2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemcheck.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/mempolicy.h>
24 #include <linux/ctype.h>
25 #include <linux/debugobjects.h>
26 #include <linux/kallsyms.h>
27 #include <linux/memory.h>
28 #include <linux/math64.h>
29 #include <linux/fault-inject.h>
31 #include <trace/events/kmem.h>
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has noone operating on it and thus there is
68 * no danger of cacheline contention.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
135 * Mininum number of partial slabs. These will be left on the partial
136 * lists even if they are empty. kmem_cache_shrink may reclaim them.
138 #define MIN_PARTIAL 5
141 * Maximum number of desirable partial slabs.
142 * The existence of more partial slabs makes kmem_cache_shrink
143 * sort the partial list by the number of objects in the.
145 #define MAX_PARTIAL 10
147 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
148 SLAB_POISON | SLAB_STORE_USER)
151 * Debugging flags that require metadata to be stored in the slab. These get
152 * disabled when slub_debug=O is used and a cache's min order increases with
155 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
158 * Set of flags that will prevent slab merging
160 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
161 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
164 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
165 SLAB_CACHE_DMA | SLAB_NOTRACK)
168 #define OO_MASK ((1 << OO_SHIFT) - 1)
169 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
171 /* Internal SLUB flags */
172 #define __OBJECT_POISON 0x80000000UL /* Poison object */
174 static int kmem_size = sizeof(struct kmem_cache);
177 static struct notifier_block slab_notifier;
181 DOWN, /* No slab functionality available */
182 PARTIAL, /* Kmem_cache_node works */
183 UP, /* Everything works but does not show up in sysfs */
187 /* A list of all slab caches on the system */
188 static DECLARE_RWSEM(slub_lock);
189 static LIST_HEAD(slab_caches);
192 * Tracking user of a slab.
195 unsigned long addr; /* Called from address */
196 int cpu; /* Was running on cpu */
197 int pid; /* Pid context */
198 unsigned long when; /* When did the operation occur */
201 enum track_item { TRACK_ALLOC, TRACK_FREE };
204 static int sysfs_slab_add(struct kmem_cache *);
205 static int sysfs_slab_alias(struct kmem_cache *, const char *);
206 static void sysfs_slab_remove(struct kmem_cache *);
209 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
210 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
212 static inline void sysfs_slab_remove(struct kmem_cache *s)
220 static inline void stat(struct kmem_cache *s, enum stat_item si)
222 #ifdef CONFIG_SLUB_STATS
223 __this_cpu_inc(s->cpu_slab->stat[si]);
227 /********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
231 int slab_is_available(void)
233 return slab_state >= UP;
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
238 return s->node[node];
241 /* Verify that a pointer has an address that is valid within a slab page */
242 static inline int check_valid_pointer(struct kmem_cache *s,
243 struct page *page, const void *object)
250 base = page_address(page);
251 if (object < base || object >= base + page->objects * s->size ||
252 (object - base) % s->size) {
259 static inline void *get_freepointer(struct kmem_cache *s, void *object)
261 return *(void **)(object + s->offset);
264 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
266 *(void **)(object + s->offset) = fp;
269 /* Loop over all objects in a slab */
270 #define for_each_object(__p, __s, __addr, __objects) \
271 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
275 #define for_each_free_object(__p, __s, __free) \
276 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
278 /* Determine object index from a given position */
279 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
281 return (p - addr) / s->size;
284 static inline size_t slab_ksize(const struct kmem_cache *s)
286 #ifdef CONFIG_SLUB_DEBUG
288 * Debugging requires use of the padding between object
289 * and whatever may come after it.
291 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
296 * If we have the need to store the freelist pointer
297 * back there or track user information then we can
298 * only use the space before that information.
300 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
303 * Else we can use all the padding etc for the allocation
308 static inline int order_objects(int order, unsigned long size, int reserved)
310 return ((PAGE_SIZE << order) - reserved) / size;
313 static inline struct kmem_cache_order_objects oo_make(int order,
314 unsigned long size, int reserved)
316 struct kmem_cache_order_objects x = {
317 (order << OO_SHIFT) + order_objects(order, size, reserved)
323 static inline int oo_order(struct kmem_cache_order_objects x)
325 return x.x >> OO_SHIFT;
328 static inline int oo_objects(struct kmem_cache_order_objects x)
330 return x.x & OO_MASK;
333 #ifdef CONFIG_SLUB_DEBUG
337 #ifdef CONFIG_SLUB_DEBUG_ON
338 static int slub_debug = DEBUG_DEFAULT_FLAGS;
340 static int slub_debug;
343 static char *slub_debug_slabs;
344 static int disable_higher_order_debug;
349 static void print_section(char *text, u8 *addr, unsigned int length)
357 for (i = 0; i < length; i++) {
359 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
362 printk(KERN_CONT " %02x", addr[i]);
364 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
366 printk(KERN_CONT " %s\n", ascii);
373 printk(KERN_CONT " ");
377 printk(KERN_CONT " %s\n", ascii);
381 static struct track *get_track(struct kmem_cache *s, void *object,
382 enum track_item alloc)
387 p = object + s->offset + sizeof(void *);
389 p = object + s->inuse;
394 static void set_track(struct kmem_cache *s, void *object,
395 enum track_item alloc, unsigned long addr)
397 struct track *p = get_track(s, object, alloc);
401 p->cpu = smp_processor_id();
402 p->pid = current->pid;
405 memset(p, 0, sizeof(struct track));
408 static void init_tracking(struct kmem_cache *s, void *object)
410 if (!(s->flags & SLAB_STORE_USER))
413 set_track(s, object, TRACK_FREE, 0UL);
414 set_track(s, object, TRACK_ALLOC, 0UL);
417 static void print_track(const char *s, struct track *t)
422 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
423 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
426 static void print_tracking(struct kmem_cache *s, void *object)
428 if (!(s->flags & SLAB_STORE_USER))
431 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
432 print_track("Freed", get_track(s, object, TRACK_FREE));
435 static void print_page_info(struct page *page)
437 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
438 page, page->objects, page->inuse, page->freelist, page->flags);
442 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
448 vsnprintf(buf, sizeof(buf), fmt, args);
450 printk(KERN_ERR "========================================"
451 "=====================================\n");
452 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
453 printk(KERN_ERR "----------------------------------------"
454 "-------------------------------------\n\n");
457 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
463 vsnprintf(buf, sizeof(buf), fmt, args);
465 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
468 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
470 unsigned int off; /* Offset of last byte */
471 u8 *addr = page_address(page);
473 print_tracking(s, p);
475 print_page_info(page);
477 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
478 p, p - addr, get_freepointer(s, p));
481 print_section("Bytes b4", p - 16, 16);
483 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
485 if (s->flags & SLAB_RED_ZONE)
486 print_section("Redzone", p + s->objsize,
487 s->inuse - s->objsize);
490 off = s->offset + sizeof(void *);
494 if (s->flags & SLAB_STORE_USER)
495 off += 2 * sizeof(struct track);
498 /* Beginning of the filler is the free pointer */
499 print_section("Padding", p + off, s->size - off);
504 static void object_err(struct kmem_cache *s, struct page *page,
505 u8 *object, char *reason)
507 slab_bug(s, "%s", reason);
508 print_trailer(s, page, object);
511 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
517 vsnprintf(buf, sizeof(buf), fmt, args);
519 slab_bug(s, "%s", buf);
520 print_page_info(page);
524 static void init_object(struct kmem_cache *s, void *object, u8 val)
528 if (s->flags & __OBJECT_POISON) {
529 memset(p, POISON_FREE, s->objsize - 1);
530 p[s->objsize - 1] = POISON_END;
533 if (s->flags & SLAB_RED_ZONE)
534 memset(p + s->objsize, val, s->inuse - s->objsize);
537 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
540 if (*start != (u8)value)
548 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
549 void *from, void *to)
551 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
552 memset(from, data, to - from);
555 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
556 u8 *object, char *what,
557 u8 *start, unsigned int value, unsigned int bytes)
562 fault = check_bytes(start, value, bytes);
567 while (end > fault && end[-1] == value)
570 slab_bug(s, "%s overwritten", what);
571 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
572 fault, end - 1, fault[0], value);
573 print_trailer(s, page, object);
575 restore_bytes(s, what, value, fault, end);
583 * Bytes of the object to be managed.
584 * If the freepointer may overlay the object then the free
585 * pointer is the first word of the object.
587 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
590 * object + s->objsize
591 * Padding to reach word boundary. This is also used for Redzoning.
592 * Padding is extended by another word if Redzoning is enabled and
595 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
596 * 0xcc (RED_ACTIVE) for objects in use.
599 * Meta data starts here.
601 * A. Free pointer (if we cannot overwrite object on free)
602 * B. Tracking data for SLAB_STORE_USER
603 * C. Padding to reach required alignment boundary or at mininum
604 * one word if debugging is on to be able to detect writes
605 * before the word boundary.
607 * Padding is done using 0x5a (POISON_INUSE)
610 * Nothing is used beyond s->size.
612 * If slabcaches are merged then the objsize and inuse boundaries are mostly
613 * ignored. And therefore no slab options that rely on these boundaries
614 * may be used with merged slabcaches.
617 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
619 unsigned long off = s->inuse; /* The end of info */
622 /* Freepointer is placed after the object. */
623 off += sizeof(void *);
625 if (s->flags & SLAB_STORE_USER)
626 /* We also have user information there */
627 off += 2 * sizeof(struct track);
632 return check_bytes_and_report(s, page, p, "Object padding",
633 p + off, POISON_INUSE, s->size - off);
636 /* Check the pad bytes at the end of a slab page */
637 static int slab_pad_check(struct kmem_cache *s, struct page *page)
645 if (!(s->flags & SLAB_POISON))
648 start = page_address(page);
649 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
650 end = start + length;
651 remainder = length % s->size;
655 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
658 while (end > fault && end[-1] == POISON_INUSE)
661 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
662 print_section("Padding", end - remainder, remainder);
664 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
668 static int check_object(struct kmem_cache *s, struct page *page,
669 void *object, u8 val)
672 u8 *endobject = object + s->objsize;
674 if (s->flags & SLAB_RED_ZONE) {
675 if (!check_bytes_and_report(s, page, object, "Redzone",
676 endobject, val, s->inuse - s->objsize))
679 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
680 check_bytes_and_report(s, page, p, "Alignment padding",
681 endobject, POISON_INUSE, s->inuse - s->objsize);
685 if (s->flags & SLAB_POISON) {
686 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
687 (!check_bytes_and_report(s, page, p, "Poison", p,
688 POISON_FREE, s->objsize - 1) ||
689 !check_bytes_and_report(s, page, p, "Poison",
690 p + s->objsize - 1, POISON_END, 1)))
693 * check_pad_bytes cleans up on its own.
695 check_pad_bytes(s, page, p);
698 if (!s->offset && val == SLUB_RED_ACTIVE)
700 * Object and freepointer overlap. Cannot check
701 * freepointer while object is allocated.
705 /* Check free pointer validity */
706 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
707 object_err(s, page, p, "Freepointer corrupt");
709 * No choice but to zap it and thus lose the remainder
710 * of the free objects in this slab. May cause
711 * another error because the object count is now wrong.
713 set_freepointer(s, p, NULL);
719 static int check_slab(struct kmem_cache *s, struct page *page)
723 VM_BUG_ON(!irqs_disabled());
725 if (!PageSlab(page)) {
726 slab_err(s, page, "Not a valid slab page");
730 maxobj = order_objects(compound_order(page), s->size, s->reserved);
731 if (page->objects > maxobj) {
732 slab_err(s, page, "objects %u > max %u",
733 s->name, page->objects, maxobj);
736 if (page->inuse > page->objects) {
737 slab_err(s, page, "inuse %u > max %u",
738 s->name, page->inuse, page->objects);
741 /* Slab_pad_check fixes things up after itself */
742 slab_pad_check(s, page);
747 * Determine if a certain object on a page is on the freelist. Must hold the
748 * slab lock to guarantee that the chains are in a consistent state.
750 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
753 void *fp = page->freelist;
755 unsigned long max_objects;
757 while (fp && nr <= page->objects) {
760 if (!check_valid_pointer(s, page, fp)) {
762 object_err(s, page, object,
763 "Freechain corrupt");
764 set_freepointer(s, object, NULL);
767 slab_err(s, page, "Freepointer corrupt");
768 page->freelist = NULL;
769 page->inuse = page->objects;
770 slab_fix(s, "Freelist cleared");
776 fp = get_freepointer(s, object);
780 max_objects = order_objects(compound_order(page), s->size, s->reserved);
781 if (max_objects > MAX_OBJS_PER_PAGE)
782 max_objects = MAX_OBJS_PER_PAGE;
784 if (page->objects != max_objects) {
785 slab_err(s, page, "Wrong number of objects. Found %d but "
786 "should be %d", page->objects, max_objects);
787 page->objects = max_objects;
788 slab_fix(s, "Number of objects adjusted.");
790 if (page->inuse != page->objects - nr) {
791 slab_err(s, page, "Wrong object count. Counter is %d but "
792 "counted were %d", page->inuse, page->objects - nr);
793 page->inuse = page->objects - nr;
794 slab_fix(s, "Object count adjusted.");
796 return search == NULL;
799 static void trace(struct kmem_cache *s, struct page *page, void *object,
802 if (s->flags & SLAB_TRACE) {
803 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
805 alloc ? "alloc" : "free",
810 print_section("Object", (void *)object, s->objsize);
817 * Hooks for other subsystems that check memory allocations. In a typical
818 * production configuration these hooks all should produce no code at all.
820 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
822 flags &= gfp_allowed_mask;
823 lockdep_trace_alloc(flags);
824 might_sleep_if(flags & __GFP_WAIT);
826 return should_failslab(s->objsize, flags, s->flags);
829 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
831 flags &= gfp_allowed_mask;
832 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
833 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
836 static inline void slab_free_hook(struct kmem_cache *s, void *x)
838 kmemleak_free_recursive(x, s->flags);
841 static inline void slab_free_hook_irq(struct kmem_cache *s, void *object)
843 kmemcheck_slab_free(s, object, s->objsize);
844 debug_check_no_locks_freed(object, s->objsize);
845 if (!(s->flags & SLAB_DEBUG_OBJECTS))
846 debug_check_no_obj_freed(object, s->objsize);
850 * Tracking of fully allocated slabs for debugging purposes.
852 static void add_full(struct kmem_cache_node *n, struct page *page)
854 spin_lock(&n->list_lock);
855 list_add(&page->lru, &n->full);
856 spin_unlock(&n->list_lock);
859 static void remove_full(struct kmem_cache *s, struct page *page)
861 struct kmem_cache_node *n;
863 if (!(s->flags & SLAB_STORE_USER))
866 n = get_node(s, page_to_nid(page));
868 spin_lock(&n->list_lock);
869 list_del(&page->lru);
870 spin_unlock(&n->list_lock);
873 /* Tracking of the number of slabs for debugging purposes */
874 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
876 struct kmem_cache_node *n = get_node(s, node);
878 return atomic_long_read(&n->nr_slabs);
881 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
883 return atomic_long_read(&n->nr_slabs);
886 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
888 struct kmem_cache_node *n = get_node(s, node);
891 * May be called early in order to allocate a slab for the
892 * kmem_cache_node structure. Solve the chicken-egg
893 * dilemma by deferring the increment of the count during
894 * bootstrap (see early_kmem_cache_node_alloc).
897 atomic_long_inc(&n->nr_slabs);
898 atomic_long_add(objects, &n->total_objects);
901 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
903 struct kmem_cache_node *n = get_node(s, node);
905 atomic_long_dec(&n->nr_slabs);
906 atomic_long_sub(objects, &n->total_objects);
909 /* Object debug checks for alloc/free paths */
910 static void setup_object_debug(struct kmem_cache *s, struct page *page,
913 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
916 init_object(s, object, SLUB_RED_INACTIVE);
917 init_tracking(s, object);
920 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
921 void *object, unsigned long addr)
923 if (!check_slab(s, page))
926 if (!on_freelist(s, page, object)) {
927 object_err(s, page, object, "Object already allocated");
931 if (!check_valid_pointer(s, page, object)) {
932 object_err(s, page, object, "Freelist Pointer check fails");
936 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
939 /* Success perform special debug activities for allocs */
940 if (s->flags & SLAB_STORE_USER)
941 set_track(s, object, TRACK_ALLOC, addr);
942 trace(s, page, object, 1);
943 init_object(s, object, SLUB_RED_ACTIVE);
947 if (PageSlab(page)) {
949 * If this is a slab page then lets do the best we can
950 * to avoid issues in the future. Marking all objects
951 * as used avoids touching the remaining objects.
953 slab_fix(s, "Marking all objects used");
954 page->inuse = page->objects;
955 page->freelist = NULL;
960 static noinline int free_debug_processing(struct kmem_cache *s,
961 struct page *page, void *object, unsigned long addr)
963 if (!check_slab(s, page))
966 if (!check_valid_pointer(s, page, object)) {
967 slab_err(s, page, "Invalid object pointer 0x%p", object);
971 if (on_freelist(s, page, object)) {
972 object_err(s, page, object, "Object already free");
976 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
979 if (unlikely(s != page->slab)) {
980 if (!PageSlab(page)) {
981 slab_err(s, page, "Attempt to free object(0x%p) "
982 "outside of slab", object);
983 } else if (!page->slab) {
985 "SLUB <none>: no slab for object 0x%p.\n",
989 object_err(s, page, object,
990 "page slab pointer corrupt.");
994 /* Special debug activities for freeing objects */
995 if (!PageSlubFrozen(page) && !page->freelist)
996 remove_full(s, page);
997 if (s->flags & SLAB_STORE_USER)
998 set_track(s, object, TRACK_FREE, addr);
999 trace(s, page, object, 0);
1000 init_object(s, object, SLUB_RED_INACTIVE);
1004 slab_fix(s, "Object at 0x%p not freed", object);
1008 static int __init setup_slub_debug(char *str)
1010 slub_debug = DEBUG_DEFAULT_FLAGS;
1011 if (*str++ != '=' || !*str)
1013 * No options specified. Switch on full debugging.
1019 * No options but restriction on slabs. This means full
1020 * debugging for slabs matching a pattern.
1024 if (tolower(*str) == 'o') {
1026 * Avoid enabling debugging on caches if its minimum order
1027 * would increase as a result.
1029 disable_higher_order_debug = 1;
1036 * Switch off all debugging measures.
1041 * Determine which debug features should be switched on
1043 for (; *str && *str != ','; str++) {
1044 switch (tolower(*str)) {
1046 slub_debug |= SLAB_DEBUG_FREE;
1049 slub_debug |= SLAB_RED_ZONE;
1052 slub_debug |= SLAB_POISON;
1055 slub_debug |= SLAB_STORE_USER;
1058 slub_debug |= SLAB_TRACE;
1061 slub_debug |= SLAB_FAILSLAB;
1064 printk(KERN_ERR "slub_debug option '%c' "
1065 "unknown. skipped\n", *str);
1071 slub_debug_slabs = str + 1;
1076 __setup("slub_debug", setup_slub_debug);
1078 static unsigned long kmem_cache_flags(unsigned long objsize,
1079 unsigned long flags, const char *name,
1080 void (*ctor)(void *))
1083 * Enable debugging if selected on the kernel commandline.
1085 if (slub_debug && (!slub_debug_slabs ||
1086 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1087 flags |= slub_debug;
1092 static inline void setup_object_debug(struct kmem_cache *s,
1093 struct page *page, void *object) {}
1095 static inline int alloc_debug_processing(struct kmem_cache *s,
1096 struct page *page, void *object, unsigned long addr) { return 0; }
1098 static inline int free_debug_processing(struct kmem_cache *s,
1099 struct page *page, void *object, unsigned long addr) { return 0; }
1101 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1103 static inline int check_object(struct kmem_cache *s, struct page *page,
1104 void *object, u8 val) { return 1; }
1105 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1106 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1107 unsigned long flags, const char *name,
1108 void (*ctor)(void *))
1112 #define slub_debug 0
1114 #define disable_higher_order_debug 0
1116 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1118 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1120 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1122 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1125 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1128 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1131 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1133 static inline void slab_free_hook_irq(struct kmem_cache *s,
1136 #endif /* CONFIG_SLUB_DEBUG */
1139 * Slab allocation and freeing
1141 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1142 struct kmem_cache_order_objects oo)
1144 int order = oo_order(oo);
1146 flags |= __GFP_NOTRACK;
1148 if (node == NUMA_NO_NODE)
1149 return alloc_pages(flags, order);
1151 return alloc_pages_exact_node(node, flags, order);
1154 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1157 struct kmem_cache_order_objects oo = s->oo;
1160 flags |= s->allocflags;
1163 * Let the initial higher-order allocation fail under memory pressure
1164 * so we fall-back to the minimum order allocation.
1166 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1168 page = alloc_slab_page(alloc_gfp, node, oo);
1169 if (unlikely(!page)) {
1172 * Allocation may have failed due to fragmentation.
1173 * Try a lower order alloc if possible
1175 page = alloc_slab_page(flags, node, oo);
1179 stat(s, ORDER_FALLBACK);
1182 if (kmemcheck_enabled
1183 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1184 int pages = 1 << oo_order(oo);
1186 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1189 * Objects from caches that have a constructor don't get
1190 * cleared when they're allocated, so we need to do it here.
1193 kmemcheck_mark_uninitialized_pages(page, pages);
1195 kmemcheck_mark_unallocated_pages(page, pages);
1198 page->objects = oo_objects(oo);
1199 mod_zone_page_state(page_zone(page),
1200 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1201 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1207 static void setup_object(struct kmem_cache *s, struct page *page,
1210 setup_object_debug(s, page, object);
1211 if (unlikely(s->ctor))
1215 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1222 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1224 page = allocate_slab(s,
1225 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1229 inc_slabs_node(s, page_to_nid(page), page->objects);
1231 page->flags |= 1 << PG_slab;
1233 start = page_address(page);
1235 if (unlikely(s->flags & SLAB_POISON))
1236 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1239 for_each_object(p, s, start, page->objects) {
1240 setup_object(s, page, last);
1241 set_freepointer(s, last, p);
1244 setup_object(s, page, last);
1245 set_freepointer(s, last, NULL);
1247 page->freelist = start;
1253 static void __free_slab(struct kmem_cache *s, struct page *page)
1255 int order = compound_order(page);
1256 int pages = 1 << order;
1258 if (kmem_cache_debug(s)) {
1261 slab_pad_check(s, page);
1262 for_each_object(p, s, page_address(page),
1264 check_object(s, page, p, SLUB_RED_INACTIVE);
1267 kmemcheck_free_shadow(page, compound_order(page));
1269 mod_zone_page_state(page_zone(page),
1270 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1271 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1274 __ClearPageSlab(page);
1275 reset_page_mapcount(page);
1276 if (current->reclaim_state)
1277 current->reclaim_state->reclaimed_slab += pages;
1278 __free_pages(page, order);
1281 #define need_reserve_slab_rcu \
1282 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1284 static void rcu_free_slab(struct rcu_head *h)
1288 if (need_reserve_slab_rcu)
1289 page = virt_to_head_page(h);
1291 page = container_of((struct list_head *)h, struct page, lru);
1293 __free_slab(page->slab, page);
1296 static void free_slab(struct kmem_cache *s, struct page *page)
1298 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1299 struct rcu_head *head;
1301 if (need_reserve_slab_rcu) {
1302 int order = compound_order(page);
1303 int offset = (PAGE_SIZE << order) - s->reserved;
1305 VM_BUG_ON(s->reserved != sizeof(*head));
1306 head = page_address(page) + offset;
1309 * RCU free overloads the RCU head over the LRU
1311 head = (void *)&page->lru;
1314 call_rcu(head, rcu_free_slab);
1316 __free_slab(s, page);
1319 static void discard_slab(struct kmem_cache *s, struct page *page)
1321 dec_slabs_node(s, page_to_nid(page), page->objects);
1326 * Per slab locking using the pagelock
1328 static __always_inline void slab_lock(struct page *page)
1330 bit_spin_lock(PG_locked, &page->flags);
1333 static __always_inline void slab_unlock(struct page *page)
1335 __bit_spin_unlock(PG_locked, &page->flags);
1338 static __always_inline int slab_trylock(struct page *page)
1342 rc = bit_spin_trylock(PG_locked, &page->flags);
1347 * Management of partially allocated slabs
1349 static void add_partial(struct kmem_cache_node *n,
1350 struct page *page, int tail)
1352 spin_lock(&n->list_lock);
1355 list_add_tail(&page->lru, &n->partial);
1357 list_add(&page->lru, &n->partial);
1358 spin_unlock(&n->list_lock);
1361 static inline void __remove_partial(struct kmem_cache_node *n,
1364 list_del(&page->lru);
1368 static void remove_partial(struct kmem_cache *s, struct page *page)
1370 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1372 spin_lock(&n->list_lock);
1373 __remove_partial(n, page);
1374 spin_unlock(&n->list_lock);
1378 * Lock slab and remove from the partial list.
1380 * Must hold list_lock.
1382 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1385 if (slab_trylock(page)) {
1386 __remove_partial(n, page);
1387 __SetPageSlubFrozen(page);
1394 * Try to allocate a partial slab from a specific node.
1396 static struct page *get_partial_node(struct kmem_cache_node *n)
1401 * Racy check. If we mistakenly see no partial slabs then we
1402 * just allocate an empty slab. If we mistakenly try to get a
1403 * partial slab and there is none available then get_partials()
1406 if (!n || !n->nr_partial)
1409 spin_lock(&n->list_lock);
1410 list_for_each_entry(page, &n->partial, lru)
1411 if (lock_and_freeze_slab(n, page))
1415 spin_unlock(&n->list_lock);
1420 * Get a page from somewhere. Search in increasing NUMA distances.
1422 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1425 struct zonelist *zonelist;
1428 enum zone_type high_zoneidx = gfp_zone(flags);
1432 * The defrag ratio allows a configuration of the tradeoffs between
1433 * inter node defragmentation and node local allocations. A lower
1434 * defrag_ratio increases the tendency to do local allocations
1435 * instead of attempting to obtain partial slabs from other nodes.
1437 * If the defrag_ratio is set to 0 then kmalloc() always
1438 * returns node local objects. If the ratio is higher then kmalloc()
1439 * may return off node objects because partial slabs are obtained
1440 * from other nodes and filled up.
1442 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1443 * defrag_ratio = 1000) then every (well almost) allocation will
1444 * first attempt to defrag slab caches on other nodes. This means
1445 * scanning over all nodes to look for partial slabs which may be
1446 * expensive if we do it every time we are trying to find a slab
1447 * with available objects.
1449 if (!s->remote_node_defrag_ratio ||
1450 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1454 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1455 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1456 struct kmem_cache_node *n;
1458 n = get_node(s, zone_to_nid(zone));
1460 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1461 n->nr_partial > s->min_partial) {
1462 page = get_partial_node(n);
1475 * Get a partial page, lock it and return it.
1477 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1480 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1482 page = get_partial_node(get_node(s, searchnode));
1483 if (page || node != -1)
1486 return get_any_partial(s, flags);
1490 * Move a page back to the lists.
1492 * Must be called with the slab lock held.
1494 * On exit the slab lock will have been dropped.
1496 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1499 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1501 __ClearPageSlubFrozen(page);
1504 if (page->freelist) {
1505 add_partial(n, page, tail);
1506 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1508 stat(s, DEACTIVATE_FULL);
1509 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1514 stat(s, DEACTIVATE_EMPTY);
1515 if (n->nr_partial < s->min_partial) {
1517 * Adding an empty slab to the partial slabs in order
1518 * to avoid page allocator overhead. This slab needs
1519 * to come after the other slabs with objects in
1520 * so that the others get filled first. That way the
1521 * size of the partial list stays small.
1523 * kmem_cache_shrink can reclaim any empty slabs from
1526 add_partial(n, page, 1);
1531 discard_slab(s, page);
1537 * Remove the cpu slab
1539 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1542 struct page *page = c->page;
1546 stat(s, DEACTIVATE_REMOTE_FREES);
1548 * Merge cpu freelist into slab freelist. Typically we get here
1549 * because both freelists are empty. So this is unlikely
1552 while (unlikely(c->freelist)) {
1555 tail = 0; /* Hot objects. Put the slab first */
1557 /* Retrieve object from cpu_freelist */
1558 object = c->freelist;
1559 c->freelist = get_freepointer(s, c->freelist);
1561 /* And put onto the regular freelist */
1562 set_freepointer(s, object, page->freelist);
1563 page->freelist = object;
1567 unfreeze_slab(s, page, tail);
1570 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1572 stat(s, CPUSLAB_FLUSH);
1574 deactivate_slab(s, c);
1580 * Called from IPI handler with interrupts disabled.
1582 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1584 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1586 if (likely(c && c->page))
1590 static void flush_cpu_slab(void *d)
1592 struct kmem_cache *s = d;
1594 __flush_cpu_slab(s, smp_processor_id());
1597 static void flush_all(struct kmem_cache *s)
1599 on_each_cpu(flush_cpu_slab, s, 1);
1603 * Check if the objects in a per cpu structure fit numa
1604 * locality expectations.
1606 static inline int node_match(struct kmem_cache_cpu *c, int node)
1609 if (node != NUMA_NO_NODE && c->node != node)
1615 static int count_free(struct page *page)
1617 return page->objects - page->inuse;
1620 static unsigned long count_partial(struct kmem_cache_node *n,
1621 int (*get_count)(struct page *))
1623 unsigned long flags;
1624 unsigned long x = 0;
1627 spin_lock_irqsave(&n->list_lock, flags);
1628 list_for_each_entry(page, &n->partial, lru)
1629 x += get_count(page);
1630 spin_unlock_irqrestore(&n->list_lock, flags);
1634 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1636 #ifdef CONFIG_SLUB_DEBUG
1637 return atomic_long_read(&n->total_objects);
1643 static noinline void
1644 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1649 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1651 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1652 "default order: %d, min order: %d\n", s->name, s->objsize,
1653 s->size, oo_order(s->oo), oo_order(s->min));
1655 if (oo_order(s->min) > get_order(s->objsize))
1656 printk(KERN_WARNING " %s debugging increased min order, use "
1657 "slub_debug=O to disable.\n", s->name);
1659 for_each_online_node(node) {
1660 struct kmem_cache_node *n = get_node(s, node);
1661 unsigned long nr_slabs;
1662 unsigned long nr_objs;
1663 unsigned long nr_free;
1668 nr_free = count_partial(n, count_free);
1669 nr_slabs = node_nr_slabs(n);
1670 nr_objs = node_nr_objs(n);
1673 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1674 node, nr_slabs, nr_objs, nr_free);
1679 * Slow path. The lockless freelist is empty or we need to perform
1682 * Interrupts are disabled.
1684 * Processing is still very fast if new objects have been freed to the
1685 * regular freelist. In that case we simply take over the regular freelist
1686 * as the lockless freelist and zap the regular freelist.
1688 * If that is not working then we fall back to the partial lists. We take the
1689 * first element of the freelist as the object to allocate now and move the
1690 * rest of the freelist to the lockless freelist.
1692 * And if we were unable to get a new slab from the partial slab lists then
1693 * we need to allocate a new slab. This is the slowest path since it involves
1694 * a call to the page allocator and the setup of a new slab.
1696 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1697 unsigned long addr, struct kmem_cache_cpu *c)
1702 /* We handle __GFP_ZERO in the caller */
1703 gfpflags &= ~__GFP_ZERO;
1709 if (unlikely(!node_match(c, node)))
1712 stat(s, ALLOC_REFILL);
1715 object = c->page->freelist;
1716 if (unlikely(!object))
1718 if (kmem_cache_debug(s))
1721 c->freelist = get_freepointer(s, object);
1722 c->page->inuse = c->page->objects;
1723 c->page->freelist = NULL;
1724 c->node = page_to_nid(c->page);
1726 slab_unlock(c->page);
1727 stat(s, ALLOC_SLOWPATH);
1731 deactivate_slab(s, c);
1734 new = get_partial(s, gfpflags, node);
1737 stat(s, ALLOC_FROM_PARTIAL);
1741 gfpflags &= gfp_allowed_mask;
1742 if (gfpflags & __GFP_WAIT)
1745 new = new_slab(s, gfpflags, node);
1747 if (gfpflags & __GFP_WAIT)
1748 local_irq_disable();
1751 c = __this_cpu_ptr(s->cpu_slab);
1752 stat(s, ALLOC_SLAB);
1756 __SetPageSlubFrozen(new);
1760 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1761 slab_out_of_memory(s, gfpflags, node);
1764 if (!alloc_debug_processing(s, c->page, object, addr))
1768 c->page->freelist = get_freepointer(s, object);
1769 c->node = NUMA_NO_NODE;
1774 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1775 * have the fastpath folded into their functions. So no function call
1776 * overhead for requests that can be satisfied on the fastpath.
1778 * The fastpath works by first checking if the lockless freelist can be used.
1779 * If not then __slab_alloc is called for slow processing.
1781 * Otherwise we can simply pick the next object from the lockless free list.
1783 static __always_inline void *slab_alloc(struct kmem_cache *s,
1784 gfp_t gfpflags, int node, unsigned long addr)
1787 struct kmem_cache_cpu *c;
1788 unsigned long flags;
1790 if (slab_pre_alloc_hook(s, gfpflags))
1793 local_irq_save(flags);
1794 c = __this_cpu_ptr(s->cpu_slab);
1795 object = c->freelist;
1796 if (unlikely(!object || !node_match(c, node)))
1798 object = __slab_alloc(s, gfpflags, node, addr, c);
1801 c->freelist = get_freepointer(s, object);
1802 stat(s, ALLOC_FASTPATH);
1804 local_irq_restore(flags);
1806 if (unlikely(gfpflags & __GFP_ZERO) && object)
1807 memset(object, 0, s->objsize);
1809 slab_post_alloc_hook(s, gfpflags, object);
1814 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1816 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1818 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1822 EXPORT_SYMBOL(kmem_cache_alloc);
1824 #ifdef CONFIG_TRACING
1825 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1827 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1828 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
1831 EXPORT_SYMBOL(kmem_cache_alloc_trace);
1833 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1835 void *ret = kmalloc_order(size, flags, order);
1836 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1839 EXPORT_SYMBOL(kmalloc_order_trace);
1843 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1845 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1847 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1848 s->objsize, s->size, gfpflags, node);
1852 EXPORT_SYMBOL(kmem_cache_alloc_node);
1854 #ifdef CONFIG_TRACING
1855 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
1857 int node, size_t size)
1859 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1861 trace_kmalloc_node(_RET_IP_, ret,
1862 size, s->size, gfpflags, node);
1865 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
1870 * Slow patch handling. This may still be called frequently since objects
1871 * have a longer lifetime than the cpu slabs in most processing loads.
1873 * So we still attempt to reduce cache line usage. Just take the slab
1874 * lock and free the item. If there is no additional partial page
1875 * handling required then we can return immediately.
1877 static void __slab_free(struct kmem_cache *s, struct page *page,
1878 void *x, unsigned long addr)
1881 void **object = (void *)x;
1883 stat(s, FREE_SLOWPATH);
1886 if (kmem_cache_debug(s))
1890 prior = page->freelist;
1891 set_freepointer(s, object, prior);
1892 page->freelist = object;
1895 if (unlikely(PageSlubFrozen(page))) {
1896 stat(s, FREE_FROZEN);
1900 if (unlikely(!page->inuse))
1904 * Objects left in the slab. If it was not on the partial list before
1907 if (unlikely(!prior)) {
1908 add_partial(get_node(s, page_to_nid(page)), page, 1);
1909 stat(s, FREE_ADD_PARTIAL);
1919 * Slab still on the partial list.
1921 remove_partial(s, page);
1922 stat(s, FREE_REMOVE_PARTIAL);
1926 discard_slab(s, page);
1930 if (!free_debug_processing(s, page, x, addr))
1936 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1937 * can perform fastpath freeing without additional function calls.
1939 * The fastpath is only possible if we are freeing to the current cpu slab
1940 * of this processor. This typically the case if we have just allocated
1943 * If fastpath is not possible then fall back to __slab_free where we deal
1944 * with all sorts of special processing.
1946 static __always_inline void slab_free(struct kmem_cache *s,
1947 struct page *page, void *x, unsigned long addr)
1949 void **object = (void *)x;
1950 struct kmem_cache_cpu *c;
1951 unsigned long flags;
1953 slab_free_hook(s, x);
1955 local_irq_save(flags);
1956 c = __this_cpu_ptr(s->cpu_slab);
1958 slab_free_hook_irq(s, x);
1960 if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
1961 set_freepointer(s, object, c->freelist);
1962 c->freelist = object;
1963 stat(s, FREE_FASTPATH);
1965 __slab_free(s, page, x, addr);
1967 local_irq_restore(flags);
1970 void kmem_cache_free(struct kmem_cache *s, void *x)
1974 page = virt_to_head_page(x);
1976 slab_free(s, page, x, _RET_IP_);
1978 trace_kmem_cache_free(_RET_IP_, x);
1980 EXPORT_SYMBOL(kmem_cache_free);
1983 * Object placement in a slab is made very easy because we always start at
1984 * offset 0. If we tune the size of the object to the alignment then we can
1985 * get the required alignment by putting one properly sized object after
1988 * Notice that the allocation order determines the sizes of the per cpu
1989 * caches. Each processor has always one slab available for allocations.
1990 * Increasing the allocation order reduces the number of times that slabs
1991 * must be moved on and off the partial lists and is therefore a factor in
1996 * Mininum / Maximum order of slab pages. This influences locking overhead
1997 * and slab fragmentation. A higher order reduces the number of partial slabs
1998 * and increases the number of allocations possible without having to
1999 * take the list_lock.
2001 static int slub_min_order;
2002 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2003 static int slub_min_objects;
2006 * Merge control. If this is set then no merging of slab caches will occur.
2007 * (Could be removed. This was introduced to pacify the merge skeptics.)
2009 static int slub_nomerge;
2012 * Calculate the order of allocation given an slab object size.
2014 * The order of allocation has significant impact on performance and other
2015 * system components. Generally order 0 allocations should be preferred since
2016 * order 0 does not cause fragmentation in the page allocator. Larger objects
2017 * be problematic to put into order 0 slabs because there may be too much
2018 * unused space left. We go to a higher order if more than 1/16th of the slab
2021 * In order to reach satisfactory performance we must ensure that a minimum
2022 * number of objects is in one slab. Otherwise we may generate too much
2023 * activity on the partial lists which requires taking the list_lock. This is
2024 * less a concern for large slabs though which are rarely used.
2026 * slub_max_order specifies the order where we begin to stop considering the
2027 * number of objects in a slab as critical. If we reach slub_max_order then
2028 * we try to keep the page order as low as possible. So we accept more waste
2029 * of space in favor of a small page order.
2031 * Higher order allocations also allow the placement of more objects in a
2032 * slab and thereby reduce object handling overhead. If the user has
2033 * requested a higher mininum order then we start with that one instead of
2034 * the smallest order which will fit the object.
2036 static inline int slab_order(int size, int min_objects,
2037 int max_order, int fract_leftover, int reserved)
2041 int min_order = slub_min_order;
2043 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2044 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2046 for (order = max(min_order,
2047 fls(min_objects * size - 1) - PAGE_SHIFT);
2048 order <= max_order; order++) {
2050 unsigned long slab_size = PAGE_SIZE << order;
2052 if (slab_size < min_objects * size + reserved)
2055 rem = (slab_size - reserved) % size;
2057 if (rem <= slab_size / fract_leftover)
2065 static inline int calculate_order(int size, int reserved)
2073 * Attempt to find best configuration for a slab. This
2074 * works by first attempting to generate a layout with
2075 * the best configuration and backing off gradually.
2077 * First we reduce the acceptable waste in a slab. Then
2078 * we reduce the minimum objects required in a slab.
2080 min_objects = slub_min_objects;
2082 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2083 max_objects = order_objects(slub_max_order, size, reserved);
2084 min_objects = min(min_objects, max_objects);
2086 while (min_objects > 1) {
2088 while (fraction >= 4) {
2089 order = slab_order(size, min_objects,
2090 slub_max_order, fraction, reserved);
2091 if (order <= slub_max_order)
2099 * We were unable to place multiple objects in a slab. Now
2100 * lets see if we can place a single object there.
2102 order = slab_order(size, 1, slub_max_order, 1, reserved);
2103 if (order <= slub_max_order)
2107 * Doh this slab cannot be placed using slub_max_order.
2109 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2110 if (order < MAX_ORDER)
2116 * Figure out what the alignment of the objects will be.
2118 static unsigned long calculate_alignment(unsigned long flags,
2119 unsigned long align, unsigned long size)
2122 * If the user wants hardware cache aligned objects then follow that
2123 * suggestion if the object is sufficiently large.
2125 * The hardware cache alignment cannot override the specified
2126 * alignment though. If that is greater then use it.
2128 if (flags & SLAB_HWCACHE_ALIGN) {
2129 unsigned long ralign = cache_line_size();
2130 while (size <= ralign / 2)
2132 align = max(align, ralign);
2135 if (align < ARCH_SLAB_MINALIGN)
2136 align = ARCH_SLAB_MINALIGN;
2138 return ALIGN(align, sizeof(void *));
2142 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2145 spin_lock_init(&n->list_lock);
2146 INIT_LIST_HEAD(&n->partial);
2147 #ifdef CONFIG_SLUB_DEBUG
2148 atomic_long_set(&n->nr_slabs, 0);
2149 atomic_long_set(&n->total_objects, 0);
2150 INIT_LIST_HEAD(&n->full);
2154 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2156 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2157 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2159 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2161 return s->cpu_slab != NULL;
2164 static struct kmem_cache *kmem_cache_node;
2167 * No kmalloc_node yet so do it by hand. We know that this is the first
2168 * slab on the node for this slabcache. There are no concurrent accesses
2171 * Note that this function only works on the kmalloc_node_cache
2172 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2173 * memory on a fresh node that has no slab structures yet.
2175 static void early_kmem_cache_node_alloc(int node)
2178 struct kmem_cache_node *n;
2179 unsigned long flags;
2181 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2183 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2186 if (page_to_nid(page) != node) {
2187 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2189 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2190 "in order to be able to continue\n");
2195 page->freelist = get_freepointer(kmem_cache_node, n);
2197 kmem_cache_node->node[node] = n;
2198 #ifdef CONFIG_SLUB_DEBUG
2199 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2200 init_tracking(kmem_cache_node, n);
2202 init_kmem_cache_node(n, kmem_cache_node);
2203 inc_slabs_node(kmem_cache_node, node, page->objects);
2206 * lockdep requires consistent irq usage for each lock
2207 * so even though there cannot be a race this early in
2208 * the boot sequence, we still disable irqs.
2210 local_irq_save(flags);
2211 add_partial(n, page, 0);
2212 local_irq_restore(flags);
2215 static void free_kmem_cache_nodes(struct kmem_cache *s)
2219 for_each_node_state(node, N_NORMAL_MEMORY) {
2220 struct kmem_cache_node *n = s->node[node];
2223 kmem_cache_free(kmem_cache_node, n);
2225 s->node[node] = NULL;
2229 static int init_kmem_cache_nodes(struct kmem_cache *s)
2233 for_each_node_state(node, N_NORMAL_MEMORY) {
2234 struct kmem_cache_node *n;
2236 if (slab_state == DOWN) {
2237 early_kmem_cache_node_alloc(node);
2240 n = kmem_cache_alloc_node(kmem_cache_node,
2244 free_kmem_cache_nodes(s);
2249 init_kmem_cache_node(n, s);
2254 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2256 if (min < MIN_PARTIAL)
2258 else if (min > MAX_PARTIAL)
2260 s->min_partial = min;
2264 * calculate_sizes() determines the order and the distribution of data within
2267 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2269 unsigned long flags = s->flags;
2270 unsigned long size = s->objsize;
2271 unsigned long align = s->align;
2275 * Round up object size to the next word boundary. We can only
2276 * place the free pointer at word boundaries and this determines
2277 * the possible location of the free pointer.
2279 size = ALIGN(size, sizeof(void *));
2281 #ifdef CONFIG_SLUB_DEBUG
2283 * Determine if we can poison the object itself. If the user of
2284 * the slab may touch the object after free or before allocation
2285 * then we should never poison the object itself.
2287 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2289 s->flags |= __OBJECT_POISON;
2291 s->flags &= ~__OBJECT_POISON;
2295 * If we are Redzoning then check if there is some space between the
2296 * end of the object and the free pointer. If not then add an
2297 * additional word to have some bytes to store Redzone information.
2299 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2300 size += sizeof(void *);
2304 * With that we have determined the number of bytes in actual use
2305 * by the object. This is the potential offset to the free pointer.
2309 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2312 * Relocate free pointer after the object if it is not
2313 * permitted to overwrite the first word of the object on
2316 * This is the case if we do RCU, have a constructor or
2317 * destructor or are poisoning the objects.
2320 size += sizeof(void *);
2323 #ifdef CONFIG_SLUB_DEBUG
2324 if (flags & SLAB_STORE_USER)
2326 * Need to store information about allocs and frees after
2329 size += 2 * sizeof(struct track);
2331 if (flags & SLAB_RED_ZONE)
2333 * Add some empty padding so that we can catch
2334 * overwrites from earlier objects rather than let
2335 * tracking information or the free pointer be
2336 * corrupted if a user writes before the start
2339 size += sizeof(void *);
2343 * Determine the alignment based on various parameters that the
2344 * user specified and the dynamic determination of cache line size
2347 align = calculate_alignment(flags, align, s->objsize);
2351 * SLUB stores one object immediately after another beginning from
2352 * offset 0. In order to align the objects we have to simply size
2353 * each object to conform to the alignment.
2355 size = ALIGN(size, align);
2357 if (forced_order >= 0)
2358 order = forced_order;
2360 order = calculate_order(size, s->reserved);
2367 s->allocflags |= __GFP_COMP;
2369 if (s->flags & SLAB_CACHE_DMA)
2370 s->allocflags |= SLUB_DMA;
2372 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2373 s->allocflags |= __GFP_RECLAIMABLE;
2376 * Determine the number of objects per slab
2378 s->oo = oo_make(order, size, s->reserved);
2379 s->min = oo_make(get_order(size), size, s->reserved);
2380 if (oo_objects(s->oo) > oo_objects(s->max))
2383 return !!oo_objects(s->oo);
2387 static int kmem_cache_open(struct kmem_cache *s,
2388 const char *name, size_t size,
2389 size_t align, unsigned long flags,
2390 void (*ctor)(void *))
2392 memset(s, 0, kmem_size);
2397 s->flags = kmem_cache_flags(size, flags, name, ctor);
2400 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2401 s->reserved = sizeof(struct rcu_head);
2403 if (!calculate_sizes(s, -1))
2405 if (disable_higher_order_debug) {
2407 * Disable debugging flags that store metadata if the min slab
2410 if (get_order(s->size) > get_order(s->objsize)) {
2411 s->flags &= ~DEBUG_METADATA_FLAGS;
2413 if (!calculate_sizes(s, -1))
2419 * The larger the object size is, the more pages we want on the partial
2420 * list to avoid pounding the page allocator excessively.
2422 set_min_partial(s, ilog2(s->size));
2425 s->remote_node_defrag_ratio = 1000;
2427 if (!init_kmem_cache_nodes(s))
2430 if (alloc_kmem_cache_cpus(s))
2433 free_kmem_cache_nodes(s);
2435 if (flags & SLAB_PANIC)
2436 panic("Cannot create slab %s size=%lu realsize=%u "
2437 "order=%u offset=%u flags=%lx\n",
2438 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2444 * Determine the size of a slab object
2446 unsigned int kmem_cache_size(struct kmem_cache *s)
2450 EXPORT_SYMBOL(kmem_cache_size);
2452 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2455 #ifdef CONFIG_SLUB_DEBUG
2456 void *addr = page_address(page);
2458 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2459 sizeof(long), GFP_ATOMIC);
2462 slab_err(s, page, "%s", text);
2464 for_each_free_object(p, s, page->freelist)
2465 set_bit(slab_index(p, s, addr), map);
2467 for_each_object(p, s, addr, page->objects) {
2469 if (!test_bit(slab_index(p, s, addr), map)) {
2470 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2472 print_tracking(s, p);
2481 * Attempt to free all partial slabs on a node.
2483 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2485 unsigned long flags;
2486 struct page *page, *h;
2488 spin_lock_irqsave(&n->list_lock, flags);
2489 list_for_each_entry_safe(page, h, &n->partial, lru) {
2491 __remove_partial(n, page);
2492 discard_slab(s, page);
2494 list_slab_objects(s, page,
2495 "Objects remaining on kmem_cache_close()");
2498 spin_unlock_irqrestore(&n->list_lock, flags);
2502 * Release all resources used by a slab cache.
2504 static inline int kmem_cache_close(struct kmem_cache *s)
2509 free_percpu(s->cpu_slab);
2510 /* Attempt to free all objects */
2511 for_each_node_state(node, N_NORMAL_MEMORY) {
2512 struct kmem_cache_node *n = get_node(s, node);
2515 if (n->nr_partial || slabs_node(s, node))
2518 free_kmem_cache_nodes(s);
2523 * Close a cache and release the kmem_cache structure
2524 * (must be used for caches created using kmem_cache_create)
2526 void kmem_cache_destroy(struct kmem_cache *s)
2528 down_write(&slub_lock);
2532 if (kmem_cache_close(s)) {
2533 printk(KERN_ERR "SLUB %s: %s called for cache that "
2534 "still has objects.\n", s->name, __func__);
2537 if (s->flags & SLAB_DESTROY_BY_RCU)
2539 sysfs_slab_remove(s);
2541 up_write(&slub_lock);
2543 EXPORT_SYMBOL(kmem_cache_destroy);
2545 /********************************************************************
2547 *******************************************************************/
2549 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2550 EXPORT_SYMBOL(kmalloc_caches);
2552 static struct kmem_cache *kmem_cache;
2554 #ifdef CONFIG_ZONE_DMA
2555 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2558 static int __init setup_slub_min_order(char *str)
2560 get_option(&str, &slub_min_order);
2565 __setup("slub_min_order=", setup_slub_min_order);
2567 static int __init setup_slub_max_order(char *str)
2569 get_option(&str, &slub_max_order);
2570 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2575 __setup("slub_max_order=", setup_slub_max_order);
2577 static int __init setup_slub_min_objects(char *str)
2579 get_option(&str, &slub_min_objects);
2584 __setup("slub_min_objects=", setup_slub_min_objects);
2586 static int __init setup_slub_nomerge(char *str)
2592 __setup("slub_nomerge", setup_slub_nomerge);
2594 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2595 int size, unsigned int flags)
2597 struct kmem_cache *s;
2599 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2602 * This function is called with IRQs disabled during early-boot on
2603 * single CPU so there's no need to take slub_lock here.
2605 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2609 list_add(&s->list, &slab_caches);
2613 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2618 * Conversion table for small slabs sizes / 8 to the index in the
2619 * kmalloc array. This is necessary for slabs < 192 since we have non power
2620 * of two cache sizes there. The size of larger slabs can be determined using
2623 static s8 size_index[24] = {
2650 static inline int size_index_elem(size_t bytes)
2652 return (bytes - 1) / 8;
2655 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2661 return ZERO_SIZE_PTR;
2663 index = size_index[size_index_elem(size)];
2665 index = fls(size - 1);
2667 #ifdef CONFIG_ZONE_DMA
2668 if (unlikely((flags & SLUB_DMA)))
2669 return kmalloc_dma_caches[index];
2672 return kmalloc_caches[index];
2675 void *__kmalloc(size_t size, gfp_t flags)
2677 struct kmem_cache *s;
2680 if (unlikely(size > SLUB_MAX_SIZE))
2681 return kmalloc_large(size, flags);
2683 s = get_slab(size, flags);
2685 if (unlikely(ZERO_OR_NULL_PTR(s)))
2688 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2690 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2694 EXPORT_SYMBOL(__kmalloc);
2697 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2702 flags |= __GFP_COMP | __GFP_NOTRACK;
2703 page = alloc_pages_node(node, flags, get_order(size));
2705 ptr = page_address(page);
2707 kmemleak_alloc(ptr, size, 1, flags);
2711 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2713 struct kmem_cache *s;
2716 if (unlikely(size > SLUB_MAX_SIZE)) {
2717 ret = kmalloc_large_node(size, flags, node);
2719 trace_kmalloc_node(_RET_IP_, ret,
2720 size, PAGE_SIZE << get_order(size),
2726 s = get_slab(size, flags);
2728 if (unlikely(ZERO_OR_NULL_PTR(s)))
2731 ret = slab_alloc(s, flags, node, _RET_IP_);
2733 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2737 EXPORT_SYMBOL(__kmalloc_node);
2740 size_t ksize(const void *object)
2744 if (unlikely(object == ZERO_SIZE_PTR))
2747 page = virt_to_head_page(object);
2749 if (unlikely(!PageSlab(page))) {
2750 WARN_ON(!PageCompound(page));
2751 return PAGE_SIZE << compound_order(page);
2754 return slab_ksize(page->slab);
2756 EXPORT_SYMBOL(ksize);
2758 void kfree(const void *x)
2761 void *object = (void *)x;
2763 trace_kfree(_RET_IP_, x);
2765 if (unlikely(ZERO_OR_NULL_PTR(x)))
2768 page = virt_to_head_page(x);
2769 if (unlikely(!PageSlab(page))) {
2770 BUG_ON(!PageCompound(page));
2775 slab_free(page->slab, page, object, _RET_IP_);
2777 EXPORT_SYMBOL(kfree);
2780 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2781 * the remaining slabs by the number of items in use. The slabs with the
2782 * most items in use come first. New allocations will then fill those up
2783 * and thus they can be removed from the partial lists.
2785 * The slabs with the least items are placed last. This results in them
2786 * being allocated from last increasing the chance that the last objects
2787 * are freed in them.
2789 int kmem_cache_shrink(struct kmem_cache *s)
2793 struct kmem_cache_node *n;
2796 int objects = oo_objects(s->max);
2797 struct list_head *slabs_by_inuse =
2798 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2799 unsigned long flags;
2801 if (!slabs_by_inuse)
2805 for_each_node_state(node, N_NORMAL_MEMORY) {
2806 n = get_node(s, node);
2811 for (i = 0; i < objects; i++)
2812 INIT_LIST_HEAD(slabs_by_inuse + i);
2814 spin_lock_irqsave(&n->list_lock, flags);
2817 * Build lists indexed by the items in use in each slab.
2819 * Note that concurrent frees may occur while we hold the
2820 * list_lock. page->inuse here is the upper limit.
2822 list_for_each_entry_safe(page, t, &n->partial, lru) {
2823 if (!page->inuse && slab_trylock(page)) {
2825 * Must hold slab lock here because slab_free
2826 * may have freed the last object and be
2827 * waiting to release the slab.
2829 __remove_partial(n, page);
2831 discard_slab(s, page);
2833 list_move(&page->lru,
2834 slabs_by_inuse + page->inuse);
2839 * Rebuild the partial list with the slabs filled up most
2840 * first and the least used slabs at the end.
2842 for (i = objects - 1; i >= 0; i--)
2843 list_splice(slabs_by_inuse + i, n->partial.prev);
2845 spin_unlock_irqrestore(&n->list_lock, flags);
2848 kfree(slabs_by_inuse);
2851 EXPORT_SYMBOL(kmem_cache_shrink);
2853 #if defined(CONFIG_MEMORY_HOTPLUG)
2854 static int slab_mem_going_offline_callback(void *arg)
2856 struct kmem_cache *s;
2858 down_read(&slub_lock);
2859 list_for_each_entry(s, &slab_caches, list)
2860 kmem_cache_shrink(s);
2861 up_read(&slub_lock);
2866 static void slab_mem_offline_callback(void *arg)
2868 struct kmem_cache_node *n;
2869 struct kmem_cache *s;
2870 struct memory_notify *marg = arg;
2873 offline_node = marg->status_change_nid;
2876 * If the node still has available memory. we need kmem_cache_node
2879 if (offline_node < 0)
2882 down_read(&slub_lock);
2883 list_for_each_entry(s, &slab_caches, list) {
2884 n = get_node(s, offline_node);
2887 * if n->nr_slabs > 0, slabs still exist on the node
2888 * that is going down. We were unable to free them,
2889 * and offline_pages() function shouldn't call this
2890 * callback. So, we must fail.
2892 BUG_ON(slabs_node(s, offline_node));
2894 s->node[offline_node] = NULL;
2895 kmem_cache_free(kmem_cache_node, n);
2898 up_read(&slub_lock);
2901 static int slab_mem_going_online_callback(void *arg)
2903 struct kmem_cache_node *n;
2904 struct kmem_cache *s;
2905 struct memory_notify *marg = arg;
2906 int nid = marg->status_change_nid;
2910 * If the node's memory is already available, then kmem_cache_node is
2911 * already created. Nothing to do.
2917 * We are bringing a node online. No memory is available yet. We must
2918 * allocate a kmem_cache_node structure in order to bring the node
2921 down_read(&slub_lock);
2922 list_for_each_entry(s, &slab_caches, list) {
2924 * XXX: kmem_cache_alloc_node will fallback to other nodes
2925 * since memory is not yet available from the node that
2928 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
2933 init_kmem_cache_node(n, s);
2937 up_read(&slub_lock);
2941 static int slab_memory_callback(struct notifier_block *self,
2942 unsigned long action, void *arg)
2947 case MEM_GOING_ONLINE:
2948 ret = slab_mem_going_online_callback(arg);
2950 case MEM_GOING_OFFLINE:
2951 ret = slab_mem_going_offline_callback(arg);
2954 case MEM_CANCEL_ONLINE:
2955 slab_mem_offline_callback(arg);
2958 case MEM_CANCEL_OFFLINE:
2962 ret = notifier_from_errno(ret);
2968 #endif /* CONFIG_MEMORY_HOTPLUG */
2970 /********************************************************************
2971 * Basic setup of slabs
2972 *******************************************************************/
2975 * Used for early kmem_cache structures that were allocated using
2976 * the page allocator
2979 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
2983 list_add(&s->list, &slab_caches);
2986 for_each_node_state(node, N_NORMAL_MEMORY) {
2987 struct kmem_cache_node *n = get_node(s, node);
2991 list_for_each_entry(p, &n->partial, lru)
2994 #ifdef CONFIG_SLAB_DEBUG
2995 list_for_each_entry(p, &n->full, lru)
3002 void __init kmem_cache_init(void)
3006 struct kmem_cache *temp_kmem_cache;
3008 struct kmem_cache *temp_kmem_cache_node;
3009 unsigned long kmalloc_size;
3011 kmem_size = offsetof(struct kmem_cache, node) +
3012 nr_node_ids * sizeof(struct kmem_cache_node *);
3014 /* Allocate two kmem_caches from the page allocator */
3015 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3016 order = get_order(2 * kmalloc_size);
3017 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3020 * Must first have the slab cache available for the allocations of the
3021 * struct kmem_cache_node's. There is special bootstrap code in
3022 * kmem_cache_open for slab_state == DOWN.
3024 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3026 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3027 sizeof(struct kmem_cache_node),
3028 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3030 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3032 /* Able to allocate the per node structures */
3033 slab_state = PARTIAL;
3035 temp_kmem_cache = kmem_cache;
3036 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3037 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3038 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3039 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3042 * Allocate kmem_cache_node properly from the kmem_cache slab.
3043 * kmem_cache_node is separately allocated so no need to
3044 * update any list pointers.
3046 temp_kmem_cache_node = kmem_cache_node;
3048 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3049 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3051 kmem_cache_bootstrap_fixup(kmem_cache_node);
3054 kmem_cache_bootstrap_fixup(kmem_cache);
3056 /* Free temporary boot structure */
3057 free_pages((unsigned long)temp_kmem_cache, order);
3059 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3062 * Patch up the size_index table if we have strange large alignment
3063 * requirements for the kmalloc array. This is only the case for
3064 * MIPS it seems. The standard arches will not generate any code here.
3066 * Largest permitted alignment is 256 bytes due to the way we
3067 * handle the index determination for the smaller caches.
3069 * Make sure that nothing crazy happens if someone starts tinkering
3070 * around with ARCH_KMALLOC_MINALIGN
3072 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3073 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3075 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3076 int elem = size_index_elem(i);
3077 if (elem >= ARRAY_SIZE(size_index))
3079 size_index[elem] = KMALLOC_SHIFT_LOW;
3082 if (KMALLOC_MIN_SIZE == 64) {
3084 * The 96 byte size cache is not used if the alignment
3087 for (i = 64 + 8; i <= 96; i += 8)
3088 size_index[size_index_elem(i)] = 7;
3089 } else if (KMALLOC_MIN_SIZE == 128) {
3091 * The 192 byte sized cache is not used if the alignment
3092 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3095 for (i = 128 + 8; i <= 192; i += 8)
3096 size_index[size_index_elem(i)] = 8;
3099 /* Caches that are not of the two-to-the-power-of size */
3100 if (KMALLOC_MIN_SIZE <= 32) {
3101 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3105 if (KMALLOC_MIN_SIZE <= 64) {
3106 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3110 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3111 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3117 /* Provide the correct kmalloc names now that the caches are up */
3118 if (KMALLOC_MIN_SIZE <= 32) {
3119 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3120 BUG_ON(!kmalloc_caches[1]->name);
3123 if (KMALLOC_MIN_SIZE <= 64) {
3124 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3125 BUG_ON(!kmalloc_caches[2]->name);
3128 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3129 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3132 kmalloc_caches[i]->name = s;
3136 register_cpu_notifier(&slab_notifier);
3139 #ifdef CONFIG_ZONE_DMA
3140 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3141 struct kmem_cache *s = kmalloc_caches[i];
3144 char *name = kasprintf(GFP_NOWAIT,
3145 "dma-kmalloc-%d", s->objsize);
3148 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3149 s->objsize, SLAB_CACHE_DMA);
3154 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3155 " CPUs=%d, Nodes=%d\n",
3156 caches, cache_line_size(),
3157 slub_min_order, slub_max_order, slub_min_objects,
3158 nr_cpu_ids, nr_node_ids);
3161 void __init kmem_cache_init_late(void)
3166 * Find a mergeable slab cache
3168 static int slab_unmergeable(struct kmem_cache *s)
3170 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3177 * We may have set a slab to be unmergeable during bootstrap.
3179 if (s->refcount < 0)
3185 static struct kmem_cache *find_mergeable(size_t size,
3186 size_t align, unsigned long flags, const char *name,
3187 void (*ctor)(void *))
3189 struct kmem_cache *s;
3191 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3197 size = ALIGN(size, sizeof(void *));
3198 align = calculate_alignment(flags, align, size);
3199 size = ALIGN(size, align);
3200 flags = kmem_cache_flags(size, flags, name, NULL);
3202 list_for_each_entry(s, &slab_caches, list) {
3203 if (slab_unmergeable(s))
3209 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3212 * Check if alignment is compatible.
3213 * Courtesy of Adrian Drzewiecki
3215 if ((s->size & ~(align - 1)) != s->size)
3218 if (s->size - size >= sizeof(void *))
3226 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3227 size_t align, unsigned long flags, void (*ctor)(void *))
3229 struct kmem_cache *s;
3235 down_write(&slub_lock);
3236 s = find_mergeable(size, align, flags, name, ctor);
3240 * Adjust the object sizes so that we clear
3241 * the complete object on kzalloc.
3243 s->objsize = max(s->objsize, (int)size);
3244 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3246 if (sysfs_slab_alias(s, name)) {
3250 up_write(&slub_lock);
3254 n = kstrdup(name, GFP_KERNEL);
3258 s = kmalloc(kmem_size, GFP_KERNEL);
3260 if (kmem_cache_open(s, n,
3261 size, align, flags, ctor)) {
3262 list_add(&s->list, &slab_caches);
3263 if (sysfs_slab_add(s)) {
3269 up_write(&slub_lock);
3276 up_write(&slub_lock);
3278 if (flags & SLAB_PANIC)
3279 panic("Cannot create slabcache %s\n", name);
3284 EXPORT_SYMBOL(kmem_cache_create);
3288 * Use the cpu notifier to insure that the cpu slabs are flushed when
3291 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3292 unsigned long action, void *hcpu)
3294 long cpu = (long)hcpu;
3295 struct kmem_cache *s;
3296 unsigned long flags;
3299 case CPU_UP_CANCELED:
3300 case CPU_UP_CANCELED_FROZEN:
3302 case CPU_DEAD_FROZEN:
3303 down_read(&slub_lock);
3304 list_for_each_entry(s, &slab_caches, list) {
3305 local_irq_save(flags);
3306 __flush_cpu_slab(s, cpu);
3307 local_irq_restore(flags);
3309 up_read(&slub_lock);
3317 static struct notifier_block __cpuinitdata slab_notifier = {
3318 .notifier_call = slab_cpuup_callback
3323 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3325 struct kmem_cache *s;
3328 if (unlikely(size > SLUB_MAX_SIZE))
3329 return kmalloc_large(size, gfpflags);
3331 s = get_slab(size, gfpflags);
3333 if (unlikely(ZERO_OR_NULL_PTR(s)))
3336 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3338 /* Honor the call site pointer we recieved. */
3339 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3345 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3346 int node, unsigned long caller)
3348 struct kmem_cache *s;
3351 if (unlikely(size > SLUB_MAX_SIZE)) {
3352 ret = kmalloc_large_node(size, gfpflags, node);
3354 trace_kmalloc_node(caller, ret,
3355 size, PAGE_SIZE << get_order(size),
3361 s = get_slab(size, gfpflags);
3363 if (unlikely(ZERO_OR_NULL_PTR(s)))
3366 ret = slab_alloc(s, gfpflags, node, caller);
3368 /* Honor the call site pointer we recieved. */
3369 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3376 static int count_inuse(struct page *page)
3381 static int count_total(struct page *page)
3383 return page->objects;
3387 #ifdef CONFIG_SLUB_DEBUG
3388 static int validate_slab(struct kmem_cache *s, struct page *page,
3392 void *addr = page_address(page);
3394 if (!check_slab(s, page) ||
3395 !on_freelist(s, page, NULL))
3398 /* Now we know that a valid freelist exists */
3399 bitmap_zero(map, page->objects);
3401 for_each_free_object(p, s, page->freelist) {
3402 set_bit(slab_index(p, s, addr), map);
3403 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3407 for_each_object(p, s, addr, page->objects)
3408 if (!test_bit(slab_index(p, s, addr), map))
3409 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3414 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3417 if (slab_trylock(page)) {
3418 validate_slab(s, page, map);
3421 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3425 static int validate_slab_node(struct kmem_cache *s,
3426 struct kmem_cache_node *n, unsigned long *map)
3428 unsigned long count = 0;
3430 unsigned long flags;
3432 spin_lock_irqsave(&n->list_lock, flags);
3434 list_for_each_entry(page, &n->partial, lru) {
3435 validate_slab_slab(s, page, map);
3438 if (count != n->nr_partial)
3439 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3440 "counter=%ld\n", s->name, count, n->nr_partial);
3442 if (!(s->flags & SLAB_STORE_USER))
3445 list_for_each_entry(page, &n->full, lru) {
3446 validate_slab_slab(s, page, map);
3449 if (count != atomic_long_read(&n->nr_slabs))
3450 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3451 "counter=%ld\n", s->name, count,
3452 atomic_long_read(&n->nr_slabs));
3455 spin_unlock_irqrestore(&n->list_lock, flags);
3459 static long validate_slab_cache(struct kmem_cache *s)
3462 unsigned long count = 0;
3463 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3464 sizeof(unsigned long), GFP_KERNEL);
3470 for_each_node_state(node, N_NORMAL_MEMORY) {
3471 struct kmem_cache_node *n = get_node(s, node);
3473 count += validate_slab_node(s, n, map);
3479 * Generate lists of code addresses where slabcache objects are allocated
3484 unsigned long count;
3491 DECLARE_BITMAP(cpus, NR_CPUS);
3497 unsigned long count;
3498 struct location *loc;
3501 static void free_loc_track(struct loc_track *t)
3504 free_pages((unsigned long)t->loc,
3505 get_order(sizeof(struct location) * t->max));
3508 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3513 order = get_order(sizeof(struct location) * max);
3515 l = (void *)__get_free_pages(flags, order);
3520 memcpy(l, t->loc, sizeof(struct location) * t->count);
3528 static int add_location(struct loc_track *t, struct kmem_cache *s,
3529 const struct track *track)
3531 long start, end, pos;
3533 unsigned long caddr;
3534 unsigned long age = jiffies - track->when;
3540 pos = start + (end - start + 1) / 2;
3543 * There is nothing at "end". If we end up there
3544 * we need to add something to before end.
3549 caddr = t->loc[pos].addr;
3550 if (track->addr == caddr) {
3556 if (age < l->min_time)
3558 if (age > l->max_time)
3561 if (track->pid < l->min_pid)
3562 l->min_pid = track->pid;
3563 if (track->pid > l->max_pid)
3564 l->max_pid = track->pid;
3566 cpumask_set_cpu(track->cpu,
3567 to_cpumask(l->cpus));
3569 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3573 if (track->addr < caddr)
3580 * Not found. Insert new tracking element.
3582 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3588 (t->count - pos) * sizeof(struct location));
3591 l->addr = track->addr;
3595 l->min_pid = track->pid;
3596 l->max_pid = track->pid;
3597 cpumask_clear(to_cpumask(l->cpus));
3598 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3599 nodes_clear(l->nodes);
3600 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3604 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3605 struct page *page, enum track_item alloc,
3608 void *addr = page_address(page);
3611 bitmap_zero(map, page->objects);
3612 for_each_free_object(p, s, page->freelist)
3613 set_bit(slab_index(p, s, addr), map);
3615 for_each_object(p, s, addr, page->objects)
3616 if (!test_bit(slab_index(p, s, addr), map))
3617 add_location(t, s, get_track(s, p, alloc));
3620 static int list_locations(struct kmem_cache *s, char *buf,
3621 enum track_item alloc)
3625 struct loc_track t = { 0, 0, NULL };
3627 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3628 sizeof(unsigned long), GFP_KERNEL);
3630 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3633 return sprintf(buf, "Out of memory\n");
3635 /* Push back cpu slabs */
3638 for_each_node_state(node, N_NORMAL_MEMORY) {
3639 struct kmem_cache_node *n = get_node(s, node);
3640 unsigned long flags;
3643 if (!atomic_long_read(&n->nr_slabs))
3646 spin_lock_irqsave(&n->list_lock, flags);
3647 list_for_each_entry(page, &n->partial, lru)
3648 process_slab(&t, s, page, alloc, map);
3649 list_for_each_entry(page, &n->full, lru)
3650 process_slab(&t, s, page, alloc, map);
3651 spin_unlock_irqrestore(&n->list_lock, flags);
3654 for (i = 0; i < t.count; i++) {
3655 struct location *l = &t.loc[i];
3657 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3659 len += sprintf(buf + len, "%7ld ", l->count);
3662 len += sprintf(buf + len, "%pS", (void *)l->addr);
3664 len += sprintf(buf + len, "<not-available>");
3666 if (l->sum_time != l->min_time) {
3667 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3669 (long)div_u64(l->sum_time, l->count),
3672 len += sprintf(buf + len, " age=%ld",
3675 if (l->min_pid != l->max_pid)
3676 len += sprintf(buf + len, " pid=%ld-%ld",
3677 l->min_pid, l->max_pid);
3679 len += sprintf(buf + len, " pid=%ld",
3682 if (num_online_cpus() > 1 &&
3683 !cpumask_empty(to_cpumask(l->cpus)) &&
3684 len < PAGE_SIZE - 60) {
3685 len += sprintf(buf + len, " cpus=");
3686 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3687 to_cpumask(l->cpus));
3690 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3691 len < PAGE_SIZE - 60) {
3692 len += sprintf(buf + len, " nodes=");
3693 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3697 len += sprintf(buf + len, "\n");
3703 len += sprintf(buf, "No data\n");
3708 #ifdef SLUB_RESILIENCY_TEST
3709 static void resiliency_test(void)
3713 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3715 printk(KERN_ERR "SLUB resiliency testing\n");
3716 printk(KERN_ERR "-----------------------\n");
3717 printk(KERN_ERR "A. Corruption after allocation\n");
3719 p = kzalloc(16, GFP_KERNEL);
3721 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3722 " 0x12->0x%p\n\n", p + 16);
3724 validate_slab_cache(kmalloc_caches[4]);
3726 /* Hmmm... The next two are dangerous */
3727 p = kzalloc(32, GFP_KERNEL);
3728 p[32 + sizeof(void *)] = 0x34;
3729 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3730 " 0x34 -> -0x%p\n", p);
3732 "If allocated object is overwritten then not detectable\n\n");
3734 validate_slab_cache(kmalloc_caches[5]);
3735 p = kzalloc(64, GFP_KERNEL);
3736 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3738 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3741 "If allocated object is overwritten then not detectable\n\n");
3742 validate_slab_cache(kmalloc_caches[6]);
3744 printk(KERN_ERR "\nB. Corruption after free\n");
3745 p = kzalloc(128, GFP_KERNEL);
3748 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3749 validate_slab_cache(kmalloc_caches[7]);
3751 p = kzalloc(256, GFP_KERNEL);
3754 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3756 validate_slab_cache(kmalloc_caches[8]);
3758 p = kzalloc(512, GFP_KERNEL);
3761 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3762 validate_slab_cache(kmalloc_caches[9]);
3766 static void resiliency_test(void) {};
3771 enum slab_stat_type {
3772 SL_ALL, /* All slabs */
3773 SL_PARTIAL, /* Only partially allocated slabs */
3774 SL_CPU, /* Only slabs used for cpu caches */
3775 SL_OBJECTS, /* Determine allocated objects not slabs */
3776 SL_TOTAL /* Determine object capacity not slabs */
3779 #define SO_ALL (1 << SL_ALL)
3780 #define SO_PARTIAL (1 << SL_PARTIAL)
3781 #define SO_CPU (1 << SL_CPU)
3782 #define SO_OBJECTS (1 << SL_OBJECTS)
3783 #define SO_TOTAL (1 << SL_TOTAL)
3785 static ssize_t show_slab_objects(struct kmem_cache *s,
3786 char *buf, unsigned long flags)
3788 unsigned long total = 0;
3791 unsigned long *nodes;
3792 unsigned long *per_cpu;
3794 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3797 per_cpu = nodes + nr_node_ids;
3799 if (flags & SO_CPU) {
3802 for_each_possible_cpu(cpu) {
3803 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3805 if (!c || c->node < 0)
3809 if (flags & SO_TOTAL)
3810 x = c->page->objects;
3811 else if (flags & SO_OBJECTS)
3817 nodes[c->node] += x;
3823 lock_memory_hotplug();
3824 #ifdef CONFIG_SLUB_DEBUG
3825 if (flags & SO_ALL) {
3826 for_each_node_state(node, N_NORMAL_MEMORY) {
3827 struct kmem_cache_node *n = get_node(s, node);
3829 if (flags & SO_TOTAL)
3830 x = atomic_long_read(&n->total_objects);
3831 else if (flags & SO_OBJECTS)
3832 x = atomic_long_read(&n->total_objects) -
3833 count_partial(n, count_free);
3836 x = atomic_long_read(&n->nr_slabs);
3843 if (flags & SO_PARTIAL) {
3844 for_each_node_state(node, N_NORMAL_MEMORY) {
3845 struct kmem_cache_node *n = get_node(s, node);
3847 if (flags & SO_TOTAL)
3848 x = count_partial(n, count_total);
3849 else if (flags & SO_OBJECTS)
3850 x = count_partial(n, count_inuse);
3857 x = sprintf(buf, "%lu", total);
3859 for_each_node_state(node, N_NORMAL_MEMORY)
3861 x += sprintf(buf + x, " N%d=%lu",
3864 unlock_memory_hotplug();
3866 return x + sprintf(buf + x, "\n");
3869 #ifdef CONFIG_SLUB_DEBUG
3870 static int any_slab_objects(struct kmem_cache *s)
3874 for_each_online_node(node) {
3875 struct kmem_cache_node *n = get_node(s, node);
3880 if (atomic_long_read(&n->total_objects))
3887 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3888 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3890 struct slab_attribute {
3891 struct attribute attr;
3892 ssize_t (*show)(struct kmem_cache *s, char *buf);
3893 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3896 #define SLAB_ATTR_RO(_name) \
3897 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3899 #define SLAB_ATTR(_name) \
3900 static struct slab_attribute _name##_attr = \
3901 __ATTR(_name, 0644, _name##_show, _name##_store)
3903 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3905 return sprintf(buf, "%d\n", s->size);
3907 SLAB_ATTR_RO(slab_size);
3909 static ssize_t align_show(struct kmem_cache *s, char *buf)
3911 return sprintf(buf, "%d\n", s->align);
3913 SLAB_ATTR_RO(align);
3915 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3917 return sprintf(buf, "%d\n", s->objsize);
3919 SLAB_ATTR_RO(object_size);
3921 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3923 return sprintf(buf, "%d\n", oo_objects(s->oo));
3925 SLAB_ATTR_RO(objs_per_slab);
3927 static ssize_t order_store(struct kmem_cache *s,
3928 const char *buf, size_t length)
3930 unsigned long order;
3933 err = strict_strtoul(buf, 10, &order);
3937 if (order > slub_max_order || order < slub_min_order)
3940 calculate_sizes(s, order);
3944 static ssize_t order_show(struct kmem_cache *s, char *buf)
3946 return sprintf(buf, "%d\n", oo_order(s->oo));
3950 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3952 return sprintf(buf, "%lu\n", s->min_partial);
3955 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3961 err = strict_strtoul(buf, 10, &min);
3965 set_min_partial(s, min);
3968 SLAB_ATTR(min_partial);
3970 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3974 return sprintf(buf, "%pS\n", s->ctor);
3978 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3980 return sprintf(buf, "%d\n", s->refcount - 1);
3982 SLAB_ATTR_RO(aliases);
3984 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3986 return show_slab_objects(s, buf, SO_PARTIAL);
3988 SLAB_ATTR_RO(partial);
3990 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3992 return show_slab_objects(s, buf, SO_CPU);
3994 SLAB_ATTR_RO(cpu_slabs);
3996 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3998 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4000 SLAB_ATTR_RO(objects);
4002 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4004 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4006 SLAB_ATTR_RO(objects_partial);
4008 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4010 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4013 static ssize_t reclaim_account_store(struct kmem_cache *s,
4014 const char *buf, size_t length)
4016 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4018 s->flags |= SLAB_RECLAIM_ACCOUNT;
4021 SLAB_ATTR(reclaim_account);
4023 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4025 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4027 SLAB_ATTR_RO(hwcache_align);
4029 #ifdef CONFIG_ZONE_DMA
4030 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4032 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4034 SLAB_ATTR_RO(cache_dma);
4037 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4039 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4041 SLAB_ATTR_RO(destroy_by_rcu);
4043 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4045 return sprintf(buf, "%d\n", s->reserved);
4047 SLAB_ATTR_RO(reserved);
4049 #ifdef CONFIG_SLUB_DEBUG
4050 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4052 return show_slab_objects(s, buf, SO_ALL);
4054 SLAB_ATTR_RO(slabs);
4056 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4058 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4060 SLAB_ATTR_RO(total_objects);
4062 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4064 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4067 static ssize_t sanity_checks_store(struct kmem_cache *s,
4068 const char *buf, size_t length)
4070 s->flags &= ~SLAB_DEBUG_FREE;
4072 s->flags |= SLAB_DEBUG_FREE;
4075 SLAB_ATTR(sanity_checks);
4077 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4079 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4082 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4085 s->flags &= ~SLAB_TRACE;
4087 s->flags |= SLAB_TRACE;
4092 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4094 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4097 static ssize_t red_zone_store(struct kmem_cache *s,
4098 const char *buf, size_t length)
4100 if (any_slab_objects(s))
4103 s->flags &= ~SLAB_RED_ZONE;
4105 s->flags |= SLAB_RED_ZONE;
4106 calculate_sizes(s, -1);
4109 SLAB_ATTR(red_zone);
4111 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4113 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4116 static ssize_t poison_store(struct kmem_cache *s,
4117 const char *buf, size_t length)
4119 if (any_slab_objects(s))
4122 s->flags &= ~SLAB_POISON;
4124 s->flags |= SLAB_POISON;
4125 calculate_sizes(s, -1);
4130 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4132 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4135 static ssize_t store_user_store(struct kmem_cache *s,
4136 const char *buf, size_t length)
4138 if (any_slab_objects(s))
4141 s->flags &= ~SLAB_STORE_USER;
4143 s->flags |= SLAB_STORE_USER;
4144 calculate_sizes(s, -1);
4147 SLAB_ATTR(store_user);
4149 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4154 static ssize_t validate_store(struct kmem_cache *s,
4155 const char *buf, size_t length)
4159 if (buf[0] == '1') {
4160 ret = validate_slab_cache(s);
4166 SLAB_ATTR(validate);
4168 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4170 if (!(s->flags & SLAB_STORE_USER))
4172 return list_locations(s, buf, TRACK_ALLOC);
4174 SLAB_ATTR_RO(alloc_calls);
4176 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4178 if (!(s->flags & SLAB_STORE_USER))
4180 return list_locations(s, buf, TRACK_FREE);
4182 SLAB_ATTR_RO(free_calls);
4183 #endif /* CONFIG_SLUB_DEBUG */
4185 #ifdef CONFIG_FAILSLAB
4186 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4188 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4191 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4194 s->flags &= ~SLAB_FAILSLAB;
4196 s->flags |= SLAB_FAILSLAB;
4199 SLAB_ATTR(failslab);
4202 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4207 static ssize_t shrink_store(struct kmem_cache *s,
4208 const char *buf, size_t length)
4210 if (buf[0] == '1') {
4211 int rc = kmem_cache_shrink(s);
4222 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4224 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4227 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4228 const char *buf, size_t length)
4230 unsigned long ratio;
4233 err = strict_strtoul(buf, 10, &ratio);
4238 s->remote_node_defrag_ratio = ratio * 10;
4242 SLAB_ATTR(remote_node_defrag_ratio);
4245 #ifdef CONFIG_SLUB_STATS
4246 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4248 unsigned long sum = 0;
4251 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4256 for_each_online_cpu(cpu) {
4257 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4263 len = sprintf(buf, "%lu", sum);
4266 for_each_online_cpu(cpu) {
4267 if (data[cpu] && len < PAGE_SIZE - 20)
4268 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4272 return len + sprintf(buf + len, "\n");
4275 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4279 for_each_online_cpu(cpu)
4280 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4283 #define STAT_ATTR(si, text) \
4284 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4286 return show_stat(s, buf, si); \
4288 static ssize_t text##_store(struct kmem_cache *s, \
4289 const char *buf, size_t length) \
4291 if (buf[0] != '0') \
4293 clear_stat(s, si); \
4298 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4299 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4300 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4301 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4302 STAT_ATTR(FREE_FROZEN, free_frozen);
4303 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4304 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4305 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4306 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4307 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4308 STAT_ATTR(FREE_SLAB, free_slab);
4309 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4310 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4311 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4312 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4313 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4314 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4315 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4318 static struct attribute *slab_attrs[] = {
4319 &slab_size_attr.attr,
4320 &object_size_attr.attr,
4321 &objs_per_slab_attr.attr,
4323 &min_partial_attr.attr,
4325 &objects_partial_attr.attr,
4327 &cpu_slabs_attr.attr,
4331 &hwcache_align_attr.attr,
4332 &reclaim_account_attr.attr,
4333 &destroy_by_rcu_attr.attr,
4335 &reserved_attr.attr,
4336 #ifdef CONFIG_SLUB_DEBUG
4337 &total_objects_attr.attr,
4339 &sanity_checks_attr.attr,
4341 &red_zone_attr.attr,
4343 &store_user_attr.attr,
4344 &validate_attr.attr,
4345 &alloc_calls_attr.attr,
4346 &free_calls_attr.attr,
4348 #ifdef CONFIG_ZONE_DMA
4349 &cache_dma_attr.attr,
4352 &remote_node_defrag_ratio_attr.attr,
4354 #ifdef CONFIG_SLUB_STATS
4355 &alloc_fastpath_attr.attr,
4356 &alloc_slowpath_attr.attr,
4357 &free_fastpath_attr.attr,
4358 &free_slowpath_attr.attr,
4359 &free_frozen_attr.attr,
4360 &free_add_partial_attr.attr,
4361 &free_remove_partial_attr.attr,
4362 &alloc_from_partial_attr.attr,
4363 &alloc_slab_attr.attr,
4364 &alloc_refill_attr.attr,
4365 &free_slab_attr.attr,
4366 &cpuslab_flush_attr.attr,
4367 &deactivate_full_attr.attr,
4368 &deactivate_empty_attr.attr,
4369 &deactivate_to_head_attr.attr,
4370 &deactivate_to_tail_attr.attr,
4371 &deactivate_remote_frees_attr.attr,
4372 &order_fallback_attr.attr,
4374 #ifdef CONFIG_FAILSLAB
4375 &failslab_attr.attr,
4381 static struct attribute_group slab_attr_group = {
4382 .attrs = slab_attrs,
4385 static ssize_t slab_attr_show(struct kobject *kobj,
4386 struct attribute *attr,
4389 struct slab_attribute *attribute;
4390 struct kmem_cache *s;
4393 attribute = to_slab_attr(attr);
4396 if (!attribute->show)
4399 err = attribute->show(s, buf);
4404 static ssize_t slab_attr_store(struct kobject *kobj,
4405 struct attribute *attr,
4406 const char *buf, size_t len)
4408 struct slab_attribute *attribute;
4409 struct kmem_cache *s;
4412 attribute = to_slab_attr(attr);
4415 if (!attribute->store)
4418 err = attribute->store(s, buf, len);
4423 static void kmem_cache_release(struct kobject *kobj)
4425 struct kmem_cache *s = to_slab(kobj);
4431 static const struct sysfs_ops slab_sysfs_ops = {
4432 .show = slab_attr_show,
4433 .store = slab_attr_store,
4436 static struct kobj_type slab_ktype = {
4437 .sysfs_ops = &slab_sysfs_ops,
4438 .release = kmem_cache_release
4441 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4443 struct kobj_type *ktype = get_ktype(kobj);
4445 if (ktype == &slab_ktype)
4450 static const struct kset_uevent_ops slab_uevent_ops = {
4451 .filter = uevent_filter,
4454 static struct kset *slab_kset;
4456 #define ID_STR_LENGTH 64
4458 /* Create a unique string id for a slab cache:
4460 * Format :[flags-]size
4462 static char *create_unique_id(struct kmem_cache *s)
4464 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4471 * First flags affecting slabcache operations. We will only
4472 * get here for aliasable slabs so we do not need to support
4473 * too many flags. The flags here must cover all flags that
4474 * are matched during merging to guarantee that the id is
4477 if (s->flags & SLAB_CACHE_DMA)
4479 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4481 if (s->flags & SLAB_DEBUG_FREE)
4483 if (!(s->flags & SLAB_NOTRACK))
4487 p += sprintf(p, "%07d", s->size);
4488 BUG_ON(p > name + ID_STR_LENGTH - 1);
4492 static int sysfs_slab_add(struct kmem_cache *s)
4498 if (slab_state < SYSFS)
4499 /* Defer until later */
4502 unmergeable = slab_unmergeable(s);
4505 * Slabcache can never be merged so we can use the name proper.
4506 * This is typically the case for debug situations. In that
4507 * case we can catch duplicate names easily.
4509 sysfs_remove_link(&slab_kset->kobj, s->name);
4513 * Create a unique name for the slab as a target
4516 name = create_unique_id(s);
4519 s->kobj.kset = slab_kset;
4520 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4522 kobject_put(&s->kobj);
4526 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4528 kobject_del(&s->kobj);
4529 kobject_put(&s->kobj);
4532 kobject_uevent(&s->kobj, KOBJ_ADD);
4534 /* Setup first alias */
4535 sysfs_slab_alias(s, s->name);
4541 static void sysfs_slab_remove(struct kmem_cache *s)
4543 if (slab_state < SYSFS)
4545 * Sysfs has not been setup yet so no need to remove the
4550 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4551 kobject_del(&s->kobj);
4552 kobject_put(&s->kobj);
4556 * Need to buffer aliases during bootup until sysfs becomes
4557 * available lest we lose that information.
4559 struct saved_alias {
4560 struct kmem_cache *s;
4562 struct saved_alias *next;
4565 static struct saved_alias *alias_list;
4567 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4569 struct saved_alias *al;
4571 if (slab_state == SYSFS) {
4573 * If we have a leftover link then remove it.
4575 sysfs_remove_link(&slab_kset->kobj, name);
4576 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4579 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4585 al->next = alias_list;
4590 static int __init slab_sysfs_init(void)
4592 struct kmem_cache *s;
4595 down_write(&slub_lock);
4597 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4599 up_write(&slub_lock);
4600 printk(KERN_ERR "Cannot register slab subsystem.\n");
4606 list_for_each_entry(s, &slab_caches, list) {
4607 err = sysfs_slab_add(s);
4609 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4610 " to sysfs\n", s->name);
4613 while (alias_list) {
4614 struct saved_alias *al = alias_list;
4616 alias_list = alias_list->next;
4617 err = sysfs_slab_alias(al->s, al->name);
4619 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4620 " %s to sysfs\n", s->name);
4624 up_write(&slub_lock);
4629 __initcall(slab_sysfs_init);
4630 #endif /* CONFIG_SYSFS */
4633 * The /proc/slabinfo ABI
4635 #ifdef CONFIG_SLABINFO
4636 static void print_slabinfo_header(struct seq_file *m)
4638 seq_puts(m, "slabinfo - version: 2.1\n");
4639 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4640 "<objperslab> <pagesperslab>");
4641 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4642 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4646 static void *s_start(struct seq_file *m, loff_t *pos)
4650 down_read(&slub_lock);
4652 print_slabinfo_header(m);
4654 return seq_list_start(&slab_caches, *pos);
4657 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4659 return seq_list_next(p, &slab_caches, pos);
4662 static void s_stop(struct seq_file *m, void *p)
4664 up_read(&slub_lock);
4667 static int s_show(struct seq_file *m, void *p)
4669 unsigned long nr_partials = 0;
4670 unsigned long nr_slabs = 0;
4671 unsigned long nr_inuse = 0;
4672 unsigned long nr_objs = 0;
4673 unsigned long nr_free = 0;
4674 struct kmem_cache *s;
4677 s = list_entry(p, struct kmem_cache, list);
4679 for_each_online_node(node) {
4680 struct kmem_cache_node *n = get_node(s, node);
4685 nr_partials += n->nr_partial;
4686 nr_slabs += atomic_long_read(&n->nr_slabs);
4687 nr_objs += atomic_long_read(&n->total_objects);
4688 nr_free += count_partial(n, count_free);
4691 nr_inuse = nr_objs - nr_free;
4693 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4694 nr_objs, s->size, oo_objects(s->oo),
4695 (1 << oo_order(s->oo)));
4696 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4697 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4703 static const struct seq_operations slabinfo_op = {
4710 static int slabinfo_open(struct inode *inode, struct file *file)
4712 return seq_open(file, &slabinfo_op);
4715 static const struct file_operations proc_slabinfo_operations = {
4716 .open = slabinfo_open,
4718 .llseek = seq_lseek,
4719 .release = seq_release,
4722 static int __init slab_proc_init(void)
4724 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4727 module_init(slab_proc_init);
4728 #endif /* CONFIG_SLABINFO */