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
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/proc_fs.h>
18 #include <linux/seq_file.h>
19 #include <linux/cpu.h>
20 #include <linux/cpuset.h>
21 #include <linux/mempolicy.h>
22 #include <linux/ctype.h>
23 #include <linux/debugobjects.h>
24 #include <linux/kallsyms.h>
25 #include <linux/memory.h>
26 #include <linux/math64.h>
33 * The slab_lock protects operations on the object of a particular
34 * slab and its metadata in the page struct. If the slab lock
35 * has been taken then no allocations nor frees can be performed
36 * on the objects in the slab nor can the slab be added or removed
37 * from the partial or full lists since this would mean modifying
38 * the page_struct of the slab.
40 * The list_lock protects the partial and full list on each node and
41 * the partial slab counter. If taken then no new slabs may be added or
42 * removed from the lists nor make the number of partial slabs be modified.
43 * (Note that the total number of slabs is an atomic value that may be
44 * modified without taking the list lock).
46 * The list_lock is a centralized lock and thus we avoid taking it as
47 * much as possible. As long as SLUB does not have to handle partial
48 * slabs, operations can continue without any centralized lock. F.e.
49 * allocating a long series of objects that fill up slabs does not require
52 * The lock order is sometimes inverted when we are trying to get a slab
53 * off a list. We take the list_lock and then look for a page on the list
54 * to use. While we do that objects in the slabs may be freed. We can
55 * only operate on the slab if we have also taken the slab_lock. So we use
56 * a slab_trylock() on the slab. If trylock was successful then no frees
57 * can occur anymore and we can use the slab for allocations etc. If the
58 * slab_trylock() does not succeed then frees are in progress in the slab and
59 * we must stay away from it for a while since we may cause a bouncing
60 * cacheline if we try to acquire the lock. So go onto the next slab.
61 * If all pages are busy then we may allocate a new slab instead of reusing
62 * a partial slab. A new slab has noone operating on it and thus there is
63 * no danger of cacheline contention.
65 * Interrupts are disabled during allocation and deallocation in order to
66 * make the slab allocator safe to use in the context of an irq. In addition
67 * interrupts are disabled to ensure that the processor does not change
68 * while handling per_cpu slabs, due to kernel preemption.
70 * SLUB assigns one slab for allocation to each processor.
71 * Allocations only occur from these slabs called cpu slabs.
73 * Slabs with free elements are kept on a partial list and during regular
74 * operations no list for full slabs is used. If an object in a full slab is
75 * freed then the slab will show up again on the partial lists.
76 * We track full slabs for debugging purposes though because otherwise we
77 * cannot scan all objects.
79 * Slabs are freed when they become empty. Teardown and setup is
80 * minimal so we rely on the page allocators per cpu caches for
81 * fast frees and allocs.
83 * Overloading of page flags that are otherwise used for LRU management.
85 * PageActive The slab is frozen and exempt from list processing.
86 * This means that the slab is dedicated to a purpose
87 * such as satisfying allocations for a specific
88 * processor. Objects may be freed in the slab while
89 * it is frozen but slab_free will then skip the usual
90 * list operations. It is up to the processor holding
91 * the slab to integrate the slab into the slab lists
92 * when the slab is no longer needed.
94 * One use of this flag is to mark slabs that are
95 * used for allocations. Then such a slab becomes a cpu
96 * slab. The cpu slab may be equipped with an additional
97 * freelist that allows lockless access to
98 * free objects in addition to the regular freelist
99 * that requires the slab lock.
101 * PageError Slab requires special handling due to debug
102 * options set. This moves slab handling out of
103 * the fast path and disables lockless freelists.
106 #ifdef CONFIG_SLUB_DEBUG
113 * Issues still to be resolved:
115 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
117 * - Variable sizing of the per node arrays
120 /* Enable to test recovery from slab corruption on boot */
121 #undef SLUB_RESILIENCY_TEST
124 * Mininum number of partial slabs. These will be left on the partial
125 * lists even if they are empty. kmem_cache_shrink may reclaim them.
127 #define MIN_PARTIAL 5
130 * Maximum number of desirable partial slabs.
131 * The existence of more partial slabs makes kmem_cache_shrink
132 * sort the partial list by the number of objects in the.
134 #define MAX_PARTIAL 10
136 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
137 SLAB_POISON | SLAB_STORE_USER)
140 * Set of flags that will prevent slab merging
142 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
143 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
145 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
148 #ifndef ARCH_KMALLOC_MINALIGN
149 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
152 #ifndef ARCH_SLAB_MINALIGN
153 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
156 /* Internal SLUB flags */
157 #define __OBJECT_POISON 0x80000000 /* Poison object */
158 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
160 static int kmem_size = sizeof(struct kmem_cache);
163 static struct notifier_block slab_notifier;
167 DOWN, /* No slab functionality available */
168 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
169 UP, /* Everything works but does not show up in sysfs */
173 /* A list of all slab caches on the system */
174 static DECLARE_RWSEM(slub_lock);
175 static LIST_HEAD(slab_caches);
178 * Tracking user of a slab.
181 void *addr; /* Called from address */
182 int cpu; /* Was running on cpu */
183 int pid; /* Pid context */
184 unsigned long when; /* When did the operation occur */
187 enum track_item { TRACK_ALLOC, TRACK_FREE };
189 #ifdef CONFIG_SLUB_DEBUG
190 static int sysfs_slab_add(struct kmem_cache *);
191 static int sysfs_slab_alias(struct kmem_cache *, const char *);
192 static void sysfs_slab_remove(struct kmem_cache *);
195 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
196 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
198 static inline void sysfs_slab_remove(struct kmem_cache *s)
205 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
207 #ifdef CONFIG_SLUB_STATS
212 /********************************************************************
213 * Core slab cache functions
214 *******************************************************************/
216 int slab_is_available(void)
218 return slab_state >= UP;
221 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
224 return s->node[node];
226 return &s->local_node;
230 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
233 return s->cpu_slab[cpu];
239 /* Verify that a pointer has an address that is valid within a slab page */
240 static inline int check_valid_pointer(struct kmem_cache *s,
241 struct page *page, const void *object)
248 base = page_address(page);
249 if (object < base || object >= base + page->objects * s->size ||
250 (object - base) % s->size) {
258 * Slow version of get and set free pointer.
260 * This version requires touching the cache lines of kmem_cache which
261 * we avoid to do in the fast alloc free paths. There we obtain the offset
262 * from the page struct.
264 static inline void *get_freepointer(struct kmem_cache *s, void *object)
266 return *(void **)(object + s->offset);
269 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
271 *(void **)(object + s->offset) = fp;
274 /* Loop over all objects in a slab */
275 #define for_each_object(__p, __s, __addr, __objects) \
276 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
280 #define for_each_free_object(__p, __s, __free) \
281 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
283 /* Determine object index from a given position */
284 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
286 return (p - addr) / s->size;
289 static inline struct kmem_cache_order_objects oo_make(int order,
292 struct kmem_cache_order_objects x = {
293 (order << 16) + (PAGE_SIZE << order) / size
299 static inline int oo_order(struct kmem_cache_order_objects x)
304 static inline int oo_objects(struct kmem_cache_order_objects x)
306 return x.x & ((1 << 16) - 1);
309 #ifdef CONFIG_SLUB_DEBUG
313 #ifdef CONFIG_SLUB_DEBUG_ON
314 static int slub_debug = DEBUG_DEFAULT_FLAGS;
316 static int slub_debug;
319 static char *slub_debug_slabs;
324 static void print_section(char *text, u8 *addr, unsigned int length)
332 for (i = 0; i < length; i++) {
334 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
337 printk(KERN_CONT " %02x", addr[i]);
339 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
341 printk(KERN_CONT " %s\n", ascii);
348 printk(KERN_CONT " ");
352 printk(KERN_CONT " %s\n", ascii);
356 static struct track *get_track(struct kmem_cache *s, void *object,
357 enum track_item alloc)
362 p = object + s->offset + sizeof(void *);
364 p = object + s->inuse;
369 static void set_track(struct kmem_cache *s, void *object,
370 enum track_item alloc, void *addr)
375 p = object + s->offset + sizeof(void *);
377 p = object + s->inuse;
382 p->cpu = smp_processor_id();
383 p->pid = current->pid;
386 memset(p, 0, sizeof(struct track));
389 static void init_tracking(struct kmem_cache *s, void *object)
391 if (!(s->flags & SLAB_STORE_USER))
394 set_track(s, object, TRACK_FREE, NULL);
395 set_track(s, object, TRACK_ALLOC, NULL);
398 static void print_track(const char *s, struct track *t)
403 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
404 s, t->addr, jiffies - t->when, t->cpu, t->pid);
407 static void print_tracking(struct kmem_cache *s, void *object)
409 if (!(s->flags & SLAB_STORE_USER))
412 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
413 print_track("Freed", get_track(s, object, TRACK_FREE));
416 static void print_page_info(struct page *page)
418 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
419 page, page->objects, page->inuse, page->freelist, page->flags);
423 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
429 vsnprintf(buf, sizeof(buf), fmt, args);
431 printk(KERN_ERR "========================================"
432 "=====================================\n");
433 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
434 printk(KERN_ERR "----------------------------------------"
435 "-------------------------------------\n\n");
438 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
444 vsnprintf(buf, sizeof(buf), fmt, args);
446 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
449 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
451 unsigned int off; /* Offset of last byte */
452 u8 *addr = page_address(page);
454 print_tracking(s, p);
456 print_page_info(page);
458 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
459 p, p - addr, get_freepointer(s, p));
462 print_section("Bytes b4", p - 16, 16);
464 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
466 if (s->flags & SLAB_RED_ZONE)
467 print_section("Redzone", p + s->objsize,
468 s->inuse - s->objsize);
471 off = s->offset + sizeof(void *);
475 if (s->flags & SLAB_STORE_USER)
476 off += 2 * sizeof(struct track);
479 /* Beginning of the filler is the free pointer */
480 print_section("Padding", p + off, s->size - off);
485 static void object_err(struct kmem_cache *s, struct page *page,
486 u8 *object, char *reason)
488 slab_bug(s, "%s", reason);
489 print_trailer(s, page, object);
492 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
498 vsnprintf(buf, sizeof(buf), fmt, args);
500 slab_bug(s, "%s", buf);
501 print_page_info(page);
505 static void init_object(struct kmem_cache *s, void *object, int active)
509 if (s->flags & __OBJECT_POISON) {
510 memset(p, POISON_FREE, s->objsize - 1);
511 p[s->objsize - 1] = POISON_END;
514 if (s->flags & SLAB_RED_ZONE)
515 memset(p + s->objsize,
516 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
517 s->inuse - s->objsize);
520 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
523 if (*start != (u8)value)
531 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
532 void *from, void *to)
534 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
535 memset(from, data, to - from);
538 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
539 u8 *object, char *what,
540 u8 *start, unsigned int value, unsigned int bytes)
545 fault = check_bytes(start, value, bytes);
550 while (end > fault && end[-1] == value)
553 slab_bug(s, "%s overwritten", what);
554 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
555 fault, end - 1, fault[0], value);
556 print_trailer(s, page, object);
558 restore_bytes(s, what, value, fault, end);
566 * Bytes of the object to be managed.
567 * If the freepointer may overlay the object then the free
568 * pointer is the first word of the object.
570 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
573 * object + s->objsize
574 * Padding to reach word boundary. This is also used for Redzoning.
575 * Padding is extended by another word if Redzoning is enabled and
578 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
579 * 0xcc (RED_ACTIVE) for objects in use.
582 * Meta data starts here.
584 * A. Free pointer (if we cannot overwrite object on free)
585 * B. Tracking data for SLAB_STORE_USER
586 * C. Padding to reach required alignment boundary or at mininum
587 * one word if debugging is on to be able to detect writes
588 * before the word boundary.
590 * Padding is done using 0x5a (POISON_INUSE)
593 * Nothing is used beyond s->size.
595 * If slabcaches are merged then the objsize and inuse boundaries are mostly
596 * ignored. And therefore no slab options that rely on these boundaries
597 * may be used with merged slabcaches.
600 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
602 unsigned long off = s->inuse; /* The end of info */
605 /* Freepointer is placed after the object. */
606 off += sizeof(void *);
608 if (s->flags & SLAB_STORE_USER)
609 /* We also have user information there */
610 off += 2 * sizeof(struct track);
615 return check_bytes_and_report(s, page, p, "Object padding",
616 p + off, POISON_INUSE, s->size - off);
619 /* Check the pad bytes at the end of a slab page */
620 static int slab_pad_check(struct kmem_cache *s, struct page *page)
628 if (!(s->flags & SLAB_POISON))
631 start = page_address(page);
632 length = (PAGE_SIZE << compound_order(page));
633 end = start + length;
634 remainder = length % s->size;
638 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
641 while (end > fault && end[-1] == POISON_INUSE)
644 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
645 print_section("Padding", end - remainder, remainder);
647 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
651 static int check_object(struct kmem_cache *s, struct page *page,
652 void *object, int active)
655 u8 *endobject = object + s->objsize;
657 if (s->flags & SLAB_RED_ZONE) {
659 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
661 if (!check_bytes_and_report(s, page, object, "Redzone",
662 endobject, red, s->inuse - s->objsize))
665 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
666 check_bytes_and_report(s, page, p, "Alignment padding",
667 endobject, POISON_INUSE, s->inuse - s->objsize);
671 if (s->flags & SLAB_POISON) {
672 if (!active && (s->flags & __OBJECT_POISON) &&
673 (!check_bytes_and_report(s, page, p, "Poison", p,
674 POISON_FREE, s->objsize - 1) ||
675 !check_bytes_and_report(s, page, p, "Poison",
676 p + s->objsize - 1, POISON_END, 1)))
679 * check_pad_bytes cleans up on its own.
681 check_pad_bytes(s, page, p);
684 if (!s->offset && active)
686 * Object and freepointer overlap. Cannot check
687 * freepointer while object is allocated.
691 /* Check free pointer validity */
692 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
693 object_err(s, page, p, "Freepointer corrupt");
695 * No choice but to zap it and thus loose the remainder
696 * of the free objects in this slab. May cause
697 * another error because the object count is now wrong.
699 set_freepointer(s, p, NULL);
705 static int check_slab(struct kmem_cache *s, struct page *page)
709 VM_BUG_ON(!irqs_disabled());
711 if (!PageSlab(page)) {
712 slab_err(s, page, "Not a valid slab page");
716 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
717 if (page->objects > maxobj) {
718 slab_err(s, page, "objects %u > max %u",
719 s->name, page->objects, maxobj);
722 if (page->inuse > page->objects) {
723 slab_err(s, page, "inuse %u > max %u",
724 s->name, page->inuse, page->objects);
727 /* Slab_pad_check fixes things up after itself */
728 slab_pad_check(s, page);
733 * Determine if a certain object on a page is on the freelist. Must hold the
734 * slab lock to guarantee that the chains are in a consistent state.
736 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
739 void *fp = page->freelist;
741 unsigned long max_objects;
743 while (fp && nr <= page->objects) {
746 if (!check_valid_pointer(s, page, fp)) {
748 object_err(s, page, object,
749 "Freechain corrupt");
750 set_freepointer(s, object, NULL);
753 slab_err(s, page, "Freepointer corrupt");
754 page->freelist = NULL;
755 page->inuse = page->objects;
756 slab_fix(s, "Freelist cleared");
762 fp = get_freepointer(s, object);
766 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
767 if (max_objects > 65535)
770 if (page->objects != max_objects) {
771 slab_err(s, page, "Wrong number of objects. Found %d but "
772 "should be %d", page->objects, max_objects);
773 page->objects = max_objects;
774 slab_fix(s, "Number of objects adjusted.");
776 if (page->inuse != page->objects - nr) {
777 slab_err(s, page, "Wrong object count. Counter is %d but "
778 "counted were %d", page->inuse, page->objects - nr);
779 page->inuse = page->objects - nr;
780 slab_fix(s, "Object count adjusted.");
782 return search == NULL;
785 static void trace(struct kmem_cache *s, struct page *page, void *object,
788 if (s->flags & SLAB_TRACE) {
789 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
791 alloc ? "alloc" : "free",
796 print_section("Object", (void *)object, s->objsize);
803 * Tracking of fully allocated slabs for debugging purposes.
805 static void add_full(struct kmem_cache_node *n, struct page *page)
807 spin_lock(&n->list_lock);
808 list_add(&page->lru, &n->full);
809 spin_unlock(&n->list_lock);
812 static void remove_full(struct kmem_cache *s, struct page *page)
814 struct kmem_cache_node *n;
816 if (!(s->flags & SLAB_STORE_USER))
819 n = get_node(s, page_to_nid(page));
821 spin_lock(&n->list_lock);
822 list_del(&page->lru);
823 spin_unlock(&n->list_lock);
826 /* Tracking of the number of slabs for debugging purposes */
827 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
829 struct kmem_cache_node *n = get_node(s, node);
831 return atomic_long_read(&n->nr_slabs);
834 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
836 struct kmem_cache_node *n = get_node(s, node);
839 * May be called early in order to allocate a slab for the
840 * kmem_cache_node structure. Solve the chicken-egg
841 * dilemma by deferring the increment of the count during
842 * bootstrap (see early_kmem_cache_node_alloc).
844 if (!NUMA_BUILD || n) {
845 atomic_long_inc(&n->nr_slabs);
846 atomic_long_add(objects, &n->total_objects);
849 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
851 struct kmem_cache_node *n = get_node(s, node);
853 atomic_long_dec(&n->nr_slabs);
854 atomic_long_sub(objects, &n->total_objects);
857 /* Object debug checks for alloc/free paths */
858 static void setup_object_debug(struct kmem_cache *s, struct page *page,
861 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
864 init_object(s, object, 0);
865 init_tracking(s, object);
868 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
869 void *object, void *addr)
871 if (!check_slab(s, page))
874 if (!on_freelist(s, page, object)) {
875 object_err(s, page, object, "Object already allocated");
879 if (!check_valid_pointer(s, page, object)) {
880 object_err(s, page, object, "Freelist Pointer check fails");
884 if (!check_object(s, page, object, 0))
887 /* Success perform special debug activities for allocs */
888 if (s->flags & SLAB_STORE_USER)
889 set_track(s, object, TRACK_ALLOC, addr);
890 trace(s, page, object, 1);
891 init_object(s, object, 1);
895 if (PageSlab(page)) {
897 * If this is a slab page then lets do the best we can
898 * to avoid issues in the future. Marking all objects
899 * as used avoids touching the remaining objects.
901 slab_fix(s, "Marking all objects used");
902 page->inuse = page->objects;
903 page->freelist = NULL;
908 static int free_debug_processing(struct kmem_cache *s, struct page *page,
909 void *object, void *addr)
911 if (!check_slab(s, page))
914 if (!check_valid_pointer(s, page, object)) {
915 slab_err(s, page, "Invalid object pointer 0x%p", object);
919 if (on_freelist(s, page, object)) {
920 object_err(s, page, object, "Object already free");
924 if (!check_object(s, page, object, 1))
927 if (unlikely(s != page->slab)) {
928 if (!PageSlab(page)) {
929 slab_err(s, page, "Attempt to free object(0x%p) "
930 "outside of slab", object);
931 } else if (!page->slab) {
933 "SLUB <none>: no slab for object 0x%p.\n",
937 object_err(s, page, object,
938 "page slab pointer corrupt.");
942 /* Special debug activities for freeing objects */
943 if (!PageSlubFrozen(page) && !page->freelist)
944 remove_full(s, page);
945 if (s->flags & SLAB_STORE_USER)
946 set_track(s, object, TRACK_FREE, addr);
947 trace(s, page, object, 0);
948 init_object(s, object, 0);
952 slab_fix(s, "Object at 0x%p not freed", object);
956 static int __init setup_slub_debug(char *str)
958 slub_debug = DEBUG_DEFAULT_FLAGS;
959 if (*str++ != '=' || !*str)
961 * No options specified. Switch on full debugging.
967 * No options but restriction on slabs. This means full
968 * debugging for slabs matching a pattern.
975 * Switch off all debugging measures.
980 * Determine which debug features should be switched on
982 for (; *str && *str != ','; str++) {
983 switch (tolower(*str)) {
985 slub_debug |= SLAB_DEBUG_FREE;
988 slub_debug |= SLAB_RED_ZONE;
991 slub_debug |= SLAB_POISON;
994 slub_debug |= SLAB_STORE_USER;
997 slub_debug |= SLAB_TRACE;
1000 printk(KERN_ERR "slub_debug option '%c' "
1001 "unknown. skipped\n", *str);
1007 slub_debug_slabs = str + 1;
1012 __setup("slub_debug", setup_slub_debug);
1014 static unsigned long kmem_cache_flags(unsigned long objsize,
1015 unsigned long flags, const char *name,
1016 void (*ctor)(void *))
1019 * Enable debugging if selected on the kernel commandline.
1021 if (slub_debug && (!slub_debug_slabs ||
1022 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1023 flags |= slub_debug;
1028 static inline void setup_object_debug(struct kmem_cache *s,
1029 struct page *page, void *object) {}
1031 static inline int alloc_debug_processing(struct kmem_cache *s,
1032 struct page *page, void *object, void *addr) { return 0; }
1034 static inline int free_debug_processing(struct kmem_cache *s,
1035 struct page *page, void *object, void *addr) { return 0; }
1037 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1039 static inline int check_object(struct kmem_cache *s, struct page *page,
1040 void *object, int active) { return 1; }
1041 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1042 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1043 unsigned long flags, const char *name,
1044 void (*ctor)(void *))
1048 #define slub_debug 0
1050 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1052 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1054 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1059 * Slab allocation and freeing
1061 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1062 struct kmem_cache_order_objects oo)
1064 int order = oo_order(oo);
1067 return alloc_pages(flags, order);
1069 return alloc_pages_node(node, flags, order);
1072 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1075 struct kmem_cache_order_objects oo = s->oo;
1077 flags |= s->allocflags;
1079 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1081 if (unlikely(!page)) {
1084 * Allocation may have failed due to fragmentation.
1085 * Try a lower order alloc if possible
1087 page = alloc_slab_page(flags, node, oo);
1091 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1093 page->objects = oo_objects(oo);
1094 mod_zone_page_state(page_zone(page),
1095 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1096 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1102 static void setup_object(struct kmem_cache *s, struct page *page,
1105 setup_object_debug(s, page, object);
1106 if (unlikely(s->ctor))
1110 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1117 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1119 page = allocate_slab(s,
1120 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1124 inc_slabs_node(s, page_to_nid(page), page->objects);
1126 page->flags |= 1 << PG_slab;
1127 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1128 SLAB_STORE_USER | SLAB_TRACE))
1129 __SetPageSlubDebug(page);
1131 start = page_address(page);
1133 if (unlikely(s->flags & SLAB_POISON))
1134 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1137 for_each_object(p, s, start, page->objects) {
1138 setup_object(s, page, last);
1139 set_freepointer(s, last, p);
1142 setup_object(s, page, last);
1143 set_freepointer(s, last, NULL);
1145 page->freelist = start;
1151 static void __free_slab(struct kmem_cache *s, struct page *page)
1153 int order = compound_order(page);
1154 int pages = 1 << order;
1156 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1159 slab_pad_check(s, page);
1160 for_each_object(p, s, page_address(page),
1162 check_object(s, page, p, 0);
1163 __ClearPageSlubDebug(page);
1166 mod_zone_page_state(page_zone(page),
1167 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1168 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1171 __ClearPageSlab(page);
1172 reset_page_mapcount(page);
1173 __free_pages(page, order);
1176 static void rcu_free_slab(struct rcu_head *h)
1180 page = container_of((struct list_head *)h, struct page, lru);
1181 __free_slab(page->slab, page);
1184 static void free_slab(struct kmem_cache *s, struct page *page)
1186 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1188 * RCU free overloads the RCU head over the LRU
1190 struct rcu_head *head = (void *)&page->lru;
1192 call_rcu(head, rcu_free_slab);
1194 __free_slab(s, page);
1197 static void discard_slab(struct kmem_cache *s, struct page *page)
1199 dec_slabs_node(s, page_to_nid(page), page->objects);
1204 * Per slab locking using the pagelock
1206 static __always_inline void slab_lock(struct page *page)
1208 bit_spin_lock(PG_locked, &page->flags);
1211 static __always_inline void slab_unlock(struct page *page)
1213 __bit_spin_unlock(PG_locked, &page->flags);
1216 static __always_inline int slab_trylock(struct page *page)
1220 rc = bit_spin_trylock(PG_locked, &page->flags);
1225 * Management of partially allocated slabs
1227 static void add_partial(struct kmem_cache_node *n,
1228 struct page *page, int tail)
1230 spin_lock(&n->list_lock);
1233 list_add_tail(&page->lru, &n->partial);
1235 list_add(&page->lru, &n->partial);
1236 spin_unlock(&n->list_lock);
1239 static void remove_partial(struct kmem_cache *s, struct page *page)
1241 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1243 spin_lock(&n->list_lock);
1244 list_del(&page->lru);
1246 spin_unlock(&n->list_lock);
1250 * Lock slab and remove from the partial list.
1252 * Must hold list_lock.
1254 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1257 if (slab_trylock(page)) {
1258 list_del(&page->lru);
1260 __SetPageSlubFrozen(page);
1267 * Try to allocate a partial slab from a specific node.
1269 static struct page *get_partial_node(struct kmem_cache_node *n)
1274 * Racy check. If we mistakenly see no partial slabs then we
1275 * just allocate an empty slab. If we mistakenly try to get a
1276 * partial slab and there is none available then get_partials()
1279 if (!n || !n->nr_partial)
1282 spin_lock(&n->list_lock);
1283 list_for_each_entry(page, &n->partial, lru)
1284 if (lock_and_freeze_slab(n, page))
1288 spin_unlock(&n->list_lock);
1293 * Get a page from somewhere. Search in increasing NUMA distances.
1295 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1298 struct zonelist *zonelist;
1301 enum zone_type high_zoneidx = gfp_zone(flags);
1305 * The defrag ratio allows a configuration of the tradeoffs between
1306 * inter node defragmentation and node local allocations. A lower
1307 * defrag_ratio increases the tendency to do local allocations
1308 * instead of attempting to obtain partial slabs from other nodes.
1310 * If the defrag_ratio is set to 0 then kmalloc() always
1311 * returns node local objects. If the ratio is higher then kmalloc()
1312 * may return off node objects because partial slabs are obtained
1313 * from other nodes and filled up.
1315 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1316 * defrag_ratio = 1000) then every (well almost) allocation will
1317 * first attempt to defrag slab caches on other nodes. This means
1318 * scanning over all nodes to look for partial slabs which may be
1319 * expensive if we do it every time we are trying to find a slab
1320 * with available objects.
1322 if (!s->remote_node_defrag_ratio ||
1323 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1326 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1327 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1328 struct kmem_cache_node *n;
1330 n = get_node(s, zone_to_nid(zone));
1332 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1333 n->nr_partial > n->min_partial) {
1334 page = get_partial_node(n);
1344 * Get a partial page, lock it and return it.
1346 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1349 int searchnode = (node == -1) ? numa_node_id() : node;
1351 page = get_partial_node(get_node(s, searchnode));
1352 if (page || (flags & __GFP_THISNODE))
1355 return get_any_partial(s, flags);
1359 * Move a page back to the lists.
1361 * Must be called with the slab lock held.
1363 * On exit the slab lock will have been dropped.
1365 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1367 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1368 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1370 __ClearPageSlubFrozen(page);
1373 if (page->freelist) {
1374 add_partial(n, page, tail);
1375 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1377 stat(c, DEACTIVATE_FULL);
1378 if (SLABDEBUG && PageSlubDebug(page) &&
1379 (s->flags & SLAB_STORE_USER))
1384 stat(c, DEACTIVATE_EMPTY);
1385 if (n->nr_partial < n->min_partial) {
1387 * Adding an empty slab to the partial slabs in order
1388 * to avoid page allocator overhead. This slab needs
1389 * to come after the other slabs with objects in
1390 * so that the others get filled first. That way the
1391 * size of the partial list stays small.
1393 * kmem_cache_shrink can reclaim any empty slabs from
1396 add_partial(n, page, 1);
1400 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1401 discard_slab(s, page);
1407 * Remove the cpu slab
1409 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1411 struct page *page = c->page;
1415 stat(c, DEACTIVATE_REMOTE_FREES);
1417 * Merge cpu freelist into slab freelist. Typically we get here
1418 * because both freelists are empty. So this is unlikely
1421 while (unlikely(c->freelist)) {
1424 tail = 0; /* Hot objects. Put the slab first */
1426 /* Retrieve object from cpu_freelist */
1427 object = c->freelist;
1428 c->freelist = c->freelist[c->offset];
1430 /* And put onto the regular freelist */
1431 object[c->offset] = page->freelist;
1432 page->freelist = object;
1436 unfreeze_slab(s, page, tail);
1439 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1441 stat(c, CPUSLAB_FLUSH);
1443 deactivate_slab(s, c);
1449 * Called from IPI handler with interrupts disabled.
1451 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1453 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1455 if (likely(c && c->page))
1459 static void flush_cpu_slab(void *d)
1461 struct kmem_cache *s = d;
1463 __flush_cpu_slab(s, smp_processor_id());
1466 static void flush_all(struct kmem_cache *s)
1468 on_each_cpu(flush_cpu_slab, s, 1);
1472 * Check if the objects in a per cpu structure fit numa
1473 * locality expectations.
1475 static inline int node_match(struct kmem_cache_cpu *c, int node)
1478 if (node != -1 && c->node != node)
1485 * Slow path. The lockless freelist is empty or we need to perform
1488 * Interrupts are disabled.
1490 * Processing is still very fast if new objects have been freed to the
1491 * regular freelist. In that case we simply take over the regular freelist
1492 * as the lockless freelist and zap the regular freelist.
1494 * If that is not working then we fall back to the partial lists. We take the
1495 * first element of the freelist as the object to allocate now and move the
1496 * rest of the freelist to the lockless freelist.
1498 * And if we were unable to get a new slab from the partial slab lists then
1499 * we need to allocate a new slab. This is the slowest path since it involves
1500 * a call to the page allocator and the setup of a new slab.
1502 static void *__slab_alloc(struct kmem_cache *s,
1503 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1508 /* We handle __GFP_ZERO in the caller */
1509 gfpflags &= ~__GFP_ZERO;
1515 if (unlikely(!node_match(c, node)))
1518 stat(c, ALLOC_REFILL);
1521 object = c->page->freelist;
1522 if (unlikely(!object))
1524 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1527 c->freelist = object[c->offset];
1528 c->page->inuse = c->page->objects;
1529 c->page->freelist = NULL;
1530 c->node = page_to_nid(c->page);
1532 slab_unlock(c->page);
1533 stat(c, ALLOC_SLOWPATH);
1537 deactivate_slab(s, c);
1540 new = get_partial(s, gfpflags, node);
1543 stat(c, ALLOC_FROM_PARTIAL);
1547 if (gfpflags & __GFP_WAIT)
1550 new = new_slab(s, gfpflags, node);
1552 if (gfpflags & __GFP_WAIT)
1553 local_irq_disable();
1556 c = get_cpu_slab(s, smp_processor_id());
1557 stat(c, ALLOC_SLAB);
1561 __SetPageSlubFrozen(new);
1567 if (!alloc_debug_processing(s, c->page, object, addr))
1571 c->page->freelist = object[c->offset];
1577 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1578 * have the fastpath folded into their functions. So no function call
1579 * overhead for requests that can be satisfied on the fastpath.
1581 * The fastpath works by first checking if the lockless freelist can be used.
1582 * If not then __slab_alloc is called for slow processing.
1584 * Otherwise we can simply pick the next object from the lockless free list.
1586 static __always_inline void *slab_alloc(struct kmem_cache *s,
1587 gfp_t gfpflags, int node, void *addr)
1590 struct kmem_cache_cpu *c;
1591 unsigned long flags;
1592 unsigned int objsize;
1594 might_sleep_if(gfpflags & __GFP_WAIT);
1595 local_irq_save(flags);
1596 c = get_cpu_slab(s, smp_processor_id());
1597 objsize = c->objsize;
1598 if (unlikely(!c->freelist || !node_match(c, node)))
1600 object = __slab_alloc(s, gfpflags, node, addr, c);
1603 object = c->freelist;
1604 c->freelist = object[c->offset];
1605 stat(c, ALLOC_FASTPATH);
1607 local_irq_restore(flags);
1609 if (unlikely((gfpflags & __GFP_ZERO) && object))
1610 memset(object, 0, objsize);
1615 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1617 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1619 EXPORT_SYMBOL(kmem_cache_alloc);
1622 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1624 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1626 EXPORT_SYMBOL(kmem_cache_alloc_node);
1630 * Slow patch handling. This may still be called frequently since objects
1631 * have a longer lifetime than the cpu slabs in most processing loads.
1633 * So we still attempt to reduce cache line usage. Just take the slab
1634 * lock and free the item. If there is no additional partial page
1635 * handling required then we can return immediately.
1637 static void __slab_free(struct kmem_cache *s, struct page *page,
1638 void *x, void *addr, unsigned int offset)
1641 void **object = (void *)x;
1642 struct kmem_cache_cpu *c;
1644 c = get_cpu_slab(s, raw_smp_processor_id());
1645 stat(c, FREE_SLOWPATH);
1648 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1652 prior = object[offset] = page->freelist;
1653 page->freelist = object;
1656 if (unlikely(PageSlubFrozen(page))) {
1657 stat(c, FREE_FROZEN);
1661 if (unlikely(!page->inuse))
1665 * Objects left in the slab. If it was not on the partial list before
1668 if (unlikely(!prior)) {
1669 add_partial(get_node(s, page_to_nid(page)), page, 1);
1670 stat(c, FREE_ADD_PARTIAL);
1680 * Slab still on the partial list.
1682 remove_partial(s, page);
1683 stat(c, FREE_REMOVE_PARTIAL);
1687 discard_slab(s, page);
1691 if (!free_debug_processing(s, page, x, addr))
1697 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1698 * can perform fastpath freeing without additional function calls.
1700 * The fastpath is only possible if we are freeing to the current cpu slab
1701 * of this processor. This typically the case if we have just allocated
1704 * If fastpath is not possible then fall back to __slab_free where we deal
1705 * with all sorts of special processing.
1707 static __always_inline void slab_free(struct kmem_cache *s,
1708 struct page *page, void *x, void *addr)
1710 void **object = (void *)x;
1711 struct kmem_cache_cpu *c;
1712 unsigned long flags;
1714 local_irq_save(flags);
1715 c = get_cpu_slab(s, smp_processor_id());
1716 debug_check_no_locks_freed(object, c->objsize);
1717 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1718 debug_check_no_obj_freed(object, s->objsize);
1719 if (likely(page == c->page && c->node >= 0)) {
1720 object[c->offset] = c->freelist;
1721 c->freelist = object;
1722 stat(c, FREE_FASTPATH);
1724 __slab_free(s, page, x, addr, c->offset);
1726 local_irq_restore(flags);
1729 void kmem_cache_free(struct kmem_cache *s, void *x)
1733 page = virt_to_head_page(x);
1735 slab_free(s, page, x, __builtin_return_address(0));
1737 EXPORT_SYMBOL(kmem_cache_free);
1739 /* Figure out on which slab object the object resides */
1740 static struct page *get_object_page(const void *x)
1742 struct page *page = virt_to_head_page(x);
1744 if (!PageSlab(page))
1751 * Object placement in a slab is made very easy because we always start at
1752 * offset 0. If we tune the size of the object to the alignment then we can
1753 * get the required alignment by putting one properly sized object after
1756 * Notice that the allocation order determines the sizes of the per cpu
1757 * caches. Each processor has always one slab available for allocations.
1758 * Increasing the allocation order reduces the number of times that slabs
1759 * must be moved on and off the partial lists and is therefore a factor in
1764 * Mininum / Maximum order of slab pages. This influences locking overhead
1765 * and slab fragmentation. A higher order reduces the number of partial slabs
1766 * and increases the number of allocations possible without having to
1767 * take the list_lock.
1769 static int slub_min_order;
1770 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1771 static int slub_min_objects;
1774 * Merge control. If this is set then no merging of slab caches will occur.
1775 * (Could be removed. This was introduced to pacify the merge skeptics.)
1777 static int slub_nomerge;
1780 * Calculate the order of allocation given an slab object size.
1782 * The order of allocation has significant impact on performance and other
1783 * system components. Generally order 0 allocations should be preferred since
1784 * order 0 does not cause fragmentation in the page allocator. Larger objects
1785 * be problematic to put into order 0 slabs because there may be too much
1786 * unused space left. We go to a higher order if more than 1/16th of the slab
1789 * In order to reach satisfactory performance we must ensure that a minimum
1790 * number of objects is in one slab. Otherwise we may generate too much
1791 * activity on the partial lists which requires taking the list_lock. This is
1792 * less a concern for large slabs though which are rarely used.
1794 * slub_max_order specifies the order where we begin to stop considering the
1795 * number of objects in a slab as critical. If we reach slub_max_order then
1796 * we try to keep the page order as low as possible. So we accept more waste
1797 * of space in favor of a small page order.
1799 * Higher order allocations also allow the placement of more objects in a
1800 * slab and thereby reduce object handling overhead. If the user has
1801 * requested a higher mininum order then we start with that one instead of
1802 * the smallest order which will fit the object.
1804 static inline int slab_order(int size, int min_objects,
1805 int max_order, int fract_leftover)
1809 int min_order = slub_min_order;
1811 if ((PAGE_SIZE << min_order) / size > 65535)
1812 return get_order(size * 65535) - 1;
1814 for (order = max(min_order,
1815 fls(min_objects * size - 1) - PAGE_SHIFT);
1816 order <= max_order; order++) {
1818 unsigned long slab_size = PAGE_SIZE << order;
1820 if (slab_size < min_objects * size)
1823 rem = slab_size % size;
1825 if (rem <= slab_size / fract_leftover)
1833 static inline int calculate_order(int size)
1840 * Attempt to find best configuration for a slab. This
1841 * works by first attempting to generate a layout with
1842 * the best configuration and backing off gradually.
1844 * First we reduce the acceptable waste in a slab. Then
1845 * we reduce the minimum objects required in a slab.
1847 min_objects = slub_min_objects;
1849 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1850 while (min_objects > 1) {
1852 while (fraction >= 4) {
1853 order = slab_order(size, min_objects,
1854 slub_max_order, fraction);
1855 if (order <= slub_max_order)
1863 * We were unable to place multiple objects in a slab. Now
1864 * lets see if we can place a single object there.
1866 order = slab_order(size, 1, slub_max_order, 1);
1867 if (order <= slub_max_order)
1871 * Doh this slab cannot be placed using slub_max_order.
1873 order = slab_order(size, 1, MAX_ORDER, 1);
1874 if (order <= MAX_ORDER)
1880 * Figure out what the alignment of the objects will be.
1882 static unsigned long calculate_alignment(unsigned long flags,
1883 unsigned long align, unsigned long size)
1886 * If the user wants hardware cache aligned objects then follow that
1887 * suggestion if the object is sufficiently large.
1889 * The hardware cache alignment cannot override the specified
1890 * alignment though. If that is greater then use it.
1892 if (flags & SLAB_HWCACHE_ALIGN) {
1893 unsigned long ralign = cache_line_size();
1894 while (size <= ralign / 2)
1896 align = max(align, ralign);
1899 if (align < ARCH_SLAB_MINALIGN)
1900 align = ARCH_SLAB_MINALIGN;
1902 return ALIGN(align, sizeof(void *));
1905 static void init_kmem_cache_cpu(struct kmem_cache *s,
1906 struct kmem_cache_cpu *c)
1911 c->offset = s->offset / sizeof(void *);
1912 c->objsize = s->objsize;
1913 #ifdef CONFIG_SLUB_STATS
1914 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1919 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1924 * The larger the object size is, the more pages we want on the partial
1925 * list to avoid pounding the page allocator excessively.
1927 n->min_partial = ilog2(s->size);
1928 if (n->min_partial < MIN_PARTIAL)
1929 n->min_partial = MIN_PARTIAL;
1930 else if (n->min_partial > MAX_PARTIAL)
1931 n->min_partial = MAX_PARTIAL;
1933 spin_lock_init(&n->list_lock);
1934 INIT_LIST_HEAD(&n->partial);
1935 #ifdef CONFIG_SLUB_DEBUG
1936 atomic_long_set(&n->nr_slabs, 0);
1937 atomic_long_set(&n->total_objects, 0);
1938 INIT_LIST_HEAD(&n->full);
1944 * Per cpu array for per cpu structures.
1946 * The per cpu array places all kmem_cache_cpu structures from one processor
1947 * close together meaning that it becomes possible that multiple per cpu
1948 * structures are contained in one cacheline. This may be particularly
1949 * beneficial for the kmalloc caches.
1951 * A desktop system typically has around 60-80 slabs. With 100 here we are
1952 * likely able to get per cpu structures for all caches from the array defined
1953 * here. We must be able to cover all kmalloc caches during bootstrap.
1955 * If the per cpu array is exhausted then fall back to kmalloc
1956 * of individual cachelines. No sharing is possible then.
1958 #define NR_KMEM_CACHE_CPU 100
1960 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1961 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1963 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1964 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1966 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1967 int cpu, gfp_t flags)
1969 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1972 per_cpu(kmem_cache_cpu_free, cpu) =
1973 (void *)c->freelist;
1975 /* Table overflow: So allocate ourselves */
1977 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1978 flags, cpu_to_node(cpu));
1983 init_kmem_cache_cpu(s, c);
1987 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1989 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1990 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1994 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1995 per_cpu(kmem_cache_cpu_free, cpu) = c;
1998 static void free_kmem_cache_cpus(struct kmem_cache *s)
2002 for_each_online_cpu(cpu) {
2003 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2006 s->cpu_slab[cpu] = NULL;
2007 free_kmem_cache_cpu(c, cpu);
2012 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2016 for_each_online_cpu(cpu) {
2017 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2022 c = alloc_kmem_cache_cpu(s, cpu, flags);
2024 free_kmem_cache_cpus(s);
2027 s->cpu_slab[cpu] = c;
2033 * Initialize the per cpu array.
2035 static void init_alloc_cpu_cpu(int cpu)
2039 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2042 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2043 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2045 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2048 static void __init init_alloc_cpu(void)
2052 for_each_online_cpu(cpu)
2053 init_alloc_cpu_cpu(cpu);
2057 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2058 static inline void init_alloc_cpu(void) {}
2060 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2062 init_kmem_cache_cpu(s, &s->cpu_slab);
2069 * No kmalloc_node yet so do it by hand. We know that this is the first
2070 * slab on the node for this slabcache. There are no concurrent accesses
2073 * Note that this function only works on the kmalloc_node_cache
2074 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2075 * memory on a fresh node that has no slab structures yet.
2077 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2081 struct kmem_cache_node *n;
2082 unsigned long flags;
2084 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2086 page = new_slab(kmalloc_caches, gfpflags, node);
2089 if (page_to_nid(page) != node) {
2090 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2092 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2093 "in order to be able to continue\n");
2098 page->freelist = get_freepointer(kmalloc_caches, n);
2100 kmalloc_caches->node[node] = n;
2101 #ifdef CONFIG_SLUB_DEBUG
2102 init_object(kmalloc_caches, n, 1);
2103 init_tracking(kmalloc_caches, n);
2105 init_kmem_cache_node(n, kmalloc_caches);
2106 inc_slabs_node(kmalloc_caches, node, page->objects);
2109 * lockdep requires consistent irq usage for each lock
2110 * so even though there cannot be a race this early in
2111 * the boot sequence, we still disable irqs.
2113 local_irq_save(flags);
2114 add_partial(n, page, 0);
2115 local_irq_restore(flags);
2119 static void free_kmem_cache_nodes(struct kmem_cache *s)
2123 for_each_node_state(node, N_NORMAL_MEMORY) {
2124 struct kmem_cache_node *n = s->node[node];
2125 if (n && n != &s->local_node)
2126 kmem_cache_free(kmalloc_caches, n);
2127 s->node[node] = NULL;
2131 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2136 if (slab_state >= UP)
2137 local_node = page_to_nid(virt_to_page(s));
2141 for_each_node_state(node, N_NORMAL_MEMORY) {
2142 struct kmem_cache_node *n;
2144 if (local_node == node)
2147 if (slab_state == DOWN) {
2148 n = early_kmem_cache_node_alloc(gfpflags,
2152 n = kmem_cache_alloc_node(kmalloc_caches,
2156 free_kmem_cache_nodes(s);
2162 init_kmem_cache_node(n, s);
2167 static void free_kmem_cache_nodes(struct kmem_cache *s)
2171 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2173 init_kmem_cache_node(&s->local_node, s);
2179 * calculate_sizes() determines the order and the distribution of data within
2182 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2184 unsigned long flags = s->flags;
2185 unsigned long size = s->objsize;
2186 unsigned long align = s->align;
2190 * Round up object size to the next word boundary. We can only
2191 * place the free pointer at word boundaries and this determines
2192 * the possible location of the free pointer.
2194 size = ALIGN(size, sizeof(void *));
2196 #ifdef CONFIG_SLUB_DEBUG
2198 * Determine if we can poison the object itself. If the user of
2199 * the slab may touch the object after free or before allocation
2200 * then we should never poison the object itself.
2202 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2204 s->flags |= __OBJECT_POISON;
2206 s->flags &= ~__OBJECT_POISON;
2210 * If we are Redzoning then check if there is some space between the
2211 * end of the object and the free pointer. If not then add an
2212 * additional word to have some bytes to store Redzone information.
2214 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2215 size += sizeof(void *);
2219 * With that we have determined the number of bytes in actual use
2220 * by the object. This is the potential offset to the free pointer.
2224 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2227 * Relocate free pointer after the object if it is not
2228 * permitted to overwrite the first word of the object on
2231 * This is the case if we do RCU, have a constructor or
2232 * destructor or are poisoning the objects.
2235 size += sizeof(void *);
2238 #ifdef CONFIG_SLUB_DEBUG
2239 if (flags & SLAB_STORE_USER)
2241 * Need to store information about allocs and frees after
2244 size += 2 * sizeof(struct track);
2246 if (flags & SLAB_RED_ZONE)
2248 * Add some empty padding so that we can catch
2249 * overwrites from earlier objects rather than let
2250 * tracking information or the free pointer be
2251 * corrupted if an user writes before the start
2254 size += sizeof(void *);
2258 * Determine the alignment based on various parameters that the
2259 * user specified and the dynamic determination of cache line size
2262 align = calculate_alignment(flags, align, s->objsize);
2265 * SLUB stores one object immediately after another beginning from
2266 * offset 0. In order to align the objects we have to simply size
2267 * each object to conform to the alignment.
2269 size = ALIGN(size, align);
2271 if (forced_order >= 0)
2272 order = forced_order;
2274 order = calculate_order(size);
2281 s->allocflags |= __GFP_COMP;
2283 if (s->flags & SLAB_CACHE_DMA)
2284 s->allocflags |= SLUB_DMA;
2286 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2287 s->allocflags |= __GFP_RECLAIMABLE;
2290 * Determine the number of objects per slab
2292 s->oo = oo_make(order, size);
2293 s->min = oo_make(get_order(size), size);
2294 if (oo_objects(s->oo) > oo_objects(s->max))
2297 return !!oo_objects(s->oo);
2301 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2302 const char *name, size_t size,
2303 size_t align, unsigned long flags,
2304 void (*ctor)(void *))
2306 memset(s, 0, kmem_size);
2311 s->flags = kmem_cache_flags(size, flags, name, ctor);
2313 if (!calculate_sizes(s, -1))
2318 s->remote_node_defrag_ratio = 1000;
2320 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2323 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2325 free_kmem_cache_nodes(s);
2327 if (flags & SLAB_PANIC)
2328 panic("Cannot create slab %s size=%lu realsize=%u "
2329 "order=%u offset=%u flags=%lx\n",
2330 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2336 * Check if a given pointer is valid
2338 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2342 page = get_object_page(object);
2344 if (!page || s != page->slab)
2345 /* No slab or wrong slab */
2348 if (!check_valid_pointer(s, page, object))
2352 * We could also check if the object is on the slabs freelist.
2353 * But this would be too expensive and it seems that the main
2354 * purpose of kmem_ptr_valid() is to check if the object belongs
2355 * to a certain slab.
2359 EXPORT_SYMBOL(kmem_ptr_validate);
2362 * Determine the size of a slab object
2364 unsigned int kmem_cache_size(struct kmem_cache *s)
2368 EXPORT_SYMBOL(kmem_cache_size);
2370 const char *kmem_cache_name(struct kmem_cache *s)
2374 EXPORT_SYMBOL(kmem_cache_name);
2376 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2379 #ifdef CONFIG_SLUB_DEBUG
2380 void *addr = page_address(page);
2382 DECLARE_BITMAP(map, page->objects);
2384 bitmap_zero(map, page->objects);
2385 slab_err(s, page, "%s", text);
2387 for_each_free_object(p, s, page->freelist)
2388 set_bit(slab_index(p, s, addr), map);
2390 for_each_object(p, s, addr, page->objects) {
2392 if (!test_bit(slab_index(p, s, addr), map)) {
2393 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2395 print_tracking(s, p);
2403 * Attempt to free all partial slabs on a node.
2405 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2407 unsigned long flags;
2408 struct page *page, *h;
2410 spin_lock_irqsave(&n->list_lock, flags);
2411 list_for_each_entry_safe(page, h, &n->partial, lru) {
2413 list_del(&page->lru);
2414 discard_slab(s, page);
2417 list_slab_objects(s, page,
2418 "Objects remaining on kmem_cache_close()");
2421 spin_unlock_irqrestore(&n->list_lock, flags);
2425 * Release all resources used by a slab cache.
2427 static inline int kmem_cache_close(struct kmem_cache *s)
2433 /* Attempt to free all objects */
2434 free_kmem_cache_cpus(s);
2435 for_each_node_state(node, N_NORMAL_MEMORY) {
2436 struct kmem_cache_node *n = get_node(s, node);
2439 if (n->nr_partial || slabs_node(s, node))
2442 free_kmem_cache_nodes(s);
2447 * Close a cache and release the kmem_cache structure
2448 * (must be used for caches created using kmem_cache_create)
2450 void kmem_cache_destroy(struct kmem_cache *s)
2452 down_write(&slub_lock);
2456 up_write(&slub_lock);
2457 if (kmem_cache_close(s)) {
2458 printk(KERN_ERR "SLUB %s: %s called for cache that "
2459 "still has objects.\n", s->name, __func__);
2462 sysfs_slab_remove(s);
2464 up_write(&slub_lock);
2466 EXPORT_SYMBOL(kmem_cache_destroy);
2468 /********************************************************************
2470 *******************************************************************/
2472 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2473 EXPORT_SYMBOL(kmalloc_caches);
2475 static int __init setup_slub_min_order(char *str)
2477 get_option(&str, &slub_min_order);
2482 __setup("slub_min_order=", setup_slub_min_order);
2484 static int __init setup_slub_max_order(char *str)
2486 get_option(&str, &slub_max_order);
2491 __setup("slub_max_order=", setup_slub_max_order);
2493 static int __init setup_slub_min_objects(char *str)
2495 get_option(&str, &slub_min_objects);
2500 __setup("slub_min_objects=", setup_slub_min_objects);
2502 static int __init setup_slub_nomerge(char *str)
2508 __setup("slub_nomerge", setup_slub_nomerge);
2510 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2511 const char *name, int size, gfp_t gfp_flags)
2513 unsigned int flags = 0;
2515 if (gfp_flags & SLUB_DMA)
2516 flags = SLAB_CACHE_DMA;
2518 down_write(&slub_lock);
2519 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2523 list_add(&s->list, &slab_caches);
2524 up_write(&slub_lock);
2525 if (sysfs_slab_add(s))
2530 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2533 #ifdef CONFIG_ZONE_DMA
2534 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2536 static void sysfs_add_func(struct work_struct *w)
2538 struct kmem_cache *s;
2540 down_write(&slub_lock);
2541 list_for_each_entry(s, &slab_caches, list) {
2542 if (s->flags & __SYSFS_ADD_DEFERRED) {
2543 s->flags &= ~__SYSFS_ADD_DEFERRED;
2547 up_write(&slub_lock);
2550 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2552 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2554 struct kmem_cache *s;
2558 s = kmalloc_caches_dma[index];
2562 /* Dynamically create dma cache */
2563 if (flags & __GFP_WAIT)
2564 down_write(&slub_lock);
2566 if (!down_write_trylock(&slub_lock))
2570 if (kmalloc_caches_dma[index])
2573 realsize = kmalloc_caches[index].objsize;
2574 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2575 (unsigned int)realsize);
2576 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2578 if (!s || !text || !kmem_cache_open(s, flags, text,
2579 realsize, ARCH_KMALLOC_MINALIGN,
2580 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2586 list_add(&s->list, &slab_caches);
2587 kmalloc_caches_dma[index] = s;
2589 schedule_work(&sysfs_add_work);
2592 up_write(&slub_lock);
2594 return kmalloc_caches_dma[index];
2599 * Conversion table for small slabs sizes / 8 to the index in the
2600 * kmalloc array. This is necessary for slabs < 192 since we have non power
2601 * of two cache sizes there. The size of larger slabs can be determined using
2604 static s8 size_index[24] = {
2631 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2637 return ZERO_SIZE_PTR;
2639 index = size_index[(size - 1) / 8];
2641 index = fls(size - 1);
2643 #ifdef CONFIG_ZONE_DMA
2644 if (unlikely((flags & SLUB_DMA)))
2645 return dma_kmalloc_cache(index, flags);
2648 return &kmalloc_caches[index];
2651 void *__kmalloc(size_t size, gfp_t flags)
2653 struct kmem_cache *s;
2655 if (unlikely(size > PAGE_SIZE))
2656 return kmalloc_large(size, flags);
2658 s = get_slab(size, flags);
2660 if (unlikely(ZERO_OR_NULL_PTR(s)))
2663 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2665 EXPORT_SYMBOL(__kmalloc);
2667 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2669 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2673 return page_address(page);
2679 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2681 struct kmem_cache *s;
2683 if (unlikely(size > PAGE_SIZE))
2684 return kmalloc_large_node(size, flags, node);
2686 s = get_slab(size, flags);
2688 if (unlikely(ZERO_OR_NULL_PTR(s)))
2691 return slab_alloc(s, flags, node, __builtin_return_address(0));
2693 EXPORT_SYMBOL(__kmalloc_node);
2696 size_t ksize(const void *object)
2699 struct kmem_cache *s;
2701 if (unlikely(object == ZERO_SIZE_PTR))
2704 page = virt_to_head_page(object);
2706 if (unlikely(!PageSlab(page))) {
2707 WARN_ON(!PageCompound(page));
2708 return PAGE_SIZE << compound_order(page);
2712 #ifdef CONFIG_SLUB_DEBUG
2714 * Debugging requires use of the padding between object
2715 * and whatever may come after it.
2717 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2722 * If we have the need to store the freelist pointer
2723 * back there or track user information then we can
2724 * only use the space before that information.
2726 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2729 * Else we can use all the padding etc for the allocation
2734 void kfree(const void *x)
2737 void *object = (void *)x;
2739 if (unlikely(ZERO_OR_NULL_PTR(x)))
2742 page = virt_to_head_page(x);
2743 if (unlikely(!PageSlab(page))) {
2744 BUG_ON(!PageCompound(page));
2748 slab_free(page->slab, page, object, __builtin_return_address(0));
2750 EXPORT_SYMBOL(kfree);
2753 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2754 * the remaining slabs by the number of items in use. The slabs with the
2755 * most items in use come first. New allocations will then fill those up
2756 * and thus they can be removed from the partial lists.
2758 * The slabs with the least items are placed last. This results in them
2759 * being allocated from last increasing the chance that the last objects
2760 * are freed in them.
2762 int kmem_cache_shrink(struct kmem_cache *s)
2766 struct kmem_cache_node *n;
2769 int objects = oo_objects(s->max);
2770 struct list_head *slabs_by_inuse =
2771 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2772 unsigned long flags;
2774 if (!slabs_by_inuse)
2778 for_each_node_state(node, N_NORMAL_MEMORY) {
2779 n = get_node(s, node);
2784 for (i = 0; i < objects; i++)
2785 INIT_LIST_HEAD(slabs_by_inuse + i);
2787 spin_lock_irqsave(&n->list_lock, flags);
2790 * Build lists indexed by the items in use in each slab.
2792 * Note that concurrent frees may occur while we hold the
2793 * list_lock. page->inuse here is the upper limit.
2795 list_for_each_entry_safe(page, t, &n->partial, lru) {
2796 if (!page->inuse && slab_trylock(page)) {
2798 * Must hold slab lock here because slab_free
2799 * may have freed the last object and be
2800 * waiting to release the slab.
2802 list_del(&page->lru);
2805 discard_slab(s, page);
2807 list_move(&page->lru,
2808 slabs_by_inuse + page->inuse);
2813 * Rebuild the partial list with the slabs filled up most
2814 * first and the least used slabs at the end.
2816 for (i = objects - 1; i >= 0; i--)
2817 list_splice(slabs_by_inuse + i, n->partial.prev);
2819 spin_unlock_irqrestore(&n->list_lock, flags);
2822 kfree(slabs_by_inuse);
2825 EXPORT_SYMBOL(kmem_cache_shrink);
2827 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2828 static int slab_mem_going_offline_callback(void *arg)
2830 struct kmem_cache *s;
2832 down_read(&slub_lock);
2833 list_for_each_entry(s, &slab_caches, list)
2834 kmem_cache_shrink(s);
2835 up_read(&slub_lock);
2840 static void slab_mem_offline_callback(void *arg)
2842 struct kmem_cache_node *n;
2843 struct kmem_cache *s;
2844 struct memory_notify *marg = arg;
2847 offline_node = marg->status_change_nid;
2850 * If the node still has available memory. we need kmem_cache_node
2853 if (offline_node < 0)
2856 down_read(&slub_lock);
2857 list_for_each_entry(s, &slab_caches, list) {
2858 n = get_node(s, offline_node);
2861 * if n->nr_slabs > 0, slabs still exist on the node
2862 * that is going down. We were unable to free them,
2863 * and offline_pages() function shoudn't call this
2864 * callback. So, we must fail.
2866 BUG_ON(slabs_node(s, offline_node));
2868 s->node[offline_node] = NULL;
2869 kmem_cache_free(kmalloc_caches, n);
2872 up_read(&slub_lock);
2875 static int slab_mem_going_online_callback(void *arg)
2877 struct kmem_cache_node *n;
2878 struct kmem_cache *s;
2879 struct memory_notify *marg = arg;
2880 int nid = marg->status_change_nid;
2884 * If the node's memory is already available, then kmem_cache_node is
2885 * already created. Nothing to do.
2891 * We are bringing a node online. No memory is available yet. We must
2892 * allocate a kmem_cache_node structure in order to bring the node
2895 down_read(&slub_lock);
2896 list_for_each_entry(s, &slab_caches, list) {
2898 * XXX: kmem_cache_alloc_node will fallback to other nodes
2899 * since memory is not yet available from the node that
2902 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2907 init_kmem_cache_node(n, s);
2911 up_read(&slub_lock);
2915 static int slab_memory_callback(struct notifier_block *self,
2916 unsigned long action, void *arg)
2921 case MEM_GOING_ONLINE:
2922 ret = slab_mem_going_online_callback(arg);
2924 case MEM_GOING_OFFLINE:
2925 ret = slab_mem_going_offline_callback(arg);
2928 case MEM_CANCEL_ONLINE:
2929 slab_mem_offline_callback(arg);
2932 case MEM_CANCEL_OFFLINE:
2936 ret = notifier_from_errno(ret);
2942 #endif /* CONFIG_MEMORY_HOTPLUG */
2944 /********************************************************************
2945 * Basic setup of slabs
2946 *******************************************************************/
2948 void __init kmem_cache_init(void)
2957 * Must first have the slab cache available for the allocations of the
2958 * struct kmem_cache_node's. There is special bootstrap code in
2959 * kmem_cache_open for slab_state == DOWN.
2961 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2962 sizeof(struct kmem_cache_node), GFP_KERNEL);
2963 kmalloc_caches[0].refcount = -1;
2966 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
2969 /* Able to allocate the per node structures */
2970 slab_state = PARTIAL;
2972 /* Caches that are not of the two-to-the-power-of size */
2973 if (KMALLOC_MIN_SIZE <= 64) {
2974 create_kmalloc_cache(&kmalloc_caches[1],
2975 "kmalloc-96", 96, GFP_KERNEL);
2977 create_kmalloc_cache(&kmalloc_caches[2],
2978 "kmalloc-192", 192, GFP_KERNEL);
2982 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
2983 create_kmalloc_cache(&kmalloc_caches[i],
2984 "kmalloc", 1 << i, GFP_KERNEL);
2990 * Patch up the size_index table if we have strange large alignment
2991 * requirements for the kmalloc array. This is only the case for
2992 * MIPS it seems. The standard arches will not generate any code here.
2994 * Largest permitted alignment is 256 bytes due to the way we
2995 * handle the index determination for the smaller caches.
2997 * Make sure that nothing crazy happens if someone starts tinkering
2998 * around with ARCH_KMALLOC_MINALIGN
3000 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3001 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3003 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3004 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3006 if (KMALLOC_MIN_SIZE == 128) {
3008 * The 192 byte sized cache is not used if the alignment
3009 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3012 for (i = 128 + 8; i <= 192; i += 8)
3013 size_index[(i - 1) / 8] = 8;
3018 /* Provide the correct kmalloc names now that the caches are up */
3019 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3020 kmalloc_caches[i]. name =
3021 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3024 register_cpu_notifier(&slab_notifier);
3025 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3026 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3028 kmem_size = sizeof(struct kmem_cache);
3032 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3033 " CPUs=%d, Nodes=%d\n",
3034 caches, cache_line_size(),
3035 slub_min_order, slub_max_order, slub_min_objects,
3036 nr_cpu_ids, nr_node_ids);
3040 * Find a mergeable slab cache
3042 static int slab_unmergeable(struct kmem_cache *s)
3044 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3051 * We may have set a slab to be unmergeable during bootstrap.
3053 if (s->refcount < 0)
3059 static struct kmem_cache *find_mergeable(size_t size,
3060 size_t align, unsigned long flags, const char *name,
3061 void (*ctor)(void *))
3063 struct kmem_cache *s;
3065 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3071 size = ALIGN(size, sizeof(void *));
3072 align = calculate_alignment(flags, align, size);
3073 size = ALIGN(size, align);
3074 flags = kmem_cache_flags(size, flags, name, NULL);
3076 list_for_each_entry(s, &slab_caches, list) {
3077 if (slab_unmergeable(s))
3083 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3086 * Check if alignment is compatible.
3087 * Courtesy of Adrian Drzewiecki
3089 if ((s->size & ~(align - 1)) != s->size)
3092 if (s->size - size >= sizeof(void *))
3100 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3101 size_t align, unsigned long flags, void (*ctor)(void *))
3103 struct kmem_cache *s;
3105 down_write(&slub_lock);
3106 s = find_mergeable(size, align, flags, name, ctor);
3112 * Adjust the object sizes so that we clear
3113 * the complete object on kzalloc.
3115 s->objsize = max(s->objsize, (int)size);
3118 * And then we need to update the object size in the
3119 * per cpu structures
3121 for_each_online_cpu(cpu)
3122 get_cpu_slab(s, cpu)->objsize = s->objsize;
3124 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3125 up_write(&slub_lock);
3127 if (sysfs_slab_alias(s, name))
3132 s = kmalloc(kmem_size, GFP_KERNEL);
3134 if (kmem_cache_open(s, GFP_KERNEL, name,
3135 size, align, flags, ctor)) {
3136 list_add(&s->list, &slab_caches);
3137 up_write(&slub_lock);
3138 if (sysfs_slab_add(s))
3144 up_write(&slub_lock);
3147 if (flags & SLAB_PANIC)
3148 panic("Cannot create slabcache %s\n", name);
3153 EXPORT_SYMBOL(kmem_cache_create);
3157 * Use the cpu notifier to insure that the cpu slabs are flushed when
3160 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3161 unsigned long action, void *hcpu)
3163 long cpu = (long)hcpu;
3164 struct kmem_cache *s;
3165 unsigned long flags;
3168 case CPU_UP_PREPARE:
3169 case CPU_UP_PREPARE_FROZEN:
3170 init_alloc_cpu_cpu(cpu);
3171 down_read(&slub_lock);
3172 list_for_each_entry(s, &slab_caches, list)
3173 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3175 up_read(&slub_lock);
3178 case CPU_UP_CANCELED:
3179 case CPU_UP_CANCELED_FROZEN:
3181 case CPU_DEAD_FROZEN:
3182 down_read(&slub_lock);
3183 list_for_each_entry(s, &slab_caches, list) {
3184 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3186 local_irq_save(flags);
3187 __flush_cpu_slab(s, cpu);
3188 local_irq_restore(flags);
3189 free_kmem_cache_cpu(c, cpu);
3190 s->cpu_slab[cpu] = NULL;
3192 up_read(&slub_lock);
3200 static struct notifier_block __cpuinitdata slab_notifier = {
3201 .notifier_call = slab_cpuup_callback
3206 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3208 struct kmem_cache *s;
3210 if (unlikely(size > PAGE_SIZE))
3211 return kmalloc_large(size, gfpflags);
3213 s = get_slab(size, gfpflags);
3215 if (unlikely(ZERO_OR_NULL_PTR(s)))
3218 return slab_alloc(s, gfpflags, -1, caller);
3221 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3222 int node, void *caller)
3224 struct kmem_cache *s;
3226 if (unlikely(size > PAGE_SIZE))
3227 return kmalloc_large_node(size, gfpflags, node);
3229 s = get_slab(size, gfpflags);
3231 if (unlikely(ZERO_OR_NULL_PTR(s)))
3234 return slab_alloc(s, gfpflags, node, caller);
3237 #ifdef CONFIG_SLUB_DEBUG
3238 static unsigned long count_partial(struct kmem_cache_node *n,
3239 int (*get_count)(struct page *))
3241 unsigned long flags;
3242 unsigned long x = 0;
3245 spin_lock_irqsave(&n->list_lock, flags);
3246 list_for_each_entry(page, &n->partial, lru)
3247 x += get_count(page);
3248 spin_unlock_irqrestore(&n->list_lock, flags);
3252 static int count_inuse(struct page *page)
3257 static int count_total(struct page *page)
3259 return page->objects;
3262 static int count_free(struct page *page)
3264 return page->objects - page->inuse;
3267 static int validate_slab(struct kmem_cache *s, struct page *page,
3271 void *addr = page_address(page);
3273 if (!check_slab(s, page) ||
3274 !on_freelist(s, page, NULL))
3277 /* Now we know that a valid freelist exists */
3278 bitmap_zero(map, page->objects);
3280 for_each_free_object(p, s, page->freelist) {
3281 set_bit(slab_index(p, s, addr), map);
3282 if (!check_object(s, page, p, 0))
3286 for_each_object(p, s, addr, page->objects)
3287 if (!test_bit(slab_index(p, s, addr), map))
3288 if (!check_object(s, page, p, 1))
3293 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3296 if (slab_trylock(page)) {
3297 validate_slab(s, page, map);
3300 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3303 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3304 if (!PageSlubDebug(page))
3305 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3306 "on slab 0x%p\n", s->name, page);
3308 if (PageSlubDebug(page))
3309 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3310 "slab 0x%p\n", s->name, page);
3314 static int validate_slab_node(struct kmem_cache *s,
3315 struct kmem_cache_node *n, unsigned long *map)
3317 unsigned long count = 0;
3319 unsigned long flags;
3321 spin_lock_irqsave(&n->list_lock, flags);
3323 list_for_each_entry(page, &n->partial, lru) {
3324 validate_slab_slab(s, page, map);
3327 if (count != n->nr_partial)
3328 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3329 "counter=%ld\n", s->name, count, n->nr_partial);
3331 if (!(s->flags & SLAB_STORE_USER))
3334 list_for_each_entry(page, &n->full, lru) {
3335 validate_slab_slab(s, page, map);
3338 if (count != atomic_long_read(&n->nr_slabs))
3339 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3340 "counter=%ld\n", s->name, count,
3341 atomic_long_read(&n->nr_slabs));
3344 spin_unlock_irqrestore(&n->list_lock, flags);
3348 static long validate_slab_cache(struct kmem_cache *s)
3351 unsigned long count = 0;
3352 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3353 sizeof(unsigned long), GFP_KERNEL);
3359 for_each_node_state(node, N_NORMAL_MEMORY) {
3360 struct kmem_cache_node *n = get_node(s, node);
3362 count += validate_slab_node(s, n, map);
3368 #ifdef SLUB_RESILIENCY_TEST
3369 static void resiliency_test(void)
3373 printk(KERN_ERR "SLUB resiliency testing\n");
3374 printk(KERN_ERR "-----------------------\n");
3375 printk(KERN_ERR "A. Corruption after allocation\n");
3377 p = kzalloc(16, GFP_KERNEL);
3379 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3380 " 0x12->0x%p\n\n", p + 16);
3382 validate_slab_cache(kmalloc_caches + 4);
3384 /* Hmmm... The next two are dangerous */
3385 p = kzalloc(32, GFP_KERNEL);
3386 p[32 + sizeof(void *)] = 0x34;
3387 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3388 " 0x34 -> -0x%p\n", p);
3390 "If allocated object is overwritten then not detectable\n\n");
3392 validate_slab_cache(kmalloc_caches + 5);
3393 p = kzalloc(64, GFP_KERNEL);
3394 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3396 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3399 "If allocated object is overwritten then not detectable\n\n");
3400 validate_slab_cache(kmalloc_caches + 6);
3402 printk(KERN_ERR "\nB. Corruption after free\n");
3403 p = kzalloc(128, GFP_KERNEL);
3406 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3407 validate_slab_cache(kmalloc_caches + 7);
3409 p = kzalloc(256, GFP_KERNEL);
3412 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3414 validate_slab_cache(kmalloc_caches + 8);
3416 p = kzalloc(512, GFP_KERNEL);
3419 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3420 validate_slab_cache(kmalloc_caches + 9);
3423 static void resiliency_test(void) {};
3427 * Generate lists of code addresses where slabcache objects are allocated
3432 unsigned long count;
3445 unsigned long count;
3446 struct location *loc;
3449 static void free_loc_track(struct loc_track *t)
3452 free_pages((unsigned long)t->loc,
3453 get_order(sizeof(struct location) * t->max));
3456 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3461 order = get_order(sizeof(struct location) * max);
3463 l = (void *)__get_free_pages(flags, order);
3468 memcpy(l, t->loc, sizeof(struct location) * t->count);
3476 static int add_location(struct loc_track *t, struct kmem_cache *s,
3477 const struct track *track)
3479 long start, end, pos;
3482 unsigned long age = jiffies - track->when;
3488 pos = start + (end - start + 1) / 2;
3491 * There is nothing at "end". If we end up there
3492 * we need to add something to before end.
3497 caddr = t->loc[pos].addr;
3498 if (track->addr == caddr) {
3504 if (age < l->min_time)
3506 if (age > l->max_time)
3509 if (track->pid < l->min_pid)
3510 l->min_pid = track->pid;
3511 if (track->pid > l->max_pid)
3512 l->max_pid = track->pid;
3514 cpu_set(track->cpu, l->cpus);
3516 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3520 if (track->addr < caddr)
3527 * Not found. Insert new tracking element.
3529 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3535 (t->count - pos) * sizeof(struct location));
3538 l->addr = track->addr;
3542 l->min_pid = track->pid;
3543 l->max_pid = track->pid;
3544 cpus_clear(l->cpus);
3545 cpu_set(track->cpu, l->cpus);
3546 nodes_clear(l->nodes);
3547 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3551 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3552 struct page *page, enum track_item alloc)
3554 void *addr = page_address(page);
3555 DECLARE_BITMAP(map, page->objects);
3558 bitmap_zero(map, page->objects);
3559 for_each_free_object(p, s, page->freelist)
3560 set_bit(slab_index(p, s, addr), map);
3562 for_each_object(p, s, addr, page->objects)
3563 if (!test_bit(slab_index(p, s, addr), map))
3564 add_location(t, s, get_track(s, p, alloc));
3567 static int list_locations(struct kmem_cache *s, char *buf,
3568 enum track_item alloc)
3572 struct loc_track t = { 0, 0, NULL };
3575 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3577 return sprintf(buf, "Out of memory\n");
3579 /* Push back cpu slabs */
3582 for_each_node_state(node, N_NORMAL_MEMORY) {
3583 struct kmem_cache_node *n = get_node(s, node);
3584 unsigned long flags;
3587 if (!atomic_long_read(&n->nr_slabs))
3590 spin_lock_irqsave(&n->list_lock, flags);
3591 list_for_each_entry(page, &n->partial, lru)
3592 process_slab(&t, s, page, alloc);
3593 list_for_each_entry(page, &n->full, lru)
3594 process_slab(&t, s, page, alloc);
3595 spin_unlock_irqrestore(&n->list_lock, flags);
3598 for (i = 0; i < t.count; i++) {
3599 struct location *l = &t.loc[i];
3601 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3603 len += sprintf(buf + len, "%7ld ", l->count);
3606 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3608 len += sprintf(buf + len, "<not-available>");
3610 if (l->sum_time != l->min_time) {
3611 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3613 (long)div_u64(l->sum_time, l->count),
3616 len += sprintf(buf + len, " age=%ld",
3619 if (l->min_pid != l->max_pid)
3620 len += sprintf(buf + len, " pid=%ld-%ld",
3621 l->min_pid, l->max_pid);
3623 len += sprintf(buf + len, " pid=%ld",
3626 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3627 len < PAGE_SIZE - 60) {
3628 len += sprintf(buf + len, " cpus=");
3629 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3633 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3634 len < PAGE_SIZE - 60) {
3635 len += sprintf(buf + len, " nodes=");
3636 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3640 len += sprintf(buf + len, "\n");
3645 len += sprintf(buf, "No data\n");
3649 enum slab_stat_type {
3650 SL_ALL, /* All slabs */
3651 SL_PARTIAL, /* Only partially allocated slabs */
3652 SL_CPU, /* Only slabs used for cpu caches */
3653 SL_OBJECTS, /* Determine allocated objects not slabs */
3654 SL_TOTAL /* Determine object capacity not slabs */
3657 #define SO_ALL (1 << SL_ALL)
3658 #define SO_PARTIAL (1 << SL_PARTIAL)
3659 #define SO_CPU (1 << SL_CPU)
3660 #define SO_OBJECTS (1 << SL_OBJECTS)
3661 #define SO_TOTAL (1 << SL_TOTAL)
3663 static ssize_t show_slab_objects(struct kmem_cache *s,
3664 char *buf, unsigned long flags)
3666 unsigned long total = 0;
3669 unsigned long *nodes;
3670 unsigned long *per_cpu;
3672 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3675 per_cpu = nodes + nr_node_ids;
3677 if (flags & SO_CPU) {
3680 for_each_possible_cpu(cpu) {
3681 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3683 if (!c || c->node < 0)
3687 if (flags & SO_TOTAL)
3688 x = c->page->objects;
3689 else if (flags & SO_OBJECTS)
3695 nodes[c->node] += x;
3701 if (flags & SO_ALL) {
3702 for_each_node_state(node, N_NORMAL_MEMORY) {
3703 struct kmem_cache_node *n = get_node(s, node);
3705 if (flags & SO_TOTAL)
3706 x = atomic_long_read(&n->total_objects);
3707 else if (flags & SO_OBJECTS)
3708 x = atomic_long_read(&n->total_objects) -
3709 count_partial(n, count_free);
3712 x = atomic_long_read(&n->nr_slabs);
3717 } else if (flags & SO_PARTIAL) {
3718 for_each_node_state(node, N_NORMAL_MEMORY) {
3719 struct kmem_cache_node *n = get_node(s, node);
3721 if (flags & SO_TOTAL)
3722 x = count_partial(n, count_total);
3723 else if (flags & SO_OBJECTS)
3724 x = count_partial(n, count_inuse);
3731 x = sprintf(buf, "%lu", total);
3733 for_each_node_state(node, N_NORMAL_MEMORY)
3735 x += sprintf(buf + x, " N%d=%lu",
3739 return x + sprintf(buf + x, "\n");
3742 static int any_slab_objects(struct kmem_cache *s)
3746 for_each_online_node(node) {
3747 struct kmem_cache_node *n = get_node(s, node);
3752 if (atomic_long_read(&n->total_objects))
3758 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3759 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3761 struct slab_attribute {
3762 struct attribute attr;
3763 ssize_t (*show)(struct kmem_cache *s, char *buf);
3764 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3767 #define SLAB_ATTR_RO(_name) \
3768 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3770 #define SLAB_ATTR(_name) \
3771 static struct slab_attribute _name##_attr = \
3772 __ATTR(_name, 0644, _name##_show, _name##_store)
3774 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3776 return sprintf(buf, "%d\n", s->size);
3778 SLAB_ATTR_RO(slab_size);
3780 static ssize_t align_show(struct kmem_cache *s, char *buf)
3782 return sprintf(buf, "%d\n", s->align);
3784 SLAB_ATTR_RO(align);
3786 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3788 return sprintf(buf, "%d\n", s->objsize);
3790 SLAB_ATTR_RO(object_size);
3792 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3794 return sprintf(buf, "%d\n", oo_objects(s->oo));
3796 SLAB_ATTR_RO(objs_per_slab);
3798 static ssize_t order_store(struct kmem_cache *s,
3799 const char *buf, size_t length)
3801 unsigned long order;
3804 err = strict_strtoul(buf, 10, &order);
3808 if (order > slub_max_order || order < slub_min_order)
3811 calculate_sizes(s, order);
3815 static ssize_t order_show(struct kmem_cache *s, char *buf)
3817 return sprintf(buf, "%d\n", oo_order(s->oo));
3821 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3824 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3826 return n + sprintf(buf + n, "\n");
3832 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3834 return sprintf(buf, "%d\n", s->refcount - 1);
3836 SLAB_ATTR_RO(aliases);
3838 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3840 return show_slab_objects(s, buf, SO_ALL);
3842 SLAB_ATTR_RO(slabs);
3844 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3846 return show_slab_objects(s, buf, SO_PARTIAL);
3848 SLAB_ATTR_RO(partial);
3850 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3852 return show_slab_objects(s, buf, SO_CPU);
3854 SLAB_ATTR_RO(cpu_slabs);
3856 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3858 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3860 SLAB_ATTR_RO(objects);
3862 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3864 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3866 SLAB_ATTR_RO(objects_partial);
3868 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3870 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3872 SLAB_ATTR_RO(total_objects);
3874 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3876 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3879 static ssize_t sanity_checks_store(struct kmem_cache *s,
3880 const char *buf, size_t length)
3882 s->flags &= ~SLAB_DEBUG_FREE;
3884 s->flags |= SLAB_DEBUG_FREE;
3887 SLAB_ATTR(sanity_checks);
3889 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3891 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3894 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3897 s->flags &= ~SLAB_TRACE;
3899 s->flags |= SLAB_TRACE;
3904 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3906 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3909 static ssize_t reclaim_account_store(struct kmem_cache *s,
3910 const char *buf, size_t length)
3912 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3914 s->flags |= SLAB_RECLAIM_ACCOUNT;
3917 SLAB_ATTR(reclaim_account);
3919 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3921 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3923 SLAB_ATTR_RO(hwcache_align);
3925 #ifdef CONFIG_ZONE_DMA
3926 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3928 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3930 SLAB_ATTR_RO(cache_dma);
3933 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3935 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3937 SLAB_ATTR_RO(destroy_by_rcu);
3939 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3941 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3944 static ssize_t red_zone_store(struct kmem_cache *s,
3945 const char *buf, size_t length)
3947 if (any_slab_objects(s))
3950 s->flags &= ~SLAB_RED_ZONE;
3952 s->flags |= SLAB_RED_ZONE;
3953 calculate_sizes(s, -1);
3956 SLAB_ATTR(red_zone);
3958 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3960 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3963 static ssize_t poison_store(struct kmem_cache *s,
3964 const char *buf, size_t length)
3966 if (any_slab_objects(s))
3969 s->flags &= ~SLAB_POISON;
3971 s->flags |= SLAB_POISON;
3972 calculate_sizes(s, -1);
3977 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3979 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3982 static ssize_t store_user_store(struct kmem_cache *s,
3983 const char *buf, size_t length)
3985 if (any_slab_objects(s))
3988 s->flags &= ~SLAB_STORE_USER;
3990 s->flags |= SLAB_STORE_USER;
3991 calculate_sizes(s, -1);
3994 SLAB_ATTR(store_user);
3996 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4001 static ssize_t validate_store(struct kmem_cache *s,
4002 const char *buf, size_t length)
4006 if (buf[0] == '1') {
4007 ret = validate_slab_cache(s);
4013 SLAB_ATTR(validate);
4015 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4020 static ssize_t shrink_store(struct kmem_cache *s,
4021 const char *buf, size_t length)
4023 if (buf[0] == '1') {
4024 int rc = kmem_cache_shrink(s);
4034 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4036 if (!(s->flags & SLAB_STORE_USER))
4038 return list_locations(s, buf, TRACK_ALLOC);
4040 SLAB_ATTR_RO(alloc_calls);
4042 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4044 if (!(s->flags & SLAB_STORE_USER))
4046 return list_locations(s, buf, TRACK_FREE);
4048 SLAB_ATTR_RO(free_calls);
4051 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4053 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4056 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4057 const char *buf, size_t length)
4059 unsigned long ratio;
4062 err = strict_strtoul(buf, 10, &ratio);
4067 s->remote_node_defrag_ratio = ratio * 10;
4071 SLAB_ATTR(remote_node_defrag_ratio);
4074 #ifdef CONFIG_SLUB_STATS
4075 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4077 unsigned long sum = 0;
4080 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4085 for_each_online_cpu(cpu) {
4086 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4092 len = sprintf(buf, "%lu", sum);
4095 for_each_online_cpu(cpu) {
4096 if (data[cpu] && len < PAGE_SIZE - 20)
4097 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4101 return len + sprintf(buf + len, "\n");
4104 #define STAT_ATTR(si, text) \
4105 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4107 return show_stat(s, buf, si); \
4109 SLAB_ATTR_RO(text); \
4111 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4112 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4113 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4114 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4115 STAT_ATTR(FREE_FROZEN, free_frozen);
4116 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4117 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4118 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4119 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4120 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4121 STAT_ATTR(FREE_SLAB, free_slab);
4122 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4123 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4124 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4125 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4126 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4127 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4128 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4131 static struct attribute *slab_attrs[] = {
4132 &slab_size_attr.attr,
4133 &object_size_attr.attr,
4134 &objs_per_slab_attr.attr,
4137 &objects_partial_attr.attr,
4138 &total_objects_attr.attr,
4141 &cpu_slabs_attr.attr,
4145 &sanity_checks_attr.attr,
4147 &hwcache_align_attr.attr,
4148 &reclaim_account_attr.attr,
4149 &destroy_by_rcu_attr.attr,
4150 &red_zone_attr.attr,
4152 &store_user_attr.attr,
4153 &validate_attr.attr,
4155 &alloc_calls_attr.attr,
4156 &free_calls_attr.attr,
4157 #ifdef CONFIG_ZONE_DMA
4158 &cache_dma_attr.attr,
4161 &remote_node_defrag_ratio_attr.attr,
4163 #ifdef CONFIG_SLUB_STATS
4164 &alloc_fastpath_attr.attr,
4165 &alloc_slowpath_attr.attr,
4166 &free_fastpath_attr.attr,
4167 &free_slowpath_attr.attr,
4168 &free_frozen_attr.attr,
4169 &free_add_partial_attr.attr,
4170 &free_remove_partial_attr.attr,
4171 &alloc_from_partial_attr.attr,
4172 &alloc_slab_attr.attr,
4173 &alloc_refill_attr.attr,
4174 &free_slab_attr.attr,
4175 &cpuslab_flush_attr.attr,
4176 &deactivate_full_attr.attr,
4177 &deactivate_empty_attr.attr,
4178 &deactivate_to_head_attr.attr,
4179 &deactivate_to_tail_attr.attr,
4180 &deactivate_remote_frees_attr.attr,
4181 &order_fallback_attr.attr,
4186 static struct attribute_group slab_attr_group = {
4187 .attrs = slab_attrs,
4190 static ssize_t slab_attr_show(struct kobject *kobj,
4191 struct attribute *attr,
4194 struct slab_attribute *attribute;
4195 struct kmem_cache *s;
4198 attribute = to_slab_attr(attr);
4201 if (!attribute->show)
4204 err = attribute->show(s, buf);
4209 static ssize_t slab_attr_store(struct kobject *kobj,
4210 struct attribute *attr,
4211 const char *buf, size_t len)
4213 struct slab_attribute *attribute;
4214 struct kmem_cache *s;
4217 attribute = to_slab_attr(attr);
4220 if (!attribute->store)
4223 err = attribute->store(s, buf, len);
4228 static void kmem_cache_release(struct kobject *kobj)
4230 struct kmem_cache *s = to_slab(kobj);
4235 static struct sysfs_ops slab_sysfs_ops = {
4236 .show = slab_attr_show,
4237 .store = slab_attr_store,
4240 static struct kobj_type slab_ktype = {
4241 .sysfs_ops = &slab_sysfs_ops,
4242 .release = kmem_cache_release
4245 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4247 struct kobj_type *ktype = get_ktype(kobj);
4249 if (ktype == &slab_ktype)
4254 static struct kset_uevent_ops slab_uevent_ops = {
4255 .filter = uevent_filter,
4258 static struct kset *slab_kset;
4260 #define ID_STR_LENGTH 64
4262 /* Create a unique string id for a slab cache:
4264 * Format :[flags-]size
4266 static char *create_unique_id(struct kmem_cache *s)
4268 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4275 * First flags affecting slabcache operations. We will only
4276 * get here for aliasable slabs so we do not need to support
4277 * too many flags. The flags here must cover all flags that
4278 * are matched during merging to guarantee that the id is
4281 if (s->flags & SLAB_CACHE_DMA)
4283 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4285 if (s->flags & SLAB_DEBUG_FREE)
4289 p += sprintf(p, "%07d", s->size);
4290 BUG_ON(p > name + ID_STR_LENGTH - 1);
4294 static int sysfs_slab_add(struct kmem_cache *s)
4300 if (slab_state < SYSFS)
4301 /* Defer until later */
4304 unmergeable = slab_unmergeable(s);
4307 * Slabcache can never be merged so we can use the name proper.
4308 * This is typically the case for debug situations. In that
4309 * case we can catch duplicate names easily.
4311 sysfs_remove_link(&slab_kset->kobj, s->name);
4315 * Create a unique name for the slab as a target
4318 name = create_unique_id(s);
4321 s->kobj.kset = slab_kset;
4322 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4324 kobject_put(&s->kobj);
4328 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4331 kobject_uevent(&s->kobj, KOBJ_ADD);
4333 /* Setup first alias */
4334 sysfs_slab_alias(s, s->name);
4340 static void sysfs_slab_remove(struct kmem_cache *s)
4342 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4343 kobject_del(&s->kobj);
4344 kobject_put(&s->kobj);
4348 * Need to buffer aliases during bootup until sysfs becomes
4349 * available lest we loose that information.
4351 struct saved_alias {
4352 struct kmem_cache *s;
4354 struct saved_alias *next;
4357 static struct saved_alias *alias_list;
4359 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4361 struct saved_alias *al;
4363 if (slab_state == SYSFS) {
4365 * If we have a leftover link then remove it.
4367 sysfs_remove_link(&slab_kset->kobj, name);
4368 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4371 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4377 al->next = alias_list;
4382 static int __init slab_sysfs_init(void)
4384 struct kmem_cache *s;
4387 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4389 printk(KERN_ERR "Cannot register slab subsystem.\n");
4395 list_for_each_entry(s, &slab_caches, list) {
4396 err = sysfs_slab_add(s);
4398 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4399 " to sysfs\n", s->name);
4402 while (alias_list) {
4403 struct saved_alias *al = alias_list;
4405 alias_list = alias_list->next;
4406 err = sysfs_slab_alias(al->s, al->name);
4408 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4409 " %s to sysfs\n", s->name);
4417 __initcall(slab_sysfs_init);
4421 * The /proc/slabinfo ABI
4423 #ifdef CONFIG_SLABINFO
4424 static void print_slabinfo_header(struct seq_file *m)
4426 seq_puts(m, "slabinfo - version: 2.1\n");
4427 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4428 "<objperslab> <pagesperslab>");
4429 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4430 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4434 static void *s_start(struct seq_file *m, loff_t *pos)
4438 down_read(&slub_lock);
4440 print_slabinfo_header(m);
4442 return seq_list_start(&slab_caches, *pos);
4445 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4447 return seq_list_next(p, &slab_caches, pos);
4450 static void s_stop(struct seq_file *m, void *p)
4452 up_read(&slub_lock);
4455 static int s_show(struct seq_file *m, void *p)
4457 unsigned long nr_partials = 0;
4458 unsigned long nr_slabs = 0;
4459 unsigned long nr_inuse = 0;
4460 unsigned long nr_objs = 0;
4461 unsigned long nr_free = 0;
4462 struct kmem_cache *s;
4465 s = list_entry(p, struct kmem_cache, list);
4467 for_each_online_node(node) {
4468 struct kmem_cache_node *n = get_node(s, node);
4473 nr_partials += n->nr_partial;
4474 nr_slabs += atomic_long_read(&n->nr_slabs);
4475 nr_objs += atomic_long_read(&n->total_objects);
4476 nr_free += count_partial(n, count_free);
4479 nr_inuse = nr_objs - nr_free;
4481 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4482 nr_objs, s->size, oo_objects(s->oo),
4483 (1 << oo_order(s->oo)));
4484 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4485 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4491 static const struct seq_operations slabinfo_op = {
4498 static int slabinfo_open(struct inode *inode, struct file *file)
4500 return seq_open(file, &slabinfo_op);
4503 static const struct file_operations proc_slabinfo_operations = {
4504 .open = slabinfo_open,
4506 .llseek = seq_lseek,
4507 .release = seq_release,
4510 static int __init slab_proc_init(void)
4512 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4515 module_init(slab_proc_init);
4516 #endif /* CONFIG_SLABINFO */