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 /* Verify that a pointer has an address that is valid within a slab page */
295 static inline int check_valid_pointer(struct kmem_cache *s,
296 struct page *page, const void *object)
303 base = page_address(page);
304 if (object < base || object >= base + s->objects * s->size ||
305 (object - base) % s->size) {
313 * Slow version of get and set free pointer.
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
319 static inline void *get_freepointer(struct kmem_cache *s, void *object)
321 return *(void **)(object + s->offset);
324 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
326 *(void **)(object + s->offset) = fp;
329 /* Loop over all objects in a slab */
330 #define for_each_object(__p, __s, __addr) \
331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
335 #define for_each_free_object(__p, __s, __free) \
336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
338 /* Determine object index from a given position */
339 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
341 return (p - addr) / s->size;
344 #ifdef CONFIG_SLUB_DEBUG
348 #ifdef CONFIG_SLUB_DEBUG_ON
349 static int slub_debug = DEBUG_DEFAULT_FLAGS;
351 static int slub_debug;
354 static char *slub_debug_slabs;
359 static void print_section(char *text, u8 *addr, unsigned int length)
367 for (i = 0; i < length; i++) {
369 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
372 printk(KERN_CONT " %02x", addr[i]);
374 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
376 printk(KERN_CONT " %s\n", ascii);
383 printk(KERN_CONT " ");
387 printk(KERN_CONT " %s\n", ascii);
391 static struct track *get_track(struct kmem_cache *s, void *object,
392 enum track_item alloc)
397 p = object + s->offset + sizeof(void *);
399 p = object + s->inuse;
404 static void set_track(struct kmem_cache *s, void *object,
405 enum track_item alloc, void *addr)
410 p = object + s->offset + sizeof(void *);
412 p = object + s->inuse;
417 p->cpu = smp_processor_id();
418 p->pid = current ? current->pid : -1;
421 memset(p, 0, sizeof(struct track));
424 static void init_tracking(struct kmem_cache *s, void *object)
426 if (!(s->flags & SLAB_STORE_USER))
429 set_track(s, object, TRACK_FREE, NULL);
430 set_track(s, object, TRACK_ALLOC, NULL);
433 static void print_track(const char *s, struct track *t)
438 printk(KERN_ERR "INFO: %s in ", s);
439 __print_symbol("%s", (unsigned long)t->addr);
440 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
443 static void print_tracking(struct kmem_cache *s, void *object)
445 if (!(s->flags & SLAB_STORE_USER))
448 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
449 print_track("Freed", get_track(s, object, TRACK_FREE));
452 static void print_page_info(struct page *page)
454 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
455 page, page->inuse, page->freelist, page->flags);
459 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
465 vsnprintf(buf, sizeof(buf), fmt, args);
467 printk(KERN_ERR "========================================"
468 "=====================================\n");
469 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
470 printk(KERN_ERR "----------------------------------------"
471 "-------------------------------------\n\n");
474 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
480 vsnprintf(buf, sizeof(buf), fmt, args);
482 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
485 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
487 unsigned int off; /* Offset of last byte */
488 u8 *addr = page_address(page);
490 print_tracking(s, p);
492 print_page_info(page);
494 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
495 p, p - addr, get_freepointer(s, p));
498 print_section("Bytes b4", p - 16, 16);
500 print_section("Object", p, min(s->objsize, 128));
502 if (s->flags & SLAB_RED_ZONE)
503 print_section("Redzone", p + s->objsize,
504 s->inuse - s->objsize);
507 off = s->offset + sizeof(void *);
511 if (s->flags & SLAB_STORE_USER)
512 off += 2 * sizeof(struct track);
515 /* Beginning of the filler is the free pointer */
516 print_section("Padding", p + off, s->size - off);
521 static void object_err(struct kmem_cache *s, struct page *page,
522 u8 *object, char *reason)
524 slab_bug(s, "%s", reason);
525 print_trailer(s, page, object);
528 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
534 vsnprintf(buf, sizeof(buf), fmt, args);
536 slab_bug(s, "%s", buf);
537 print_page_info(page);
541 static void init_object(struct kmem_cache *s, void *object, int active)
545 if (s->flags & __OBJECT_POISON) {
546 memset(p, POISON_FREE, s->objsize - 1);
547 p[s->objsize - 1] = POISON_END;
550 if (s->flags & SLAB_RED_ZONE)
551 memset(p + s->objsize,
552 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
553 s->inuse - s->objsize);
556 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
559 if (*start != (u8)value)
567 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
568 void *from, void *to)
570 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
571 memset(from, data, to - from);
574 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
575 u8 *object, char *what,
576 u8 *start, unsigned int value, unsigned int bytes)
581 fault = check_bytes(start, value, bytes);
586 while (end > fault && end[-1] == value)
589 slab_bug(s, "%s overwritten", what);
590 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
591 fault, end - 1, fault[0], value);
592 print_trailer(s, page, object);
594 restore_bytes(s, what, value, fault, end);
602 * Bytes of the object to be managed.
603 * If the freepointer may overlay the object then the free
604 * pointer is the first word of the object.
606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
609 * object + s->objsize
610 * Padding to reach word boundary. This is also used for Redzoning.
611 * Padding is extended by another word if Redzoning is enabled and
614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
615 * 0xcc (RED_ACTIVE) for objects in use.
618 * Meta data starts here.
620 * A. Free pointer (if we cannot overwrite object on free)
621 * B. Tracking data for SLAB_STORE_USER
622 * C. Padding to reach required alignment boundary or at mininum
623 * one word if debugging is on to be able to detect writes
624 * before the word boundary.
626 * Padding is done using 0x5a (POISON_INUSE)
629 * Nothing is used beyond s->size.
631 * If slabcaches are merged then the objsize and inuse boundaries are mostly
632 * ignored. And therefore no slab options that rely on these boundaries
633 * may be used with merged slabcaches.
636 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
638 unsigned long off = s->inuse; /* The end of info */
641 /* Freepointer is placed after the object. */
642 off += sizeof(void *);
644 if (s->flags & SLAB_STORE_USER)
645 /* We also have user information there */
646 off += 2 * sizeof(struct track);
651 return check_bytes_and_report(s, page, p, "Object padding",
652 p + off, POISON_INUSE, s->size - off);
655 static int slab_pad_check(struct kmem_cache *s, struct page *page)
663 if (!(s->flags & SLAB_POISON))
666 start = page_address(page);
667 end = start + (PAGE_SIZE << s->order);
668 length = s->objects * s->size;
669 remainder = end - (start + length);
673 fault = check_bytes(start + length, POISON_INUSE, remainder);
676 while (end > fault && end[-1] == POISON_INUSE)
679 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
680 print_section("Padding", start, length);
682 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
686 static int check_object(struct kmem_cache *s, struct page *page,
687 void *object, int active)
690 u8 *endobject = object + s->objsize;
692 if (s->flags & SLAB_RED_ZONE) {
694 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
696 if (!check_bytes_and_report(s, page, object, "Redzone",
697 endobject, red, s->inuse - s->objsize))
700 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
701 check_bytes_and_report(s, page, p, "Alignment padding",
702 endobject, POISON_INUSE, s->inuse - s->objsize);
706 if (s->flags & SLAB_POISON) {
707 if (!active && (s->flags & __OBJECT_POISON) &&
708 (!check_bytes_and_report(s, page, p, "Poison", p,
709 POISON_FREE, s->objsize - 1) ||
710 !check_bytes_and_report(s, page, p, "Poison",
711 p + s->objsize - 1, POISON_END, 1)))
714 * check_pad_bytes cleans up on its own.
716 check_pad_bytes(s, page, p);
719 if (!s->offset && active)
721 * Object and freepointer overlap. Cannot check
722 * freepointer while object is allocated.
726 /* Check free pointer validity */
727 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
728 object_err(s, page, p, "Freepointer corrupt");
730 * No choice but to zap it and thus loose the remainder
731 * of the free objects in this slab. May cause
732 * another error because the object count is now wrong.
734 set_freepointer(s, p, NULL);
740 static int check_slab(struct kmem_cache *s, struct page *page)
742 VM_BUG_ON(!irqs_disabled());
744 if (!PageSlab(page)) {
745 slab_err(s, page, "Not a valid slab page");
748 if (page->inuse > s->objects) {
749 slab_err(s, page, "inuse %u > max %u",
750 s->name, page->inuse, s->objects);
753 /* Slab_pad_check fixes things up after itself */
754 slab_pad_check(s, page);
759 * Determine if a certain object on a page is on the freelist. Must hold the
760 * slab lock to guarantee that the chains are in a consistent state.
762 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
765 void *fp = page->freelist;
768 while (fp && nr <= s->objects) {
771 if (!check_valid_pointer(s, page, fp)) {
773 object_err(s, page, object,
774 "Freechain corrupt");
775 set_freepointer(s, object, NULL);
778 slab_err(s, page, "Freepointer corrupt");
779 page->freelist = NULL;
780 page->inuse = s->objects;
781 slab_fix(s, "Freelist cleared");
787 fp = get_freepointer(s, object);
791 if (page->inuse != s->objects - nr) {
792 slab_err(s, page, "Wrong object count. Counter is %d but "
793 "counted were %d", page->inuse, s->objects - nr);
794 page->inuse = s->objects - nr;
795 slab_fix(s, "Object count adjusted.");
797 return search == NULL;
800 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
802 if (s->flags & SLAB_TRACE) {
803 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
805 alloc ? "alloc" : "free",
810 print_section("Object", (void *)object, s->objsize);
817 * Tracking of fully allocated slabs for debugging purposes.
819 static void add_full(struct kmem_cache_node *n, struct page *page)
821 spin_lock(&n->list_lock);
822 list_add(&page->lru, &n->full);
823 spin_unlock(&n->list_lock);
826 static void remove_full(struct kmem_cache *s, struct page *page)
828 struct kmem_cache_node *n;
830 if (!(s->flags & SLAB_STORE_USER))
833 n = get_node(s, page_to_nid(page));
835 spin_lock(&n->list_lock);
836 list_del(&page->lru);
837 spin_unlock(&n->list_lock);
840 /* Tracking of the number of slabs for debugging purposes */
841 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
843 struct kmem_cache_node *n = get_node(s, node);
845 return atomic_long_read(&n->nr_slabs);
848 static inline void inc_slabs_node(struct kmem_cache *s, int node)
850 struct kmem_cache_node *n = get_node(s, node);
853 * May be called early in order to allocate a slab for the
854 * kmem_cache_node structure. Solve the chicken-egg
855 * dilemma by deferring the increment of the count during
856 * bootstrap (see early_kmem_cache_node_alloc).
858 if (!NUMA_BUILD || n)
859 atomic_long_inc(&n->nr_slabs);
861 static inline void dec_slabs_node(struct kmem_cache *s, int node)
863 struct kmem_cache_node *n = get_node(s, node);
865 atomic_long_dec(&n->nr_slabs);
868 /* Object debug checks for alloc/free paths */
869 static void setup_object_debug(struct kmem_cache *s, struct page *page,
872 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
875 init_object(s, object, 0);
876 init_tracking(s, object);
879 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
880 void *object, void *addr)
882 if (!check_slab(s, page))
885 if (!on_freelist(s, page, object)) {
886 object_err(s, page, object, "Object already allocated");
890 if (!check_valid_pointer(s, page, object)) {
891 object_err(s, page, object, "Freelist Pointer check fails");
895 if (!check_object(s, page, object, 0))
898 /* Success perform special debug activities for allocs */
899 if (s->flags & SLAB_STORE_USER)
900 set_track(s, object, TRACK_ALLOC, addr);
901 trace(s, page, object, 1);
902 init_object(s, object, 1);
906 if (PageSlab(page)) {
908 * If this is a slab page then lets do the best we can
909 * to avoid issues in the future. Marking all objects
910 * as used avoids touching the remaining objects.
912 slab_fix(s, "Marking all objects used");
913 page->inuse = s->objects;
914 page->freelist = NULL;
919 static int free_debug_processing(struct kmem_cache *s, struct page *page,
920 void *object, void *addr)
922 if (!check_slab(s, page))
925 if (!check_valid_pointer(s, page, object)) {
926 slab_err(s, page, "Invalid object pointer 0x%p", object);
930 if (on_freelist(s, page, object)) {
931 object_err(s, page, object, "Object already free");
935 if (!check_object(s, page, object, 1))
938 if (unlikely(s != page->slab)) {
939 if (!PageSlab(page)) {
940 slab_err(s, page, "Attempt to free object(0x%p) "
941 "outside of slab", object);
942 } else if (!page->slab) {
944 "SLUB <none>: no slab for object 0x%p.\n",
948 object_err(s, page, object,
949 "page slab pointer corrupt.");
953 /* Special debug activities for freeing objects */
954 if (!SlabFrozen(page) && !page->freelist)
955 remove_full(s, page);
956 if (s->flags & SLAB_STORE_USER)
957 set_track(s, object, TRACK_FREE, addr);
958 trace(s, page, object, 0);
959 init_object(s, object, 0);
963 slab_fix(s, "Object at 0x%p not freed", object);
967 static int __init setup_slub_debug(char *str)
969 slub_debug = DEBUG_DEFAULT_FLAGS;
970 if (*str++ != '=' || !*str)
972 * No options specified. Switch on full debugging.
978 * No options but restriction on slabs. This means full
979 * debugging for slabs matching a pattern.
986 * Switch off all debugging measures.
991 * Determine which debug features should be switched on
993 for (; *str && *str != ','; str++) {
994 switch (tolower(*str)) {
996 slub_debug |= SLAB_DEBUG_FREE;
999 slub_debug |= SLAB_RED_ZONE;
1002 slub_debug |= SLAB_POISON;
1005 slub_debug |= SLAB_STORE_USER;
1008 slub_debug |= SLAB_TRACE;
1011 printk(KERN_ERR "slub_debug option '%c' "
1012 "unknown. skipped\n", *str);
1018 slub_debug_slabs = str + 1;
1023 __setup("slub_debug", setup_slub_debug);
1025 static unsigned long kmem_cache_flags(unsigned long objsize,
1026 unsigned long flags, const char *name,
1027 void (*ctor)(struct kmem_cache *, void *))
1030 * Enable debugging if selected on the kernel commandline.
1032 if (slub_debug && (!slub_debug_slabs ||
1033 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1034 flags |= slub_debug;
1039 static inline void setup_object_debug(struct kmem_cache *s,
1040 struct page *page, void *object) {}
1042 static inline int alloc_debug_processing(struct kmem_cache *s,
1043 struct page *page, void *object, void *addr) { return 0; }
1045 static inline int free_debug_processing(struct kmem_cache *s,
1046 struct page *page, void *object, void *addr) { return 0; }
1048 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1050 static inline int check_object(struct kmem_cache *s, struct page *page,
1051 void *object, int active) { return 1; }
1052 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1053 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1054 unsigned long flags, const char *name,
1055 void (*ctor)(struct kmem_cache *, void *))
1059 #define slub_debug 0
1061 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1063 static inline void inc_slabs_node(struct kmem_cache *s, int node) {}
1064 static inline void dec_slabs_node(struct kmem_cache *s, int node) {}
1067 * Slab allocation and freeing
1069 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1072 int pages = 1 << s->order;
1074 flags |= s->allocflags;
1077 page = alloc_pages(flags, s->order);
1079 page = alloc_pages_node(node, flags, s->order);
1084 mod_zone_page_state(page_zone(page),
1085 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1086 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1092 static void setup_object(struct kmem_cache *s, struct page *page,
1095 setup_object_debug(s, page, object);
1096 if (unlikely(s->ctor))
1100 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1107 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1109 page = allocate_slab(s,
1110 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1114 inc_slabs_node(s, page_to_nid(page));
1116 page->flags |= 1 << PG_slab;
1117 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1118 SLAB_STORE_USER | SLAB_TRACE))
1121 start = page_address(page);
1123 if (unlikely(s->flags & SLAB_POISON))
1124 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1127 for_each_object(p, s, start) {
1128 setup_object(s, page, last);
1129 set_freepointer(s, last, p);
1132 setup_object(s, page, last);
1133 set_freepointer(s, last, NULL);
1135 page->freelist = start;
1141 static void __free_slab(struct kmem_cache *s, struct page *page)
1143 int pages = 1 << s->order;
1145 if (unlikely(SlabDebug(page))) {
1148 slab_pad_check(s, page);
1149 for_each_object(p, s, page_address(page))
1150 check_object(s, page, p, 0);
1151 ClearSlabDebug(page);
1154 mod_zone_page_state(page_zone(page),
1155 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1156 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1159 __ClearPageSlab(page);
1160 reset_page_mapcount(page);
1161 __free_pages(page, s->order);
1164 static void rcu_free_slab(struct rcu_head *h)
1168 page = container_of((struct list_head *)h, struct page, lru);
1169 __free_slab(page->slab, page);
1172 static void free_slab(struct kmem_cache *s, struct page *page)
1174 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1176 * RCU free overloads the RCU head over the LRU
1178 struct rcu_head *head = (void *)&page->lru;
1180 call_rcu(head, rcu_free_slab);
1182 __free_slab(s, page);
1185 static void discard_slab(struct kmem_cache *s, struct page *page)
1187 dec_slabs_node(s, page_to_nid(page));
1192 * Per slab locking using the pagelock
1194 static __always_inline void slab_lock(struct page *page)
1196 bit_spin_lock(PG_locked, &page->flags);
1199 static __always_inline void slab_unlock(struct page *page)
1201 __bit_spin_unlock(PG_locked, &page->flags);
1204 static __always_inline int slab_trylock(struct page *page)
1208 rc = bit_spin_trylock(PG_locked, &page->flags);
1213 * Management of partially allocated slabs
1215 static void add_partial(struct kmem_cache_node *n,
1216 struct page *page, int tail)
1218 spin_lock(&n->list_lock);
1221 list_add_tail(&page->lru, &n->partial);
1223 list_add(&page->lru, &n->partial);
1224 spin_unlock(&n->list_lock);
1227 static void remove_partial(struct kmem_cache *s,
1230 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1232 spin_lock(&n->list_lock);
1233 list_del(&page->lru);
1235 spin_unlock(&n->list_lock);
1239 * Lock slab and remove from the partial list.
1241 * Must hold list_lock.
1243 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1245 if (slab_trylock(page)) {
1246 list_del(&page->lru);
1248 SetSlabFrozen(page);
1255 * Try to allocate a partial slab from a specific node.
1257 static struct page *get_partial_node(struct kmem_cache_node *n)
1262 * Racy check. If we mistakenly see no partial slabs then we
1263 * just allocate an empty slab. If we mistakenly try to get a
1264 * partial slab and there is none available then get_partials()
1267 if (!n || !n->nr_partial)
1270 spin_lock(&n->list_lock);
1271 list_for_each_entry(page, &n->partial, lru)
1272 if (lock_and_freeze_slab(n, page))
1276 spin_unlock(&n->list_lock);
1281 * Get a page from somewhere. Search in increasing NUMA distances.
1283 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1286 struct zonelist *zonelist;
1289 enum zone_type high_zoneidx = gfp_zone(flags);
1293 * The defrag ratio allows a configuration of the tradeoffs between
1294 * inter node defragmentation and node local allocations. A lower
1295 * defrag_ratio increases the tendency to do local allocations
1296 * instead of attempting to obtain partial slabs from other nodes.
1298 * If the defrag_ratio is set to 0 then kmalloc() always
1299 * returns node local objects. If the ratio is higher then kmalloc()
1300 * may return off node objects because partial slabs are obtained
1301 * from other nodes and filled up.
1303 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1304 * defrag_ratio = 1000) then every (well almost) allocation will
1305 * first attempt to defrag slab caches on other nodes. This means
1306 * scanning over all nodes to look for partial slabs which may be
1307 * expensive if we do it every time we are trying to find a slab
1308 * with available objects.
1310 if (!s->remote_node_defrag_ratio ||
1311 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1314 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1315 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1316 struct kmem_cache_node *n;
1318 n = get_node(s, zone_to_nid(zone));
1320 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1321 n->nr_partial > MIN_PARTIAL) {
1322 page = get_partial_node(n);
1332 * Get a partial page, lock it and return it.
1334 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1337 int searchnode = (node == -1) ? numa_node_id() : node;
1339 page = get_partial_node(get_node(s, searchnode));
1340 if (page || (flags & __GFP_THISNODE))
1343 return get_any_partial(s, flags);
1347 * Move a page back to the lists.
1349 * Must be called with the slab lock held.
1351 * On exit the slab lock will have been dropped.
1353 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1355 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1356 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1358 ClearSlabFrozen(page);
1361 if (page->freelist) {
1362 add_partial(n, page, tail);
1363 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1365 stat(c, DEACTIVATE_FULL);
1366 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1371 stat(c, DEACTIVATE_EMPTY);
1372 if (n->nr_partial < MIN_PARTIAL) {
1374 * Adding an empty slab to the partial slabs in order
1375 * to avoid page allocator overhead. This slab needs
1376 * to come after the other slabs with objects in
1377 * so that the others get filled first. That way the
1378 * size of the partial list stays small.
1380 * kmem_cache_shrink can reclaim any empty slabs from the
1383 add_partial(n, page, 1);
1387 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1388 discard_slab(s, page);
1394 * Remove the cpu slab
1396 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1398 struct page *page = c->page;
1402 stat(c, DEACTIVATE_REMOTE_FREES);
1404 * Merge cpu freelist into slab freelist. Typically we get here
1405 * because both freelists are empty. So this is unlikely
1408 while (unlikely(c->freelist)) {
1411 tail = 0; /* Hot objects. Put the slab first */
1413 /* Retrieve object from cpu_freelist */
1414 object = c->freelist;
1415 c->freelist = c->freelist[c->offset];
1417 /* And put onto the regular freelist */
1418 object[c->offset] = page->freelist;
1419 page->freelist = object;
1423 unfreeze_slab(s, page, tail);
1426 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1428 stat(c, CPUSLAB_FLUSH);
1430 deactivate_slab(s, c);
1436 * Called from IPI handler with interrupts disabled.
1438 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1440 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1442 if (likely(c && c->page))
1446 static void flush_cpu_slab(void *d)
1448 struct kmem_cache *s = d;
1450 __flush_cpu_slab(s, smp_processor_id());
1453 static void flush_all(struct kmem_cache *s)
1456 on_each_cpu(flush_cpu_slab, s, 1, 1);
1458 unsigned long flags;
1460 local_irq_save(flags);
1462 local_irq_restore(flags);
1467 * Check if the objects in a per cpu structure fit numa
1468 * locality expectations.
1470 static inline int node_match(struct kmem_cache_cpu *c, int node)
1473 if (node != -1 && c->node != node)
1480 * Slow path. The lockless freelist is empty or we need to perform
1483 * Interrupts are disabled.
1485 * Processing is still very fast if new objects have been freed to the
1486 * regular freelist. In that case we simply take over the regular freelist
1487 * as the lockless freelist and zap the regular freelist.
1489 * If that is not working then we fall back to the partial lists. We take the
1490 * first element of the freelist as the object to allocate now and move the
1491 * rest of the freelist to the lockless freelist.
1493 * And if we were unable to get a new slab from the partial slab lists then
1494 * we need to allocate a new slab. This is the slowest path since it involves
1495 * a call to the page allocator and the setup of a new slab.
1497 static void *__slab_alloc(struct kmem_cache *s,
1498 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1503 /* We handle __GFP_ZERO in the caller */
1504 gfpflags &= ~__GFP_ZERO;
1510 if (unlikely(!node_match(c, node)))
1513 stat(c, ALLOC_REFILL);
1516 object = c->page->freelist;
1517 if (unlikely(!object))
1519 if (unlikely(SlabDebug(c->page)))
1522 c->freelist = object[c->offset];
1523 c->page->inuse = s->objects;
1524 c->page->freelist = NULL;
1525 c->node = page_to_nid(c->page);
1527 slab_unlock(c->page);
1528 stat(c, ALLOC_SLOWPATH);
1532 deactivate_slab(s, c);
1535 new = get_partial(s, gfpflags, node);
1538 stat(c, ALLOC_FROM_PARTIAL);
1542 if (gfpflags & __GFP_WAIT)
1545 new = new_slab(s, gfpflags, node);
1547 if (gfpflags & __GFP_WAIT)
1548 local_irq_disable();
1551 c = get_cpu_slab(s, smp_processor_id());
1552 stat(c, ALLOC_SLAB);
1562 * No memory available.
1564 * If the slab uses higher order allocs but the object is
1565 * smaller than a page size then we can fallback in emergencies
1566 * to the page allocator via kmalloc_large. The page allocator may
1567 * have failed to obtain a higher order page and we can try to
1568 * allocate a single page if the object fits into a single page.
1569 * That is only possible if certain conditions are met that are being
1570 * checked when a slab is created.
1572 if (!(gfpflags & __GFP_NORETRY) &&
1573 (s->flags & __PAGE_ALLOC_FALLBACK)) {
1574 if (gfpflags & __GFP_WAIT)
1576 object = kmalloc_large(s->objsize, gfpflags);
1577 if (gfpflags & __GFP_WAIT)
1578 local_irq_disable();
1583 if (!alloc_debug_processing(s, c->page, object, addr))
1587 c->page->freelist = object[c->offset];
1593 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1594 * have the fastpath folded into their functions. So no function call
1595 * overhead for requests that can be satisfied on the fastpath.
1597 * The fastpath works by first checking if the lockless freelist can be used.
1598 * If not then __slab_alloc is called for slow processing.
1600 * Otherwise we can simply pick the next object from the lockless free list.
1602 static __always_inline void *slab_alloc(struct kmem_cache *s,
1603 gfp_t gfpflags, int node, void *addr)
1606 struct kmem_cache_cpu *c;
1607 unsigned long flags;
1609 local_irq_save(flags);
1610 c = get_cpu_slab(s, smp_processor_id());
1611 if (unlikely(!c->freelist || !node_match(c, node)))
1613 object = __slab_alloc(s, gfpflags, node, addr, c);
1616 object = c->freelist;
1617 c->freelist = object[c->offset];
1618 stat(c, ALLOC_FASTPATH);
1620 local_irq_restore(flags);
1622 if (unlikely((gfpflags & __GFP_ZERO) && object))
1623 memset(object, 0, c->objsize);
1628 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1630 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1632 EXPORT_SYMBOL(kmem_cache_alloc);
1635 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1637 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1639 EXPORT_SYMBOL(kmem_cache_alloc_node);
1643 * Slow patch handling. This may still be called frequently since objects
1644 * have a longer lifetime than the cpu slabs in most processing loads.
1646 * So we still attempt to reduce cache line usage. Just take the slab
1647 * lock and free the item. If there is no additional partial page
1648 * handling required then we can return immediately.
1650 static void __slab_free(struct kmem_cache *s, struct page *page,
1651 void *x, void *addr, unsigned int offset)
1654 void **object = (void *)x;
1655 struct kmem_cache_cpu *c;
1657 c = get_cpu_slab(s, raw_smp_processor_id());
1658 stat(c, FREE_SLOWPATH);
1661 if (unlikely(SlabDebug(page)))
1665 prior = object[offset] = page->freelist;
1666 page->freelist = object;
1669 if (unlikely(SlabFrozen(page))) {
1670 stat(c, FREE_FROZEN);
1674 if (unlikely(!page->inuse))
1678 * Objects left in the slab. If it was not on the partial list before
1681 if (unlikely(!prior)) {
1682 add_partial(get_node(s, page_to_nid(page)), page, 1);
1683 stat(c, FREE_ADD_PARTIAL);
1693 * Slab still on the partial list.
1695 remove_partial(s, page);
1696 stat(c, FREE_REMOVE_PARTIAL);
1700 discard_slab(s, page);
1704 if (!free_debug_processing(s, page, x, addr))
1710 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1711 * can perform fastpath freeing without additional function calls.
1713 * The fastpath is only possible if we are freeing to the current cpu slab
1714 * of this processor. This typically the case if we have just allocated
1717 * If fastpath is not possible then fall back to __slab_free where we deal
1718 * with all sorts of special processing.
1720 static __always_inline void slab_free(struct kmem_cache *s,
1721 struct page *page, void *x, void *addr)
1723 void **object = (void *)x;
1724 struct kmem_cache_cpu *c;
1725 unsigned long flags;
1727 local_irq_save(flags);
1728 c = get_cpu_slab(s, smp_processor_id());
1729 debug_check_no_locks_freed(object, c->objsize);
1730 if (likely(page == c->page && c->node >= 0)) {
1731 object[c->offset] = c->freelist;
1732 c->freelist = object;
1733 stat(c, FREE_FASTPATH);
1735 __slab_free(s, page, x, addr, c->offset);
1737 local_irq_restore(flags);
1740 void kmem_cache_free(struct kmem_cache *s, void *x)
1744 page = virt_to_head_page(x);
1746 slab_free(s, page, x, __builtin_return_address(0));
1748 EXPORT_SYMBOL(kmem_cache_free);
1750 /* Figure out on which slab object the object resides */
1751 static struct page *get_object_page(const void *x)
1753 struct page *page = virt_to_head_page(x);
1755 if (!PageSlab(page))
1762 * Object placement in a slab is made very easy because we always start at
1763 * offset 0. If we tune the size of the object to the alignment then we can
1764 * get the required alignment by putting one properly sized object after
1767 * Notice that the allocation order determines the sizes of the per cpu
1768 * caches. Each processor has always one slab available for allocations.
1769 * Increasing the allocation order reduces the number of times that slabs
1770 * must be moved on and off the partial lists and is therefore a factor in
1775 * Mininum / Maximum order of slab pages. This influences locking overhead
1776 * and slab fragmentation. A higher order reduces the number of partial slabs
1777 * and increases the number of allocations possible without having to
1778 * take the list_lock.
1780 static int slub_min_order;
1781 static int slub_max_order = DEFAULT_MAX_ORDER;
1782 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1785 * Merge control. If this is set then no merging of slab caches will occur.
1786 * (Could be removed. This was introduced to pacify the merge skeptics.)
1788 static int slub_nomerge;
1791 * Calculate the order of allocation given an slab object size.
1793 * The order of allocation has significant impact on performance and other
1794 * system components. Generally order 0 allocations should be preferred since
1795 * order 0 does not cause fragmentation in the page allocator. Larger objects
1796 * be problematic to put into order 0 slabs because there may be too much
1797 * unused space left. We go to a higher order if more than 1/8th of the slab
1800 * In order to reach satisfactory performance we must ensure that a minimum
1801 * number of objects is in one slab. Otherwise we may generate too much
1802 * activity on the partial lists which requires taking the list_lock. This is
1803 * less a concern for large slabs though which are rarely used.
1805 * slub_max_order specifies the order where we begin to stop considering the
1806 * number of objects in a slab as critical. If we reach slub_max_order then
1807 * we try to keep the page order as low as possible. So we accept more waste
1808 * of space in favor of a small page order.
1810 * Higher order allocations also allow the placement of more objects in a
1811 * slab and thereby reduce object handling overhead. If the user has
1812 * requested a higher mininum order then we start with that one instead of
1813 * the smallest order which will fit the object.
1815 static inline int slab_order(int size, int min_objects,
1816 int max_order, int fract_leftover)
1820 int min_order = slub_min_order;
1822 for (order = max(min_order,
1823 fls(min_objects * size - 1) - PAGE_SHIFT);
1824 order <= max_order; order++) {
1826 unsigned long slab_size = PAGE_SIZE << order;
1828 if (slab_size < min_objects * size)
1831 rem = slab_size % size;
1833 if (rem <= slab_size / fract_leftover)
1841 static inline int calculate_order(int size)
1848 * Attempt to find best configuration for a slab. This
1849 * works by first attempting to generate a layout with
1850 * the best configuration and backing off gradually.
1852 * First we reduce the acceptable waste in a slab. Then
1853 * we reduce the minimum objects required in a slab.
1855 min_objects = slub_min_objects;
1856 while (min_objects > 1) {
1858 while (fraction >= 4) {
1859 order = slab_order(size, min_objects,
1860 slub_max_order, fraction);
1861 if (order <= slub_max_order)
1869 * We were unable to place multiple objects in a slab. Now
1870 * lets see if we can place a single object there.
1872 order = slab_order(size, 1, slub_max_order, 1);
1873 if (order <= slub_max_order)
1877 * Doh this slab cannot be placed using slub_max_order.
1879 order = slab_order(size, 1, MAX_ORDER, 1);
1880 if (order <= MAX_ORDER)
1886 * Figure out what the alignment of the objects will be.
1888 static unsigned long calculate_alignment(unsigned long flags,
1889 unsigned long align, unsigned long size)
1892 * If the user wants hardware cache aligned objects then follow that
1893 * suggestion if the object is sufficiently large.
1895 * The hardware cache alignment cannot override the specified
1896 * alignment though. If that is greater then use it.
1898 if (flags & SLAB_HWCACHE_ALIGN) {
1899 unsigned long ralign = cache_line_size();
1900 while (size <= ralign / 2)
1902 align = max(align, ralign);
1905 if (align < ARCH_SLAB_MINALIGN)
1906 align = ARCH_SLAB_MINALIGN;
1908 return ALIGN(align, sizeof(void *));
1911 static void init_kmem_cache_cpu(struct kmem_cache *s,
1912 struct kmem_cache_cpu *c)
1917 c->offset = s->offset / sizeof(void *);
1918 c->objsize = s->objsize;
1919 #ifdef CONFIG_SLUB_STATS
1920 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1924 static void init_kmem_cache_node(struct kmem_cache_node *n)
1927 spin_lock_init(&n->list_lock);
1928 INIT_LIST_HEAD(&n->partial);
1929 #ifdef CONFIG_SLUB_DEBUG
1930 atomic_long_set(&n->nr_slabs, 0);
1931 INIT_LIST_HEAD(&n->full);
1937 * Per cpu array for per cpu structures.
1939 * The per cpu array places all kmem_cache_cpu structures from one processor
1940 * close together meaning that it becomes possible that multiple per cpu
1941 * structures are contained in one cacheline. This may be particularly
1942 * beneficial for the kmalloc caches.
1944 * A desktop system typically has around 60-80 slabs. With 100 here we are
1945 * likely able to get per cpu structures for all caches from the array defined
1946 * here. We must be able to cover all kmalloc caches during bootstrap.
1948 * If the per cpu array is exhausted then fall back to kmalloc
1949 * of individual cachelines. No sharing is possible then.
1951 #define NR_KMEM_CACHE_CPU 100
1953 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1954 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1956 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1957 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1959 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1960 int cpu, gfp_t flags)
1962 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1965 per_cpu(kmem_cache_cpu_free, cpu) =
1966 (void *)c->freelist;
1968 /* Table overflow: So allocate ourselves */
1970 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1971 flags, cpu_to_node(cpu));
1976 init_kmem_cache_cpu(s, c);
1980 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1982 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1983 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1987 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1988 per_cpu(kmem_cache_cpu_free, cpu) = c;
1991 static void free_kmem_cache_cpus(struct kmem_cache *s)
1995 for_each_online_cpu(cpu) {
1996 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1999 s->cpu_slab[cpu] = NULL;
2000 free_kmem_cache_cpu(c, cpu);
2005 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2009 for_each_online_cpu(cpu) {
2010 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2015 c = alloc_kmem_cache_cpu(s, cpu, flags);
2017 free_kmem_cache_cpus(s);
2020 s->cpu_slab[cpu] = c;
2026 * Initialize the per cpu array.
2028 static void init_alloc_cpu_cpu(int cpu)
2032 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2035 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2036 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2038 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2041 static void __init init_alloc_cpu(void)
2045 for_each_online_cpu(cpu)
2046 init_alloc_cpu_cpu(cpu);
2050 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2051 static inline void init_alloc_cpu(void) {}
2053 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2055 init_kmem_cache_cpu(s, &s->cpu_slab);
2062 * No kmalloc_node yet so do it by hand. We know that this is the first
2063 * slab on the node for this slabcache. There are no concurrent accesses
2066 * Note that this function only works on the kmalloc_node_cache
2067 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2068 * memory on a fresh node that has no slab structures yet.
2070 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2074 struct kmem_cache_node *n;
2075 unsigned long flags;
2077 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2079 page = new_slab(kmalloc_caches, gfpflags, node);
2082 if (page_to_nid(page) != node) {
2083 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2085 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2086 "in order to be able to continue\n");
2091 page->freelist = get_freepointer(kmalloc_caches, n);
2093 kmalloc_caches->node[node] = n;
2094 #ifdef CONFIG_SLUB_DEBUG
2095 init_object(kmalloc_caches, n, 1);
2096 init_tracking(kmalloc_caches, n);
2098 init_kmem_cache_node(n);
2099 inc_slabs_node(kmalloc_caches, node);
2102 * lockdep requires consistent irq usage for each lock
2103 * so even though there cannot be a race this early in
2104 * the boot sequence, we still disable irqs.
2106 local_irq_save(flags);
2107 add_partial(n, page, 0);
2108 local_irq_restore(flags);
2112 static void free_kmem_cache_nodes(struct kmem_cache *s)
2116 for_each_node_state(node, N_NORMAL_MEMORY) {
2117 struct kmem_cache_node *n = s->node[node];
2118 if (n && n != &s->local_node)
2119 kmem_cache_free(kmalloc_caches, n);
2120 s->node[node] = NULL;
2124 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2129 if (slab_state >= UP)
2130 local_node = page_to_nid(virt_to_page(s));
2134 for_each_node_state(node, N_NORMAL_MEMORY) {
2135 struct kmem_cache_node *n;
2137 if (local_node == node)
2140 if (slab_state == DOWN) {
2141 n = early_kmem_cache_node_alloc(gfpflags,
2145 n = kmem_cache_alloc_node(kmalloc_caches,
2149 free_kmem_cache_nodes(s);
2155 init_kmem_cache_node(n);
2160 static void free_kmem_cache_nodes(struct kmem_cache *s)
2164 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2166 init_kmem_cache_node(&s->local_node);
2172 * calculate_sizes() determines the order and the distribution of data within
2175 static int calculate_sizes(struct kmem_cache *s)
2177 unsigned long flags = s->flags;
2178 unsigned long size = s->objsize;
2179 unsigned long align = s->align;
2182 * Round up object size to the next word boundary. We can only
2183 * place the free pointer at word boundaries and this determines
2184 * the possible location of the free pointer.
2186 size = ALIGN(size, sizeof(void *));
2188 #ifdef CONFIG_SLUB_DEBUG
2190 * Determine if we can poison the object itself. If the user of
2191 * the slab may touch the object after free or before allocation
2192 * then we should never poison the object itself.
2194 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2196 s->flags |= __OBJECT_POISON;
2198 s->flags &= ~__OBJECT_POISON;
2202 * If we are Redzoning then check if there is some space between the
2203 * end of the object and the free pointer. If not then add an
2204 * additional word to have some bytes to store Redzone information.
2206 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2207 size += sizeof(void *);
2211 * With that we have determined the number of bytes in actual use
2212 * by the object. This is the potential offset to the free pointer.
2216 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2219 * Relocate free pointer after the object if it is not
2220 * permitted to overwrite the first word of the object on
2223 * This is the case if we do RCU, have a constructor or
2224 * destructor or are poisoning the objects.
2227 size += sizeof(void *);
2230 #ifdef CONFIG_SLUB_DEBUG
2231 if (flags & SLAB_STORE_USER)
2233 * Need to store information about allocs and frees after
2236 size += 2 * sizeof(struct track);
2238 if (flags & SLAB_RED_ZONE)
2240 * Add some empty padding so that we can catch
2241 * overwrites from earlier objects rather than let
2242 * tracking information or the free pointer be
2243 * corrupted if an user writes before the start
2246 size += sizeof(void *);
2250 * Determine the alignment based on various parameters that the
2251 * user specified and the dynamic determination of cache line size
2254 align = calculate_alignment(flags, align, s->objsize);
2257 * SLUB stores one object immediately after another beginning from
2258 * offset 0. In order to align the objects we have to simply size
2259 * each object to conform to the alignment.
2261 size = ALIGN(size, align);
2264 if ((flags & __KMALLOC_CACHE) &&
2265 PAGE_SIZE / size < slub_min_objects) {
2267 * Kmalloc cache that would not have enough objects in
2268 * an order 0 page. Kmalloc slabs can fallback to
2269 * page allocator order 0 allocs so take a reasonably large
2270 * order that will allows us a good number of objects.
2272 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2273 s->flags |= __PAGE_ALLOC_FALLBACK;
2274 s->allocflags |= __GFP_NOWARN;
2276 s->order = calculate_order(size);
2283 s->allocflags |= __GFP_COMP;
2285 if (s->flags & SLAB_CACHE_DMA)
2286 s->allocflags |= SLUB_DMA;
2288 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2289 s->allocflags |= __GFP_RECLAIMABLE;
2292 * Determine the number of objects per slab
2294 s->objects = (PAGE_SIZE << s->order) / size;
2296 return !!s->objects;
2300 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2301 const char *name, size_t size,
2302 size_t align, unsigned long flags,
2303 void (*ctor)(struct kmem_cache *, void *))
2305 memset(s, 0, kmem_size);
2310 s->flags = kmem_cache_flags(size, flags, name, ctor);
2312 if (!calculate_sizes(s))
2317 s->remote_node_defrag_ratio = 100;
2319 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2322 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2324 free_kmem_cache_nodes(s);
2326 if (flags & SLAB_PANIC)
2327 panic("Cannot create slab %s size=%lu realsize=%u "
2328 "order=%u offset=%u flags=%lx\n",
2329 s->name, (unsigned long)size, s->size, s->order,
2335 * Check if a given pointer is valid
2337 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2341 page = get_object_page(object);
2343 if (!page || s != page->slab)
2344 /* No slab or wrong slab */
2347 if (!check_valid_pointer(s, page, object))
2351 * We could also check if the object is on the slabs freelist.
2352 * But this would be too expensive and it seems that the main
2353 * purpose of kmem_ptr_valid() is to check if the object belongs
2354 * to a certain slab.
2358 EXPORT_SYMBOL(kmem_ptr_validate);
2361 * Determine the size of a slab object
2363 unsigned int kmem_cache_size(struct kmem_cache *s)
2367 EXPORT_SYMBOL(kmem_cache_size);
2369 const char *kmem_cache_name(struct kmem_cache *s)
2373 EXPORT_SYMBOL(kmem_cache_name);
2376 * Attempt to free all slabs on a node. Return the number of slabs we
2377 * were unable to free.
2379 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2380 struct list_head *list)
2382 int slabs_inuse = 0;
2383 unsigned long flags;
2384 struct page *page, *h;
2386 spin_lock_irqsave(&n->list_lock, flags);
2387 list_for_each_entry_safe(page, h, list, lru)
2389 list_del(&page->lru);
2390 discard_slab(s, page);
2393 spin_unlock_irqrestore(&n->list_lock, flags);
2398 * Release all resources used by a slab cache.
2400 static inline int kmem_cache_close(struct kmem_cache *s)
2406 /* Attempt to free all objects */
2407 free_kmem_cache_cpus(s);
2408 for_each_node_state(node, N_NORMAL_MEMORY) {
2409 struct kmem_cache_node *n = get_node(s, node);
2411 n->nr_partial -= free_list(s, n, &n->partial);
2412 if (slabs_node(s, node))
2415 free_kmem_cache_nodes(s);
2420 * Close a cache and release the kmem_cache structure
2421 * (must be used for caches created using kmem_cache_create)
2423 void kmem_cache_destroy(struct kmem_cache *s)
2425 down_write(&slub_lock);
2429 up_write(&slub_lock);
2430 if (kmem_cache_close(s))
2432 sysfs_slab_remove(s);
2434 up_write(&slub_lock);
2436 EXPORT_SYMBOL(kmem_cache_destroy);
2438 /********************************************************************
2440 *******************************************************************/
2442 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2443 EXPORT_SYMBOL(kmalloc_caches);
2445 static int __init setup_slub_min_order(char *str)
2447 get_option(&str, &slub_min_order);
2452 __setup("slub_min_order=", setup_slub_min_order);
2454 static int __init setup_slub_max_order(char *str)
2456 get_option(&str, &slub_max_order);
2461 __setup("slub_max_order=", setup_slub_max_order);
2463 static int __init setup_slub_min_objects(char *str)
2465 get_option(&str, &slub_min_objects);
2470 __setup("slub_min_objects=", setup_slub_min_objects);
2472 static int __init setup_slub_nomerge(char *str)
2478 __setup("slub_nomerge", setup_slub_nomerge);
2480 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2481 const char *name, int size, gfp_t gfp_flags)
2483 unsigned int flags = 0;
2485 if (gfp_flags & SLUB_DMA)
2486 flags = SLAB_CACHE_DMA;
2488 down_write(&slub_lock);
2489 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2490 flags | __KMALLOC_CACHE, NULL))
2493 list_add(&s->list, &slab_caches);
2494 up_write(&slub_lock);
2495 if (sysfs_slab_add(s))
2500 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2503 #ifdef CONFIG_ZONE_DMA
2504 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2506 static void sysfs_add_func(struct work_struct *w)
2508 struct kmem_cache *s;
2510 down_write(&slub_lock);
2511 list_for_each_entry(s, &slab_caches, list) {
2512 if (s->flags & __SYSFS_ADD_DEFERRED) {
2513 s->flags &= ~__SYSFS_ADD_DEFERRED;
2517 up_write(&slub_lock);
2520 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2522 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2524 struct kmem_cache *s;
2528 s = kmalloc_caches_dma[index];
2532 /* Dynamically create dma cache */
2533 if (flags & __GFP_WAIT)
2534 down_write(&slub_lock);
2536 if (!down_write_trylock(&slub_lock))
2540 if (kmalloc_caches_dma[index])
2543 realsize = kmalloc_caches[index].objsize;
2544 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2545 (unsigned int)realsize);
2546 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2548 if (!s || !text || !kmem_cache_open(s, flags, text,
2549 realsize, ARCH_KMALLOC_MINALIGN,
2550 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2556 list_add(&s->list, &slab_caches);
2557 kmalloc_caches_dma[index] = s;
2559 schedule_work(&sysfs_add_work);
2562 up_write(&slub_lock);
2564 return kmalloc_caches_dma[index];
2569 * Conversion table for small slabs sizes / 8 to the index in the
2570 * kmalloc array. This is necessary for slabs < 192 since we have non power
2571 * of two cache sizes there. The size of larger slabs can be determined using
2574 static s8 size_index[24] = {
2601 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2607 return ZERO_SIZE_PTR;
2609 index = size_index[(size - 1) / 8];
2611 index = fls(size - 1);
2613 #ifdef CONFIG_ZONE_DMA
2614 if (unlikely((flags & SLUB_DMA)))
2615 return dma_kmalloc_cache(index, flags);
2618 return &kmalloc_caches[index];
2621 void *__kmalloc(size_t size, gfp_t flags)
2623 struct kmem_cache *s;
2625 if (unlikely(size > PAGE_SIZE))
2626 return kmalloc_large(size, flags);
2628 s = get_slab(size, flags);
2630 if (unlikely(ZERO_OR_NULL_PTR(s)))
2633 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2635 EXPORT_SYMBOL(__kmalloc);
2637 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2639 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2643 return page_address(page);
2649 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2651 struct kmem_cache *s;
2653 if (unlikely(size > PAGE_SIZE))
2654 return kmalloc_large_node(size, flags, node);
2656 s = get_slab(size, flags);
2658 if (unlikely(ZERO_OR_NULL_PTR(s)))
2661 return slab_alloc(s, flags, node, __builtin_return_address(0));
2663 EXPORT_SYMBOL(__kmalloc_node);
2666 size_t ksize(const void *object)
2669 struct kmem_cache *s;
2671 if (unlikely(object == ZERO_SIZE_PTR))
2674 page = virt_to_head_page(object);
2676 if (unlikely(!PageSlab(page)))
2677 return PAGE_SIZE << compound_order(page);
2681 #ifdef CONFIG_SLUB_DEBUG
2683 * Debugging requires use of the padding between object
2684 * and whatever may come after it.
2686 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2691 * If we have the need to store the freelist pointer
2692 * back there or track user information then we can
2693 * only use the space before that information.
2695 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2698 * Else we can use all the padding etc for the allocation
2702 EXPORT_SYMBOL(ksize);
2704 void kfree(const void *x)
2707 void *object = (void *)x;
2709 if (unlikely(ZERO_OR_NULL_PTR(x)))
2712 page = virt_to_head_page(x);
2713 if (unlikely(!PageSlab(page))) {
2717 slab_free(page->slab, page, object, __builtin_return_address(0));
2719 EXPORT_SYMBOL(kfree);
2722 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2723 * the remaining slabs by the number of items in use. The slabs with the
2724 * most items in use come first. New allocations will then fill those up
2725 * and thus they can be removed from the partial lists.
2727 * The slabs with the least items are placed last. This results in them
2728 * being allocated from last increasing the chance that the last objects
2729 * are freed in them.
2731 int kmem_cache_shrink(struct kmem_cache *s)
2735 struct kmem_cache_node *n;
2738 struct list_head *slabs_by_inuse =
2739 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2740 unsigned long flags;
2742 if (!slabs_by_inuse)
2746 for_each_node_state(node, N_NORMAL_MEMORY) {
2747 n = get_node(s, node);
2752 for (i = 0; i < s->objects; i++)
2753 INIT_LIST_HEAD(slabs_by_inuse + i);
2755 spin_lock_irqsave(&n->list_lock, flags);
2758 * Build lists indexed by the items in use in each slab.
2760 * Note that concurrent frees may occur while we hold the
2761 * list_lock. page->inuse here is the upper limit.
2763 list_for_each_entry_safe(page, t, &n->partial, lru) {
2764 if (!page->inuse && slab_trylock(page)) {
2766 * Must hold slab lock here because slab_free
2767 * may have freed the last object and be
2768 * waiting to release the slab.
2770 list_del(&page->lru);
2773 discard_slab(s, page);
2775 list_move(&page->lru,
2776 slabs_by_inuse + page->inuse);
2781 * Rebuild the partial list with the slabs filled up most
2782 * first and the least used slabs at the end.
2784 for (i = s->objects - 1; i >= 0; i--)
2785 list_splice(slabs_by_inuse + i, n->partial.prev);
2787 spin_unlock_irqrestore(&n->list_lock, flags);
2790 kfree(slabs_by_inuse);
2793 EXPORT_SYMBOL(kmem_cache_shrink);
2795 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2796 static int slab_mem_going_offline_callback(void *arg)
2798 struct kmem_cache *s;
2800 down_read(&slub_lock);
2801 list_for_each_entry(s, &slab_caches, list)
2802 kmem_cache_shrink(s);
2803 up_read(&slub_lock);
2808 static void slab_mem_offline_callback(void *arg)
2810 struct kmem_cache_node *n;
2811 struct kmem_cache *s;
2812 struct memory_notify *marg = arg;
2815 offline_node = marg->status_change_nid;
2818 * If the node still has available memory. we need kmem_cache_node
2821 if (offline_node < 0)
2824 down_read(&slub_lock);
2825 list_for_each_entry(s, &slab_caches, list) {
2826 n = get_node(s, offline_node);
2829 * if n->nr_slabs > 0, slabs still exist on the node
2830 * that is going down. We were unable to free them,
2831 * and offline_pages() function shoudn't call this
2832 * callback. So, we must fail.
2834 BUG_ON(slabs_node(s, offline_node));
2836 s->node[offline_node] = NULL;
2837 kmem_cache_free(kmalloc_caches, n);
2840 up_read(&slub_lock);
2843 static int slab_mem_going_online_callback(void *arg)
2845 struct kmem_cache_node *n;
2846 struct kmem_cache *s;
2847 struct memory_notify *marg = arg;
2848 int nid = marg->status_change_nid;
2852 * If the node's memory is already available, then kmem_cache_node is
2853 * already created. Nothing to do.
2859 * We are bringing a node online. No memory is availabe yet. We must
2860 * allocate a kmem_cache_node structure in order to bring the node
2863 down_read(&slub_lock);
2864 list_for_each_entry(s, &slab_caches, list) {
2866 * XXX: kmem_cache_alloc_node will fallback to other nodes
2867 * since memory is not yet available from the node that
2870 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2875 init_kmem_cache_node(n);
2879 up_read(&slub_lock);
2883 static int slab_memory_callback(struct notifier_block *self,
2884 unsigned long action, void *arg)
2889 case MEM_GOING_ONLINE:
2890 ret = slab_mem_going_online_callback(arg);
2892 case MEM_GOING_OFFLINE:
2893 ret = slab_mem_going_offline_callback(arg);
2896 case MEM_CANCEL_ONLINE:
2897 slab_mem_offline_callback(arg);
2900 case MEM_CANCEL_OFFLINE:
2904 ret = notifier_from_errno(ret);
2908 #endif /* CONFIG_MEMORY_HOTPLUG */
2910 /********************************************************************
2911 * Basic setup of slabs
2912 *******************************************************************/
2914 void __init kmem_cache_init(void)
2923 * Must first have the slab cache available for the allocations of the
2924 * struct kmem_cache_node's. There is special bootstrap code in
2925 * kmem_cache_open for slab_state == DOWN.
2927 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2928 sizeof(struct kmem_cache_node), GFP_KERNEL);
2929 kmalloc_caches[0].refcount = -1;
2932 hotplug_memory_notifier(slab_memory_callback, 1);
2935 /* Able to allocate the per node structures */
2936 slab_state = PARTIAL;
2938 /* Caches that are not of the two-to-the-power-of size */
2939 if (KMALLOC_MIN_SIZE <= 64) {
2940 create_kmalloc_cache(&kmalloc_caches[1],
2941 "kmalloc-96", 96, GFP_KERNEL);
2944 if (KMALLOC_MIN_SIZE <= 128) {
2945 create_kmalloc_cache(&kmalloc_caches[2],
2946 "kmalloc-192", 192, GFP_KERNEL);
2950 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
2951 create_kmalloc_cache(&kmalloc_caches[i],
2952 "kmalloc", 1 << i, GFP_KERNEL);
2958 * Patch up the size_index table if we have strange large alignment
2959 * requirements for the kmalloc array. This is only the case for
2960 * MIPS it seems. The standard arches will not generate any code here.
2962 * Largest permitted alignment is 256 bytes due to the way we
2963 * handle the index determination for the smaller caches.
2965 * Make sure that nothing crazy happens if someone starts tinkering
2966 * around with ARCH_KMALLOC_MINALIGN
2968 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2969 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2971 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2972 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2976 /* Provide the correct kmalloc names now that the caches are up */
2977 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
2978 kmalloc_caches[i]. name =
2979 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2982 register_cpu_notifier(&slab_notifier);
2983 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2984 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2986 kmem_size = sizeof(struct kmem_cache);
2990 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2991 " CPUs=%d, Nodes=%d\n",
2992 caches, cache_line_size(),
2993 slub_min_order, slub_max_order, slub_min_objects,
2994 nr_cpu_ids, nr_node_ids);
2998 * Find a mergeable slab cache
3000 static int slab_unmergeable(struct kmem_cache *s)
3002 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3005 if ((s->flags & __PAGE_ALLOC_FALLBACK))
3012 * We may have set a slab to be unmergeable during bootstrap.
3014 if (s->refcount < 0)
3020 static struct kmem_cache *find_mergeable(size_t size,
3021 size_t align, unsigned long flags, const char *name,
3022 void (*ctor)(struct kmem_cache *, void *))
3024 struct kmem_cache *s;
3026 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3032 size = ALIGN(size, sizeof(void *));
3033 align = calculate_alignment(flags, align, size);
3034 size = ALIGN(size, align);
3035 flags = kmem_cache_flags(size, flags, name, NULL);
3037 list_for_each_entry(s, &slab_caches, list) {
3038 if (slab_unmergeable(s))
3044 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3047 * Check if alignment is compatible.
3048 * Courtesy of Adrian Drzewiecki
3050 if ((s->size & ~(align - 1)) != s->size)
3053 if (s->size - size >= sizeof(void *))
3061 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3062 size_t align, unsigned long flags,
3063 void (*ctor)(struct kmem_cache *, void *))
3065 struct kmem_cache *s;
3067 down_write(&slub_lock);
3068 s = find_mergeable(size, align, flags, name, ctor);
3074 * Adjust the object sizes so that we clear
3075 * the complete object on kzalloc.
3077 s->objsize = max(s->objsize, (int)size);
3080 * And then we need to update the object size in the
3081 * per cpu structures
3083 for_each_online_cpu(cpu)
3084 get_cpu_slab(s, cpu)->objsize = s->objsize;
3086 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3087 up_write(&slub_lock);
3089 if (sysfs_slab_alias(s, name))
3094 s = kmalloc(kmem_size, GFP_KERNEL);
3096 if (kmem_cache_open(s, GFP_KERNEL, name,
3097 size, align, flags, ctor)) {
3098 list_add(&s->list, &slab_caches);
3099 up_write(&slub_lock);
3100 if (sysfs_slab_add(s))
3106 up_write(&slub_lock);
3109 if (flags & SLAB_PANIC)
3110 panic("Cannot create slabcache %s\n", name);
3115 EXPORT_SYMBOL(kmem_cache_create);
3119 * Use the cpu notifier to insure that the cpu slabs are flushed when
3122 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3123 unsigned long action, void *hcpu)
3125 long cpu = (long)hcpu;
3126 struct kmem_cache *s;
3127 unsigned long flags;
3130 case CPU_UP_PREPARE:
3131 case CPU_UP_PREPARE_FROZEN:
3132 init_alloc_cpu_cpu(cpu);
3133 down_read(&slub_lock);
3134 list_for_each_entry(s, &slab_caches, list)
3135 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3137 up_read(&slub_lock);
3140 case CPU_UP_CANCELED:
3141 case CPU_UP_CANCELED_FROZEN:
3143 case CPU_DEAD_FROZEN:
3144 down_read(&slub_lock);
3145 list_for_each_entry(s, &slab_caches, list) {
3146 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3148 local_irq_save(flags);
3149 __flush_cpu_slab(s, cpu);
3150 local_irq_restore(flags);
3151 free_kmem_cache_cpu(c, cpu);
3152 s->cpu_slab[cpu] = NULL;
3154 up_read(&slub_lock);
3162 static struct notifier_block __cpuinitdata slab_notifier = {
3163 .notifier_call = slab_cpuup_callback
3168 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3170 struct kmem_cache *s;
3172 if (unlikely(size > PAGE_SIZE))
3173 return kmalloc_large(size, gfpflags);
3175 s = get_slab(size, gfpflags);
3177 if (unlikely(ZERO_OR_NULL_PTR(s)))
3180 return slab_alloc(s, gfpflags, -1, caller);
3183 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3184 int node, void *caller)
3186 struct kmem_cache *s;
3188 if (unlikely(size > PAGE_SIZE))
3189 return kmalloc_large_node(size, gfpflags, node);
3191 s = get_slab(size, gfpflags);
3193 if (unlikely(ZERO_OR_NULL_PTR(s)))
3196 return slab_alloc(s, gfpflags, node, caller);
3199 #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3200 static unsigned long count_partial(struct kmem_cache_node *n)
3202 unsigned long flags;
3203 unsigned long x = 0;
3206 spin_lock_irqsave(&n->list_lock, flags);
3207 list_for_each_entry(page, &n->partial, lru)
3209 spin_unlock_irqrestore(&n->list_lock, flags);
3214 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3215 static int validate_slab(struct kmem_cache *s, struct page *page,
3219 void *addr = page_address(page);
3221 if (!check_slab(s, page) ||
3222 !on_freelist(s, page, NULL))
3225 /* Now we know that a valid freelist exists */
3226 bitmap_zero(map, s->objects);
3228 for_each_free_object(p, s, page->freelist) {
3229 set_bit(slab_index(p, s, addr), map);
3230 if (!check_object(s, page, p, 0))
3234 for_each_object(p, s, addr)
3235 if (!test_bit(slab_index(p, s, addr), map))
3236 if (!check_object(s, page, p, 1))
3241 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3244 if (slab_trylock(page)) {
3245 validate_slab(s, page, map);
3248 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3251 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3252 if (!SlabDebug(page))
3253 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3254 "on slab 0x%p\n", s->name, page);
3256 if (SlabDebug(page))
3257 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3258 "slab 0x%p\n", s->name, page);
3262 static int validate_slab_node(struct kmem_cache *s,
3263 struct kmem_cache_node *n, unsigned long *map)
3265 unsigned long count = 0;
3267 unsigned long flags;
3269 spin_lock_irqsave(&n->list_lock, flags);
3271 list_for_each_entry(page, &n->partial, lru) {
3272 validate_slab_slab(s, page, map);
3275 if (count != n->nr_partial)
3276 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3277 "counter=%ld\n", s->name, count, n->nr_partial);
3279 if (!(s->flags & SLAB_STORE_USER))
3282 list_for_each_entry(page, &n->full, lru) {
3283 validate_slab_slab(s, page, map);
3286 if (count != atomic_long_read(&n->nr_slabs))
3287 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3288 "counter=%ld\n", s->name, count,
3289 atomic_long_read(&n->nr_slabs));
3292 spin_unlock_irqrestore(&n->list_lock, flags);
3296 static long validate_slab_cache(struct kmem_cache *s)
3299 unsigned long count = 0;
3300 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3301 sizeof(unsigned long), GFP_KERNEL);
3307 for_each_node_state(node, N_NORMAL_MEMORY) {
3308 struct kmem_cache_node *n = get_node(s, node);
3310 count += validate_slab_node(s, n, map);
3316 #ifdef SLUB_RESILIENCY_TEST
3317 static void resiliency_test(void)
3321 printk(KERN_ERR "SLUB resiliency testing\n");
3322 printk(KERN_ERR "-----------------------\n");
3323 printk(KERN_ERR "A. Corruption after allocation\n");
3325 p = kzalloc(16, GFP_KERNEL);
3327 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3328 " 0x12->0x%p\n\n", p + 16);
3330 validate_slab_cache(kmalloc_caches + 4);
3332 /* Hmmm... The next two are dangerous */
3333 p = kzalloc(32, GFP_KERNEL);
3334 p[32 + sizeof(void *)] = 0x34;
3335 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3336 " 0x34 -> -0x%p\n", p);
3338 "If allocated object is overwritten then not detectable\n\n");
3340 validate_slab_cache(kmalloc_caches + 5);
3341 p = kzalloc(64, GFP_KERNEL);
3342 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3344 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3347 "If allocated object is overwritten then not detectable\n\n");
3348 validate_slab_cache(kmalloc_caches + 6);
3350 printk(KERN_ERR "\nB. Corruption after free\n");
3351 p = kzalloc(128, GFP_KERNEL);
3354 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3355 validate_slab_cache(kmalloc_caches + 7);
3357 p = kzalloc(256, GFP_KERNEL);
3360 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3362 validate_slab_cache(kmalloc_caches + 8);
3364 p = kzalloc(512, GFP_KERNEL);
3367 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3368 validate_slab_cache(kmalloc_caches + 9);
3371 static void resiliency_test(void) {};
3375 * Generate lists of code addresses where slabcache objects are allocated
3380 unsigned long count;
3393 unsigned long count;
3394 struct location *loc;
3397 static void free_loc_track(struct loc_track *t)
3400 free_pages((unsigned long)t->loc,
3401 get_order(sizeof(struct location) * t->max));
3404 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3409 order = get_order(sizeof(struct location) * max);
3411 l = (void *)__get_free_pages(flags, order);
3416 memcpy(l, t->loc, sizeof(struct location) * t->count);
3424 static int add_location(struct loc_track *t, struct kmem_cache *s,
3425 const struct track *track)
3427 long start, end, pos;
3430 unsigned long age = jiffies - track->when;
3436 pos = start + (end - start + 1) / 2;
3439 * There is nothing at "end". If we end up there
3440 * we need to add something to before end.
3445 caddr = t->loc[pos].addr;
3446 if (track->addr == caddr) {
3452 if (age < l->min_time)
3454 if (age > l->max_time)
3457 if (track->pid < l->min_pid)
3458 l->min_pid = track->pid;
3459 if (track->pid > l->max_pid)
3460 l->max_pid = track->pid;
3462 cpu_set(track->cpu, l->cpus);
3464 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3468 if (track->addr < caddr)
3475 * Not found. Insert new tracking element.
3477 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3483 (t->count - pos) * sizeof(struct location));
3486 l->addr = track->addr;
3490 l->min_pid = track->pid;
3491 l->max_pid = track->pid;
3492 cpus_clear(l->cpus);
3493 cpu_set(track->cpu, l->cpus);
3494 nodes_clear(l->nodes);
3495 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3499 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3500 struct page *page, enum track_item alloc)
3502 void *addr = page_address(page);
3503 DECLARE_BITMAP(map, s->objects);
3506 bitmap_zero(map, s->objects);
3507 for_each_free_object(p, s, page->freelist)
3508 set_bit(slab_index(p, s, addr), map);
3510 for_each_object(p, s, addr)
3511 if (!test_bit(slab_index(p, s, addr), map))
3512 add_location(t, s, get_track(s, p, alloc));
3515 static int list_locations(struct kmem_cache *s, char *buf,
3516 enum track_item alloc)
3520 struct loc_track t = { 0, 0, NULL };
3523 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3525 return sprintf(buf, "Out of memory\n");
3527 /* Push back cpu slabs */
3530 for_each_node_state(node, N_NORMAL_MEMORY) {
3531 struct kmem_cache_node *n = get_node(s, node);
3532 unsigned long flags;
3535 if (!atomic_long_read(&n->nr_slabs))
3538 spin_lock_irqsave(&n->list_lock, flags);
3539 list_for_each_entry(page, &n->partial, lru)
3540 process_slab(&t, s, page, alloc);
3541 list_for_each_entry(page, &n->full, lru)
3542 process_slab(&t, s, page, alloc);
3543 spin_unlock_irqrestore(&n->list_lock, flags);
3546 for (i = 0; i < t.count; i++) {
3547 struct location *l = &t.loc[i];
3549 if (len > PAGE_SIZE - 100)
3551 len += sprintf(buf + len, "%7ld ", l->count);
3554 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3556 len += sprintf(buf + len, "<not-available>");
3558 if (l->sum_time != l->min_time) {
3559 unsigned long remainder;
3561 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3563 div_long_long_rem(l->sum_time, l->count, &remainder),
3566 len += sprintf(buf + len, " age=%ld",
3569 if (l->min_pid != l->max_pid)
3570 len += sprintf(buf + len, " pid=%ld-%ld",
3571 l->min_pid, l->max_pid);
3573 len += sprintf(buf + len, " pid=%ld",
3576 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3577 len < PAGE_SIZE - 60) {
3578 len += sprintf(buf + len, " cpus=");
3579 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3583 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3584 len < PAGE_SIZE - 60) {
3585 len += sprintf(buf + len, " nodes=");
3586 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3590 len += sprintf(buf + len, "\n");
3595 len += sprintf(buf, "No data\n");
3599 enum slab_stat_type {
3606 #define SO_FULL (1 << SL_FULL)
3607 #define SO_PARTIAL (1 << SL_PARTIAL)
3608 #define SO_CPU (1 << SL_CPU)
3609 #define SO_OBJECTS (1 << SL_OBJECTS)
3611 static ssize_t show_slab_objects(struct kmem_cache *s,
3612 char *buf, unsigned long flags)
3614 unsigned long total = 0;
3618 unsigned long *nodes;
3619 unsigned long *per_cpu;
3621 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3624 per_cpu = nodes + nr_node_ids;
3626 for_each_possible_cpu(cpu) {
3628 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3638 if (flags & SO_CPU) {
3639 if (flags & SO_OBJECTS)
3650 for_each_node_state(node, N_NORMAL_MEMORY) {
3651 struct kmem_cache_node *n = get_node(s, node);
3653 if (flags & SO_PARTIAL) {
3654 if (flags & SO_OBJECTS)
3655 x = count_partial(n);
3662 if (flags & SO_FULL) {
3663 int full_slabs = atomic_long_read(&n->nr_slabs)
3667 if (flags & SO_OBJECTS)
3668 x = full_slabs * s->objects;
3676 x = sprintf(buf, "%lu", total);
3678 for_each_node_state(node, N_NORMAL_MEMORY)
3680 x += sprintf(buf + x, " N%d=%lu",
3684 return x + sprintf(buf + x, "\n");
3687 static int any_slab_objects(struct kmem_cache *s)
3692 for_each_possible_cpu(cpu) {
3693 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3699 for_each_online_node(node) {
3700 struct kmem_cache_node *n = get_node(s, node);
3705 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3711 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3712 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3714 struct slab_attribute {
3715 struct attribute attr;
3716 ssize_t (*show)(struct kmem_cache *s, char *buf);
3717 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3720 #define SLAB_ATTR_RO(_name) \
3721 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3723 #define SLAB_ATTR(_name) \
3724 static struct slab_attribute _name##_attr = \
3725 __ATTR(_name, 0644, _name##_show, _name##_store)
3727 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3729 return sprintf(buf, "%d\n", s->size);
3731 SLAB_ATTR_RO(slab_size);
3733 static ssize_t align_show(struct kmem_cache *s, char *buf)
3735 return sprintf(buf, "%d\n", s->align);
3737 SLAB_ATTR_RO(align);
3739 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3741 return sprintf(buf, "%d\n", s->objsize);
3743 SLAB_ATTR_RO(object_size);
3745 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3747 return sprintf(buf, "%d\n", s->objects);
3749 SLAB_ATTR_RO(objs_per_slab);
3751 static ssize_t order_show(struct kmem_cache *s, char *buf)
3753 return sprintf(buf, "%d\n", s->order);
3755 SLAB_ATTR_RO(order);
3757 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3760 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3762 return n + sprintf(buf + n, "\n");
3768 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3770 return sprintf(buf, "%d\n", s->refcount - 1);
3772 SLAB_ATTR_RO(aliases);
3774 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3776 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3778 SLAB_ATTR_RO(slabs);
3780 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3782 return show_slab_objects(s, buf, SO_PARTIAL);
3784 SLAB_ATTR_RO(partial);
3786 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3788 return show_slab_objects(s, buf, SO_CPU);
3790 SLAB_ATTR_RO(cpu_slabs);
3792 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3794 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3796 SLAB_ATTR_RO(objects);
3798 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3800 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3803 static ssize_t sanity_checks_store(struct kmem_cache *s,
3804 const char *buf, size_t length)
3806 s->flags &= ~SLAB_DEBUG_FREE;
3808 s->flags |= SLAB_DEBUG_FREE;
3811 SLAB_ATTR(sanity_checks);
3813 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3815 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3818 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3821 s->flags &= ~SLAB_TRACE;
3823 s->flags |= SLAB_TRACE;
3828 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3830 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3833 static ssize_t reclaim_account_store(struct kmem_cache *s,
3834 const char *buf, size_t length)
3836 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3838 s->flags |= SLAB_RECLAIM_ACCOUNT;
3841 SLAB_ATTR(reclaim_account);
3843 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3845 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3847 SLAB_ATTR_RO(hwcache_align);
3849 #ifdef CONFIG_ZONE_DMA
3850 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3852 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3854 SLAB_ATTR_RO(cache_dma);
3857 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3859 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3861 SLAB_ATTR_RO(destroy_by_rcu);
3863 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3865 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3868 static ssize_t red_zone_store(struct kmem_cache *s,
3869 const char *buf, size_t length)
3871 if (any_slab_objects(s))
3874 s->flags &= ~SLAB_RED_ZONE;
3876 s->flags |= SLAB_RED_ZONE;
3880 SLAB_ATTR(red_zone);
3882 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3884 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3887 static ssize_t poison_store(struct kmem_cache *s,
3888 const char *buf, size_t length)
3890 if (any_slab_objects(s))
3893 s->flags &= ~SLAB_POISON;
3895 s->flags |= SLAB_POISON;
3901 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3903 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3906 static ssize_t store_user_store(struct kmem_cache *s,
3907 const char *buf, size_t length)
3909 if (any_slab_objects(s))
3912 s->flags &= ~SLAB_STORE_USER;
3914 s->flags |= SLAB_STORE_USER;
3918 SLAB_ATTR(store_user);
3920 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3925 static ssize_t validate_store(struct kmem_cache *s,
3926 const char *buf, size_t length)
3930 if (buf[0] == '1') {
3931 ret = validate_slab_cache(s);
3937 SLAB_ATTR(validate);
3939 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3944 static ssize_t shrink_store(struct kmem_cache *s,
3945 const char *buf, size_t length)
3947 if (buf[0] == '1') {
3948 int rc = kmem_cache_shrink(s);
3958 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3960 if (!(s->flags & SLAB_STORE_USER))
3962 return list_locations(s, buf, TRACK_ALLOC);
3964 SLAB_ATTR_RO(alloc_calls);
3966 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3968 if (!(s->flags & SLAB_STORE_USER))
3970 return list_locations(s, buf, TRACK_FREE);
3972 SLAB_ATTR_RO(free_calls);
3975 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
3977 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
3980 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
3981 const char *buf, size_t length)
3983 int n = simple_strtoul(buf, NULL, 10);
3986 s->remote_node_defrag_ratio = n * 10;
3989 SLAB_ATTR(remote_node_defrag_ratio);
3992 #ifdef CONFIG_SLUB_STATS
3993 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3995 unsigned long sum = 0;
3998 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4003 for_each_online_cpu(cpu) {
4004 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4010 len = sprintf(buf, "%lu", sum);
4013 for_each_online_cpu(cpu) {
4014 if (data[cpu] && len < PAGE_SIZE - 20)
4015 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4019 return len + sprintf(buf + len, "\n");
4022 #define STAT_ATTR(si, text) \
4023 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4025 return show_stat(s, buf, si); \
4027 SLAB_ATTR_RO(text); \
4029 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4030 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4031 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4032 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4033 STAT_ATTR(FREE_FROZEN, free_frozen);
4034 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4035 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4036 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4037 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4038 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4039 STAT_ATTR(FREE_SLAB, free_slab);
4040 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4041 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4042 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4043 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4044 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4045 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4049 static struct attribute *slab_attrs[] = {
4050 &slab_size_attr.attr,
4051 &object_size_attr.attr,
4052 &objs_per_slab_attr.attr,
4057 &cpu_slabs_attr.attr,
4061 &sanity_checks_attr.attr,
4063 &hwcache_align_attr.attr,
4064 &reclaim_account_attr.attr,
4065 &destroy_by_rcu_attr.attr,
4066 &red_zone_attr.attr,
4068 &store_user_attr.attr,
4069 &validate_attr.attr,
4071 &alloc_calls_attr.attr,
4072 &free_calls_attr.attr,
4073 #ifdef CONFIG_ZONE_DMA
4074 &cache_dma_attr.attr,
4077 &remote_node_defrag_ratio_attr.attr,
4079 #ifdef CONFIG_SLUB_STATS
4080 &alloc_fastpath_attr.attr,
4081 &alloc_slowpath_attr.attr,
4082 &free_fastpath_attr.attr,
4083 &free_slowpath_attr.attr,
4084 &free_frozen_attr.attr,
4085 &free_add_partial_attr.attr,
4086 &free_remove_partial_attr.attr,
4087 &alloc_from_partial_attr.attr,
4088 &alloc_slab_attr.attr,
4089 &alloc_refill_attr.attr,
4090 &free_slab_attr.attr,
4091 &cpuslab_flush_attr.attr,
4092 &deactivate_full_attr.attr,
4093 &deactivate_empty_attr.attr,
4094 &deactivate_to_head_attr.attr,
4095 &deactivate_to_tail_attr.attr,
4096 &deactivate_remote_frees_attr.attr,
4101 static struct attribute_group slab_attr_group = {
4102 .attrs = slab_attrs,
4105 static ssize_t slab_attr_show(struct kobject *kobj,
4106 struct attribute *attr,
4109 struct slab_attribute *attribute;
4110 struct kmem_cache *s;
4113 attribute = to_slab_attr(attr);
4116 if (!attribute->show)
4119 err = attribute->show(s, buf);
4124 static ssize_t slab_attr_store(struct kobject *kobj,
4125 struct attribute *attr,
4126 const char *buf, size_t len)
4128 struct slab_attribute *attribute;
4129 struct kmem_cache *s;
4132 attribute = to_slab_attr(attr);
4135 if (!attribute->store)
4138 err = attribute->store(s, buf, len);
4143 static void kmem_cache_release(struct kobject *kobj)
4145 struct kmem_cache *s = to_slab(kobj);
4150 static struct sysfs_ops slab_sysfs_ops = {
4151 .show = slab_attr_show,
4152 .store = slab_attr_store,
4155 static struct kobj_type slab_ktype = {
4156 .sysfs_ops = &slab_sysfs_ops,
4157 .release = kmem_cache_release
4160 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4162 struct kobj_type *ktype = get_ktype(kobj);
4164 if (ktype == &slab_ktype)
4169 static struct kset_uevent_ops slab_uevent_ops = {
4170 .filter = uevent_filter,
4173 static struct kset *slab_kset;
4175 #define ID_STR_LENGTH 64
4177 /* Create a unique string id for a slab cache:
4179 * Format :[flags-]size
4181 static char *create_unique_id(struct kmem_cache *s)
4183 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4190 * First flags affecting slabcache operations. We will only
4191 * get here for aliasable slabs so we do not need to support
4192 * too many flags. The flags here must cover all flags that
4193 * are matched during merging to guarantee that the id is
4196 if (s->flags & SLAB_CACHE_DMA)
4198 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4200 if (s->flags & SLAB_DEBUG_FREE)
4204 p += sprintf(p, "%07d", s->size);
4205 BUG_ON(p > name + ID_STR_LENGTH - 1);
4209 static int sysfs_slab_add(struct kmem_cache *s)
4215 if (slab_state < SYSFS)
4216 /* Defer until later */
4219 unmergeable = slab_unmergeable(s);
4222 * Slabcache can never be merged so we can use the name proper.
4223 * This is typically the case for debug situations. In that
4224 * case we can catch duplicate names easily.
4226 sysfs_remove_link(&slab_kset->kobj, s->name);
4230 * Create a unique name for the slab as a target
4233 name = create_unique_id(s);
4236 s->kobj.kset = slab_kset;
4237 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4239 kobject_put(&s->kobj);
4243 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4246 kobject_uevent(&s->kobj, KOBJ_ADD);
4248 /* Setup first alias */
4249 sysfs_slab_alias(s, s->name);
4255 static void sysfs_slab_remove(struct kmem_cache *s)
4257 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4258 kobject_del(&s->kobj);
4259 kobject_put(&s->kobj);
4263 * Need to buffer aliases during bootup until sysfs becomes
4264 * available lest we loose that information.
4266 struct saved_alias {
4267 struct kmem_cache *s;
4269 struct saved_alias *next;
4272 static struct saved_alias *alias_list;
4274 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4276 struct saved_alias *al;
4278 if (slab_state == SYSFS) {
4280 * If we have a leftover link then remove it.
4282 sysfs_remove_link(&slab_kset->kobj, name);
4283 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4286 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4292 al->next = alias_list;
4297 static int __init slab_sysfs_init(void)
4299 struct kmem_cache *s;
4302 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4304 printk(KERN_ERR "Cannot register slab subsystem.\n");
4310 list_for_each_entry(s, &slab_caches, list) {
4311 err = sysfs_slab_add(s);
4313 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4314 " to sysfs\n", s->name);
4317 while (alias_list) {
4318 struct saved_alias *al = alias_list;
4320 alias_list = alias_list->next;
4321 err = sysfs_slab_alias(al->s, al->name);
4323 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4324 " %s to sysfs\n", s->name);
4332 __initcall(slab_sysfs_init);
4336 * The /proc/slabinfo ABI
4338 #ifdef CONFIG_SLABINFO
4340 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4341 size_t count, loff_t *ppos)
4347 static void print_slabinfo_header(struct seq_file *m)
4349 seq_puts(m, "slabinfo - version: 2.1\n");
4350 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4351 "<objperslab> <pagesperslab>");
4352 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4353 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4357 static void *s_start(struct seq_file *m, loff_t *pos)
4361 down_read(&slub_lock);
4363 print_slabinfo_header(m);
4365 return seq_list_start(&slab_caches, *pos);
4368 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4370 return seq_list_next(p, &slab_caches, pos);
4373 static void s_stop(struct seq_file *m, void *p)
4375 up_read(&slub_lock);
4378 static int s_show(struct seq_file *m, void *p)
4380 unsigned long nr_partials = 0;
4381 unsigned long nr_slabs = 0;
4382 unsigned long nr_inuse = 0;
4383 unsigned long nr_objs;
4384 struct kmem_cache *s;
4387 s = list_entry(p, struct kmem_cache, list);
4389 for_each_online_node(node) {
4390 struct kmem_cache_node *n = get_node(s, node);
4395 nr_partials += n->nr_partial;
4396 nr_slabs += atomic_long_read(&n->nr_slabs);
4397 nr_inuse += count_partial(n);
4400 nr_objs = nr_slabs * s->objects;
4401 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4403 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4404 nr_objs, s->size, s->objects, (1 << s->order));
4405 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4406 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4412 const struct seq_operations slabinfo_op = {
4419 #endif /* CONFIG_SLABINFO */