2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
23 #include <linux/memory.h>
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
72 * freed then the slab will show up again on the partial lists.
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
80 * Overloading of page flags that are otherwise used for LRU management.
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
94 * freelist that allows lockless access to
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
100 * the fast path and disables lockless freelists.
103 #define FROZEN (1 << PG_active)
105 #ifdef CONFIG_SLUB_DEBUG
106 #define SLABDEBUG (1 << PG_error)
111 static inline int SlabFrozen(struct page *page)
113 return page->flags & FROZEN;
116 static inline void SetSlabFrozen(struct page *page)
118 page->flags |= FROZEN;
121 static inline void ClearSlabFrozen(struct page *page)
123 page->flags &= ~FROZEN;
126 static inline int SlabDebug(struct page *page)
128 return page->flags & SLABDEBUG;
131 static inline void SetSlabDebug(struct page *page)
133 page->flags |= SLABDEBUG;
136 static inline void ClearSlabDebug(struct page *page)
138 page->flags &= ~SLABDEBUG;
142 * Issues still to be resolved:
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
146 * - Variable sizing of the per node arrays
149 /* Enable to test recovery from slab corruption on boot */
150 #undef SLUB_RESILIENCY_TEST
155 * Small page size. Make sure that we do not fragment memory
157 #define DEFAULT_MAX_ORDER 1
158 #define DEFAULT_MIN_OBJECTS 4
163 * Large page machines are customarily able to handle larger
166 #define DEFAULT_MAX_ORDER 2
167 #define DEFAULT_MIN_OBJECTS 8
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
175 #define MIN_PARTIAL 5
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
182 #define MAX_PARTIAL 10
184 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
188 * Set of flags that will prevent slab merging
190 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
193 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
196 #ifndef ARCH_KMALLOC_MINALIGN
197 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
200 #ifndef ARCH_SLAB_MINALIGN
201 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
204 /* Internal SLUB flags */
205 #define __OBJECT_POISON 0x80000000 /* Poison object */
206 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
208 /* Not all arches define cache_line_size */
209 #ifndef cache_line_size
210 #define cache_line_size() L1_CACHE_BYTES
213 static int kmem_size = sizeof(struct kmem_cache);
216 static struct notifier_block slab_notifier;
220 DOWN, /* No slab functionality available */
221 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
222 UP, /* Everything works but does not show up in sysfs */
226 /* A list of all slab caches on the system */
227 static DECLARE_RWSEM(slub_lock);
228 static LIST_HEAD(slab_caches);
231 * Tracking user of a slab.
234 void *addr; /* Called from address */
235 int cpu; /* Was running on cpu */
236 int pid; /* Pid context */
237 unsigned long when; /* When did the operation occur */
240 enum track_item { TRACK_ALLOC, TRACK_FREE };
242 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
243 static int sysfs_slab_add(struct kmem_cache *);
244 static int sysfs_slab_alias(struct kmem_cache *, const char *);
245 static void sysfs_slab_remove(struct kmem_cache *);
248 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
249 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
251 static inline void sysfs_slab_remove(struct kmem_cache *s)
258 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
260 #ifdef CONFIG_SLUB_STATS
265 /********************************************************************
266 * Core slab cache functions
267 *******************************************************************/
269 int slab_is_available(void)
271 return slab_state >= UP;
274 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
277 return s->node[node];
279 return &s->local_node;
283 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
286 return s->cpu_slab[cpu];
292 /* Verify that a pointer has an address that is valid within a slab page */
293 static inline int check_valid_pointer(struct kmem_cache *s,
294 struct page *page, const void *object)
301 base = page_address(page);
302 if (object < base || object >= base + page->objects * s->size ||
303 (object - base) % s->size) {
311 * Slow version of get and set free pointer.
313 * This version requires touching the cache lines of kmem_cache which
314 * we avoid to do in the fast alloc free paths. There we obtain the offset
315 * from the page struct.
317 static inline void *get_freepointer(struct kmem_cache *s, void *object)
319 return *(void **)(object + s->offset);
322 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
324 *(void **)(object + s->offset) = fp;
327 /* Loop over all objects in a slab */
328 #define for_each_object(__p, __s, __addr, __objects) \
329 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
333 #define for_each_free_object(__p, __s, __free) \
334 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
336 /* Determine object index from a given position */
337 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
339 return (p - addr) / s->size;
342 static inline struct kmem_cache_order_objects oo_make(int order,
345 struct kmem_cache_order_objects x = {
346 (order << 16) + (PAGE_SIZE << order) / size
352 static inline int oo_order(struct kmem_cache_order_objects x)
357 static inline int oo_objects(struct kmem_cache_order_objects x)
359 return x.x & ((1 << 16) - 1);
362 #ifdef CONFIG_SLUB_DEBUG
366 #ifdef CONFIG_SLUB_DEBUG_ON
367 static int slub_debug = DEBUG_DEFAULT_FLAGS;
369 static int slub_debug;
372 static char *slub_debug_slabs;
377 static void print_section(char *text, u8 *addr, unsigned int length)
385 for (i = 0; i < length; i++) {
387 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
390 printk(KERN_CONT " %02x", addr[i]);
392 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
394 printk(KERN_CONT " %s\n", ascii);
401 printk(KERN_CONT " ");
405 printk(KERN_CONT " %s\n", ascii);
409 static struct track *get_track(struct kmem_cache *s, void *object,
410 enum track_item alloc)
415 p = object + s->offset + sizeof(void *);
417 p = object + s->inuse;
422 static void set_track(struct kmem_cache *s, void *object,
423 enum track_item alloc, void *addr)
428 p = object + s->offset + sizeof(void *);
430 p = object + s->inuse;
435 p->cpu = smp_processor_id();
436 p->pid = current ? current->pid : -1;
439 memset(p, 0, sizeof(struct track));
442 static void init_tracking(struct kmem_cache *s, void *object)
444 if (!(s->flags & SLAB_STORE_USER))
447 set_track(s, object, TRACK_FREE, NULL);
448 set_track(s, object, TRACK_ALLOC, NULL);
451 static void print_track(const char *s, struct track *t)
456 printk(KERN_ERR "INFO: %s in ", s);
457 __print_symbol("%s", (unsigned long)t->addr);
458 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
461 static void print_tracking(struct kmem_cache *s, void *object)
463 if (!(s->flags & SLAB_STORE_USER))
466 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
467 print_track("Freed", get_track(s, object, TRACK_FREE));
470 static void print_page_info(struct page *page)
472 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
473 page, page->objects, page->inuse, page->freelist, page->flags);
477 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
483 vsnprintf(buf, sizeof(buf), fmt, args);
485 printk(KERN_ERR "========================================"
486 "=====================================\n");
487 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
488 printk(KERN_ERR "----------------------------------------"
489 "-------------------------------------\n\n");
492 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
498 vsnprintf(buf, sizeof(buf), fmt, args);
500 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
503 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
505 unsigned int off; /* Offset of last byte */
506 u8 *addr = page_address(page);
508 print_tracking(s, p);
510 print_page_info(page);
512 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
513 p, p - addr, get_freepointer(s, p));
516 print_section("Bytes b4", p - 16, 16);
518 print_section("Object", p, min(s->objsize, 128));
520 if (s->flags & SLAB_RED_ZONE)
521 print_section("Redzone", p + s->objsize,
522 s->inuse - s->objsize);
525 off = s->offset + sizeof(void *);
529 if (s->flags & SLAB_STORE_USER)
530 off += 2 * sizeof(struct track);
533 /* Beginning of the filler is the free pointer */
534 print_section("Padding", p + off, s->size - off);
539 static void object_err(struct kmem_cache *s, struct page *page,
540 u8 *object, char *reason)
542 slab_bug(s, "%s", reason);
543 print_trailer(s, page, object);
546 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
552 vsnprintf(buf, sizeof(buf), fmt, args);
554 slab_bug(s, "%s", buf);
555 print_page_info(page);
559 static void init_object(struct kmem_cache *s, void *object, int active)
563 if (s->flags & __OBJECT_POISON) {
564 memset(p, POISON_FREE, s->objsize - 1);
565 p[s->objsize - 1] = POISON_END;
568 if (s->flags & SLAB_RED_ZONE)
569 memset(p + s->objsize,
570 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
571 s->inuse - s->objsize);
574 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
577 if (*start != (u8)value)
585 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
586 void *from, void *to)
588 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
589 memset(from, data, to - from);
592 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
593 u8 *object, char *what,
594 u8 *start, unsigned int value, unsigned int bytes)
599 fault = check_bytes(start, value, bytes);
604 while (end > fault && end[-1] == value)
607 slab_bug(s, "%s overwritten", what);
608 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
609 fault, end - 1, fault[0], value);
610 print_trailer(s, page, object);
612 restore_bytes(s, what, value, fault, end);
620 * Bytes of the object to be managed.
621 * If the freepointer may overlay the object then the free
622 * pointer is the first word of the object.
624 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
627 * object + s->objsize
628 * Padding to reach word boundary. This is also used for Redzoning.
629 * Padding is extended by another word if Redzoning is enabled and
632 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
633 * 0xcc (RED_ACTIVE) for objects in use.
636 * Meta data starts here.
638 * A. Free pointer (if we cannot overwrite object on free)
639 * B. Tracking data for SLAB_STORE_USER
640 * C. Padding to reach required alignment boundary or at mininum
641 * one word if debugging is on to be able to detect writes
642 * before the word boundary.
644 * Padding is done using 0x5a (POISON_INUSE)
647 * Nothing is used beyond s->size.
649 * If slabcaches are merged then the objsize and inuse boundaries are mostly
650 * ignored. And therefore no slab options that rely on these boundaries
651 * may be used with merged slabcaches.
654 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
656 unsigned long off = s->inuse; /* The end of info */
659 /* Freepointer is placed after the object. */
660 off += sizeof(void *);
662 if (s->flags & SLAB_STORE_USER)
663 /* We also have user information there */
664 off += 2 * sizeof(struct track);
669 return check_bytes_and_report(s, page, p, "Object padding",
670 p + off, POISON_INUSE, s->size - off);
673 /* Check the pad bytes at the end of a slab page */
674 static int slab_pad_check(struct kmem_cache *s, struct page *page)
682 if (!(s->flags & SLAB_POISON))
685 start = page_address(page);
686 length = (PAGE_SIZE << compound_order(page));
687 end = start + length;
688 remainder = length % s->size;
692 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
695 while (end > fault && end[-1] == POISON_INUSE)
698 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
699 print_section("Padding", end - remainder, remainder);
701 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
705 static int check_object(struct kmem_cache *s, struct page *page,
706 void *object, int active)
709 u8 *endobject = object + s->objsize;
711 if (s->flags & SLAB_RED_ZONE) {
713 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
715 if (!check_bytes_and_report(s, page, object, "Redzone",
716 endobject, red, s->inuse - s->objsize))
719 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
720 check_bytes_and_report(s, page, p, "Alignment padding",
721 endobject, POISON_INUSE, s->inuse - s->objsize);
725 if (s->flags & SLAB_POISON) {
726 if (!active && (s->flags & __OBJECT_POISON) &&
727 (!check_bytes_and_report(s, page, p, "Poison", p,
728 POISON_FREE, s->objsize - 1) ||
729 !check_bytes_and_report(s, page, p, "Poison",
730 p + s->objsize - 1, POISON_END, 1)))
733 * check_pad_bytes cleans up on its own.
735 check_pad_bytes(s, page, p);
738 if (!s->offset && active)
740 * Object and freepointer overlap. Cannot check
741 * freepointer while object is allocated.
745 /* Check free pointer validity */
746 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
747 object_err(s, page, p, "Freepointer corrupt");
749 * No choice but to zap it and thus loose the remainder
750 * of the free objects in this slab. May cause
751 * another error because the object count is now wrong.
753 set_freepointer(s, p, NULL);
759 static int check_slab(struct kmem_cache *s, struct page *page)
763 VM_BUG_ON(!irqs_disabled());
765 if (!PageSlab(page)) {
766 slab_err(s, page, "Not a valid slab page");
770 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
771 if (page->objects > maxobj) {
772 slab_err(s, page, "objects %u > max %u",
773 s->name, page->objects, maxobj);
776 if (page->inuse > page->objects) {
777 slab_err(s, page, "inuse %u > max %u",
778 s->name, page->inuse, page->objects);
781 /* Slab_pad_check fixes things up after itself */
782 slab_pad_check(s, page);
787 * Determine if a certain object on a page is on the freelist. Must hold the
788 * slab lock to guarantee that the chains are in a consistent state.
790 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
793 void *fp = page->freelist;
795 unsigned long max_objects;
797 while (fp && nr <= page->objects) {
800 if (!check_valid_pointer(s, page, fp)) {
802 object_err(s, page, object,
803 "Freechain corrupt");
804 set_freepointer(s, object, NULL);
807 slab_err(s, page, "Freepointer corrupt");
808 page->freelist = NULL;
809 page->inuse = page->objects;
810 slab_fix(s, "Freelist cleared");
816 fp = get_freepointer(s, object);
820 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
821 if (max_objects > 65535)
824 if (page->objects != max_objects) {
825 slab_err(s, page, "Wrong number of objects. Found %d but "
826 "should be %d", page->objects, max_objects);
827 page->objects = max_objects;
828 slab_fix(s, "Number of objects adjusted.");
830 if (page->inuse != page->objects - nr) {
831 slab_err(s, page, "Wrong object count. Counter is %d but "
832 "counted were %d", page->inuse, page->objects - nr);
833 page->inuse = page->objects - nr;
834 slab_fix(s, "Object count adjusted.");
836 return search == NULL;
839 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
841 if (s->flags & SLAB_TRACE) {
842 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
844 alloc ? "alloc" : "free",
849 print_section("Object", (void *)object, s->objsize);
856 * Tracking of fully allocated slabs for debugging purposes.
858 static void add_full(struct kmem_cache_node *n, struct page *page)
860 spin_lock(&n->list_lock);
861 list_add(&page->lru, &n->full);
862 spin_unlock(&n->list_lock);
865 static void remove_full(struct kmem_cache *s, struct page *page)
867 struct kmem_cache_node *n;
869 if (!(s->flags & SLAB_STORE_USER))
872 n = get_node(s, page_to_nid(page));
874 spin_lock(&n->list_lock);
875 list_del(&page->lru);
876 spin_unlock(&n->list_lock);
879 /* Tracking of the number of slabs for debugging purposes */
880 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
882 struct kmem_cache_node *n = get_node(s, node);
884 return atomic_long_read(&n->nr_slabs);
887 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
889 struct kmem_cache_node *n = get_node(s, node);
892 * May be called early in order to allocate a slab for the
893 * kmem_cache_node structure. Solve the chicken-egg
894 * dilemma by deferring the increment of the count during
895 * bootstrap (see early_kmem_cache_node_alloc).
897 if (!NUMA_BUILD || n) {
898 atomic_long_inc(&n->nr_slabs);
899 atomic_long_add(objects, &n->total_objects);
902 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
904 struct kmem_cache_node *n = get_node(s, node);
906 atomic_long_dec(&n->nr_slabs);
907 atomic_long_sub(objects, &n->total_objects);
910 /* Object debug checks for alloc/free paths */
911 static void setup_object_debug(struct kmem_cache *s, struct page *page,
914 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
917 init_object(s, object, 0);
918 init_tracking(s, object);
921 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
922 void *object, void *addr)
924 if (!check_slab(s, page))
927 if (!on_freelist(s, page, object)) {
928 object_err(s, page, object, "Object already allocated");
932 if (!check_valid_pointer(s, page, object)) {
933 object_err(s, page, object, "Freelist Pointer check fails");
937 if (!check_object(s, page, object, 0))
940 /* Success perform special debug activities for allocs */
941 if (s->flags & SLAB_STORE_USER)
942 set_track(s, object, TRACK_ALLOC, addr);
943 trace(s, page, object, 1);
944 init_object(s, object, 1);
948 if (PageSlab(page)) {
950 * If this is a slab page then lets do the best we can
951 * to avoid issues in the future. Marking all objects
952 * as used avoids touching the remaining objects.
954 slab_fix(s, "Marking all objects used");
955 page->inuse = page->objects;
956 page->freelist = NULL;
961 static int free_debug_processing(struct kmem_cache *s, struct page *page,
962 void *object, void *addr)
964 if (!check_slab(s, page))
967 if (!check_valid_pointer(s, page, object)) {
968 slab_err(s, page, "Invalid object pointer 0x%p", object);
972 if (on_freelist(s, page, object)) {
973 object_err(s, page, object, "Object already free");
977 if (!check_object(s, page, object, 1))
980 if (unlikely(s != page->slab)) {
981 if (!PageSlab(page)) {
982 slab_err(s, page, "Attempt to free object(0x%p) "
983 "outside of slab", object);
984 } else if (!page->slab) {
986 "SLUB <none>: no slab for object 0x%p.\n",
990 object_err(s, page, object,
991 "page slab pointer corrupt.");
995 /* Special debug activities for freeing objects */
996 if (!SlabFrozen(page) && !page->freelist)
997 remove_full(s, page);
998 if (s->flags & SLAB_STORE_USER)
999 set_track(s, object, TRACK_FREE, addr);
1000 trace(s, page, object, 0);
1001 init_object(s, object, 0);
1005 slab_fix(s, "Object at 0x%p not freed", object);
1009 static int __init setup_slub_debug(char *str)
1011 slub_debug = DEBUG_DEFAULT_FLAGS;
1012 if (*str++ != '=' || !*str)
1014 * No options specified. Switch on full debugging.
1020 * No options but restriction on slabs. This means full
1021 * debugging for slabs matching a pattern.
1028 * Switch off all debugging measures.
1033 * Determine which debug features should be switched on
1035 for (; *str && *str != ','; str++) {
1036 switch (tolower(*str)) {
1038 slub_debug |= SLAB_DEBUG_FREE;
1041 slub_debug |= SLAB_RED_ZONE;
1044 slub_debug |= SLAB_POISON;
1047 slub_debug |= SLAB_STORE_USER;
1050 slub_debug |= SLAB_TRACE;
1053 printk(KERN_ERR "slub_debug option '%c' "
1054 "unknown. skipped\n", *str);
1060 slub_debug_slabs = str + 1;
1065 __setup("slub_debug", setup_slub_debug);
1067 static unsigned long kmem_cache_flags(unsigned long objsize,
1068 unsigned long flags, const char *name,
1069 void (*ctor)(struct kmem_cache *, void *))
1072 * Enable debugging if selected on the kernel commandline.
1074 if (slub_debug && (!slub_debug_slabs ||
1075 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1076 flags |= slub_debug;
1081 static inline void setup_object_debug(struct kmem_cache *s,
1082 struct page *page, void *object) {}
1084 static inline int alloc_debug_processing(struct kmem_cache *s,
1085 struct page *page, void *object, void *addr) { return 0; }
1087 static inline int free_debug_processing(struct kmem_cache *s,
1088 struct page *page, void *object, void *addr) { return 0; }
1090 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1092 static inline int check_object(struct kmem_cache *s, struct page *page,
1093 void *object, int active) { return 1; }
1094 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1095 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1096 unsigned long flags, const char *name,
1097 void (*ctor)(struct kmem_cache *, void *))
1101 #define slub_debug 0
1103 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1105 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1107 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1112 * Slab allocation and freeing
1114 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1115 struct kmem_cache_order_objects oo)
1117 int order = oo_order(oo);
1120 return alloc_pages(flags, order);
1122 return alloc_pages_node(node, flags, order);
1125 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1128 struct kmem_cache_order_objects oo = s->oo;
1130 flags |= s->allocflags;
1132 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1134 if (unlikely(!page)) {
1137 * Allocation may have failed due to fragmentation.
1138 * Try a lower order alloc if possible
1140 page = alloc_slab_page(flags, node, oo);
1144 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1146 page->objects = oo_objects(oo);
1147 mod_zone_page_state(page_zone(page),
1148 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1149 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1155 static void setup_object(struct kmem_cache *s, struct page *page,
1158 setup_object_debug(s, page, object);
1159 if (unlikely(s->ctor))
1163 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1170 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1172 page = allocate_slab(s,
1173 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1177 inc_slabs_node(s, page_to_nid(page), page->objects);
1179 page->flags |= 1 << PG_slab;
1180 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1181 SLAB_STORE_USER | SLAB_TRACE))
1184 start = page_address(page);
1186 if (unlikely(s->flags & SLAB_POISON))
1187 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1190 for_each_object(p, s, start, page->objects) {
1191 setup_object(s, page, last);
1192 set_freepointer(s, last, p);
1195 setup_object(s, page, last);
1196 set_freepointer(s, last, NULL);
1198 page->freelist = start;
1204 static void __free_slab(struct kmem_cache *s, struct page *page)
1206 int order = compound_order(page);
1207 int pages = 1 << order;
1209 if (unlikely(SlabDebug(page))) {
1212 slab_pad_check(s, page);
1213 for_each_object(p, s, page_address(page),
1215 check_object(s, page, p, 0);
1216 ClearSlabDebug(page);
1219 mod_zone_page_state(page_zone(page),
1220 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1221 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1224 __ClearPageSlab(page);
1225 reset_page_mapcount(page);
1226 __free_pages(page, order);
1229 static void rcu_free_slab(struct rcu_head *h)
1233 page = container_of((struct list_head *)h, struct page, lru);
1234 __free_slab(page->slab, page);
1237 static void free_slab(struct kmem_cache *s, struct page *page)
1239 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1241 * RCU free overloads the RCU head over the LRU
1243 struct rcu_head *head = (void *)&page->lru;
1245 call_rcu(head, rcu_free_slab);
1247 __free_slab(s, page);
1250 static void discard_slab(struct kmem_cache *s, struct page *page)
1252 dec_slabs_node(s, page_to_nid(page), page->objects);
1257 * Per slab locking using the pagelock
1259 static __always_inline void slab_lock(struct page *page)
1261 bit_spin_lock(PG_locked, &page->flags);
1264 static __always_inline void slab_unlock(struct page *page)
1266 __bit_spin_unlock(PG_locked, &page->flags);
1269 static __always_inline int slab_trylock(struct page *page)
1273 rc = bit_spin_trylock(PG_locked, &page->flags);
1278 * Management of partially allocated slabs
1280 static void add_partial(struct kmem_cache_node *n,
1281 struct page *page, int tail)
1283 spin_lock(&n->list_lock);
1286 list_add_tail(&page->lru, &n->partial);
1288 list_add(&page->lru, &n->partial);
1289 spin_unlock(&n->list_lock);
1292 static void remove_partial(struct kmem_cache *s,
1295 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1297 spin_lock(&n->list_lock);
1298 list_del(&page->lru);
1300 spin_unlock(&n->list_lock);
1304 * Lock slab and remove from the partial list.
1306 * Must hold list_lock.
1308 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1310 if (slab_trylock(page)) {
1311 list_del(&page->lru);
1313 SetSlabFrozen(page);
1320 * Try to allocate a partial slab from a specific node.
1322 static struct page *get_partial_node(struct kmem_cache_node *n)
1327 * Racy check. If we mistakenly see no partial slabs then we
1328 * just allocate an empty slab. If we mistakenly try to get a
1329 * partial slab and there is none available then get_partials()
1332 if (!n || !n->nr_partial)
1335 spin_lock(&n->list_lock);
1336 list_for_each_entry(page, &n->partial, lru)
1337 if (lock_and_freeze_slab(n, page))
1341 spin_unlock(&n->list_lock);
1346 * Get a page from somewhere. Search in increasing NUMA distances.
1348 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1351 struct zonelist *zonelist;
1356 * The defrag ratio allows a configuration of the tradeoffs between
1357 * inter node defragmentation and node local allocations. A lower
1358 * defrag_ratio increases the tendency to do local allocations
1359 * instead of attempting to obtain partial slabs from other nodes.
1361 * If the defrag_ratio is set to 0 then kmalloc() always
1362 * returns node local objects. If the ratio is higher then kmalloc()
1363 * may return off node objects because partial slabs are obtained
1364 * from other nodes and filled up.
1366 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1367 * defrag_ratio = 1000) then every (well almost) allocation will
1368 * first attempt to defrag slab caches on other nodes. This means
1369 * scanning over all nodes to look for partial slabs which may be
1370 * expensive if we do it every time we are trying to find a slab
1371 * with available objects.
1373 if (!s->remote_node_defrag_ratio ||
1374 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1377 zonelist = &NODE_DATA(
1378 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
1379 for (z = zonelist->zones; *z; z++) {
1380 struct kmem_cache_node *n;
1382 n = get_node(s, zone_to_nid(*z));
1384 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1385 n->nr_partial > MIN_PARTIAL) {
1386 page = get_partial_node(n);
1396 * Get a partial page, lock it and return it.
1398 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1401 int searchnode = (node == -1) ? numa_node_id() : node;
1403 page = get_partial_node(get_node(s, searchnode));
1404 if (page || (flags & __GFP_THISNODE))
1407 return get_any_partial(s, flags);
1411 * Move a page back to the lists.
1413 * Must be called with the slab lock held.
1415 * On exit the slab lock will have been dropped.
1417 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1419 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1420 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1422 ClearSlabFrozen(page);
1425 if (page->freelist) {
1426 add_partial(n, page, tail);
1427 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1429 stat(c, DEACTIVATE_FULL);
1430 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1435 stat(c, DEACTIVATE_EMPTY);
1436 if (n->nr_partial < MIN_PARTIAL) {
1438 * Adding an empty slab to the partial slabs in order
1439 * to avoid page allocator overhead. This slab needs
1440 * to come after the other slabs with objects in
1441 * so that the others get filled first. That way the
1442 * size of the partial list stays small.
1444 * kmem_cache_shrink can reclaim any empty slabs from the
1447 add_partial(n, page, 1);
1451 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1452 discard_slab(s, page);
1458 * Remove the cpu slab
1460 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1462 struct page *page = c->page;
1466 stat(c, DEACTIVATE_REMOTE_FREES);
1468 * Merge cpu freelist into slab freelist. Typically we get here
1469 * because both freelists are empty. So this is unlikely
1472 while (unlikely(c->freelist)) {
1475 tail = 0; /* Hot objects. Put the slab first */
1477 /* Retrieve object from cpu_freelist */
1478 object = c->freelist;
1479 c->freelist = c->freelist[c->offset];
1481 /* And put onto the regular freelist */
1482 object[c->offset] = page->freelist;
1483 page->freelist = object;
1487 unfreeze_slab(s, page, tail);
1490 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1492 stat(c, CPUSLAB_FLUSH);
1494 deactivate_slab(s, c);
1500 * Called from IPI handler with interrupts disabled.
1502 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1504 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1506 if (likely(c && c->page))
1510 static void flush_cpu_slab(void *d)
1512 struct kmem_cache *s = d;
1514 __flush_cpu_slab(s, smp_processor_id());
1517 static void flush_all(struct kmem_cache *s)
1520 on_each_cpu(flush_cpu_slab, s, 1, 1);
1522 unsigned long flags;
1524 local_irq_save(flags);
1526 local_irq_restore(flags);
1531 * Check if the objects in a per cpu structure fit numa
1532 * locality expectations.
1534 static inline int node_match(struct kmem_cache_cpu *c, int node)
1537 if (node != -1 && c->node != node)
1544 * Slow path. The lockless freelist is empty or we need to perform
1547 * Interrupts are disabled.
1549 * Processing is still very fast if new objects have been freed to the
1550 * regular freelist. In that case we simply take over the regular freelist
1551 * as the lockless freelist and zap the regular freelist.
1553 * If that is not working then we fall back to the partial lists. We take the
1554 * first element of the freelist as the object to allocate now and move the
1555 * rest of the freelist to the lockless freelist.
1557 * And if we were unable to get a new slab from the partial slab lists then
1558 * we need to allocate a new slab. This is the slowest path since it involves
1559 * a call to the page allocator and the setup of a new slab.
1561 static void *__slab_alloc(struct kmem_cache *s,
1562 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1567 /* We handle __GFP_ZERO in the caller */
1568 gfpflags &= ~__GFP_ZERO;
1574 if (unlikely(!node_match(c, node)))
1577 stat(c, ALLOC_REFILL);
1580 object = c->page->freelist;
1581 if (unlikely(!object))
1583 if (unlikely(SlabDebug(c->page)))
1586 c->freelist = object[c->offset];
1587 c->page->inuse = c->page->objects;
1588 c->page->freelist = NULL;
1589 c->node = page_to_nid(c->page);
1591 slab_unlock(c->page);
1592 stat(c, ALLOC_SLOWPATH);
1596 deactivate_slab(s, c);
1599 new = get_partial(s, gfpflags, node);
1602 stat(c, ALLOC_FROM_PARTIAL);
1606 if (gfpflags & __GFP_WAIT)
1609 new = new_slab(s, gfpflags, node);
1611 if (gfpflags & __GFP_WAIT)
1612 local_irq_disable();
1615 c = get_cpu_slab(s, smp_processor_id());
1616 stat(c, ALLOC_SLAB);
1626 if (!alloc_debug_processing(s, c->page, object, addr))
1630 c->page->freelist = object[c->offset];
1636 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1637 * have the fastpath folded into their functions. So no function call
1638 * overhead for requests that can be satisfied on the fastpath.
1640 * The fastpath works by first checking if the lockless freelist can be used.
1641 * If not then __slab_alloc is called for slow processing.
1643 * Otherwise we can simply pick the next object from the lockless free list.
1645 static __always_inline void *slab_alloc(struct kmem_cache *s,
1646 gfp_t gfpflags, int node, void *addr)
1649 struct kmem_cache_cpu *c;
1650 unsigned long flags;
1652 local_irq_save(flags);
1653 c = get_cpu_slab(s, smp_processor_id());
1654 if (unlikely(!c->freelist || !node_match(c, node)))
1656 object = __slab_alloc(s, gfpflags, node, addr, c);
1659 object = c->freelist;
1660 c->freelist = object[c->offset];
1661 stat(c, ALLOC_FASTPATH);
1663 local_irq_restore(flags);
1665 if (unlikely((gfpflags & __GFP_ZERO) && object))
1666 memset(object, 0, c->objsize);
1671 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1673 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1675 EXPORT_SYMBOL(kmem_cache_alloc);
1678 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1680 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1682 EXPORT_SYMBOL(kmem_cache_alloc_node);
1686 * Slow patch handling. This may still be called frequently since objects
1687 * have a longer lifetime than the cpu slabs in most processing loads.
1689 * So we still attempt to reduce cache line usage. Just take the slab
1690 * lock and free the item. If there is no additional partial page
1691 * handling required then we can return immediately.
1693 static void __slab_free(struct kmem_cache *s, struct page *page,
1694 void *x, void *addr, unsigned int offset)
1697 void **object = (void *)x;
1698 struct kmem_cache_cpu *c;
1700 c = get_cpu_slab(s, raw_smp_processor_id());
1701 stat(c, FREE_SLOWPATH);
1704 if (unlikely(SlabDebug(page)))
1708 prior = object[offset] = page->freelist;
1709 page->freelist = object;
1712 if (unlikely(SlabFrozen(page))) {
1713 stat(c, FREE_FROZEN);
1717 if (unlikely(!page->inuse))
1721 * Objects left in the slab. If it was not on the partial list before
1724 if (unlikely(!prior)) {
1725 add_partial(get_node(s, page_to_nid(page)), page, 1);
1726 stat(c, FREE_ADD_PARTIAL);
1736 * Slab still on the partial list.
1738 remove_partial(s, page);
1739 stat(c, FREE_REMOVE_PARTIAL);
1743 discard_slab(s, page);
1747 if (!free_debug_processing(s, page, x, addr))
1753 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1754 * can perform fastpath freeing without additional function calls.
1756 * The fastpath is only possible if we are freeing to the current cpu slab
1757 * of this processor. This typically the case if we have just allocated
1760 * If fastpath is not possible then fall back to __slab_free where we deal
1761 * with all sorts of special processing.
1763 static __always_inline void slab_free(struct kmem_cache *s,
1764 struct page *page, void *x, void *addr)
1766 void **object = (void *)x;
1767 struct kmem_cache_cpu *c;
1768 unsigned long flags;
1770 local_irq_save(flags);
1771 c = get_cpu_slab(s, smp_processor_id());
1772 debug_check_no_locks_freed(object, c->objsize);
1773 if (likely(page == c->page && c->node >= 0)) {
1774 object[c->offset] = c->freelist;
1775 c->freelist = object;
1776 stat(c, FREE_FASTPATH);
1778 __slab_free(s, page, x, addr, c->offset);
1780 local_irq_restore(flags);
1783 void kmem_cache_free(struct kmem_cache *s, void *x)
1787 page = virt_to_head_page(x);
1789 slab_free(s, page, x, __builtin_return_address(0));
1791 EXPORT_SYMBOL(kmem_cache_free);
1793 /* Figure out on which slab object the object resides */
1794 static struct page *get_object_page(const void *x)
1796 struct page *page = virt_to_head_page(x);
1798 if (!PageSlab(page))
1805 * Object placement in a slab is made very easy because we always start at
1806 * offset 0. If we tune the size of the object to the alignment then we can
1807 * get the required alignment by putting one properly sized object after
1810 * Notice that the allocation order determines the sizes of the per cpu
1811 * caches. Each processor has always one slab available for allocations.
1812 * Increasing the allocation order reduces the number of times that slabs
1813 * must be moved on and off the partial lists and is therefore a factor in
1818 * Mininum / Maximum order of slab pages. This influences locking overhead
1819 * and slab fragmentation. A higher order reduces the number of partial slabs
1820 * and increases the number of allocations possible without having to
1821 * take the list_lock.
1823 static int slub_min_order;
1824 static int slub_max_order = DEFAULT_MAX_ORDER;
1825 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1828 * Merge control. If this is set then no merging of slab caches will occur.
1829 * (Could be removed. This was introduced to pacify the merge skeptics.)
1831 static int slub_nomerge;
1834 * Calculate the order of allocation given an slab object size.
1836 * The order of allocation has significant impact on performance and other
1837 * system components. Generally order 0 allocations should be preferred since
1838 * order 0 does not cause fragmentation in the page allocator. Larger objects
1839 * be problematic to put into order 0 slabs because there may be too much
1840 * unused space left. We go to a higher order if more than 1/8th of the slab
1843 * In order to reach satisfactory performance we must ensure that a minimum
1844 * number of objects is in one slab. Otherwise we may generate too much
1845 * activity on the partial lists which requires taking the list_lock. This is
1846 * less a concern for large slabs though which are rarely used.
1848 * slub_max_order specifies the order where we begin to stop considering the
1849 * number of objects in a slab as critical. If we reach slub_max_order then
1850 * we try to keep the page order as low as possible. So we accept more waste
1851 * of space in favor of a small page order.
1853 * Higher order allocations also allow the placement of more objects in a
1854 * slab and thereby reduce object handling overhead. If the user has
1855 * requested a higher mininum order then we start with that one instead of
1856 * the smallest order which will fit the object.
1858 static inline int slab_order(int size, int min_objects,
1859 int max_order, int fract_leftover)
1863 int min_order = slub_min_order;
1865 if ((PAGE_SIZE << min_order) / size > 65535)
1866 return get_order(size * 65535) - 1;
1868 for (order = max(min_order,
1869 fls(min_objects * size - 1) - PAGE_SHIFT);
1870 order <= max_order; order++) {
1872 unsigned long slab_size = PAGE_SIZE << order;
1874 if (slab_size < min_objects * size)
1877 rem = slab_size % size;
1879 if (rem <= slab_size / fract_leftover)
1887 static inline int calculate_order(int size)
1894 * Attempt to find best configuration for a slab. This
1895 * works by first attempting to generate a layout with
1896 * the best configuration and backing off gradually.
1898 * First we reduce the acceptable waste in a slab. Then
1899 * we reduce the minimum objects required in a slab.
1901 min_objects = slub_min_objects;
1902 while (min_objects > 1) {
1904 while (fraction >= 4) {
1905 order = slab_order(size, min_objects,
1906 slub_max_order, fraction);
1907 if (order <= slub_max_order)
1915 * We were unable to place multiple objects in a slab. Now
1916 * lets see if we can place a single object there.
1918 order = slab_order(size, 1, slub_max_order, 1);
1919 if (order <= slub_max_order)
1923 * Doh this slab cannot be placed using slub_max_order.
1925 order = slab_order(size, 1, MAX_ORDER, 1);
1926 if (order <= MAX_ORDER)
1932 * Figure out what the alignment of the objects will be.
1934 static unsigned long calculate_alignment(unsigned long flags,
1935 unsigned long align, unsigned long size)
1938 * If the user wants hardware cache aligned objects then follow that
1939 * suggestion if the object is sufficiently large.
1941 * The hardware cache alignment cannot override the specified
1942 * alignment though. If that is greater then use it.
1944 if (flags & SLAB_HWCACHE_ALIGN) {
1945 unsigned long ralign = cache_line_size();
1946 while (size <= ralign / 2)
1948 align = max(align, ralign);
1951 if (align < ARCH_SLAB_MINALIGN)
1952 align = ARCH_SLAB_MINALIGN;
1954 return ALIGN(align, sizeof(void *));
1957 static void init_kmem_cache_cpu(struct kmem_cache *s,
1958 struct kmem_cache_cpu *c)
1963 c->offset = s->offset / sizeof(void *);
1964 c->objsize = s->objsize;
1965 #ifdef CONFIG_SLUB_STATS
1966 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1970 static void init_kmem_cache_node(struct kmem_cache_node *n)
1973 spin_lock_init(&n->list_lock);
1974 INIT_LIST_HEAD(&n->partial);
1975 #ifdef CONFIG_SLUB_DEBUG
1976 atomic_long_set(&n->nr_slabs, 0);
1977 INIT_LIST_HEAD(&n->full);
1983 * Per cpu array for per cpu structures.
1985 * The per cpu array places all kmem_cache_cpu structures from one processor
1986 * close together meaning that it becomes possible that multiple per cpu
1987 * structures are contained in one cacheline. This may be particularly
1988 * beneficial for the kmalloc caches.
1990 * A desktop system typically has around 60-80 slabs. With 100 here we are
1991 * likely able to get per cpu structures for all caches from the array defined
1992 * here. We must be able to cover all kmalloc caches during bootstrap.
1994 * If the per cpu array is exhausted then fall back to kmalloc
1995 * of individual cachelines. No sharing is possible then.
1997 #define NR_KMEM_CACHE_CPU 100
1999 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2000 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2002 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2003 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
2005 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2006 int cpu, gfp_t flags)
2008 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2011 per_cpu(kmem_cache_cpu_free, cpu) =
2012 (void *)c->freelist;
2014 /* Table overflow: So allocate ourselves */
2016 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2017 flags, cpu_to_node(cpu));
2022 init_kmem_cache_cpu(s, c);
2026 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2028 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2029 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2033 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2034 per_cpu(kmem_cache_cpu_free, cpu) = c;
2037 static void free_kmem_cache_cpus(struct kmem_cache *s)
2041 for_each_online_cpu(cpu) {
2042 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2045 s->cpu_slab[cpu] = NULL;
2046 free_kmem_cache_cpu(c, cpu);
2051 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2055 for_each_online_cpu(cpu) {
2056 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2061 c = alloc_kmem_cache_cpu(s, cpu, flags);
2063 free_kmem_cache_cpus(s);
2066 s->cpu_slab[cpu] = c;
2072 * Initialize the per cpu array.
2074 static void init_alloc_cpu_cpu(int cpu)
2078 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2081 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2082 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2084 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2087 static void __init init_alloc_cpu(void)
2091 for_each_online_cpu(cpu)
2092 init_alloc_cpu_cpu(cpu);
2096 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2097 static inline void init_alloc_cpu(void) {}
2099 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2101 init_kmem_cache_cpu(s, &s->cpu_slab);
2108 * No kmalloc_node yet so do it by hand. We know that this is the first
2109 * slab on the node for this slabcache. There are no concurrent accesses
2112 * Note that this function only works on the kmalloc_node_cache
2113 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2114 * memory on a fresh node that has no slab structures yet.
2116 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2120 struct kmem_cache_node *n;
2121 unsigned long flags;
2123 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2125 page = new_slab(kmalloc_caches, gfpflags, node);
2128 if (page_to_nid(page) != node) {
2129 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2131 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2132 "in order to be able to continue\n");
2137 page->freelist = get_freepointer(kmalloc_caches, n);
2139 kmalloc_caches->node[node] = n;
2140 #ifdef CONFIG_SLUB_DEBUG
2141 init_object(kmalloc_caches, n, 1);
2142 init_tracking(kmalloc_caches, n);
2144 init_kmem_cache_node(n);
2145 inc_slabs_node(kmalloc_caches, node, page->objects);
2148 * lockdep requires consistent irq usage for each lock
2149 * so even though there cannot be a race this early in
2150 * the boot sequence, we still disable irqs.
2152 local_irq_save(flags);
2153 add_partial(n, page, 0);
2154 local_irq_restore(flags);
2158 static void free_kmem_cache_nodes(struct kmem_cache *s)
2162 for_each_node_state(node, N_NORMAL_MEMORY) {
2163 struct kmem_cache_node *n = s->node[node];
2164 if (n && n != &s->local_node)
2165 kmem_cache_free(kmalloc_caches, n);
2166 s->node[node] = NULL;
2170 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2175 if (slab_state >= UP)
2176 local_node = page_to_nid(virt_to_page(s));
2180 for_each_node_state(node, N_NORMAL_MEMORY) {
2181 struct kmem_cache_node *n;
2183 if (local_node == node)
2186 if (slab_state == DOWN) {
2187 n = early_kmem_cache_node_alloc(gfpflags,
2191 n = kmem_cache_alloc_node(kmalloc_caches,
2195 free_kmem_cache_nodes(s);
2201 init_kmem_cache_node(n);
2206 static void free_kmem_cache_nodes(struct kmem_cache *s)
2210 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2212 init_kmem_cache_node(&s->local_node);
2218 * calculate_sizes() determines the order and the distribution of data within
2221 static int calculate_sizes(struct kmem_cache *s)
2223 unsigned long flags = s->flags;
2224 unsigned long size = s->objsize;
2225 unsigned long align = s->align;
2229 * Round up object size to the next word boundary. We can only
2230 * place the free pointer at word boundaries and this determines
2231 * the possible location of the free pointer.
2233 size = ALIGN(size, sizeof(void *));
2235 #ifdef CONFIG_SLUB_DEBUG
2237 * Determine if we can poison the object itself. If the user of
2238 * the slab may touch the object after free or before allocation
2239 * then we should never poison the object itself.
2241 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2243 s->flags |= __OBJECT_POISON;
2245 s->flags &= ~__OBJECT_POISON;
2249 * If we are Redzoning then check if there is some space between the
2250 * end of the object and the free pointer. If not then add an
2251 * additional word to have some bytes to store Redzone information.
2253 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2254 size += sizeof(void *);
2258 * With that we have determined the number of bytes in actual use
2259 * by the object. This is the potential offset to the free pointer.
2263 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2266 * Relocate free pointer after the object if it is not
2267 * permitted to overwrite the first word of the object on
2270 * This is the case if we do RCU, have a constructor or
2271 * destructor or are poisoning the objects.
2274 size += sizeof(void *);
2277 #ifdef CONFIG_SLUB_DEBUG
2278 if (flags & SLAB_STORE_USER)
2280 * Need to store information about allocs and frees after
2283 size += 2 * sizeof(struct track);
2285 if (flags & SLAB_RED_ZONE)
2287 * Add some empty padding so that we can catch
2288 * overwrites from earlier objects rather than let
2289 * tracking information or the free pointer be
2290 * corrupted if an user writes before the start
2293 size += sizeof(void *);
2297 * Determine the alignment based on various parameters that the
2298 * user specified and the dynamic determination of cache line size
2301 align = calculate_alignment(flags, align, s->objsize);
2304 * SLUB stores one object immediately after another beginning from
2305 * offset 0. In order to align the objects we have to simply size
2306 * each object to conform to the alignment.
2308 size = ALIGN(size, align);
2310 order = calculate_order(size);
2317 s->allocflags |= __GFP_COMP;
2319 if (s->flags & SLAB_CACHE_DMA)
2320 s->allocflags |= SLUB_DMA;
2322 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2323 s->allocflags |= __GFP_RECLAIMABLE;
2326 * Determine the number of objects per slab
2328 s->oo = oo_make(order, size);
2329 s->min = oo_make(get_order(size), size);
2330 if (oo_objects(s->oo) > oo_objects(s->max))
2333 return !!oo_objects(s->oo);
2337 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2338 const char *name, size_t size,
2339 size_t align, unsigned long flags,
2340 void (*ctor)(struct kmem_cache *, void *))
2342 memset(s, 0, kmem_size);
2347 s->flags = kmem_cache_flags(size, flags, name, ctor);
2349 if (!calculate_sizes(s))
2354 s->remote_node_defrag_ratio = 100;
2356 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2359 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2361 free_kmem_cache_nodes(s);
2363 if (flags & SLAB_PANIC)
2364 panic("Cannot create slab %s size=%lu realsize=%u "
2365 "order=%u offset=%u flags=%lx\n",
2366 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2372 * Check if a given pointer is valid
2374 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2378 page = get_object_page(object);
2380 if (!page || s != page->slab)
2381 /* No slab or wrong slab */
2384 if (!check_valid_pointer(s, page, object))
2388 * We could also check if the object is on the slabs freelist.
2389 * But this would be too expensive and it seems that the main
2390 * purpose of kmem_ptr_valid() is to check if the object belongs
2391 * to a certain slab.
2395 EXPORT_SYMBOL(kmem_ptr_validate);
2398 * Determine the size of a slab object
2400 unsigned int kmem_cache_size(struct kmem_cache *s)
2404 EXPORT_SYMBOL(kmem_cache_size);
2406 const char *kmem_cache_name(struct kmem_cache *s)
2410 EXPORT_SYMBOL(kmem_cache_name);
2412 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2415 #ifdef CONFIG_SLUB_DEBUG
2416 void *addr = page_address(page);
2418 DECLARE_BITMAP(map, page->objects);
2420 bitmap_zero(map, page->objects);
2421 slab_err(s, page, "%s", text);
2423 for_each_free_object(p, s, page->freelist)
2424 set_bit(slab_index(p, s, addr), map);
2426 for_each_object(p, s, addr, page->objects) {
2428 if (!test_bit(slab_index(p, s, addr), map)) {
2429 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2431 print_tracking(s, p);
2439 * Attempt to free all partial slabs on a node.
2441 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2443 unsigned long flags;
2444 struct page *page, *h;
2446 spin_lock_irqsave(&n->list_lock, flags);
2447 list_for_each_entry_safe(page, h, &n->partial, lru) {
2449 list_del(&page->lru);
2450 discard_slab(s, page);
2453 list_slab_objects(s, page,
2454 "Objects remaining on kmem_cache_close()");
2457 spin_unlock_irqrestore(&n->list_lock, flags);
2461 * Release all resources used by a slab cache.
2463 static inline int kmem_cache_close(struct kmem_cache *s)
2469 /* Attempt to free all objects */
2470 free_kmem_cache_cpus(s);
2471 for_each_node_state(node, N_NORMAL_MEMORY) {
2472 struct kmem_cache_node *n = get_node(s, node);
2475 if (n->nr_partial || slabs_node(s, node))
2478 free_kmem_cache_nodes(s);
2483 * Close a cache and release the kmem_cache structure
2484 * (must be used for caches created using kmem_cache_create)
2486 void kmem_cache_destroy(struct kmem_cache *s)
2488 down_write(&slub_lock);
2492 up_write(&slub_lock);
2493 if (kmem_cache_close(s)) {
2494 printk(KERN_ERR "SLUB %s: %s called for cache that "
2495 "still has objects.\n", s->name, __func__);
2498 sysfs_slab_remove(s);
2500 up_write(&slub_lock);
2502 EXPORT_SYMBOL(kmem_cache_destroy);
2504 /********************************************************************
2506 *******************************************************************/
2508 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2509 EXPORT_SYMBOL(kmalloc_caches);
2511 static int __init setup_slub_min_order(char *str)
2513 get_option(&str, &slub_min_order);
2518 __setup("slub_min_order=", setup_slub_min_order);
2520 static int __init setup_slub_max_order(char *str)
2522 get_option(&str, &slub_max_order);
2527 __setup("slub_max_order=", setup_slub_max_order);
2529 static int __init setup_slub_min_objects(char *str)
2531 get_option(&str, &slub_min_objects);
2536 __setup("slub_min_objects=", setup_slub_min_objects);
2538 static int __init setup_slub_nomerge(char *str)
2544 __setup("slub_nomerge", setup_slub_nomerge);
2546 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2547 const char *name, int size, gfp_t gfp_flags)
2549 unsigned int flags = 0;
2551 if (gfp_flags & SLUB_DMA)
2552 flags = SLAB_CACHE_DMA;
2554 down_write(&slub_lock);
2555 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2559 list_add(&s->list, &slab_caches);
2560 up_write(&slub_lock);
2561 if (sysfs_slab_add(s))
2566 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2569 #ifdef CONFIG_ZONE_DMA
2570 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2572 static void sysfs_add_func(struct work_struct *w)
2574 struct kmem_cache *s;
2576 down_write(&slub_lock);
2577 list_for_each_entry(s, &slab_caches, list) {
2578 if (s->flags & __SYSFS_ADD_DEFERRED) {
2579 s->flags &= ~__SYSFS_ADD_DEFERRED;
2583 up_write(&slub_lock);
2586 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2588 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2590 struct kmem_cache *s;
2594 s = kmalloc_caches_dma[index];
2598 /* Dynamically create dma cache */
2599 if (flags & __GFP_WAIT)
2600 down_write(&slub_lock);
2602 if (!down_write_trylock(&slub_lock))
2606 if (kmalloc_caches_dma[index])
2609 realsize = kmalloc_caches[index].objsize;
2610 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2611 (unsigned int)realsize);
2612 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2614 if (!s || !text || !kmem_cache_open(s, flags, text,
2615 realsize, ARCH_KMALLOC_MINALIGN,
2616 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2622 list_add(&s->list, &slab_caches);
2623 kmalloc_caches_dma[index] = s;
2625 schedule_work(&sysfs_add_work);
2628 up_write(&slub_lock);
2630 return kmalloc_caches_dma[index];
2635 * Conversion table for small slabs sizes / 8 to the index in the
2636 * kmalloc array. This is necessary for slabs < 192 since we have non power
2637 * of two cache sizes there. The size of larger slabs can be determined using
2640 static s8 size_index[24] = {
2667 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2673 return ZERO_SIZE_PTR;
2675 index = size_index[(size - 1) / 8];
2677 index = fls(size - 1);
2679 #ifdef CONFIG_ZONE_DMA
2680 if (unlikely((flags & SLUB_DMA)))
2681 return dma_kmalloc_cache(index, flags);
2684 return &kmalloc_caches[index];
2687 void *__kmalloc(size_t size, gfp_t flags)
2689 struct kmem_cache *s;
2691 if (unlikely(size > PAGE_SIZE))
2692 return kmalloc_large(size, flags);
2694 s = get_slab(size, flags);
2696 if (unlikely(ZERO_OR_NULL_PTR(s)))
2699 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2701 EXPORT_SYMBOL(__kmalloc);
2703 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2705 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2709 return page_address(page);
2715 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2717 struct kmem_cache *s;
2719 if (unlikely(size > PAGE_SIZE))
2720 return kmalloc_large_node(size, flags, node);
2722 s = get_slab(size, flags);
2724 if (unlikely(ZERO_OR_NULL_PTR(s)))
2727 return slab_alloc(s, flags, node, __builtin_return_address(0));
2729 EXPORT_SYMBOL(__kmalloc_node);
2732 size_t ksize(const void *object)
2735 struct kmem_cache *s;
2737 if (unlikely(object == ZERO_SIZE_PTR))
2740 page = virt_to_head_page(object);
2742 if (unlikely(!PageSlab(page)))
2743 return PAGE_SIZE << compound_order(page);
2747 #ifdef CONFIG_SLUB_DEBUG
2749 * Debugging requires use of the padding between object
2750 * and whatever may come after it.
2752 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2757 * If we have the need to store the freelist pointer
2758 * back there or track user information then we can
2759 * only use the space before that information.
2761 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2764 * Else we can use all the padding etc for the allocation
2768 EXPORT_SYMBOL(ksize);
2770 void kfree(const void *x)
2773 void *object = (void *)x;
2775 if (unlikely(ZERO_OR_NULL_PTR(x)))
2778 page = virt_to_head_page(x);
2779 if (unlikely(!PageSlab(page))) {
2783 slab_free(page->slab, page, object, __builtin_return_address(0));
2785 EXPORT_SYMBOL(kfree);
2788 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2789 * the remaining slabs by the number of items in use. The slabs with the
2790 * most items in use come first. New allocations will then fill those up
2791 * and thus they can be removed from the partial lists.
2793 * The slabs with the least items are placed last. This results in them
2794 * being allocated from last increasing the chance that the last objects
2795 * are freed in them.
2797 int kmem_cache_shrink(struct kmem_cache *s)
2801 struct kmem_cache_node *n;
2804 int objects = oo_objects(s->max);
2805 struct list_head *slabs_by_inuse =
2806 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2807 unsigned long flags;
2809 if (!slabs_by_inuse)
2813 for_each_node_state(node, N_NORMAL_MEMORY) {
2814 n = get_node(s, node);
2819 for (i = 0; i < objects; i++)
2820 INIT_LIST_HEAD(slabs_by_inuse + i);
2822 spin_lock_irqsave(&n->list_lock, flags);
2825 * Build lists indexed by the items in use in each slab.
2827 * Note that concurrent frees may occur while we hold the
2828 * list_lock. page->inuse here is the upper limit.
2830 list_for_each_entry_safe(page, t, &n->partial, lru) {
2831 if (!page->inuse && slab_trylock(page)) {
2833 * Must hold slab lock here because slab_free
2834 * may have freed the last object and be
2835 * waiting to release the slab.
2837 list_del(&page->lru);
2840 discard_slab(s, page);
2842 list_move(&page->lru,
2843 slabs_by_inuse + page->inuse);
2848 * Rebuild the partial list with the slabs filled up most
2849 * first and the least used slabs at the end.
2851 for (i = objects - 1; i >= 0; i--)
2852 list_splice(slabs_by_inuse + i, n->partial.prev);
2854 spin_unlock_irqrestore(&n->list_lock, flags);
2857 kfree(slabs_by_inuse);
2860 EXPORT_SYMBOL(kmem_cache_shrink);
2862 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2863 static int slab_mem_going_offline_callback(void *arg)
2865 struct kmem_cache *s;
2867 down_read(&slub_lock);
2868 list_for_each_entry(s, &slab_caches, list)
2869 kmem_cache_shrink(s);
2870 up_read(&slub_lock);
2875 static void slab_mem_offline_callback(void *arg)
2877 struct kmem_cache_node *n;
2878 struct kmem_cache *s;
2879 struct memory_notify *marg = arg;
2882 offline_node = marg->status_change_nid;
2885 * If the node still has available memory. we need kmem_cache_node
2888 if (offline_node < 0)
2891 down_read(&slub_lock);
2892 list_for_each_entry(s, &slab_caches, list) {
2893 n = get_node(s, offline_node);
2896 * if n->nr_slabs > 0, slabs still exist on the node
2897 * that is going down. We were unable to free them,
2898 * and offline_pages() function shoudn't call this
2899 * callback. So, we must fail.
2901 BUG_ON(slabs_node(s, offline_node));
2903 s->node[offline_node] = NULL;
2904 kmem_cache_free(kmalloc_caches, n);
2907 up_read(&slub_lock);
2910 static int slab_mem_going_online_callback(void *arg)
2912 struct kmem_cache_node *n;
2913 struct kmem_cache *s;
2914 struct memory_notify *marg = arg;
2915 int nid = marg->status_change_nid;
2919 * If the node's memory is already available, then kmem_cache_node is
2920 * already created. Nothing to do.
2926 * We are bringing a node online. No memory is availabe yet. We must
2927 * allocate a kmem_cache_node structure in order to bring the node
2930 down_read(&slub_lock);
2931 list_for_each_entry(s, &slab_caches, list) {
2933 * XXX: kmem_cache_alloc_node will fallback to other nodes
2934 * since memory is not yet available from the node that
2937 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2942 init_kmem_cache_node(n);
2946 up_read(&slub_lock);
2950 static int slab_memory_callback(struct notifier_block *self,
2951 unsigned long action, void *arg)
2956 case MEM_GOING_ONLINE:
2957 ret = slab_mem_going_online_callback(arg);
2959 case MEM_GOING_OFFLINE:
2960 ret = slab_mem_going_offline_callback(arg);
2963 case MEM_CANCEL_ONLINE:
2964 slab_mem_offline_callback(arg);
2967 case MEM_CANCEL_OFFLINE:
2971 ret = notifier_from_errno(ret);
2975 #endif /* CONFIG_MEMORY_HOTPLUG */
2977 /********************************************************************
2978 * Basic setup of slabs
2979 *******************************************************************/
2981 void __init kmem_cache_init(void)
2990 * Must first have the slab cache available for the allocations of the
2991 * struct kmem_cache_node's. There is special bootstrap code in
2992 * kmem_cache_open for slab_state == DOWN.
2994 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2995 sizeof(struct kmem_cache_node), GFP_KERNEL);
2996 kmalloc_caches[0].refcount = -1;
2999 hotplug_memory_notifier(slab_memory_callback, 1);
3002 /* Able to allocate the per node structures */
3003 slab_state = PARTIAL;
3005 /* Caches that are not of the two-to-the-power-of size */
3006 if (KMALLOC_MIN_SIZE <= 64) {
3007 create_kmalloc_cache(&kmalloc_caches[1],
3008 "kmalloc-96", 96, GFP_KERNEL);
3011 if (KMALLOC_MIN_SIZE <= 128) {
3012 create_kmalloc_cache(&kmalloc_caches[2],
3013 "kmalloc-192", 192, GFP_KERNEL);
3017 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
3018 create_kmalloc_cache(&kmalloc_caches[i],
3019 "kmalloc", 1 << i, GFP_KERNEL);
3025 * Patch up the size_index table if we have strange large alignment
3026 * requirements for the kmalloc array. This is only the case for
3027 * MIPS it seems. The standard arches will not generate any code here.
3029 * Largest permitted alignment is 256 bytes due to the way we
3030 * handle the index determination for the smaller caches.
3032 * Make sure that nothing crazy happens if someone starts tinkering
3033 * around with ARCH_KMALLOC_MINALIGN
3035 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3036 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3038 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3039 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3043 /* Provide the correct kmalloc names now that the caches are up */
3044 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3045 kmalloc_caches[i]. name =
3046 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3049 register_cpu_notifier(&slab_notifier);
3050 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3051 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3053 kmem_size = sizeof(struct kmem_cache);
3057 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3058 " CPUs=%d, Nodes=%d\n",
3059 caches, cache_line_size(),
3060 slub_min_order, slub_max_order, slub_min_objects,
3061 nr_cpu_ids, nr_node_ids);
3065 * Find a mergeable slab cache
3067 static int slab_unmergeable(struct kmem_cache *s)
3069 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3076 * We may have set a slab to be unmergeable during bootstrap.
3078 if (s->refcount < 0)
3084 static struct kmem_cache *find_mergeable(size_t size,
3085 size_t align, unsigned long flags, const char *name,
3086 void (*ctor)(struct kmem_cache *, void *))
3088 struct kmem_cache *s;
3090 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3096 size = ALIGN(size, sizeof(void *));
3097 align = calculate_alignment(flags, align, size);
3098 size = ALIGN(size, align);
3099 flags = kmem_cache_flags(size, flags, name, NULL);
3101 list_for_each_entry(s, &slab_caches, list) {
3102 if (slab_unmergeable(s))
3108 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3111 * Check if alignment is compatible.
3112 * Courtesy of Adrian Drzewiecki
3114 if ((s->size & ~(align - 1)) != s->size)
3117 if (s->size - size >= sizeof(void *))
3125 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3126 size_t align, unsigned long flags,
3127 void (*ctor)(struct kmem_cache *, void *))
3129 struct kmem_cache *s;
3131 down_write(&slub_lock);
3132 s = find_mergeable(size, align, flags, name, ctor);
3138 * Adjust the object sizes so that we clear
3139 * the complete object on kzalloc.
3141 s->objsize = max(s->objsize, (int)size);
3144 * And then we need to update the object size in the
3145 * per cpu structures
3147 for_each_online_cpu(cpu)
3148 get_cpu_slab(s, cpu)->objsize = s->objsize;
3150 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3151 up_write(&slub_lock);
3153 if (sysfs_slab_alias(s, name))
3158 s = kmalloc(kmem_size, GFP_KERNEL);
3160 if (kmem_cache_open(s, GFP_KERNEL, name,
3161 size, align, flags, ctor)) {
3162 list_add(&s->list, &slab_caches);
3163 up_write(&slub_lock);
3164 if (sysfs_slab_add(s))
3170 up_write(&slub_lock);
3173 if (flags & SLAB_PANIC)
3174 panic("Cannot create slabcache %s\n", name);
3179 EXPORT_SYMBOL(kmem_cache_create);
3183 * Use the cpu notifier to insure that the cpu slabs are flushed when
3186 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3187 unsigned long action, void *hcpu)
3189 long cpu = (long)hcpu;
3190 struct kmem_cache *s;
3191 unsigned long flags;
3194 case CPU_UP_PREPARE:
3195 case CPU_UP_PREPARE_FROZEN:
3196 init_alloc_cpu_cpu(cpu);
3197 down_read(&slub_lock);
3198 list_for_each_entry(s, &slab_caches, list)
3199 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3201 up_read(&slub_lock);
3204 case CPU_UP_CANCELED:
3205 case CPU_UP_CANCELED_FROZEN:
3207 case CPU_DEAD_FROZEN:
3208 down_read(&slub_lock);
3209 list_for_each_entry(s, &slab_caches, list) {
3210 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3212 local_irq_save(flags);
3213 __flush_cpu_slab(s, cpu);
3214 local_irq_restore(flags);
3215 free_kmem_cache_cpu(c, cpu);
3216 s->cpu_slab[cpu] = NULL;
3218 up_read(&slub_lock);
3226 static struct notifier_block __cpuinitdata slab_notifier = {
3227 .notifier_call = slab_cpuup_callback
3232 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3234 struct kmem_cache *s;
3236 if (unlikely(size > PAGE_SIZE))
3237 return kmalloc_large(size, gfpflags);
3239 s = get_slab(size, gfpflags);
3241 if (unlikely(ZERO_OR_NULL_PTR(s)))
3244 return slab_alloc(s, gfpflags, -1, caller);
3247 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3248 int node, void *caller)
3250 struct kmem_cache *s;
3252 if (unlikely(size > PAGE_SIZE))
3253 return kmalloc_large_node(size, gfpflags, node);
3255 s = get_slab(size, gfpflags);
3257 if (unlikely(ZERO_OR_NULL_PTR(s)))
3260 return slab_alloc(s, gfpflags, node, caller);
3263 #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3264 static unsigned long count_partial(struct kmem_cache_node *n,
3265 int (*get_count)(struct page *))
3267 unsigned long flags;
3268 unsigned long x = 0;
3271 spin_lock_irqsave(&n->list_lock, flags);
3272 list_for_each_entry(page, &n->partial, lru)
3273 x += get_count(page);
3274 spin_unlock_irqrestore(&n->list_lock, flags);
3278 static int count_inuse(struct page *page)
3283 static int count_total(struct page *page)
3285 return page->objects;
3288 static int count_free(struct page *page)
3290 return page->objects - page->inuse;
3294 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3295 static int validate_slab(struct kmem_cache *s, struct page *page,
3299 void *addr = page_address(page);
3301 if (!check_slab(s, page) ||
3302 !on_freelist(s, page, NULL))
3305 /* Now we know that a valid freelist exists */
3306 bitmap_zero(map, page->objects);
3308 for_each_free_object(p, s, page->freelist) {
3309 set_bit(slab_index(p, s, addr), map);
3310 if (!check_object(s, page, p, 0))
3314 for_each_object(p, s, addr, page->objects)
3315 if (!test_bit(slab_index(p, s, addr), map))
3316 if (!check_object(s, page, p, 1))
3321 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3324 if (slab_trylock(page)) {
3325 validate_slab(s, page, map);
3328 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3331 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3332 if (!SlabDebug(page))
3333 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3334 "on slab 0x%p\n", s->name, page);
3336 if (SlabDebug(page))
3337 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3338 "slab 0x%p\n", s->name, page);
3342 static int validate_slab_node(struct kmem_cache *s,
3343 struct kmem_cache_node *n, unsigned long *map)
3345 unsigned long count = 0;
3347 unsigned long flags;
3349 spin_lock_irqsave(&n->list_lock, flags);
3351 list_for_each_entry(page, &n->partial, lru) {
3352 validate_slab_slab(s, page, map);
3355 if (count != n->nr_partial)
3356 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3357 "counter=%ld\n", s->name, count, n->nr_partial);
3359 if (!(s->flags & SLAB_STORE_USER))
3362 list_for_each_entry(page, &n->full, lru) {
3363 validate_slab_slab(s, page, map);
3366 if (count != atomic_long_read(&n->nr_slabs))
3367 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3368 "counter=%ld\n", s->name, count,
3369 atomic_long_read(&n->nr_slabs));
3372 spin_unlock_irqrestore(&n->list_lock, flags);
3376 static long validate_slab_cache(struct kmem_cache *s)
3379 unsigned long count = 0;
3380 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3381 sizeof(unsigned long), GFP_KERNEL);
3387 for_each_node_state(node, N_NORMAL_MEMORY) {
3388 struct kmem_cache_node *n = get_node(s, node);
3390 count += validate_slab_node(s, n, map);
3396 #ifdef SLUB_RESILIENCY_TEST
3397 static void resiliency_test(void)
3401 printk(KERN_ERR "SLUB resiliency testing\n");
3402 printk(KERN_ERR "-----------------------\n");
3403 printk(KERN_ERR "A. Corruption after allocation\n");
3405 p = kzalloc(16, GFP_KERNEL);
3407 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3408 " 0x12->0x%p\n\n", p + 16);
3410 validate_slab_cache(kmalloc_caches + 4);
3412 /* Hmmm... The next two are dangerous */
3413 p = kzalloc(32, GFP_KERNEL);
3414 p[32 + sizeof(void *)] = 0x34;
3415 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3416 " 0x34 -> -0x%p\n", p);
3418 "If allocated object is overwritten then not detectable\n\n");
3420 validate_slab_cache(kmalloc_caches + 5);
3421 p = kzalloc(64, GFP_KERNEL);
3422 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3424 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3427 "If allocated object is overwritten then not detectable\n\n");
3428 validate_slab_cache(kmalloc_caches + 6);
3430 printk(KERN_ERR "\nB. Corruption after free\n");
3431 p = kzalloc(128, GFP_KERNEL);
3434 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3435 validate_slab_cache(kmalloc_caches + 7);
3437 p = kzalloc(256, GFP_KERNEL);
3440 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3442 validate_slab_cache(kmalloc_caches + 8);
3444 p = kzalloc(512, GFP_KERNEL);
3447 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3448 validate_slab_cache(kmalloc_caches + 9);
3451 static void resiliency_test(void) {};
3455 * Generate lists of code addresses where slabcache objects are allocated
3460 unsigned long count;
3473 unsigned long count;
3474 struct location *loc;
3477 static void free_loc_track(struct loc_track *t)
3480 free_pages((unsigned long)t->loc,
3481 get_order(sizeof(struct location) * t->max));
3484 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3489 order = get_order(sizeof(struct location) * max);
3491 l = (void *)__get_free_pages(flags, order);
3496 memcpy(l, t->loc, sizeof(struct location) * t->count);
3504 static int add_location(struct loc_track *t, struct kmem_cache *s,
3505 const struct track *track)
3507 long start, end, pos;
3510 unsigned long age = jiffies - track->when;
3516 pos = start + (end - start + 1) / 2;
3519 * There is nothing at "end". If we end up there
3520 * we need to add something to before end.
3525 caddr = t->loc[pos].addr;
3526 if (track->addr == caddr) {
3532 if (age < l->min_time)
3534 if (age > l->max_time)
3537 if (track->pid < l->min_pid)
3538 l->min_pid = track->pid;
3539 if (track->pid > l->max_pid)
3540 l->max_pid = track->pid;
3542 cpu_set(track->cpu, l->cpus);
3544 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3548 if (track->addr < caddr)
3555 * Not found. Insert new tracking element.
3557 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3563 (t->count - pos) * sizeof(struct location));
3566 l->addr = track->addr;
3570 l->min_pid = track->pid;
3571 l->max_pid = track->pid;
3572 cpus_clear(l->cpus);
3573 cpu_set(track->cpu, l->cpus);
3574 nodes_clear(l->nodes);
3575 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3579 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3580 struct page *page, enum track_item alloc)
3582 void *addr = page_address(page);
3583 DECLARE_BITMAP(map, page->objects);
3586 bitmap_zero(map, page->objects);
3587 for_each_free_object(p, s, page->freelist)
3588 set_bit(slab_index(p, s, addr), map);
3590 for_each_object(p, s, addr, page->objects)
3591 if (!test_bit(slab_index(p, s, addr), map))
3592 add_location(t, s, get_track(s, p, alloc));
3595 static int list_locations(struct kmem_cache *s, char *buf,
3596 enum track_item alloc)
3600 struct loc_track t = { 0, 0, NULL };
3603 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3605 return sprintf(buf, "Out of memory\n");
3607 /* Push back cpu slabs */
3610 for_each_node_state(node, N_NORMAL_MEMORY) {
3611 struct kmem_cache_node *n = get_node(s, node);
3612 unsigned long flags;
3615 if (!atomic_long_read(&n->nr_slabs))
3618 spin_lock_irqsave(&n->list_lock, flags);
3619 list_for_each_entry(page, &n->partial, lru)
3620 process_slab(&t, s, page, alloc);
3621 list_for_each_entry(page, &n->full, lru)
3622 process_slab(&t, s, page, alloc);
3623 spin_unlock_irqrestore(&n->list_lock, flags);
3626 for (i = 0; i < t.count; i++) {
3627 struct location *l = &t.loc[i];
3629 if (len > PAGE_SIZE - 100)
3631 len += sprintf(buf + len, "%7ld ", l->count);
3634 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3636 len += sprintf(buf + len, "<not-available>");
3638 if (l->sum_time != l->min_time) {
3639 unsigned long remainder;
3641 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3643 div_long_long_rem(l->sum_time, l->count, &remainder),
3646 len += sprintf(buf + len, " age=%ld",
3649 if (l->min_pid != l->max_pid)
3650 len += sprintf(buf + len, " pid=%ld-%ld",
3651 l->min_pid, l->max_pid);
3653 len += sprintf(buf + len, " pid=%ld",
3656 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3657 len < PAGE_SIZE - 60) {
3658 len += sprintf(buf + len, " cpus=");
3659 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3663 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3664 len < PAGE_SIZE - 60) {
3665 len += sprintf(buf + len, " nodes=");
3666 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3670 len += sprintf(buf + len, "\n");
3675 len += sprintf(buf, "No data\n");
3679 enum slab_stat_type {
3680 SL_ALL, /* All slabs */
3681 SL_PARTIAL, /* Only partially allocated slabs */
3682 SL_CPU, /* Only slabs used for cpu caches */
3683 SL_OBJECTS, /* Determine allocated objects not slabs */
3684 SL_TOTAL /* Determine object capacity not slabs */
3687 #define SO_ALL (1 << SL_ALL)
3688 #define SO_PARTIAL (1 << SL_PARTIAL)
3689 #define SO_CPU (1 << SL_CPU)
3690 #define SO_OBJECTS (1 << SL_OBJECTS)
3691 #define SO_TOTAL (1 << SL_TOTAL)
3693 static ssize_t show_slab_objects(struct kmem_cache *s,
3694 char *buf, unsigned long flags)
3696 unsigned long total = 0;
3699 unsigned long *nodes;
3700 unsigned long *per_cpu;
3702 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3705 per_cpu = nodes + nr_node_ids;
3707 if (flags & SO_CPU) {
3710 for_each_possible_cpu(cpu) {
3711 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3713 if (!c || c->node < 0)
3717 if (flags & SO_TOTAL)
3718 x = c->page->objects;
3719 else if (flags & SO_OBJECTS)
3725 nodes[c->node] += x;
3731 if (flags & SO_ALL) {
3732 for_each_node_state(node, N_NORMAL_MEMORY) {
3733 struct kmem_cache_node *n = get_node(s, node);
3735 if (flags & SO_TOTAL)
3736 x = atomic_long_read(&n->total_objects);
3737 else if (flags & SO_OBJECTS)
3738 x = atomic_long_read(&n->total_objects) -
3739 count_partial(n, count_free);
3742 x = atomic_long_read(&n->nr_slabs);
3747 } else if (flags & SO_PARTIAL) {
3748 for_each_node_state(node, N_NORMAL_MEMORY) {
3749 struct kmem_cache_node *n = get_node(s, node);
3751 if (flags & SO_TOTAL)
3752 x = count_partial(n, count_total);
3753 else if (flags & SO_OBJECTS)
3754 x = count_partial(n, count_inuse);
3761 x = sprintf(buf, "%lu", total);
3763 for_each_node_state(node, N_NORMAL_MEMORY)
3765 x += sprintf(buf + x, " N%d=%lu",
3769 return x + sprintf(buf + x, "\n");
3772 static int any_slab_objects(struct kmem_cache *s)
3777 for_each_possible_cpu(cpu) {
3778 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3784 for_each_online_node(node) {
3785 struct kmem_cache_node *n = get_node(s, node);
3790 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3796 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3797 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3799 struct slab_attribute {
3800 struct attribute attr;
3801 ssize_t (*show)(struct kmem_cache *s, char *buf);
3802 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3805 #define SLAB_ATTR_RO(_name) \
3806 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3808 #define SLAB_ATTR(_name) \
3809 static struct slab_attribute _name##_attr = \
3810 __ATTR(_name, 0644, _name##_show, _name##_store)
3812 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3814 return sprintf(buf, "%d\n", s->size);
3816 SLAB_ATTR_RO(slab_size);
3818 static ssize_t align_show(struct kmem_cache *s, char *buf)
3820 return sprintf(buf, "%d\n", s->align);
3822 SLAB_ATTR_RO(align);
3824 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3826 return sprintf(buf, "%d\n", s->objsize);
3828 SLAB_ATTR_RO(object_size);
3830 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3832 return sprintf(buf, "%d\n", oo_objects(s->oo));
3834 SLAB_ATTR_RO(objs_per_slab);
3836 static ssize_t order_show(struct kmem_cache *s, char *buf)
3838 return sprintf(buf, "%d\n", oo_order(s->oo));
3840 SLAB_ATTR_RO(order);
3842 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3845 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3847 return n + sprintf(buf + n, "\n");
3853 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3855 return sprintf(buf, "%d\n", s->refcount - 1);
3857 SLAB_ATTR_RO(aliases);
3859 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3861 return show_slab_objects(s, buf, SO_ALL);
3863 SLAB_ATTR_RO(slabs);
3865 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3867 return show_slab_objects(s, buf, SO_PARTIAL);
3869 SLAB_ATTR_RO(partial);
3871 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3873 return show_slab_objects(s, buf, SO_CPU);
3875 SLAB_ATTR_RO(cpu_slabs);
3877 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3879 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3881 SLAB_ATTR_RO(objects);
3883 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3885 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3887 SLAB_ATTR_RO(objects_partial);
3889 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3891 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3893 SLAB_ATTR_RO(total_objects);
3895 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3897 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3900 static ssize_t sanity_checks_store(struct kmem_cache *s,
3901 const char *buf, size_t length)
3903 s->flags &= ~SLAB_DEBUG_FREE;
3905 s->flags |= SLAB_DEBUG_FREE;
3908 SLAB_ATTR(sanity_checks);
3910 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3912 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3915 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3918 s->flags &= ~SLAB_TRACE;
3920 s->flags |= SLAB_TRACE;
3925 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3927 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3930 static ssize_t reclaim_account_store(struct kmem_cache *s,
3931 const char *buf, size_t length)
3933 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3935 s->flags |= SLAB_RECLAIM_ACCOUNT;
3938 SLAB_ATTR(reclaim_account);
3940 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3942 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3944 SLAB_ATTR_RO(hwcache_align);
3946 #ifdef CONFIG_ZONE_DMA
3947 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3949 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3951 SLAB_ATTR_RO(cache_dma);
3954 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3956 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3958 SLAB_ATTR_RO(destroy_by_rcu);
3960 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3962 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3965 static ssize_t red_zone_store(struct kmem_cache *s,
3966 const char *buf, size_t length)
3968 if (any_slab_objects(s))
3971 s->flags &= ~SLAB_RED_ZONE;
3973 s->flags |= SLAB_RED_ZONE;
3977 SLAB_ATTR(red_zone);
3979 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3981 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3984 static ssize_t poison_store(struct kmem_cache *s,
3985 const char *buf, size_t length)
3987 if (any_slab_objects(s))
3990 s->flags &= ~SLAB_POISON;
3992 s->flags |= SLAB_POISON;
3998 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4000 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4003 static ssize_t store_user_store(struct kmem_cache *s,
4004 const char *buf, size_t length)
4006 if (any_slab_objects(s))
4009 s->flags &= ~SLAB_STORE_USER;
4011 s->flags |= SLAB_STORE_USER;
4015 SLAB_ATTR(store_user);
4017 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4022 static ssize_t validate_store(struct kmem_cache *s,
4023 const char *buf, size_t length)
4027 if (buf[0] == '1') {
4028 ret = validate_slab_cache(s);
4034 SLAB_ATTR(validate);
4036 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4041 static ssize_t shrink_store(struct kmem_cache *s,
4042 const char *buf, size_t length)
4044 if (buf[0] == '1') {
4045 int rc = kmem_cache_shrink(s);
4055 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4057 if (!(s->flags & SLAB_STORE_USER))
4059 return list_locations(s, buf, TRACK_ALLOC);
4061 SLAB_ATTR_RO(alloc_calls);
4063 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4065 if (!(s->flags & SLAB_STORE_USER))
4067 return list_locations(s, buf, TRACK_FREE);
4069 SLAB_ATTR_RO(free_calls);
4072 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4074 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4077 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4078 const char *buf, size_t length)
4080 int n = simple_strtoul(buf, NULL, 10);
4083 s->remote_node_defrag_ratio = n * 10;
4086 SLAB_ATTR(remote_node_defrag_ratio);
4089 #ifdef CONFIG_SLUB_STATS
4090 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4092 unsigned long sum = 0;
4095 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4100 for_each_online_cpu(cpu) {
4101 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4107 len = sprintf(buf, "%lu", sum);
4110 for_each_online_cpu(cpu) {
4111 if (data[cpu] && len < PAGE_SIZE - 20)
4112 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4116 return len + sprintf(buf + len, "\n");
4119 #define STAT_ATTR(si, text) \
4120 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4122 return show_stat(s, buf, si); \
4124 SLAB_ATTR_RO(text); \
4126 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4127 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4128 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4129 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4130 STAT_ATTR(FREE_FROZEN, free_frozen);
4131 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4132 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4133 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4134 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4135 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4136 STAT_ATTR(FREE_SLAB, free_slab);
4137 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4138 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4139 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4140 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4141 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4142 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4143 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4146 static struct attribute *slab_attrs[] = {
4147 &slab_size_attr.attr,
4148 &object_size_attr.attr,
4149 &objs_per_slab_attr.attr,
4152 &objects_partial_attr.attr,
4153 &total_objects_attr.attr,
4156 &cpu_slabs_attr.attr,
4160 &sanity_checks_attr.attr,
4162 &hwcache_align_attr.attr,
4163 &reclaim_account_attr.attr,
4164 &destroy_by_rcu_attr.attr,
4165 &red_zone_attr.attr,
4167 &store_user_attr.attr,
4168 &validate_attr.attr,
4170 &alloc_calls_attr.attr,
4171 &free_calls_attr.attr,
4172 #ifdef CONFIG_ZONE_DMA
4173 &cache_dma_attr.attr,
4176 &remote_node_defrag_ratio_attr.attr,
4178 #ifdef CONFIG_SLUB_STATS
4179 &alloc_fastpath_attr.attr,
4180 &alloc_slowpath_attr.attr,
4181 &free_fastpath_attr.attr,
4182 &free_slowpath_attr.attr,
4183 &free_frozen_attr.attr,
4184 &free_add_partial_attr.attr,
4185 &free_remove_partial_attr.attr,
4186 &alloc_from_partial_attr.attr,
4187 &alloc_slab_attr.attr,
4188 &alloc_refill_attr.attr,
4189 &free_slab_attr.attr,
4190 &cpuslab_flush_attr.attr,
4191 &deactivate_full_attr.attr,
4192 &deactivate_empty_attr.attr,
4193 &deactivate_to_head_attr.attr,
4194 &deactivate_to_tail_attr.attr,
4195 &deactivate_remote_frees_attr.attr,
4196 &order_fallback_attr.attr,
4201 static struct attribute_group slab_attr_group = {
4202 .attrs = slab_attrs,
4205 static ssize_t slab_attr_show(struct kobject *kobj,
4206 struct attribute *attr,
4209 struct slab_attribute *attribute;
4210 struct kmem_cache *s;
4213 attribute = to_slab_attr(attr);
4216 if (!attribute->show)
4219 err = attribute->show(s, buf);
4224 static ssize_t slab_attr_store(struct kobject *kobj,
4225 struct attribute *attr,
4226 const char *buf, size_t len)
4228 struct slab_attribute *attribute;
4229 struct kmem_cache *s;
4232 attribute = to_slab_attr(attr);
4235 if (!attribute->store)
4238 err = attribute->store(s, buf, len);
4243 static void kmem_cache_release(struct kobject *kobj)
4245 struct kmem_cache *s = to_slab(kobj);
4250 static struct sysfs_ops slab_sysfs_ops = {
4251 .show = slab_attr_show,
4252 .store = slab_attr_store,
4255 static struct kobj_type slab_ktype = {
4256 .sysfs_ops = &slab_sysfs_ops,
4257 .release = kmem_cache_release
4260 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4262 struct kobj_type *ktype = get_ktype(kobj);
4264 if (ktype == &slab_ktype)
4269 static struct kset_uevent_ops slab_uevent_ops = {
4270 .filter = uevent_filter,
4273 static struct kset *slab_kset;
4275 #define ID_STR_LENGTH 64
4277 /* Create a unique string id for a slab cache:
4279 * Format :[flags-]size
4281 static char *create_unique_id(struct kmem_cache *s)
4283 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4290 * First flags affecting slabcache operations. We will only
4291 * get here for aliasable slabs so we do not need to support
4292 * too many flags. The flags here must cover all flags that
4293 * are matched during merging to guarantee that the id is
4296 if (s->flags & SLAB_CACHE_DMA)
4298 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4300 if (s->flags & SLAB_DEBUG_FREE)
4304 p += sprintf(p, "%07d", s->size);
4305 BUG_ON(p > name + ID_STR_LENGTH - 1);
4309 static int sysfs_slab_add(struct kmem_cache *s)
4315 if (slab_state < SYSFS)
4316 /* Defer until later */
4319 unmergeable = slab_unmergeable(s);
4322 * Slabcache can never be merged so we can use the name proper.
4323 * This is typically the case for debug situations. In that
4324 * case we can catch duplicate names easily.
4326 sysfs_remove_link(&slab_kset->kobj, s->name);
4330 * Create a unique name for the slab as a target
4333 name = create_unique_id(s);
4336 s->kobj.kset = slab_kset;
4337 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4339 kobject_put(&s->kobj);
4343 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4346 kobject_uevent(&s->kobj, KOBJ_ADD);
4348 /* Setup first alias */
4349 sysfs_slab_alias(s, s->name);
4355 static void sysfs_slab_remove(struct kmem_cache *s)
4357 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4358 kobject_del(&s->kobj);
4359 kobject_put(&s->kobj);
4363 * Need to buffer aliases during bootup until sysfs becomes
4364 * available lest we loose that information.
4366 struct saved_alias {
4367 struct kmem_cache *s;
4369 struct saved_alias *next;
4372 static struct saved_alias *alias_list;
4374 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4376 struct saved_alias *al;
4378 if (slab_state == SYSFS) {
4380 * If we have a leftover link then remove it.
4382 sysfs_remove_link(&slab_kset->kobj, name);
4383 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4386 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4392 al->next = alias_list;
4397 static int __init slab_sysfs_init(void)
4399 struct kmem_cache *s;
4402 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4404 printk(KERN_ERR "Cannot register slab subsystem.\n");
4410 list_for_each_entry(s, &slab_caches, list) {
4411 err = sysfs_slab_add(s);
4413 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4414 " to sysfs\n", s->name);
4417 while (alias_list) {
4418 struct saved_alias *al = alias_list;
4420 alias_list = alias_list->next;
4421 err = sysfs_slab_alias(al->s, al->name);
4423 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4424 " %s to sysfs\n", s->name);
4432 __initcall(slab_sysfs_init);
4436 * The /proc/slabinfo ABI
4438 #ifdef CONFIG_SLABINFO
4440 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4441 size_t count, loff_t *ppos)
4447 static void print_slabinfo_header(struct seq_file *m)
4449 seq_puts(m, "slabinfo - version: 2.1\n");
4450 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4451 "<objperslab> <pagesperslab>");
4452 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4453 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4457 static void *s_start(struct seq_file *m, loff_t *pos)
4461 down_read(&slub_lock);
4463 print_slabinfo_header(m);
4465 return seq_list_start(&slab_caches, *pos);
4468 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4470 return seq_list_next(p, &slab_caches, pos);
4473 static void s_stop(struct seq_file *m, void *p)
4475 up_read(&slub_lock);
4478 static int s_show(struct seq_file *m, void *p)
4480 unsigned long nr_partials = 0;
4481 unsigned long nr_slabs = 0;
4482 unsigned long nr_inuse = 0;
4483 unsigned long nr_objs = 0;
4484 unsigned long nr_free = 0;
4485 struct kmem_cache *s;
4488 s = list_entry(p, struct kmem_cache, list);
4490 for_each_online_node(node) {
4491 struct kmem_cache_node *n = get_node(s, node);
4496 nr_partials += n->nr_partial;
4497 nr_slabs += atomic_long_read(&n->nr_slabs);
4498 nr_objs += atomic_long_read(&n->total_objects);
4499 nr_free += count_partial(n, count_free);
4502 nr_inuse = nr_objs - nr_free;
4504 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4505 nr_objs, s->size, oo_objects(s->oo),
4506 (1 << oo_order(s->oo)));
4507 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4508 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4514 const struct seq_operations slabinfo_op = {
4521 #endif /* CONFIG_SLABINFO */