2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
23 #include <linux/memory.h>
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
72 * freed then the slab will show up again on the partial lists.
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
80 * Overloading of page flags that are otherwise used for LRU management.
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
94 * freelist that allows lockless access to
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
100 * the fast path and disables lockless freelists.
103 #define FROZEN (1 << PG_active)
105 #ifdef CONFIG_SLUB_DEBUG
106 #define SLABDEBUG (1 << PG_error)
111 static inline int SlabFrozen(struct page *page)
113 return page->flags & FROZEN;
116 static inline void SetSlabFrozen(struct page *page)
118 page->flags |= FROZEN;
121 static inline void ClearSlabFrozen(struct page *page)
123 page->flags &= ~FROZEN;
126 static inline int SlabDebug(struct page *page)
128 return page->flags & SLABDEBUG;
131 static inline void SetSlabDebug(struct page *page)
133 page->flags |= SLABDEBUG;
136 static inline void ClearSlabDebug(struct page *page)
138 page->flags &= ~SLABDEBUG;
142 * Issues still to be resolved:
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
146 * - Variable sizing of the per node arrays
149 /* Enable to test recovery from slab corruption on boot */
150 #undef SLUB_RESILIENCY_TEST
155 * Small page size. Make sure that we do not fragment memory
157 #define DEFAULT_MAX_ORDER 1
158 #define DEFAULT_MIN_OBJECTS 4
163 * Large page machines are customarily able to handle larger
166 #define DEFAULT_MAX_ORDER 2
167 #define DEFAULT_MIN_OBJECTS 8
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
175 #define MIN_PARTIAL 5
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
182 #define MAX_PARTIAL 10
184 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
188 * Set of flags that will prevent slab merging
190 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
193 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
196 #ifndef ARCH_KMALLOC_MINALIGN
197 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
200 #ifndef ARCH_SLAB_MINALIGN
201 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 /* Internal SLUB flags */
205 #define __OBJECT_POISON 0x80000000 /* Poison object */
206 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
207 #define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208 #define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
210 /* Not all arches define cache_line_size */
211 #ifndef cache_line_size
212 #define cache_line_size() L1_CACHE_BYTES
215 static int kmem_size = sizeof(struct kmem_cache);
218 static struct notifier_block slab_notifier;
222 DOWN, /* No slab functionality available */
223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
224 UP, /* Everything works but does not show up in sysfs */
228 /* A list of all slab caches on the system */
229 static DECLARE_RWSEM(slub_lock);
230 static LIST_HEAD(slab_caches);
233 * Tracking user of a slab.
236 void *addr; /* Called from address */
237 int cpu; /* Was running on cpu */
238 int pid; /* Pid context */
239 unsigned long when; /* When did the operation occur */
242 enum track_item { TRACK_ALLOC, TRACK_FREE };
244 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
245 static int sysfs_slab_add(struct kmem_cache *);
246 static int sysfs_slab_alias(struct kmem_cache *, const char *);
247 static void sysfs_slab_remove(struct kmem_cache *);
250 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
253 static inline void sysfs_slab_remove(struct kmem_cache *s)
260 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
262 #ifdef CONFIG_SLUB_STATS
267 /********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
271 int slab_is_available(void)
273 return slab_state >= UP;
276 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
279 return s->node[node];
281 return &s->local_node;
285 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
288 return s->cpu_slab[cpu];
294 static inline int check_valid_pointer(struct kmem_cache *s,
295 struct page *page, const void *object)
302 base = page_address(page);
303 if (object < base || object >= base + s->objects * s->size ||
304 (object - base) % s->size) {
312 * Slow version of get and set free pointer.
314 * This version requires touching the cache lines of kmem_cache which
315 * we avoid to do in the fast alloc free paths. There we obtain the offset
316 * from the page struct.
318 static inline void *get_freepointer(struct kmem_cache *s, void *object)
320 return *(void **)(object + s->offset);
323 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
325 *(void **)(object + s->offset) = fp;
328 /* Loop over all objects in a slab */
329 #define for_each_object(__p, __s, __addr) \
330 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
334 #define for_each_free_object(__p, __s, __free) \
335 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
337 /* Determine object index from a given position */
338 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
340 return (p - addr) / s->size;
343 #ifdef CONFIG_SLUB_DEBUG
347 #ifdef CONFIG_SLUB_DEBUG_ON
348 static int slub_debug = DEBUG_DEFAULT_FLAGS;
350 static int slub_debug;
353 static char *slub_debug_slabs;
358 static void print_section(char *text, u8 *addr, unsigned int length)
366 for (i = 0; i < length; i++) {
368 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
371 printk(KERN_CONT " %02x", addr[i]);
373 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
375 printk(KERN_CONT " %s\n", ascii);
382 printk(KERN_CONT " ");
386 printk(KERN_CONT " %s\n", ascii);
390 static struct track *get_track(struct kmem_cache *s, void *object,
391 enum track_item alloc)
396 p = object + s->offset + sizeof(void *);
398 p = object + s->inuse;
403 static void set_track(struct kmem_cache *s, void *object,
404 enum track_item alloc, void *addr)
409 p = object + s->offset + sizeof(void *);
411 p = object + s->inuse;
416 p->cpu = smp_processor_id();
417 p->pid = current ? current->pid : -1;
420 memset(p, 0, sizeof(struct track));
423 static void init_tracking(struct kmem_cache *s, void *object)
425 if (!(s->flags & SLAB_STORE_USER))
428 set_track(s, object, TRACK_FREE, NULL);
429 set_track(s, object, TRACK_ALLOC, NULL);
432 static void print_track(const char *s, struct track *t)
437 printk(KERN_ERR "INFO: %s in ", s);
438 __print_symbol("%s", (unsigned long)t->addr);
439 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
442 static void print_tracking(struct kmem_cache *s, void *object)
444 if (!(s->flags & SLAB_STORE_USER))
447 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
448 print_track("Freed", get_track(s, object, TRACK_FREE));
451 static void print_page_info(struct page *page)
453 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
454 page, page->inuse, page->freelist, page->flags);
458 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
464 vsnprintf(buf, sizeof(buf), fmt, args);
466 printk(KERN_ERR "========================================"
467 "=====================================\n");
468 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
469 printk(KERN_ERR "----------------------------------------"
470 "-------------------------------------\n\n");
473 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
479 vsnprintf(buf, sizeof(buf), fmt, args);
481 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
484 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
486 unsigned int off; /* Offset of last byte */
487 u8 *addr = page_address(page);
489 print_tracking(s, p);
491 print_page_info(page);
493 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
494 p, p - addr, get_freepointer(s, p));
497 print_section("Bytes b4", p - 16, 16);
499 print_section("Object", p, min(s->objsize, 128));
501 if (s->flags & SLAB_RED_ZONE)
502 print_section("Redzone", p + s->objsize,
503 s->inuse - s->objsize);
506 off = s->offset + sizeof(void *);
510 if (s->flags & SLAB_STORE_USER)
511 off += 2 * sizeof(struct track);
514 /* Beginning of the filler is the free pointer */
515 print_section("Padding", p + off, s->size - off);
520 static void object_err(struct kmem_cache *s, struct page *page,
521 u8 *object, char *reason)
524 print_trailer(s, page, object);
527 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
533 vsnprintf(buf, sizeof(buf), fmt, args);
536 print_page_info(page);
540 static void init_object(struct kmem_cache *s, void *object, int active)
544 if (s->flags & __OBJECT_POISON) {
545 memset(p, POISON_FREE, s->objsize - 1);
546 p[s->objsize - 1] = POISON_END;
549 if (s->flags & SLAB_RED_ZONE)
550 memset(p + s->objsize,
551 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
552 s->inuse - s->objsize);
555 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
558 if (*start != (u8)value)
566 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
567 void *from, void *to)
569 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
570 memset(from, data, to - from);
573 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
574 u8 *object, char *what,
575 u8 *start, unsigned int value, unsigned int bytes)
580 fault = check_bytes(start, value, bytes);
585 while (end > fault && end[-1] == value)
588 slab_bug(s, "%s overwritten", what);
589 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
590 fault, end - 1, fault[0], value);
591 print_trailer(s, page, object);
593 restore_bytes(s, what, value, fault, end);
601 * Bytes of the object to be managed.
602 * If the freepointer may overlay the object then the free
603 * pointer is the first word of the object.
605 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
608 * object + s->objsize
609 * Padding to reach word boundary. This is also used for Redzoning.
610 * Padding is extended by another word if Redzoning is enabled and
613 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
614 * 0xcc (RED_ACTIVE) for objects in use.
617 * Meta data starts here.
619 * A. Free pointer (if we cannot overwrite object on free)
620 * B. Tracking data for SLAB_STORE_USER
621 * C. Padding to reach required alignment boundary or at mininum
622 * one word if debuggin is on to be able to detect writes
623 * before the word boundary.
625 * Padding is done using 0x5a (POISON_INUSE)
628 * Nothing is used beyond s->size.
630 * If slabcaches are merged then the objsize and inuse boundaries are mostly
631 * ignored. And therefore no slab options that rely on these boundaries
632 * may be used with merged slabcaches.
635 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
637 unsigned long off = s->inuse; /* The end of info */
640 /* Freepointer is placed after the object. */
641 off += sizeof(void *);
643 if (s->flags & SLAB_STORE_USER)
644 /* We also have user information there */
645 off += 2 * sizeof(struct track);
650 return check_bytes_and_report(s, page, p, "Object padding",
651 p + off, POISON_INUSE, s->size - off);
654 static int slab_pad_check(struct kmem_cache *s, struct page *page)
662 if (!(s->flags & SLAB_POISON))
665 start = page_address(page);
666 end = start + (PAGE_SIZE << s->order);
667 length = s->objects * s->size;
668 remainder = end - (start + length);
672 fault = check_bytes(start + length, POISON_INUSE, remainder);
675 while (end > fault && end[-1] == POISON_INUSE)
678 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
679 print_section("Padding", start, length);
681 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
685 static int check_object(struct kmem_cache *s, struct page *page,
686 void *object, int active)
689 u8 *endobject = object + s->objsize;
691 if (s->flags & SLAB_RED_ZONE) {
693 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
695 if (!check_bytes_and_report(s, page, object, "Redzone",
696 endobject, red, s->inuse - s->objsize))
699 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
700 check_bytes_and_report(s, page, p, "Alignment padding",
701 endobject, POISON_INUSE, s->inuse - s->objsize);
705 if (s->flags & SLAB_POISON) {
706 if (!active && (s->flags & __OBJECT_POISON) &&
707 (!check_bytes_and_report(s, page, p, "Poison", p,
708 POISON_FREE, s->objsize - 1) ||
709 !check_bytes_and_report(s, page, p, "Poison",
710 p + s->objsize - 1, POISON_END, 1)))
713 * check_pad_bytes cleans up on its own.
715 check_pad_bytes(s, page, p);
718 if (!s->offset && active)
720 * Object and freepointer overlap. Cannot check
721 * freepointer while object is allocated.
725 /* Check free pointer validity */
726 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
727 object_err(s, page, p, "Freepointer corrupt");
729 * No choice but to zap it and thus loose the remainder
730 * of the free objects in this slab. May cause
731 * another error because the object count is now wrong.
733 set_freepointer(s, p, NULL);
739 static int check_slab(struct kmem_cache *s, struct page *page)
741 VM_BUG_ON(!irqs_disabled());
743 if (!PageSlab(page)) {
744 slab_err(s, page, "Not a valid slab page");
747 if (page->inuse > s->objects) {
748 slab_err(s, page, "inuse %u > max %u",
749 s->name, page->inuse, s->objects);
752 /* Slab_pad_check fixes things up after itself */
753 slab_pad_check(s, page);
758 * Determine if a certain object on a page is on the freelist. Must hold the
759 * slab lock to guarantee that the chains are in a consistent state.
761 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
764 void *fp = page->freelist;
767 while (fp && nr <= s->objects) {
770 if (!check_valid_pointer(s, page, fp)) {
772 object_err(s, page, object,
773 "Freechain corrupt");
774 set_freepointer(s, object, NULL);
777 slab_err(s, page, "Freepointer corrupt");
778 page->freelist = NULL;
779 page->inuse = s->objects;
780 slab_fix(s, "Freelist cleared");
786 fp = get_freepointer(s, object);
790 if (page->inuse != s->objects - nr) {
791 slab_err(s, page, "Wrong object count. Counter is %d but "
792 "counted were %d", page->inuse, s->objects - nr);
793 page->inuse = s->objects - nr;
794 slab_fix(s, "Object count adjusted.");
796 return search == NULL;
799 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
801 if (s->flags & SLAB_TRACE) {
802 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
804 alloc ? "alloc" : "free",
809 print_section("Object", (void *)object, s->objsize);
816 * Tracking of fully allocated slabs for debugging purposes.
818 static void add_full(struct kmem_cache_node *n, struct page *page)
820 spin_lock(&n->list_lock);
821 list_add(&page->lru, &n->full);
822 spin_unlock(&n->list_lock);
825 static void remove_full(struct kmem_cache *s, struct page *page)
827 struct kmem_cache_node *n;
829 if (!(s->flags & SLAB_STORE_USER))
832 n = get_node(s, page_to_nid(page));
834 spin_lock(&n->list_lock);
835 list_del(&page->lru);
836 spin_unlock(&n->list_lock);
839 static void setup_object_debug(struct kmem_cache *s, struct page *page,
842 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
845 init_object(s, object, 0);
846 init_tracking(s, object);
849 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
850 void *object, void *addr)
852 if (!check_slab(s, page))
855 if (object && !on_freelist(s, page, object)) {
856 object_err(s, page, object, "Object already allocated");
860 if (!check_valid_pointer(s, page, object)) {
861 object_err(s, page, object, "Freelist Pointer check fails");
865 if (object && !check_object(s, page, object, 0))
868 /* Success perform special debug activities for allocs */
869 if (s->flags & SLAB_STORE_USER)
870 set_track(s, object, TRACK_ALLOC, addr);
871 trace(s, page, object, 1);
872 init_object(s, object, 1);
876 if (PageSlab(page)) {
878 * If this is a slab page then lets do the best we can
879 * to avoid issues in the future. Marking all objects
880 * as used avoids touching the remaining objects.
882 slab_fix(s, "Marking all objects used");
883 page->inuse = s->objects;
884 page->freelist = NULL;
889 static int free_debug_processing(struct kmem_cache *s, struct page *page,
890 void *object, void *addr)
892 if (!check_slab(s, page))
895 if (!check_valid_pointer(s, page, object)) {
896 slab_err(s, page, "Invalid object pointer 0x%p", object);
900 if (on_freelist(s, page, object)) {
901 object_err(s, page, object, "Object already free");
905 if (!check_object(s, page, object, 1))
908 if (unlikely(s != page->slab)) {
909 if (!PageSlab(page)) {
910 slab_err(s, page, "Attempt to free object(0x%p) "
911 "outside of slab", object);
912 } else if (!page->slab) {
914 "SLUB <none>: no slab for object 0x%p.\n",
918 object_err(s, page, object,
919 "page slab pointer corrupt.");
923 /* Special debug activities for freeing objects */
924 if (!SlabFrozen(page) && !page->freelist)
925 remove_full(s, page);
926 if (s->flags & SLAB_STORE_USER)
927 set_track(s, object, TRACK_FREE, addr);
928 trace(s, page, object, 0);
929 init_object(s, object, 0);
933 slab_fix(s, "Object at 0x%p not freed", object);
937 static int __init setup_slub_debug(char *str)
939 slub_debug = DEBUG_DEFAULT_FLAGS;
940 if (*str++ != '=' || !*str)
942 * No options specified. Switch on full debugging.
948 * No options but restriction on slabs. This means full
949 * debugging for slabs matching a pattern.
956 * Switch off all debugging measures.
961 * Determine which debug features should be switched on
963 for (; *str && *str != ','; str++) {
964 switch (tolower(*str)) {
966 slub_debug |= SLAB_DEBUG_FREE;
969 slub_debug |= SLAB_RED_ZONE;
972 slub_debug |= SLAB_POISON;
975 slub_debug |= SLAB_STORE_USER;
978 slub_debug |= SLAB_TRACE;
981 printk(KERN_ERR "slub_debug option '%c' "
982 "unknown. skipped\n", *str);
988 slub_debug_slabs = str + 1;
993 __setup("slub_debug", setup_slub_debug);
995 static unsigned long kmem_cache_flags(unsigned long objsize,
996 unsigned long flags, const char *name,
997 void (*ctor)(struct kmem_cache *, void *))
1000 * The page->offset field is only 16 bit wide. This is an offset
1001 * in units of words from the beginning of an object. If the slab
1002 * size is bigger then we cannot move the free pointer behind the
1005 * On 32 bit platforms the limit is 256k. On 64bit platforms
1006 * the limit is 512k.
1008 * Debugging or ctor may create a need to move the free
1009 * pointer. Fail if this happens.
1011 if (objsize >= 65535 * sizeof(void *)) {
1012 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1013 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1017 * Enable debugging if selected on the kernel commandline.
1019 if (slub_debug && (!slub_debug_slabs ||
1020 strncmp(slub_debug_slabs, name,
1021 strlen(slub_debug_slabs)) == 0))
1022 flags |= slub_debug;
1028 static inline void setup_object_debug(struct kmem_cache *s,
1029 struct page *page, void *object) {}
1031 static inline int alloc_debug_processing(struct kmem_cache *s,
1032 struct page *page, void *object, void *addr) { return 0; }
1034 static inline int free_debug_processing(struct kmem_cache *s,
1035 struct page *page, void *object, void *addr) { return 0; }
1037 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1039 static inline int check_object(struct kmem_cache *s, struct page *page,
1040 void *object, int active) { return 1; }
1041 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1042 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1043 unsigned long flags, const char *name,
1044 void (*ctor)(struct kmem_cache *, void *))
1048 #define slub_debug 0
1051 * Slab allocation and freeing
1053 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1056 int pages = 1 << s->order;
1058 flags |= s->allocflags;
1061 page = alloc_pages(flags, s->order);
1063 page = alloc_pages_node(node, flags, s->order);
1068 mod_zone_page_state(page_zone(page),
1069 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1070 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1076 static void setup_object(struct kmem_cache *s, struct page *page,
1079 setup_object_debug(s, page, object);
1080 if (unlikely(s->ctor))
1084 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1087 struct kmem_cache_node *n;
1092 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1094 page = allocate_slab(s,
1095 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1099 n = get_node(s, page_to_nid(page));
1101 atomic_long_inc(&n->nr_slabs);
1103 page->flags |= 1 << PG_slab;
1104 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1105 SLAB_STORE_USER | SLAB_TRACE))
1108 start = page_address(page);
1110 if (unlikely(s->flags & SLAB_POISON))
1111 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1114 for_each_object(p, s, start) {
1115 setup_object(s, page, last);
1116 set_freepointer(s, last, p);
1119 setup_object(s, page, last);
1120 set_freepointer(s, last, NULL);
1122 page->freelist = start;
1128 static void __free_slab(struct kmem_cache *s, struct page *page)
1130 int pages = 1 << s->order;
1132 if (unlikely(SlabDebug(page))) {
1135 slab_pad_check(s, page);
1136 for_each_object(p, s, page_address(page))
1137 check_object(s, page, p, 0);
1138 ClearSlabDebug(page);
1141 mod_zone_page_state(page_zone(page),
1142 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1143 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1146 __free_pages(page, s->order);
1149 static void rcu_free_slab(struct rcu_head *h)
1153 page = container_of((struct list_head *)h, struct page, lru);
1154 __free_slab(page->slab, page);
1157 static void free_slab(struct kmem_cache *s, struct page *page)
1159 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1161 * RCU free overloads the RCU head over the LRU
1163 struct rcu_head *head = (void *)&page->lru;
1165 call_rcu(head, rcu_free_slab);
1167 __free_slab(s, page);
1170 static void discard_slab(struct kmem_cache *s, struct page *page)
1172 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1174 atomic_long_dec(&n->nr_slabs);
1175 reset_page_mapcount(page);
1176 __ClearPageSlab(page);
1181 * Per slab locking using the pagelock
1183 static __always_inline void slab_lock(struct page *page)
1185 bit_spin_lock(PG_locked, &page->flags);
1188 static __always_inline void slab_unlock(struct page *page)
1190 __bit_spin_unlock(PG_locked, &page->flags);
1193 static __always_inline int slab_trylock(struct page *page)
1197 rc = bit_spin_trylock(PG_locked, &page->flags);
1202 * Management of partially allocated slabs
1204 static void add_partial(struct kmem_cache_node *n,
1205 struct page *page, int tail)
1207 spin_lock(&n->list_lock);
1210 list_add_tail(&page->lru, &n->partial);
1212 list_add(&page->lru, &n->partial);
1213 spin_unlock(&n->list_lock);
1216 static void remove_partial(struct kmem_cache *s,
1219 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1221 spin_lock(&n->list_lock);
1222 list_del(&page->lru);
1224 spin_unlock(&n->list_lock);
1228 * Lock slab and remove from the partial list.
1230 * Must hold list_lock.
1232 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1234 if (slab_trylock(page)) {
1235 list_del(&page->lru);
1237 SetSlabFrozen(page);
1244 * Try to allocate a partial slab from a specific node.
1246 static struct page *get_partial_node(struct kmem_cache_node *n)
1251 * Racy check. If we mistakenly see no partial slabs then we
1252 * just allocate an empty slab. If we mistakenly try to get a
1253 * partial slab and there is none available then get_partials()
1256 if (!n || !n->nr_partial)
1259 spin_lock(&n->list_lock);
1260 list_for_each_entry(page, &n->partial, lru)
1261 if (lock_and_freeze_slab(n, page))
1265 spin_unlock(&n->list_lock);
1270 * Get a page from somewhere. Search in increasing NUMA distances.
1272 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1275 struct zonelist *zonelist;
1280 * The defrag ratio allows a configuration of the tradeoffs between
1281 * inter node defragmentation and node local allocations. A lower
1282 * defrag_ratio increases the tendency to do local allocations
1283 * instead of attempting to obtain partial slabs from other nodes.
1285 * If the defrag_ratio is set to 0 then kmalloc() always
1286 * returns node local objects. If the ratio is higher then kmalloc()
1287 * may return off node objects because partial slabs are obtained
1288 * from other nodes and filled up.
1290 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1291 * defrag_ratio = 1000) then every (well almost) allocation will
1292 * first attempt to defrag slab caches on other nodes. This means
1293 * scanning over all nodes to look for partial slabs which may be
1294 * expensive if we do it every time we are trying to find a slab
1295 * with available objects.
1297 if (!s->remote_node_defrag_ratio ||
1298 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1301 zonelist = &NODE_DATA(
1302 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
1303 for (z = zonelist->zones; *z; z++) {
1304 struct kmem_cache_node *n;
1306 n = get_node(s, zone_to_nid(*z));
1308 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1309 n->nr_partial > MIN_PARTIAL) {
1310 page = get_partial_node(n);
1320 * Get a partial page, lock it and return it.
1322 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1325 int searchnode = (node == -1) ? numa_node_id() : node;
1327 page = get_partial_node(get_node(s, searchnode));
1328 if (page || (flags & __GFP_THISNODE))
1331 return get_any_partial(s, flags);
1335 * Move a page back to the lists.
1337 * Must be called with the slab lock held.
1339 * On exit the slab lock will have been dropped.
1341 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1343 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1344 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1346 ClearSlabFrozen(page);
1349 if (page->freelist) {
1350 add_partial(n, page, tail);
1351 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1353 stat(c, DEACTIVATE_FULL);
1354 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1359 stat(c, DEACTIVATE_EMPTY);
1360 if (n->nr_partial < MIN_PARTIAL) {
1362 * Adding an empty slab to the partial slabs in order
1363 * to avoid page allocator overhead. This slab needs
1364 * to come after the other slabs with objects in
1365 * order to fill them up. That way the size of the
1366 * partial list stays small. kmem_cache_shrink can
1367 * reclaim empty slabs from the partial list.
1369 add_partial(n, page, 1);
1373 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1374 discard_slab(s, page);
1380 * Remove the cpu slab
1382 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1384 struct page *page = c->page;
1388 stat(c, DEACTIVATE_REMOTE_FREES);
1390 * Merge cpu freelist into freelist. Typically we get here
1391 * because both freelists are empty. So this is unlikely
1394 while (unlikely(c->freelist)) {
1397 tail = 0; /* Hot objects. Put the slab first */
1399 /* Retrieve object from cpu_freelist */
1400 object = c->freelist;
1401 c->freelist = c->freelist[c->offset];
1403 /* And put onto the regular freelist */
1404 object[c->offset] = page->freelist;
1405 page->freelist = object;
1409 unfreeze_slab(s, page, tail);
1412 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1414 stat(c, CPUSLAB_FLUSH);
1416 deactivate_slab(s, c);
1421 * Called from IPI handler with interrupts disabled.
1423 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1425 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1427 if (likely(c && c->page))
1431 static void flush_cpu_slab(void *d)
1433 struct kmem_cache *s = d;
1435 __flush_cpu_slab(s, smp_processor_id());
1438 static void flush_all(struct kmem_cache *s)
1441 on_each_cpu(flush_cpu_slab, s, 1, 1);
1443 unsigned long flags;
1445 local_irq_save(flags);
1447 local_irq_restore(flags);
1452 * Check if the objects in a per cpu structure fit numa
1453 * locality expectations.
1455 static inline int node_match(struct kmem_cache_cpu *c, int node)
1458 if (node != -1 && c->node != node)
1465 * Slow path. The lockless freelist is empty or we need to perform
1468 * Interrupts are disabled.
1470 * Processing is still very fast if new objects have been freed to the
1471 * regular freelist. In that case we simply take over the regular freelist
1472 * as the lockless freelist and zap the regular freelist.
1474 * If that is not working then we fall back to the partial lists. We take the
1475 * first element of the freelist as the object to allocate now and move the
1476 * rest of the freelist to the lockless freelist.
1478 * And if we were unable to get a new slab from the partial slab lists then
1479 * we need to allocate a new slab. This is slowest path since we may sleep.
1481 static void *__slab_alloc(struct kmem_cache *s,
1482 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1491 if (unlikely(!node_match(c, node)))
1493 stat(c, ALLOC_REFILL);
1495 object = c->page->freelist;
1496 if (unlikely(!object))
1498 if (unlikely(SlabDebug(c->page)))
1501 object = c->page->freelist;
1502 c->freelist = object[c->offset];
1503 c->page->inuse = s->objects;
1504 c->page->freelist = NULL;
1505 c->node = page_to_nid(c->page);
1507 slab_unlock(c->page);
1508 stat(c, ALLOC_SLOWPATH);
1512 deactivate_slab(s, c);
1515 new = get_partial(s, gfpflags, node);
1518 stat(c, ALLOC_FROM_PARTIAL);
1522 if (gfpflags & __GFP_WAIT)
1525 new = new_slab(s, gfpflags, node);
1527 if (gfpflags & __GFP_WAIT)
1528 local_irq_disable();
1531 c = get_cpu_slab(s, smp_processor_id());
1532 stat(c, ALLOC_SLAB);
1542 * No memory available.
1544 * If the slab uses higher order allocs but the object is
1545 * smaller than a page size then we can fallback in emergencies
1546 * to the page allocator via kmalloc_large. The page allocator may
1547 * have failed to obtain a higher order page and we can try to
1548 * allocate a single page if the object fits into a single page.
1549 * That is only possible if certain conditions are met that are being
1550 * checked when a slab is created.
1552 if (!(gfpflags & __GFP_NORETRY) && (s->flags & __PAGE_ALLOC_FALLBACK))
1553 return kmalloc_large(s->objsize, gfpflags);
1557 object = c->page->freelist;
1558 if (!alloc_debug_processing(s, c->page, object, addr))
1562 c->page->freelist = object[c->offset];
1568 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1569 * have the fastpath folded into their functions. So no function call
1570 * overhead for requests that can be satisfied on the fastpath.
1572 * The fastpath works by first checking if the lockless freelist can be used.
1573 * If not then __slab_alloc is called for slow processing.
1575 * Otherwise we can simply pick the next object from the lockless free list.
1577 static __always_inline void *slab_alloc(struct kmem_cache *s,
1578 gfp_t gfpflags, int node, void *addr)
1581 struct kmem_cache_cpu *c;
1582 unsigned long flags;
1584 local_irq_save(flags);
1585 c = get_cpu_slab(s, smp_processor_id());
1586 if (unlikely(!c->freelist || !node_match(c, node)))
1588 object = __slab_alloc(s, gfpflags, node, addr, c);
1591 object = c->freelist;
1592 c->freelist = object[c->offset];
1593 stat(c, ALLOC_FASTPATH);
1595 local_irq_restore(flags);
1597 if (unlikely((gfpflags & __GFP_ZERO) && object))
1598 memset(object, 0, c->objsize);
1603 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1605 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1607 EXPORT_SYMBOL(kmem_cache_alloc);
1610 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1612 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1614 EXPORT_SYMBOL(kmem_cache_alloc_node);
1618 * Slow patch handling. This may still be called frequently since objects
1619 * have a longer lifetime than the cpu slabs in most processing loads.
1621 * So we still attempt to reduce cache line usage. Just take the slab
1622 * lock and free the item. If there is no additional partial page
1623 * handling required then we can return immediately.
1625 static void __slab_free(struct kmem_cache *s, struct page *page,
1626 void *x, void *addr, unsigned int offset)
1629 void **object = (void *)x;
1630 struct kmem_cache_cpu *c;
1632 c = get_cpu_slab(s, raw_smp_processor_id());
1633 stat(c, FREE_SLOWPATH);
1636 if (unlikely(SlabDebug(page)))
1639 prior = object[offset] = page->freelist;
1640 page->freelist = object;
1643 if (unlikely(SlabFrozen(page))) {
1644 stat(c, FREE_FROZEN);
1648 if (unlikely(!page->inuse))
1652 * Objects left in the slab. If it
1653 * was not on the partial list before
1656 if (unlikely(!prior)) {
1657 add_partial(get_node(s, page_to_nid(page)), page, 1);
1658 stat(c, FREE_ADD_PARTIAL);
1668 * Slab still on the partial list.
1670 remove_partial(s, page);
1671 stat(c, FREE_REMOVE_PARTIAL);
1675 discard_slab(s, page);
1679 if (!free_debug_processing(s, page, x, addr))
1685 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1686 * can perform fastpath freeing without additional function calls.
1688 * The fastpath is only possible if we are freeing to the current cpu slab
1689 * of this processor. This typically the case if we have just allocated
1692 * If fastpath is not possible then fall back to __slab_free where we deal
1693 * with all sorts of special processing.
1695 static __always_inline void slab_free(struct kmem_cache *s,
1696 struct page *page, void *x, void *addr)
1698 void **object = (void *)x;
1699 struct kmem_cache_cpu *c;
1700 unsigned long flags;
1702 local_irq_save(flags);
1703 debug_check_no_locks_freed(object, s->objsize);
1704 c = get_cpu_slab(s, smp_processor_id());
1705 if (likely(page == c->page && c->node >= 0)) {
1706 object[c->offset] = c->freelist;
1707 c->freelist = object;
1708 stat(c, FREE_FASTPATH);
1710 __slab_free(s, page, x, addr, c->offset);
1712 local_irq_restore(flags);
1715 void kmem_cache_free(struct kmem_cache *s, void *x)
1719 page = virt_to_head_page(x);
1721 slab_free(s, page, x, __builtin_return_address(0));
1723 EXPORT_SYMBOL(kmem_cache_free);
1725 /* Figure out on which slab object the object resides */
1726 static struct page *get_object_page(const void *x)
1728 struct page *page = virt_to_head_page(x);
1730 if (!PageSlab(page))
1737 * Object placement in a slab is made very easy because we always start at
1738 * offset 0. If we tune the size of the object to the alignment then we can
1739 * get the required alignment by putting one properly sized object after
1742 * Notice that the allocation order determines the sizes of the per cpu
1743 * caches. Each processor has always one slab available for allocations.
1744 * Increasing the allocation order reduces the number of times that slabs
1745 * must be moved on and off the partial lists and is therefore a factor in
1750 * Mininum / Maximum order of slab pages. This influences locking overhead
1751 * and slab fragmentation. A higher order reduces the number of partial slabs
1752 * and increases the number of allocations possible without having to
1753 * take the list_lock.
1755 static int slub_min_order;
1756 static int slub_max_order = DEFAULT_MAX_ORDER;
1757 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1760 * Merge control. If this is set then no merging of slab caches will occur.
1761 * (Could be removed. This was introduced to pacify the merge skeptics.)
1763 static int slub_nomerge;
1766 * Calculate the order of allocation given an slab object size.
1768 * The order of allocation has significant impact on performance and other
1769 * system components. Generally order 0 allocations should be preferred since
1770 * order 0 does not cause fragmentation in the page allocator. Larger objects
1771 * be problematic to put into order 0 slabs because there may be too much
1772 * unused space left. We go to a higher order if more than 1/8th of the slab
1775 * In order to reach satisfactory performance we must ensure that a minimum
1776 * number of objects is in one slab. Otherwise we may generate too much
1777 * activity on the partial lists which requires taking the list_lock. This is
1778 * less a concern for large slabs though which are rarely used.
1780 * slub_max_order specifies the order where we begin to stop considering the
1781 * number of objects in a slab as critical. If we reach slub_max_order then
1782 * we try to keep the page order as low as possible. So we accept more waste
1783 * of space in favor of a small page order.
1785 * Higher order allocations also allow the placement of more objects in a
1786 * slab and thereby reduce object handling overhead. If the user has
1787 * requested a higher mininum order then we start with that one instead of
1788 * the smallest order which will fit the object.
1790 static inline int slab_order(int size, int min_objects,
1791 int max_order, int fract_leftover)
1795 int min_order = slub_min_order;
1797 for (order = max(min_order,
1798 fls(min_objects * size - 1) - PAGE_SHIFT);
1799 order <= max_order; order++) {
1801 unsigned long slab_size = PAGE_SIZE << order;
1803 if (slab_size < min_objects * size)
1806 rem = slab_size % size;
1808 if (rem <= slab_size / fract_leftover)
1816 static inline int calculate_order(int size)
1823 * Attempt to find best configuration for a slab. This
1824 * works by first attempting to generate a layout with
1825 * the best configuration and backing off gradually.
1827 * First we reduce the acceptable waste in a slab. Then
1828 * we reduce the minimum objects required in a slab.
1830 min_objects = slub_min_objects;
1831 while (min_objects > 1) {
1833 while (fraction >= 4) {
1834 order = slab_order(size, min_objects,
1835 slub_max_order, fraction);
1836 if (order <= slub_max_order)
1844 * We were unable to place multiple objects in a slab. Now
1845 * lets see if we can place a single object there.
1847 order = slab_order(size, 1, slub_max_order, 1);
1848 if (order <= slub_max_order)
1852 * Doh this slab cannot be placed using slub_max_order.
1854 order = slab_order(size, 1, MAX_ORDER, 1);
1855 if (order <= MAX_ORDER)
1861 * Figure out what the alignment of the objects will be.
1863 static unsigned long calculate_alignment(unsigned long flags,
1864 unsigned long align, unsigned long size)
1867 * If the user wants hardware cache aligned objects then
1868 * follow that suggestion if the object is sufficiently
1871 * The hardware cache alignment cannot override the
1872 * specified alignment though. If that is greater
1875 if ((flags & SLAB_HWCACHE_ALIGN) &&
1876 size > cache_line_size() / 2)
1877 return max_t(unsigned long, align, cache_line_size());
1879 if (align < ARCH_SLAB_MINALIGN)
1880 return ARCH_SLAB_MINALIGN;
1882 return ALIGN(align, sizeof(void *));
1885 static void init_kmem_cache_cpu(struct kmem_cache *s,
1886 struct kmem_cache_cpu *c)
1891 c->offset = s->offset / sizeof(void *);
1892 c->objsize = s->objsize;
1895 static void init_kmem_cache_node(struct kmem_cache_node *n)
1898 atomic_long_set(&n->nr_slabs, 0);
1899 spin_lock_init(&n->list_lock);
1900 INIT_LIST_HEAD(&n->partial);
1901 #ifdef CONFIG_SLUB_DEBUG
1902 INIT_LIST_HEAD(&n->full);
1908 * Per cpu array for per cpu structures.
1910 * The per cpu array places all kmem_cache_cpu structures from one processor
1911 * close together meaning that it becomes possible that multiple per cpu
1912 * structures are contained in one cacheline. This may be particularly
1913 * beneficial for the kmalloc caches.
1915 * A desktop system typically has around 60-80 slabs. With 100 here we are
1916 * likely able to get per cpu structures for all caches from the array defined
1917 * here. We must be able to cover all kmalloc caches during bootstrap.
1919 * If the per cpu array is exhausted then fall back to kmalloc
1920 * of individual cachelines. No sharing is possible then.
1922 #define NR_KMEM_CACHE_CPU 100
1924 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1925 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1927 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1928 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1930 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1931 int cpu, gfp_t flags)
1933 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1936 per_cpu(kmem_cache_cpu_free, cpu) =
1937 (void *)c->freelist;
1939 /* Table overflow: So allocate ourselves */
1941 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1942 flags, cpu_to_node(cpu));
1947 init_kmem_cache_cpu(s, c);
1951 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1953 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1954 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1958 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1959 per_cpu(kmem_cache_cpu_free, cpu) = c;
1962 static void free_kmem_cache_cpus(struct kmem_cache *s)
1966 for_each_online_cpu(cpu) {
1967 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1970 s->cpu_slab[cpu] = NULL;
1971 free_kmem_cache_cpu(c, cpu);
1976 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1980 for_each_online_cpu(cpu) {
1981 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1986 c = alloc_kmem_cache_cpu(s, cpu, flags);
1988 free_kmem_cache_cpus(s);
1991 s->cpu_slab[cpu] = c;
1997 * Initialize the per cpu array.
1999 static void init_alloc_cpu_cpu(int cpu)
2003 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2006 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2007 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2009 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2012 static void __init init_alloc_cpu(void)
2016 for_each_online_cpu(cpu)
2017 init_alloc_cpu_cpu(cpu);
2021 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2022 static inline void init_alloc_cpu(void) {}
2024 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2026 init_kmem_cache_cpu(s, &s->cpu_slab);
2033 * No kmalloc_node yet so do it by hand. We know that this is the first
2034 * slab on the node for this slabcache. There are no concurrent accesses
2037 * Note that this function only works on the kmalloc_node_cache
2038 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2039 * memory on a fresh node that has no slab structures yet.
2041 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2045 struct kmem_cache_node *n;
2046 unsigned long flags;
2048 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2050 page = new_slab(kmalloc_caches, gfpflags, node);
2053 if (page_to_nid(page) != node) {
2054 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2056 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2057 "in order to be able to continue\n");
2062 page->freelist = get_freepointer(kmalloc_caches, n);
2064 kmalloc_caches->node[node] = n;
2065 #ifdef CONFIG_SLUB_DEBUG
2066 init_object(kmalloc_caches, n, 1);
2067 init_tracking(kmalloc_caches, n);
2069 init_kmem_cache_node(n);
2070 atomic_long_inc(&n->nr_slabs);
2072 * lockdep requires consistent irq usage for each lock
2073 * so even though there cannot be a race this early in
2074 * the boot sequence, we still disable irqs.
2076 local_irq_save(flags);
2077 add_partial(n, page, 0);
2078 local_irq_restore(flags);
2082 static void free_kmem_cache_nodes(struct kmem_cache *s)
2086 for_each_node_state(node, N_NORMAL_MEMORY) {
2087 struct kmem_cache_node *n = s->node[node];
2088 if (n && n != &s->local_node)
2089 kmem_cache_free(kmalloc_caches, n);
2090 s->node[node] = NULL;
2094 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2099 if (slab_state >= UP)
2100 local_node = page_to_nid(virt_to_page(s));
2104 for_each_node_state(node, N_NORMAL_MEMORY) {
2105 struct kmem_cache_node *n;
2107 if (local_node == node)
2110 if (slab_state == DOWN) {
2111 n = early_kmem_cache_node_alloc(gfpflags,
2115 n = kmem_cache_alloc_node(kmalloc_caches,
2119 free_kmem_cache_nodes(s);
2125 init_kmem_cache_node(n);
2130 static void free_kmem_cache_nodes(struct kmem_cache *s)
2134 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2136 init_kmem_cache_node(&s->local_node);
2142 * calculate_sizes() determines the order and the distribution of data within
2145 static int calculate_sizes(struct kmem_cache *s)
2147 unsigned long flags = s->flags;
2148 unsigned long size = s->objsize;
2149 unsigned long align = s->align;
2152 * Determine if we can poison the object itself. If the user of
2153 * the slab may touch the object after free or before allocation
2154 * then we should never poison the object itself.
2156 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2158 s->flags |= __OBJECT_POISON;
2160 s->flags &= ~__OBJECT_POISON;
2163 * Round up object size to the next word boundary. We can only
2164 * place the free pointer at word boundaries and this determines
2165 * the possible location of the free pointer.
2167 size = ALIGN(size, sizeof(void *));
2169 #ifdef CONFIG_SLUB_DEBUG
2171 * If we are Redzoning then check if there is some space between the
2172 * end of the object and the free pointer. If not then add an
2173 * additional word to have some bytes to store Redzone information.
2175 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2176 size += sizeof(void *);
2180 * With that we have determined the number of bytes in actual use
2181 * by the object. This is the potential offset to the free pointer.
2185 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2188 * Relocate free pointer after the object if it is not
2189 * permitted to overwrite the first word of the object on
2192 * This is the case if we do RCU, have a constructor or
2193 * destructor or are poisoning the objects.
2196 size += sizeof(void *);
2199 #ifdef CONFIG_SLUB_DEBUG
2200 if (flags & SLAB_STORE_USER)
2202 * Need to store information about allocs and frees after
2205 size += 2 * sizeof(struct track);
2207 if (flags & SLAB_RED_ZONE)
2209 * Add some empty padding so that we can catch
2210 * overwrites from earlier objects rather than let
2211 * tracking information or the free pointer be
2212 * corrupted if an user writes before the start
2215 size += sizeof(void *);
2219 * Determine the alignment based on various parameters that the
2220 * user specified and the dynamic determination of cache line size
2223 align = calculate_alignment(flags, align, s->objsize);
2226 * SLUB stores one object immediately after another beginning from
2227 * offset 0. In order to align the objects we have to simply size
2228 * each object to conform to the alignment.
2230 size = ALIGN(size, align);
2233 if ((flags & __KMALLOC_CACHE) &&
2234 PAGE_SIZE / size < slub_min_objects) {
2236 * Kmalloc cache that would not have enough objects in
2237 * an order 0 page. Kmalloc slabs can fallback to
2238 * page allocator order 0 allocs so take a reasonably large
2239 * order that will allows us a good number of objects.
2241 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2242 s->flags |= __PAGE_ALLOC_FALLBACK;
2243 s->allocflags |= __GFP_NOWARN;
2245 s->order = calculate_order(size);
2252 s->allocflags |= __GFP_COMP;
2254 if (s->flags & SLAB_CACHE_DMA)
2255 s->allocflags |= SLUB_DMA;
2257 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2258 s->allocflags |= __GFP_RECLAIMABLE;
2261 * Determine the number of objects per slab
2263 s->objects = (PAGE_SIZE << s->order) / size;
2265 return !!s->objects;
2269 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2270 const char *name, size_t size,
2271 size_t align, unsigned long flags,
2272 void (*ctor)(struct kmem_cache *, void *))
2274 memset(s, 0, kmem_size);
2279 s->flags = kmem_cache_flags(size, flags, name, ctor);
2281 if (!calculate_sizes(s))
2286 s->remote_node_defrag_ratio = 100;
2288 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2291 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2293 free_kmem_cache_nodes(s);
2295 if (flags & SLAB_PANIC)
2296 panic("Cannot create slab %s size=%lu realsize=%u "
2297 "order=%u offset=%u flags=%lx\n",
2298 s->name, (unsigned long)size, s->size, s->order,
2304 * Check if a given pointer is valid
2306 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2310 page = get_object_page(object);
2312 if (!page || s != page->slab)
2313 /* No slab or wrong slab */
2316 if (!check_valid_pointer(s, page, object))
2320 * We could also check if the object is on the slabs freelist.
2321 * But this would be too expensive and it seems that the main
2322 * purpose of kmem_ptr_valid is to check if the object belongs
2323 * to a certain slab.
2327 EXPORT_SYMBOL(kmem_ptr_validate);
2330 * Determine the size of a slab object
2332 unsigned int kmem_cache_size(struct kmem_cache *s)
2336 EXPORT_SYMBOL(kmem_cache_size);
2338 const char *kmem_cache_name(struct kmem_cache *s)
2342 EXPORT_SYMBOL(kmem_cache_name);
2345 * Attempt to free all slabs on a node. Return the number of slabs we
2346 * were unable to free.
2348 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2349 struct list_head *list)
2351 int slabs_inuse = 0;
2352 unsigned long flags;
2353 struct page *page, *h;
2355 spin_lock_irqsave(&n->list_lock, flags);
2356 list_for_each_entry_safe(page, h, list, lru)
2358 list_del(&page->lru);
2359 discard_slab(s, page);
2362 spin_unlock_irqrestore(&n->list_lock, flags);
2367 * Release all resources used by a slab cache.
2369 static inline int kmem_cache_close(struct kmem_cache *s)
2375 /* Attempt to free all objects */
2376 free_kmem_cache_cpus(s);
2377 for_each_node_state(node, N_NORMAL_MEMORY) {
2378 struct kmem_cache_node *n = get_node(s, node);
2380 n->nr_partial -= free_list(s, n, &n->partial);
2381 if (atomic_long_read(&n->nr_slabs))
2384 free_kmem_cache_nodes(s);
2389 * Close a cache and release the kmem_cache structure
2390 * (must be used for caches created using kmem_cache_create)
2392 void kmem_cache_destroy(struct kmem_cache *s)
2394 down_write(&slub_lock);
2398 up_write(&slub_lock);
2399 if (kmem_cache_close(s))
2401 sysfs_slab_remove(s);
2403 up_write(&slub_lock);
2405 EXPORT_SYMBOL(kmem_cache_destroy);
2407 /********************************************************************
2409 *******************************************************************/
2411 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2412 EXPORT_SYMBOL(kmalloc_caches);
2414 #ifdef CONFIG_ZONE_DMA
2415 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2418 static int __init setup_slub_min_order(char *str)
2420 get_option(&str, &slub_min_order);
2425 __setup("slub_min_order=", setup_slub_min_order);
2427 static int __init setup_slub_max_order(char *str)
2429 get_option(&str, &slub_max_order);
2434 __setup("slub_max_order=", setup_slub_max_order);
2436 static int __init setup_slub_min_objects(char *str)
2438 get_option(&str, &slub_min_objects);
2443 __setup("slub_min_objects=", setup_slub_min_objects);
2445 static int __init setup_slub_nomerge(char *str)
2451 __setup("slub_nomerge", setup_slub_nomerge);
2453 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2454 const char *name, int size, gfp_t gfp_flags)
2456 unsigned int flags = 0;
2458 if (gfp_flags & SLUB_DMA)
2459 flags = SLAB_CACHE_DMA;
2461 down_write(&slub_lock);
2462 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2463 flags | __KMALLOC_CACHE, NULL))
2466 list_add(&s->list, &slab_caches);
2467 up_write(&slub_lock);
2468 if (sysfs_slab_add(s))
2473 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2476 #ifdef CONFIG_ZONE_DMA
2478 static void sysfs_add_func(struct work_struct *w)
2480 struct kmem_cache *s;
2482 down_write(&slub_lock);
2483 list_for_each_entry(s, &slab_caches, list) {
2484 if (s->flags & __SYSFS_ADD_DEFERRED) {
2485 s->flags &= ~__SYSFS_ADD_DEFERRED;
2489 up_write(&slub_lock);
2492 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2494 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2496 struct kmem_cache *s;
2500 s = kmalloc_caches_dma[index];
2504 /* Dynamically create dma cache */
2505 if (flags & __GFP_WAIT)
2506 down_write(&slub_lock);
2508 if (!down_write_trylock(&slub_lock))
2512 if (kmalloc_caches_dma[index])
2515 realsize = kmalloc_caches[index].objsize;
2516 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2517 (unsigned int)realsize);
2518 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2520 if (!s || !text || !kmem_cache_open(s, flags, text,
2521 realsize, ARCH_KMALLOC_MINALIGN,
2522 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2528 list_add(&s->list, &slab_caches);
2529 kmalloc_caches_dma[index] = s;
2531 schedule_work(&sysfs_add_work);
2534 up_write(&slub_lock);
2536 return kmalloc_caches_dma[index];
2541 * Conversion table for small slabs sizes / 8 to the index in the
2542 * kmalloc array. This is necessary for slabs < 192 since we have non power
2543 * of two cache sizes there. The size of larger slabs can be determined using
2546 static s8 size_index[24] = {
2573 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2579 return ZERO_SIZE_PTR;
2581 index = size_index[(size - 1) / 8];
2583 index = fls(size - 1);
2585 #ifdef CONFIG_ZONE_DMA
2586 if (unlikely((flags & SLUB_DMA)))
2587 return dma_kmalloc_cache(index, flags);
2590 return &kmalloc_caches[index];
2593 void *__kmalloc(size_t size, gfp_t flags)
2595 struct kmem_cache *s;
2597 if (unlikely(size > PAGE_SIZE))
2598 return kmalloc_large(size, flags);
2600 s = get_slab(size, flags);
2602 if (unlikely(ZERO_OR_NULL_PTR(s)))
2605 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2607 EXPORT_SYMBOL(__kmalloc);
2610 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2612 struct kmem_cache *s;
2614 if (unlikely(size > PAGE_SIZE))
2615 return kmalloc_large(size, flags);
2617 s = get_slab(size, flags);
2619 if (unlikely(ZERO_OR_NULL_PTR(s)))
2622 return slab_alloc(s, flags, node, __builtin_return_address(0));
2624 EXPORT_SYMBOL(__kmalloc_node);
2627 size_t ksize(const void *object)
2630 struct kmem_cache *s;
2633 if (unlikely(object == ZERO_SIZE_PTR))
2636 page = virt_to_head_page(object);
2639 if (unlikely(!PageSlab(page)))
2640 return PAGE_SIZE << compound_order(page);
2646 * Debugging requires use of the padding between object
2647 * and whatever may come after it.
2649 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2653 * If we have the need to store the freelist pointer
2654 * back there or track user information then we can
2655 * only use the space before that information.
2657 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2661 * Else we can use all the padding etc for the allocation
2665 EXPORT_SYMBOL(ksize);
2667 void kfree(const void *x)
2670 void *object = (void *)x;
2672 if (unlikely(ZERO_OR_NULL_PTR(x)))
2675 page = virt_to_head_page(x);
2676 if (unlikely(!PageSlab(page))) {
2680 slab_free(page->slab, page, object, __builtin_return_address(0));
2682 EXPORT_SYMBOL(kfree);
2684 static unsigned long count_partial(struct kmem_cache_node *n)
2686 unsigned long flags;
2687 unsigned long x = 0;
2690 spin_lock_irqsave(&n->list_lock, flags);
2691 list_for_each_entry(page, &n->partial, lru)
2693 spin_unlock_irqrestore(&n->list_lock, flags);
2698 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2699 * the remaining slabs by the number of items in use. The slabs with the
2700 * most items in use come first. New allocations will then fill those up
2701 * and thus they can be removed from the partial lists.
2703 * The slabs with the least items are placed last. This results in them
2704 * being allocated from last increasing the chance that the last objects
2705 * are freed in them.
2707 int kmem_cache_shrink(struct kmem_cache *s)
2711 struct kmem_cache_node *n;
2714 struct list_head *slabs_by_inuse =
2715 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2716 unsigned long flags;
2718 if (!slabs_by_inuse)
2722 for_each_node_state(node, N_NORMAL_MEMORY) {
2723 n = get_node(s, node);
2728 for (i = 0; i < s->objects; i++)
2729 INIT_LIST_HEAD(slabs_by_inuse + i);
2731 spin_lock_irqsave(&n->list_lock, flags);
2734 * Build lists indexed by the items in use in each slab.
2736 * Note that concurrent frees may occur while we hold the
2737 * list_lock. page->inuse here is the upper limit.
2739 list_for_each_entry_safe(page, t, &n->partial, lru) {
2740 if (!page->inuse && slab_trylock(page)) {
2742 * Must hold slab lock here because slab_free
2743 * may have freed the last object and be
2744 * waiting to release the slab.
2746 list_del(&page->lru);
2749 discard_slab(s, page);
2751 list_move(&page->lru,
2752 slabs_by_inuse + page->inuse);
2757 * Rebuild the partial list with the slabs filled up most
2758 * first and the least used slabs at the end.
2760 for (i = s->objects - 1; i >= 0; i--)
2761 list_splice(slabs_by_inuse + i, n->partial.prev);
2763 spin_unlock_irqrestore(&n->list_lock, flags);
2766 kfree(slabs_by_inuse);
2769 EXPORT_SYMBOL(kmem_cache_shrink);
2771 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2772 static int slab_mem_going_offline_callback(void *arg)
2774 struct kmem_cache *s;
2776 down_read(&slub_lock);
2777 list_for_each_entry(s, &slab_caches, list)
2778 kmem_cache_shrink(s);
2779 up_read(&slub_lock);
2784 static void slab_mem_offline_callback(void *arg)
2786 struct kmem_cache_node *n;
2787 struct kmem_cache *s;
2788 struct memory_notify *marg = arg;
2791 offline_node = marg->status_change_nid;
2794 * If the node still has available memory. we need kmem_cache_node
2797 if (offline_node < 0)
2800 down_read(&slub_lock);
2801 list_for_each_entry(s, &slab_caches, list) {
2802 n = get_node(s, offline_node);
2805 * if n->nr_slabs > 0, slabs still exist on the node
2806 * that is going down. We were unable to free them,
2807 * and offline_pages() function shoudn't call this
2808 * callback. So, we must fail.
2810 BUG_ON(atomic_long_read(&n->nr_slabs));
2812 s->node[offline_node] = NULL;
2813 kmem_cache_free(kmalloc_caches, n);
2816 up_read(&slub_lock);
2819 static int slab_mem_going_online_callback(void *arg)
2821 struct kmem_cache_node *n;
2822 struct kmem_cache *s;
2823 struct memory_notify *marg = arg;
2824 int nid = marg->status_change_nid;
2828 * If the node's memory is already available, then kmem_cache_node is
2829 * already created. Nothing to do.
2835 * We are bringing a node online. No memory is availabe yet. We must
2836 * allocate a kmem_cache_node structure in order to bring the node
2839 down_read(&slub_lock);
2840 list_for_each_entry(s, &slab_caches, list) {
2842 * XXX: kmem_cache_alloc_node will fallback to other nodes
2843 * since memory is not yet available from the node that
2846 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2851 init_kmem_cache_node(n);
2855 up_read(&slub_lock);
2859 static int slab_memory_callback(struct notifier_block *self,
2860 unsigned long action, void *arg)
2865 case MEM_GOING_ONLINE:
2866 ret = slab_mem_going_online_callback(arg);
2868 case MEM_GOING_OFFLINE:
2869 ret = slab_mem_going_offline_callback(arg);
2872 case MEM_CANCEL_ONLINE:
2873 slab_mem_offline_callback(arg);
2876 case MEM_CANCEL_OFFLINE:
2880 ret = notifier_from_errno(ret);
2884 #endif /* CONFIG_MEMORY_HOTPLUG */
2886 /********************************************************************
2887 * Basic setup of slabs
2888 *******************************************************************/
2890 void __init kmem_cache_init(void)
2899 * Must first have the slab cache available for the allocations of the
2900 * struct kmem_cache_node's. There is special bootstrap code in
2901 * kmem_cache_open for slab_state == DOWN.
2903 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2904 sizeof(struct kmem_cache_node), GFP_KERNEL);
2905 kmalloc_caches[0].refcount = -1;
2908 hotplug_memory_notifier(slab_memory_callback, 1);
2911 /* Able to allocate the per node structures */
2912 slab_state = PARTIAL;
2914 /* Caches that are not of the two-to-the-power-of size */
2915 if (KMALLOC_MIN_SIZE <= 64) {
2916 create_kmalloc_cache(&kmalloc_caches[1],
2917 "kmalloc-96", 96, GFP_KERNEL);
2920 if (KMALLOC_MIN_SIZE <= 128) {
2921 create_kmalloc_cache(&kmalloc_caches[2],
2922 "kmalloc-192", 192, GFP_KERNEL);
2926 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
2927 create_kmalloc_cache(&kmalloc_caches[i],
2928 "kmalloc", 1 << i, GFP_KERNEL);
2934 * Patch up the size_index table if we have strange large alignment
2935 * requirements for the kmalloc array. This is only the case for
2936 * mips it seems. The standard arches will not generate any code here.
2938 * Largest permitted alignment is 256 bytes due to the way we
2939 * handle the index determination for the smaller caches.
2941 * Make sure that nothing crazy happens if someone starts tinkering
2942 * around with ARCH_KMALLOC_MINALIGN
2944 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2945 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2947 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2948 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2952 /* Provide the correct kmalloc names now that the caches are up */
2953 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
2954 kmalloc_caches[i]. name =
2955 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2958 register_cpu_notifier(&slab_notifier);
2959 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2960 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2962 kmem_size = sizeof(struct kmem_cache);
2967 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2968 " CPUs=%d, Nodes=%d\n",
2969 caches, cache_line_size(),
2970 slub_min_order, slub_max_order, slub_min_objects,
2971 nr_cpu_ids, nr_node_ids);
2975 * Find a mergeable slab cache
2977 static int slab_unmergeable(struct kmem_cache *s)
2979 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2982 if ((s->flags & __PAGE_ALLOC_FALLBACK))
2989 * We may have set a slab to be unmergeable during bootstrap.
2991 if (s->refcount < 0)
2997 static struct kmem_cache *find_mergeable(size_t size,
2998 size_t align, unsigned long flags, const char *name,
2999 void (*ctor)(struct kmem_cache *, void *))
3001 struct kmem_cache *s;
3003 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3009 size = ALIGN(size, sizeof(void *));
3010 align = calculate_alignment(flags, align, size);
3011 size = ALIGN(size, align);
3012 flags = kmem_cache_flags(size, flags, name, NULL);
3014 list_for_each_entry(s, &slab_caches, list) {
3015 if (slab_unmergeable(s))
3021 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3024 * Check if alignment is compatible.
3025 * Courtesy of Adrian Drzewiecki
3027 if ((s->size & ~(align - 1)) != s->size)
3030 if (s->size - size >= sizeof(void *))
3038 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3039 size_t align, unsigned long flags,
3040 void (*ctor)(struct kmem_cache *, void *))
3042 struct kmem_cache *s;
3044 down_write(&slub_lock);
3045 s = find_mergeable(size, align, flags, name, ctor);
3051 * Adjust the object sizes so that we clear
3052 * the complete object on kzalloc.
3054 s->objsize = max(s->objsize, (int)size);
3057 * And then we need to update the object size in the
3058 * per cpu structures
3060 for_each_online_cpu(cpu)
3061 get_cpu_slab(s, cpu)->objsize = s->objsize;
3062 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3063 up_write(&slub_lock);
3064 if (sysfs_slab_alias(s, name))
3068 s = kmalloc(kmem_size, GFP_KERNEL);
3070 if (kmem_cache_open(s, GFP_KERNEL, name,
3071 size, align, flags, ctor)) {
3072 list_add(&s->list, &slab_caches);
3073 up_write(&slub_lock);
3074 if (sysfs_slab_add(s))
3080 up_write(&slub_lock);
3083 if (flags & SLAB_PANIC)
3084 panic("Cannot create slabcache %s\n", name);
3089 EXPORT_SYMBOL(kmem_cache_create);
3093 * Use the cpu notifier to insure that the cpu slabs are flushed when
3096 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3097 unsigned long action, void *hcpu)
3099 long cpu = (long)hcpu;
3100 struct kmem_cache *s;
3101 unsigned long flags;
3104 case CPU_UP_PREPARE:
3105 case CPU_UP_PREPARE_FROZEN:
3106 init_alloc_cpu_cpu(cpu);
3107 down_read(&slub_lock);
3108 list_for_each_entry(s, &slab_caches, list)
3109 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3111 up_read(&slub_lock);
3114 case CPU_UP_CANCELED:
3115 case CPU_UP_CANCELED_FROZEN:
3117 case CPU_DEAD_FROZEN:
3118 down_read(&slub_lock);
3119 list_for_each_entry(s, &slab_caches, list) {
3120 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3122 local_irq_save(flags);
3123 __flush_cpu_slab(s, cpu);
3124 local_irq_restore(flags);
3125 free_kmem_cache_cpu(c, cpu);
3126 s->cpu_slab[cpu] = NULL;
3128 up_read(&slub_lock);
3136 static struct notifier_block __cpuinitdata slab_notifier = {
3137 .notifier_call = slab_cpuup_callback
3142 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3144 struct kmem_cache *s;
3146 if (unlikely(size > PAGE_SIZE))
3147 return kmalloc_large(size, gfpflags);
3149 s = get_slab(size, gfpflags);
3151 if (unlikely(ZERO_OR_NULL_PTR(s)))
3154 return slab_alloc(s, gfpflags, -1, caller);
3157 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3158 int node, void *caller)
3160 struct kmem_cache *s;
3162 if (unlikely(size > PAGE_SIZE))
3163 return kmalloc_large(size, gfpflags);
3165 s = get_slab(size, gfpflags);
3167 if (unlikely(ZERO_OR_NULL_PTR(s)))
3170 return slab_alloc(s, gfpflags, node, caller);
3173 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3174 static int validate_slab(struct kmem_cache *s, struct page *page,
3178 void *addr = page_address(page);
3180 if (!check_slab(s, page) ||
3181 !on_freelist(s, page, NULL))
3184 /* Now we know that a valid freelist exists */
3185 bitmap_zero(map, s->objects);
3187 for_each_free_object(p, s, page->freelist) {
3188 set_bit(slab_index(p, s, addr), map);
3189 if (!check_object(s, page, p, 0))
3193 for_each_object(p, s, addr)
3194 if (!test_bit(slab_index(p, s, addr), map))
3195 if (!check_object(s, page, p, 1))
3200 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3203 if (slab_trylock(page)) {
3204 validate_slab(s, page, map);
3207 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3210 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3211 if (!SlabDebug(page))
3212 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3213 "on slab 0x%p\n", s->name, page);
3215 if (SlabDebug(page))
3216 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3217 "slab 0x%p\n", s->name, page);
3221 static int validate_slab_node(struct kmem_cache *s,
3222 struct kmem_cache_node *n, unsigned long *map)
3224 unsigned long count = 0;
3226 unsigned long flags;
3228 spin_lock_irqsave(&n->list_lock, flags);
3230 list_for_each_entry(page, &n->partial, lru) {
3231 validate_slab_slab(s, page, map);
3234 if (count != n->nr_partial)
3235 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3236 "counter=%ld\n", s->name, count, n->nr_partial);
3238 if (!(s->flags & SLAB_STORE_USER))
3241 list_for_each_entry(page, &n->full, lru) {
3242 validate_slab_slab(s, page, map);
3245 if (count != atomic_long_read(&n->nr_slabs))
3246 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3247 "counter=%ld\n", s->name, count,
3248 atomic_long_read(&n->nr_slabs));
3251 spin_unlock_irqrestore(&n->list_lock, flags);
3255 static long validate_slab_cache(struct kmem_cache *s)
3258 unsigned long count = 0;
3259 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3260 sizeof(unsigned long), GFP_KERNEL);
3266 for_each_node_state(node, N_NORMAL_MEMORY) {
3267 struct kmem_cache_node *n = get_node(s, node);
3269 count += validate_slab_node(s, n, map);
3275 #ifdef SLUB_RESILIENCY_TEST
3276 static void resiliency_test(void)
3280 printk(KERN_ERR "SLUB resiliency testing\n");
3281 printk(KERN_ERR "-----------------------\n");
3282 printk(KERN_ERR "A. Corruption after allocation\n");
3284 p = kzalloc(16, GFP_KERNEL);
3286 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3287 " 0x12->0x%p\n\n", p + 16);
3289 validate_slab_cache(kmalloc_caches + 4);
3291 /* Hmmm... The next two are dangerous */
3292 p = kzalloc(32, GFP_KERNEL);
3293 p[32 + sizeof(void *)] = 0x34;
3294 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3295 " 0x34 -> -0x%p\n", p);
3297 "If allocated object is overwritten then not detectable\n\n");
3299 validate_slab_cache(kmalloc_caches + 5);
3300 p = kzalloc(64, GFP_KERNEL);
3301 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3303 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3306 "If allocated object is overwritten then not detectable\n\n");
3307 validate_slab_cache(kmalloc_caches + 6);
3309 printk(KERN_ERR "\nB. Corruption after free\n");
3310 p = kzalloc(128, GFP_KERNEL);
3313 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3314 validate_slab_cache(kmalloc_caches + 7);
3316 p = kzalloc(256, GFP_KERNEL);
3319 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3321 validate_slab_cache(kmalloc_caches + 8);
3323 p = kzalloc(512, GFP_KERNEL);
3326 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3327 validate_slab_cache(kmalloc_caches + 9);
3330 static void resiliency_test(void) {};
3334 * Generate lists of code addresses where slabcache objects are allocated
3339 unsigned long count;
3352 unsigned long count;
3353 struct location *loc;
3356 static void free_loc_track(struct loc_track *t)
3359 free_pages((unsigned long)t->loc,
3360 get_order(sizeof(struct location) * t->max));
3363 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3368 order = get_order(sizeof(struct location) * max);
3370 l = (void *)__get_free_pages(flags, order);
3375 memcpy(l, t->loc, sizeof(struct location) * t->count);
3383 static int add_location(struct loc_track *t, struct kmem_cache *s,
3384 const struct track *track)
3386 long start, end, pos;
3389 unsigned long age = jiffies - track->when;
3395 pos = start + (end - start + 1) / 2;
3398 * There is nothing at "end". If we end up there
3399 * we need to add something to before end.
3404 caddr = t->loc[pos].addr;
3405 if (track->addr == caddr) {
3411 if (age < l->min_time)
3413 if (age > l->max_time)
3416 if (track->pid < l->min_pid)
3417 l->min_pid = track->pid;
3418 if (track->pid > l->max_pid)
3419 l->max_pid = track->pid;
3421 cpu_set(track->cpu, l->cpus);
3423 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3427 if (track->addr < caddr)
3434 * Not found. Insert new tracking element.
3436 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3442 (t->count - pos) * sizeof(struct location));
3445 l->addr = track->addr;
3449 l->min_pid = track->pid;
3450 l->max_pid = track->pid;
3451 cpus_clear(l->cpus);
3452 cpu_set(track->cpu, l->cpus);
3453 nodes_clear(l->nodes);
3454 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3458 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3459 struct page *page, enum track_item alloc)
3461 void *addr = page_address(page);
3462 DECLARE_BITMAP(map, s->objects);
3465 bitmap_zero(map, s->objects);
3466 for_each_free_object(p, s, page->freelist)
3467 set_bit(slab_index(p, s, addr), map);
3469 for_each_object(p, s, addr)
3470 if (!test_bit(slab_index(p, s, addr), map))
3471 add_location(t, s, get_track(s, p, alloc));
3474 static int list_locations(struct kmem_cache *s, char *buf,
3475 enum track_item alloc)
3479 struct loc_track t = { 0, 0, NULL };
3482 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3484 return sprintf(buf, "Out of memory\n");
3486 /* Push back cpu slabs */
3489 for_each_node_state(node, N_NORMAL_MEMORY) {
3490 struct kmem_cache_node *n = get_node(s, node);
3491 unsigned long flags;
3494 if (!atomic_long_read(&n->nr_slabs))
3497 spin_lock_irqsave(&n->list_lock, flags);
3498 list_for_each_entry(page, &n->partial, lru)
3499 process_slab(&t, s, page, alloc);
3500 list_for_each_entry(page, &n->full, lru)
3501 process_slab(&t, s, page, alloc);
3502 spin_unlock_irqrestore(&n->list_lock, flags);
3505 for (i = 0; i < t.count; i++) {
3506 struct location *l = &t.loc[i];
3508 if (len > PAGE_SIZE - 100)
3510 len += sprintf(buf + len, "%7ld ", l->count);
3513 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3515 len += sprintf(buf + len, "<not-available>");
3517 if (l->sum_time != l->min_time) {
3518 unsigned long remainder;
3520 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3522 div_long_long_rem(l->sum_time, l->count, &remainder),
3525 len += sprintf(buf + len, " age=%ld",
3528 if (l->min_pid != l->max_pid)
3529 len += sprintf(buf + len, " pid=%ld-%ld",
3530 l->min_pid, l->max_pid);
3532 len += sprintf(buf + len, " pid=%ld",
3535 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3536 len < PAGE_SIZE - 60) {
3537 len += sprintf(buf + len, " cpus=");
3538 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3542 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3543 len < PAGE_SIZE - 60) {
3544 len += sprintf(buf + len, " nodes=");
3545 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3549 len += sprintf(buf + len, "\n");
3554 len += sprintf(buf, "No data\n");
3558 enum slab_stat_type {
3565 #define SO_FULL (1 << SL_FULL)
3566 #define SO_PARTIAL (1 << SL_PARTIAL)
3567 #define SO_CPU (1 << SL_CPU)
3568 #define SO_OBJECTS (1 << SL_OBJECTS)
3570 static unsigned long show_slab_objects(struct kmem_cache *s,
3571 char *buf, unsigned long flags)
3573 unsigned long total = 0;
3577 unsigned long *nodes;
3578 unsigned long *per_cpu;
3580 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3581 per_cpu = nodes + nr_node_ids;
3583 for_each_possible_cpu(cpu) {
3585 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3595 if (flags & SO_CPU) {
3596 if (flags & SO_OBJECTS)
3607 for_each_node_state(node, N_NORMAL_MEMORY) {
3608 struct kmem_cache_node *n = get_node(s, node);
3610 if (flags & SO_PARTIAL) {
3611 if (flags & SO_OBJECTS)
3612 x = count_partial(n);
3619 if (flags & SO_FULL) {
3620 int full_slabs = atomic_long_read(&n->nr_slabs)
3624 if (flags & SO_OBJECTS)
3625 x = full_slabs * s->objects;
3633 x = sprintf(buf, "%lu", total);
3635 for_each_node_state(node, N_NORMAL_MEMORY)
3637 x += sprintf(buf + x, " N%d=%lu",
3641 return x + sprintf(buf + x, "\n");
3644 static int any_slab_objects(struct kmem_cache *s)
3649 for_each_possible_cpu(cpu) {
3650 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3656 for_each_online_node(node) {
3657 struct kmem_cache_node *n = get_node(s, node);
3662 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3668 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3669 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3671 struct slab_attribute {
3672 struct attribute attr;
3673 ssize_t (*show)(struct kmem_cache *s, char *buf);
3674 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3677 #define SLAB_ATTR_RO(_name) \
3678 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3680 #define SLAB_ATTR(_name) \
3681 static struct slab_attribute _name##_attr = \
3682 __ATTR(_name, 0644, _name##_show, _name##_store)
3684 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3686 return sprintf(buf, "%d\n", s->size);
3688 SLAB_ATTR_RO(slab_size);
3690 static ssize_t align_show(struct kmem_cache *s, char *buf)
3692 return sprintf(buf, "%d\n", s->align);
3694 SLAB_ATTR_RO(align);
3696 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3698 return sprintf(buf, "%d\n", s->objsize);
3700 SLAB_ATTR_RO(object_size);
3702 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3704 return sprintf(buf, "%d\n", s->objects);
3706 SLAB_ATTR_RO(objs_per_slab);
3708 static ssize_t order_show(struct kmem_cache *s, char *buf)
3710 return sprintf(buf, "%d\n", s->order);
3712 SLAB_ATTR_RO(order);
3714 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3717 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3719 return n + sprintf(buf + n, "\n");
3725 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3727 return sprintf(buf, "%d\n", s->refcount - 1);
3729 SLAB_ATTR_RO(aliases);
3731 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3733 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3735 SLAB_ATTR_RO(slabs);
3737 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3739 return show_slab_objects(s, buf, SO_PARTIAL);
3741 SLAB_ATTR_RO(partial);
3743 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3745 return show_slab_objects(s, buf, SO_CPU);
3747 SLAB_ATTR_RO(cpu_slabs);
3749 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3751 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3753 SLAB_ATTR_RO(objects);
3755 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3757 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3760 static ssize_t sanity_checks_store(struct kmem_cache *s,
3761 const char *buf, size_t length)
3763 s->flags &= ~SLAB_DEBUG_FREE;
3765 s->flags |= SLAB_DEBUG_FREE;
3768 SLAB_ATTR(sanity_checks);
3770 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3772 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3775 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3778 s->flags &= ~SLAB_TRACE;
3780 s->flags |= SLAB_TRACE;
3785 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3787 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3790 static ssize_t reclaim_account_store(struct kmem_cache *s,
3791 const char *buf, size_t length)
3793 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3795 s->flags |= SLAB_RECLAIM_ACCOUNT;
3798 SLAB_ATTR(reclaim_account);
3800 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3802 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3804 SLAB_ATTR_RO(hwcache_align);
3806 #ifdef CONFIG_ZONE_DMA
3807 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3809 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3811 SLAB_ATTR_RO(cache_dma);
3814 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3816 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3818 SLAB_ATTR_RO(destroy_by_rcu);
3820 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3822 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3825 static ssize_t red_zone_store(struct kmem_cache *s,
3826 const char *buf, size_t length)
3828 if (any_slab_objects(s))
3831 s->flags &= ~SLAB_RED_ZONE;
3833 s->flags |= SLAB_RED_ZONE;
3837 SLAB_ATTR(red_zone);
3839 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3841 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3844 static ssize_t poison_store(struct kmem_cache *s,
3845 const char *buf, size_t length)
3847 if (any_slab_objects(s))
3850 s->flags &= ~SLAB_POISON;
3852 s->flags |= SLAB_POISON;
3858 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3860 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3863 static ssize_t store_user_store(struct kmem_cache *s,
3864 const char *buf, size_t length)
3866 if (any_slab_objects(s))
3869 s->flags &= ~SLAB_STORE_USER;
3871 s->flags |= SLAB_STORE_USER;
3875 SLAB_ATTR(store_user);
3877 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3882 static ssize_t validate_store(struct kmem_cache *s,
3883 const char *buf, size_t length)
3887 if (buf[0] == '1') {
3888 ret = validate_slab_cache(s);
3894 SLAB_ATTR(validate);
3896 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3901 static ssize_t shrink_store(struct kmem_cache *s,
3902 const char *buf, size_t length)
3904 if (buf[0] == '1') {
3905 int rc = kmem_cache_shrink(s);
3915 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3917 if (!(s->flags & SLAB_STORE_USER))
3919 return list_locations(s, buf, TRACK_ALLOC);
3921 SLAB_ATTR_RO(alloc_calls);
3923 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3925 if (!(s->flags & SLAB_STORE_USER))
3927 return list_locations(s, buf, TRACK_FREE);
3929 SLAB_ATTR_RO(free_calls);
3932 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
3934 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
3937 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
3938 const char *buf, size_t length)
3940 int n = simple_strtoul(buf, NULL, 10);
3943 s->remote_node_defrag_ratio = n * 10;
3946 SLAB_ATTR(remote_node_defrag_ratio);
3949 #ifdef CONFIG_SLUB_STATS
3951 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3953 unsigned long sum = 0;
3956 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3961 for_each_online_cpu(cpu) {
3962 unsigned x = get_cpu_slab(s, cpu)->stat[si];
3968 len = sprintf(buf, "%lu", sum);
3970 for_each_online_cpu(cpu) {
3971 if (data[cpu] && len < PAGE_SIZE - 20)
3972 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
3975 return len + sprintf(buf + len, "\n");
3978 #define STAT_ATTR(si, text) \
3979 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
3981 return show_stat(s, buf, si); \
3983 SLAB_ATTR_RO(text); \
3985 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
3986 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
3987 STAT_ATTR(FREE_FASTPATH, free_fastpath);
3988 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
3989 STAT_ATTR(FREE_FROZEN, free_frozen);
3990 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
3991 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
3992 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
3993 STAT_ATTR(ALLOC_SLAB, alloc_slab);
3994 STAT_ATTR(ALLOC_REFILL, alloc_refill);
3995 STAT_ATTR(FREE_SLAB, free_slab);
3996 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
3997 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
3998 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
3999 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4000 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4001 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4005 static struct attribute *slab_attrs[] = {
4006 &slab_size_attr.attr,
4007 &object_size_attr.attr,
4008 &objs_per_slab_attr.attr,
4013 &cpu_slabs_attr.attr,
4017 &sanity_checks_attr.attr,
4019 &hwcache_align_attr.attr,
4020 &reclaim_account_attr.attr,
4021 &destroy_by_rcu_attr.attr,
4022 &red_zone_attr.attr,
4024 &store_user_attr.attr,
4025 &validate_attr.attr,
4027 &alloc_calls_attr.attr,
4028 &free_calls_attr.attr,
4029 #ifdef CONFIG_ZONE_DMA
4030 &cache_dma_attr.attr,
4033 &remote_node_defrag_ratio_attr.attr,
4035 #ifdef CONFIG_SLUB_STATS
4036 &alloc_fastpath_attr.attr,
4037 &alloc_slowpath_attr.attr,
4038 &free_fastpath_attr.attr,
4039 &free_slowpath_attr.attr,
4040 &free_frozen_attr.attr,
4041 &free_add_partial_attr.attr,
4042 &free_remove_partial_attr.attr,
4043 &alloc_from_partial_attr.attr,
4044 &alloc_slab_attr.attr,
4045 &alloc_refill_attr.attr,
4046 &free_slab_attr.attr,
4047 &cpuslab_flush_attr.attr,
4048 &deactivate_full_attr.attr,
4049 &deactivate_empty_attr.attr,
4050 &deactivate_to_head_attr.attr,
4051 &deactivate_to_tail_attr.attr,
4052 &deactivate_remote_frees_attr.attr,
4057 static struct attribute_group slab_attr_group = {
4058 .attrs = slab_attrs,
4061 static ssize_t slab_attr_show(struct kobject *kobj,
4062 struct attribute *attr,
4065 struct slab_attribute *attribute;
4066 struct kmem_cache *s;
4069 attribute = to_slab_attr(attr);
4072 if (!attribute->show)
4075 err = attribute->show(s, buf);
4080 static ssize_t slab_attr_store(struct kobject *kobj,
4081 struct attribute *attr,
4082 const char *buf, size_t len)
4084 struct slab_attribute *attribute;
4085 struct kmem_cache *s;
4088 attribute = to_slab_attr(attr);
4091 if (!attribute->store)
4094 err = attribute->store(s, buf, len);
4099 static void kmem_cache_release(struct kobject *kobj)
4101 struct kmem_cache *s = to_slab(kobj);
4106 static struct sysfs_ops slab_sysfs_ops = {
4107 .show = slab_attr_show,
4108 .store = slab_attr_store,
4111 static struct kobj_type slab_ktype = {
4112 .sysfs_ops = &slab_sysfs_ops,
4113 .release = kmem_cache_release
4116 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4118 struct kobj_type *ktype = get_ktype(kobj);
4120 if (ktype == &slab_ktype)
4125 static struct kset_uevent_ops slab_uevent_ops = {
4126 .filter = uevent_filter,
4129 static struct kset *slab_kset;
4131 #define ID_STR_LENGTH 64
4133 /* Create a unique string id for a slab cache:
4135 * :[flags-]size:[memory address of kmemcache]
4137 static char *create_unique_id(struct kmem_cache *s)
4139 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4146 * First flags affecting slabcache operations. We will only
4147 * get here for aliasable slabs so we do not need to support
4148 * too many flags. The flags here must cover all flags that
4149 * are matched during merging to guarantee that the id is
4152 if (s->flags & SLAB_CACHE_DMA)
4154 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4156 if (s->flags & SLAB_DEBUG_FREE)
4160 p += sprintf(p, "%07d", s->size);
4161 BUG_ON(p > name + ID_STR_LENGTH - 1);
4165 static int sysfs_slab_add(struct kmem_cache *s)
4171 if (slab_state < SYSFS)
4172 /* Defer until later */
4175 unmergeable = slab_unmergeable(s);
4178 * Slabcache can never be merged so we can use the name proper.
4179 * This is typically the case for debug situations. In that
4180 * case we can catch duplicate names easily.
4182 sysfs_remove_link(&slab_kset->kobj, s->name);
4186 * Create a unique name for the slab as a target
4189 name = create_unique_id(s);
4192 s->kobj.kset = slab_kset;
4193 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4195 kobject_put(&s->kobj);
4199 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4202 kobject_uevent(&s->kobj, KOBJ_ADD);
4204 /* Setup first alias */
4205 sysfs_slab_alias(s, s->name);
4211 static void sysfs_slab_remove(struct kmem_cache *s)
4213 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4214 kobject_del(&s->kobj);
4215 kobject_put(&s->kobj);
4219 * Need to buffer aliases during bootup until sysfs becomes
4220 * available lest we loose that information.
4222 struct saved_alias {
4223 struct kmem_cache *s;
4225 struct saved_alias *next;
4228 static struct saved_alias *alias_list;
4230 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4232 struct saved_alias *al;
4234 if (slab_state == SYSFS) {
4236 * If we have a leftover link then remove it.
4238 sysfs_remove_link(&slab_kset->kobj, name);
4239 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4242 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4248 al->next = alias_list;
4253 static int __init slab_sysfs_init(void)
4255 struct kmem_cache *s;
4258 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4260 printk(KERN_ERR "Cannot register slab subsystem.\n");
4266 list_for_each_entry(s, &slab_caches, list) {
4267 err = sysfs_slab_add(s);
4269 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4270 " to sysfs\n", s->name);
4273 while (alias_list) {
4274 struct saved_alias *al = alias_list;
4276 alias_list = alias_list->next;
4277 err = sysfs_slab_alias(al->s, al->name);
4279 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4280 " %s to sysfs\n", s->name);
4288 __initcall(slab_sysfs_init);
4292 * The /proc/slabinfo ABI
4294 #ifdef CONFIG_SLABINFO
4296 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4297 size_t count, loff_t *ppos)
4303 static void print_slabinfo_header(struct seq_file *m)
4305 seq_puts(m, "slabinfo - version: 2.1\n");
4306 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4307 "<objperslab> <pagesperslab>");
4308 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4309 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4313 static void *s_start(struct seq_file *m, loff_t *pos)
4317 down_read(&slub_lock);
4319 print_slabinfo_header(m);
4321 return seq_list_start(&slab_caches, *pos);
4324 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4326 return seq_list_next(p, &slab_caches, pos);
4329 static void s_stop(struct seq_file *m, void *p)
4331 up_read(&slub_lock);
4334 static int s_show(struct seq_file *m, void *p)
4336 unsigned long nr_partials = 0;
4337 unsigned long nr_slabs = 0;
4338 unsigned long nr_inuse = 0;
4339 unsigned long nr_objs;
4340 struct kmem_cache *s;
4343 s = list_entry(p, struct kmem_cache, list);
4345 for_each_online_node(node) {
4346 struct kmem_cache_node *n = get_node(s, node);
4351 nr_partials += n->nr_partial;
4352 nr_slabs += atomic_long_read(&n->nr_slabs);
4353 nr_inuse += count_partial(n);
4356 nr_objs = nr_slabs * s->objects;
4357 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4359 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4360 nr_objs, s->size, s->objects, (1 << s->order));
4361 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4362 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4368 const struct seq_operations slabinfo_op = {
4375 #endif /* CONFIG_SLABINFO */