456 SLAB_ATTR(defrag_ratio);
3459 static struct attribute * slab_attrs[] = {
3460 &slab_size_attr.attr,
3461 &object_size_attr.attr,
3462 &objs_per_slab_attr.attr,
3467 &cpu_slabs_attr.attr,
3471 &sanity_checks_attr.attr,
3473 &hwcache_align_attr.attr,
3474 &reclaim_account_attr.attr,
3475 &destroy_by_rcu_attr.attr,
3476 &red_zone_attr.attr,
3478 &store_user_attr.attr,
3479 &validate_attr.attr,
3481 &alloc_calls_attr.attr,
3482 &free_calls_attr.attr,
3483 #ifdef CONFIG_ZONE_DMA
3484 &cache_dma_attr.attr,
3487 &defrag_ratio_attr.attr,
3492 static struct attribute_group slab_attr_group = {
3493 .attrs = slab_attrs,
3496 static ssize_t slab_attr_show(struct kobject *kobj,
3497 struct attribute *attr,
3500 struct slab_attribute *attribute;
3501 struct kmem_cache *s;
3504 attribute = to_slab_attr(attr);
3507 if (!attribute->show)
3510 err = attribute->show(s, buf);
3515 static ssize_t slab_attr_store(struct kobject *kobj,
3516 struct attribute *attr,
3517 const char *buf, size_t len)
3519 struct slab_attribute *attribute;
3520 struct kmem_cache *s;
3523 attribute = to_slab_attr(attr);
3526 if (!attribute->store)
3529 err = attribute->store(s, buf, len);
3534 static struct sysfs_ops slab_sysfs_ops = {
3535 .show = slab_attr_show,
3536 .store = slab_attr_store,
3539 static struct kobj_type slab_ktype = {
3540 .sysfs_ops = &slab_sysfs_ops,
3543 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3545 struct kobj_type *ktype = get_ktype(kobj);
3547 if (ktype == &slab_ktype)
3552 static struct kset_uevent_ops slab_uevent_ops = {
3553 .filter = uevent_filter,
3556 decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3558 #define ID_STR_LENGTH 64
3560 /* Create a unique string id for a slab cache:
3562 * :[flags-]size:[memory address of kmemcache]
3564 static char *create_unique_id(struct kmem_cache *s)
3566 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3573 * First flags affecting slabcache operations. We will only
3574 * get here for aliasable slabs so we do not need to support
3575 * too many flags. The flags here must cover all flags that
3576 * are matched during merging to guarantee that the id is
3579 if (s->flags & SLAB_CACHE_DMA)
3581 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3583 if (s->flags & SLAB_DEBUG_FREE)
3587 p += sprintf(p, "%07d", s->size);
3588 BUG_ON(p > name + ID_STR_LENGTH - 1);
3592 static int sysfs_slab_add(struct kmem_cache *s)
3598 if (slab_state < SYSFS)
3599 /* Defer until later */
3602 unmergeable = slab_unmergeable(s);
3605 * Slabcache can never be merged so we can use the name proper.
3606 * This is typically the case for debug situations. In that
3607 * case we can catch duplicate names easily.
3609 sysfs_remove_link(&slab_subsys.kobj, s->name);
3613 * Create a unique name for the slab as a target
3616 name = create_unique_id(s);
3619 kobj_set_kset_s(s, slab_subsys);
3620 kobject_set_name(&s->kobj, name);
3621 kobject_init(&s->kobj);
3622 err = kobject_add(&s->kobj);
3626 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3629 kobject_uevent(&s->kobj, KOBJ_ADD);
3631 /* Setup first alias */
3632 sysfs_slab_alias(s, s->name);
3638 static void sysfs_slab_remove(struct kmem_cache *s)
3640 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3641 kobject_del(&s->kobj);
3645 * Need to buffer aliases during bootup until sysfs becomes
3646 * available lest we loose that information.
3648 struct saved_alias {
3649 struct kmem_cache *s;
3651 struct saved_alias *next;
3654 struct saved_alias *alias_list;
3656 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3658 struct saved_alias *al;
3660 if (slab_state == SYSFS) {
3662 * If we have a leftover link then remove it.
3664 sysfs_remove_link(&slab_subsys.kobj, name);
3665 return sysfs_create_link(&slab_subsys.kobj,
3669 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3675 al->next = alias_list;
3680 static int __init slab_sysfs_init(void)
3682 struct list_head *h;
3685 err = subsystem_register(&slab_subsys);
3687 printk(KERN_ERR "Cannot register slab subsystem.\n");
3693 list_for_each(h, &slab_caches) {
3694 struct kmem_cache *s =
3695 container_of(h, struct kmem_cache, list);
3697 err = sysfs_slab_add(s);
3701 while (alias_list) {
3702 struct saved_alias *al = alias_list;
3704 alias_list = alias_list->next;
3705 err = sysfs_slab_alias(al->s, al->name);
3714 __initcall(slab_sysfs_init);