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 int order_objects(int order, unsigned long size, int reserved)
286 return ((PAGE_SIZE << order) - reserved) / size;
289 static inline struct kmem_cache_order_objects oo_make(int order,
290 unsigned long size, int reserved)
292 struct kmem_cache_order_objects x = {
293 (order << OO_SHIFT) + order_objects(order, size, reserved)
299 static inline int oo_order(struct kmem_cache_order_objects x)
301 return x.x >> OO_SHIFT;
304 static inline int oo_objects(struct kmem_cache_order_objects x)
306 return x.x & OO_MASK;
309 #ifdef CONFIG_SLUB_DEBUG
313 #ifdef CONFIG_SLUB_DEBUG_ON
314 static int slub_debug = DEBUG_DEFAULT_FLAGS;
316 static int slub_debug;
319 static char *slub_debug_slabs;
320 static int disable_higher_order_debug;
325 static void print_section(char *text, u8 *addr, unsigned int length)
333 for (i = 0; i < length; i++) {
335 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
338 printk(KERN_CONT " %02x", addr[i]);
340 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
342 printk(KERN_CONT " %s\n", ascii);
349 printk(KERN_CONT " ");
353 printk(KERN_CONT " %s\n", ascii);
357 static struct track *get_track(struct kmem_cache *s, void *object,
358 enum track_item alloc)
363 p = object + s->offset + sizeof(void *);
365 p = object + s->inuse;
370 static void set_track(struct kmem_cache *s, void *object,
371 enum track_item alloc, unsigned long addr)
373 struct track *p = get_track(s, object, alloc);
377 p->cpu = smp_processor_id();
378 p->pid = current->pid;
381 memset(p, 0, sizeof(struct track));
384 static void init_tracking(struct kmem_cache *s, void *object)
386 if (!(s->flags & SLAB_STORE_USER))
389 set_track(s, object, TRACK_FREE, 0UL);
390 set_track(s, object, TRACK_ALLOC, 0UL);
393 static void print_track(const char *s, struct track *t)
398 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
399 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
402 static void print_tracking(struct kmem_cache *s, void *object)
404 if (!(s->flags & SLAB_STORE_USER))
407 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
408 print_track("Freed", get_track(s, object, TRACK_FREE));
411 static void print_page_info(struct page *page)
413 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
414 page, page->objects, page->inuse, page->freelist, page->flags);
418 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
424 vsnprintf(buf, sizeof(buf), fmt, args);
426 printk(KERN_ERR "========================================"
427 "=====================================\n");
428 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
429 printk(KERN_ERR "----------------------------------------"
430 "-------------------------------------\n\n");
433 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
439 vsnprintf(buf, sizeof(buf), fmt, args);
441 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
444 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
446 unsigned int off; /* Offset of last byte */
447 u8 *addr = page_address(page);
449 print_tracking(s, p);
451 print_page_info(page);
453 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
454 p, p - addr, get_freepointer(s, p));
457 print_section("Bytes b4", p - 16, 16);
459 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
461 if (s->flags & SLAB_RED_ZONE)
462 print_section("Redzone", p + s->objsize,
463 s->inuse - s->objsize);
466 off = s->offset + sizeof(void *);
470 if (s->flags & SLAB_STORE_USER)
471 off += 2 * sizeof(struct track);
474 /* Beginning of the filler is the free pointer */
475 print_section("Padding", p + off, s->size - off);
480 static void object_err(struct kmem_cache *s, struct page *page,
481 u8 *object, char *reason)
483 slab_bug(s, "%s", reason);
484 print_trailer(s, page, object);
487 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
493 vsnprintf(buf, sizeof(buf), fmt, args);
495 slab_bug(s, "%s", buf);
496 print_page_info(page);
500 static void init_object(struct kmem_cache *s, void *object, u8 val)
504 if (s->flags & __OBJECT_POISON) {
505 memset(p, POISON_FREE, s->objsize - 1);
506 p[s->objsize - 1] = POISON_END;
509 if (s->flags & SLAB_RED_ZONE)
510 memset(p + s->objsize, val, s->inuse - s->objsize);
513 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
516 if (*start != (u8)value)
524 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
525 void *from, void *to)
527 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
528 memset(from, data, to - from);
531 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
532 u8 *object, char *what,
533 u8 *start, unsigned int value, unsigned int bytes)
538 fault = check_bytes(start, value, bytes);
543 while (end > fault && end[-1] == value)
546 slab_bug(s, "%s overwritten", what);
547 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
548 fault, end - 1, fault[0], value);
549 print_trailer(s, page, object);
551 restore_bytes(s, what, value, fault, end);
559 * Bytes of the object to be managed.
560 * If the freepointer may overlay the object then the free
561 * pointer is the first word of the object.
563 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
566 * object + s->objsize
567 * Padding to reach word boundary. This is also used for Redzoning.
568 * Padding is extended by another word if Redzoning is enabled and
571 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
572 * 0xcc (RED_ACTIVE) for objects in use.
575 * Meta data starts here.
577 * A. Free pointer (if we cannot overwrite object on free)
578 * B. Tracking data for SLAB_STORE_USER
579 * C. Padding to reach required alignment boundary or at mininum
580 * one word if debugging is on to be able to detect writes
581 * before the word boundary.
583 * Padding is done using 0x5a (POISON_INUSE)
586 * Nothing is used beyond s->size.
588 * If slabcaches are merged then the objsize and inuse boundaries are mostly
589 * ignored. And therefore no slab options that rely on these boundaries
590 * may be used with merged slabcaches.
593 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
595 unsigned long off = s->inuse; /* The end of info */
598 /* Freepointer is placed after the object. */
599 off += sizeof(void *);
601 if (s->flags & SLAB_STORE_USER)
602 /* We also have user information there */
603 off += 2 * sizeof(struct track);
608 return check_bytes_and_report(s, page, p, "Object padding",
609 p + off, POISON_INUSE, s->size - off);
612 /* Check the pad bytes at the end of a slab page */
613 static int slab_pad_check(struct kmem_cache *s, struct page *page)
621 if (!(s->flags & SLAB_POISON))
624 start = page_address(page);
625 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
626 end = start + length;
627 remainder = length % s->size;
631 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
634 while (end > fault && end[-1] == POISON_INUSE)
637 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
638 print_section("Padding", end - remainder, remainder);
640 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
644 static int check_object(struct kmem_cache *s, struct page *page,
645 void *object, u8 val)
648 u8 *endobject = object + s->objsize;
650 if (s->flags & SLAB_RED_ZONE) {
651 if (!check_bytes_and_report(s, page, object, "Redzone",
652 endobject, val, s->inuse - s->objsize))
655 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
656 check_bytes_and_report(s, page, p, "Alignment padding",
657 endobject, POISON_INUSE, s->inuse - s->objsize);
661 if (s->flags & SLAB_POISON) {
662 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
663 (!check_bytes_and_report(s, page, p, "Poison", p,
664 POISON_FREE, s->objsize - 1) ||
665 !check_bytes_and_report(s, page, p, "Poison",
666 p + s->objsize - 1, POISON_END, 1)))
669 * check_pad_bytes cleans up on its own.
671 check_pad_bytes(s, page, p);
674 if (!s->offset && val == SLUB_RED_ACTIVE)
676 * Object and freepointer overlap. Cannot check
677 * freepointer while object is allocated.
681 /* Check free pointer validity */
682 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
683 object_err(s, page, p, "Freepointer corrupt");
685 * No choice but to zap it and thus lose the remainder
686 * of the free objects in this slab. May cause
687 * another error because the object count is now wrong.
689 set_freepointer(s, p, NULL);
695 static int check_slab(struct kmem_cache *s, struct page *page)
699 VM_BUG_ON(!irqs_disabled());
701 if (!PageSlab(page)) {
702 slab_err(s, page, "Not a valid slab page");
706 maxobj = order_objects(compound_order(page), s->size, s->reserved);
707 if (page->objects > maxobj) {
708 slab_err(s, page, "objects %u > max %u",
709 s->name, page->objects, maxobj);
712 if (page->inuse > page->objects) {
713 slab_err(s, page, "inuse %u > max %u",
714 s->name, page->inuse, page->objects);
717 /* Slab_pad_check fixes things up after itself */
718 slab_pad_check(s, page);
723 * Determine if a certain object on a page is on the freelist. Must hold the
724 * slab lock to guarantee that the chains are in a consistent state.
726 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
729 void *fp = page->freelist;
731 unsigned long max_objects;
733 while (fp && nr <= page->objects) {
736 if (!check_valid_pointer(s, page, fp)) {
738 object_err(s, page, object,
739 "Freechain corrupt");
740 set_freepointer(s, object, NULL);
743 slab_err(s, page, "Freepointer corrupt");
744 page->freelist = NULL;
745 page->inuse = page->objects;
746 slab_fix(s, "Freelist cleared");
752 fp = get_freepointer(s, object);
756 max_objects = order_objects(compound_order(page), s->size, s->reserved);
757 if (max_objects > MAX_OBJS_PER_PAGE)
758 max_objects = MAX_OBJS_PER_PAGE;
760 if (page->objects != max_objects) {
761 slab_err(s, page, "Wrong number of objects. Found %d but "
762 "should be %d", page->objects, max_objects);
763 page->objects = max_objects;
764 slab_fix(s, "Number of objects adjusted.");
766 if (page->inuse != page->objects - nr) {
767 slab_err(s, page, "Wrong object count. Counter is %d but "
768 "counted were %d", page->inuse, page->objects - nr);
769 page->inuse = page->objects - nr;
770 slab_fix(s, "Object count adjusted.");
772 return search == NULL;
775 static void trace(struct kmem_cache *s, struct page *page, void *object,
778 if (s->flags & SLAB_TRACE) {
779 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
781 alloc ? "alloc" : "free",
786 print_section("Object", (void *)object, s->objsize);
793 * Hooks for other subsystems that check memory allocations. In a typical
794 * production configuration these hooks all should produce no code at all.
796 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
798 flags &= gfp_allowed_mask;
799 lockdep_trace_alloc(flags);
800 might_sleep_if(flags & __GFP_WAIT);
802 return should_failslab(s->objsize, flags, s->flags);
805 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
807 flags &= gfp_allowed_mask;
808 kmemcheck_slab_alloc(s, flags, object, s->objsize);
809 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
812 static inline void slab_free_hook(struct kmem_cache *s, void *x)
814 kmemleak_free_recursive(x, s->flags);
817 static inline void slab_free_hook_irq(struct kmem_cache *s, void *object)
819 kmemcheck_slab_free(s, object, s->objsize);
820 debug_check_no_locks_freed(object, s->objsize);
821 if (!(s->flags & SLAB_DEBUG_OBJECTS))
822 debug_check_no_obj_freed(object, s->objsize);
826 * Tracking of fully allocated slabs for debugging purposes.
828 static void add_full(struct kmem_cache_node *n, struct page *page)
830 spin_lock(&n->list_lock);
831 list_add(&page->lru, &n->full);
832 spin_unlock(&n->list_lock);
835 static void remove_full(struct kmem_cache *s, struct page *page)
837 struct kmem_cache_node *n;
839 if (!(s->flags & SLAB_STORE_USER))
842 n = get_node(s, page_to_nid(page));
844 spin_lock(&n->list_lock);
845 list_del(&page->lru);
846 spin_unlock(&n->list_lock);
849 /* Tracking of the number of slabs for debugging purposes */
850 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
852 struct kmem_cache_node *n = get_node(s, node);
854 return atomic_long_read(&n->nr_slabs);
857 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
859 return atomic_long_read(&n->nr_slabs);
862 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
864 struct kmem_cache_node *n = get_node(s, node);
867 * May be called early in order to allocate a slab for the
868 * kmem_cache_node structure. Solve the chicken-egg
869 * dilemma by deferring the increment of the count during
870 * bootstrap (see early_kmem_cache_node_alloc).
873 atomic_long_inc(&n->nr_slabs);
874 atomic_long_add(objects, &n->total_objects);
877 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
879 struct kmem_cache_node *n = get_node(s, node);
881 atomic_long_dec(&n->nr_slabs);
882 atomic_long_sub(objects, &n->total_objects);
885 /* Object debug checks for alloc/free paths */
886 static void setup_object_debug(struct kmem_cache *s, struct page *page,
889 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
892 init_object(s, object, SLUB_RED_INACTIVE);
893 init_tracking(s, object);
896 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
897 void *object, unsigned long addr)
899 if (!check_slab(s, page))
902 if (!on_freelist(s, page, object)) {
903 object_err(s, page, object, "Object already allocated");
907 if (!check_valid_pointer(s, page, object)) {
908 object_err(s, page, object, "Freelist Pointer check fails");
912 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
915 /* Success perform special debug activities for allocs */
916 if (s->flags & SLAB_STORE_USER)
917 set_track(s, object, TRACK_ALLOC, addr);
918 trace(s, page, object, 1);
919 init_object(s, object, SLUB_RED_ACTIVE);
923 if (PageSlab(page)) {
925 * If this is a slab page then lets do the best we can
926 * to avoid issues in the future. Marking all objects
927 * as used avoids touching the remaining objects.
929 slab_fix(s, "Marking all objects used");
930 page->inuse = page->objects;
931 page->freelist = NULL;
936 static noinline int free_debug_processing(struct kmem_cache *s,
937 struct page *page, void *object, unsigned long addr)
939 if (!check_slab(s, page))
942 if (!check_valid_pointer(s, page, object)) {
943 slab_err(s, page, "Invalid object pointer 0x%p", object);
947 if (on_freelist(s, page, object)) {
948 object_err(s, page, object, "Object already free");
952 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
955 if (unlikely(s != page->slab)) {
956 if (!PageSlab(page)) {
957 slab_err(s, page, "Attempt to free object(0x%p) "
958 "outside of slab", object);
959 } else if (!page->slab) {
961 "SLUB <none>: no slab for object 0x%p.\n",
965 object_err(s, page, object,
966 "page slab pointer corrupt.");
970 /* Special debug activities for freeing objects */
971 if (!PageSlubFrozen(page) && !page->freelist)
972 remove_full(s, page);
973 if (s->flags & SLAB_STORE_USER)
974 set_track(s, object, TRACK_FREE, addr);
975 trace(s, page, object, 0);
976 init_object(s, object, SLUB_RED_INACTIVE);
980 slab_fix(s, "Object at 0x%p not freed", object);
984 static int __init setup_slub_debug(char *str)
986 slub_debug = DEBUG_DEFAULT_FLAGS;
987 if (*str++ != '=' || !*str)
989 * No options specified. Switch on full debugging.
995 * No options but restriction on slabs. This means full
996 * debugging for slabs matching a pattern.
1000 if (tolower(*str) == 'o') {
1002 * Avoid enabling debugging on caches if its minimum order
1003 * would increase as a result.
1005 disable_higher_order_debug = 1;
1012 * Switch off all debugging measures.
1017 * Determine which debug features should be switched on
1019 for (; *str && *str != ','; str++) {
1020 switch (tolower(*str)) {
1022 slub_debug |= SLAB_DEBUG_FREE;
1025 slub_debug |= SLAB_RED_ZONE;
1028 slub_debug |= SLAB_POISON;
1031 slub_debug |= SLAB_STORE_USER;
1034 slub_debug |= SLAB_TRACE;
1037 slub_debug |= SLAB_FAILSLAB;
1040 printk(KERN_ERR "slub_debug option '%c' "
1041 "unknown. skipped\n", *str);
1047 slub_debug_slabs = str + 1;
1052 __setup("slub_debug", setup_slub_debug);
1054 static unsigned long kmem_cache_flags(unsigned long objsize,
1055 unsigned long flags, const char *name,
1056 void (*ctor)(void *))
1059 * Enable debugging if selected on the kernel commandline.
1061 if (slub_debug && (!slub_debug_slabs ||
1062 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1063 flags |= slub_debug;
1068 static inline void setup_object_debug(struct kmem_cache *s,
1069 struct page *page, void *object) {}
1071 static inline int alloc_debug_processing(struct kmem_cache *s,
1072 struct page *page, void *object, unsigned long addr) { return 0; }
1074 static inline int free_debug_processing(struct kmem_cache *s,
1075 struct page *page, void *object, unsigned long addr) { return 0; }
1077 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1079 static inline int check_object(struct kmem_cache *s, struct page *page,
1080 void *object, u8 val) { return 1; }
1081 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1082 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1083 unsigned long flags, const char *name,
1084 void (*ctor)(void *))
1088 #define slub_debug 0
1090 #define disable_higher_order_debug 0
1092 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1094 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1096 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1098 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1101 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1104 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1107 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1109 static inline void slab_free_hook_irq(struct kmem_cache *s,
1112 #endif /* CONFIG_SLUB_DEBUG */
1115 * Slab allocation and freeing
1117 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1118 struct kmem_cache_order_objects oo)
1120 int order = oo_order(oo);
1122 flags |= __GFP_NOTRACK;
1124 if (node == NUMA_NO_NODE)
1125 return alloc_pages(flags, order);
1127 return alloc_pages_exact_node(node, flags, order);
1130 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1133 struct kmem_cache_order_objects oo = s->oo;
1136 flags |= s->allocflags;
1139 * Let the initial higher-order allocation fail under memory pressure
1140 * so we fall-back to the minimum order allocation.
1142 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1144 page = alloc_slab_page(alloc_gfp, node, oo);
1145 if (unlikely(!page)) {
1148 * Allocation may have failed due to fragmentation.
1149 * Try a lower order alloc if possible
1151 page = alloc_slab_page(flags, node, oo);
1155 stat(s, ORDER_FALLBACK);
1158 if (kmemcheck_enabled
1159 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1160 int pages = 1 << oo_order(oo);
1162 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1165 * Objects from caches that have a constructor don't get
1166 * cleared when they're allocated, so we need to do it here.
1169 kmemcheck_mark_uninitialized_pages(page, pages);
1171 kmemcheck_mark_unallocated_pages(page, pages);
1174 page->objects = oo_objects(oo);
1175 mod_zone_page_state(page_zone(page),
1176 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1177 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1183 static void setup_object(struct kmem_cache *s, struct page *page,
1186 setup_object_debug(s, page, object);
1187 if (unlikely(s->ctor))
1191 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1198 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1200 page = allocate_slab(s,
1201 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1205 inc_slabs_node(s, page_to_nid(page), page->objects);
1207 page->flags |= 1 << PG_slab;
1209 start = page_address(page);
1211 if (unlikely(s->flags & SLAB_POISON))
1212 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1215 for_each_object(p, s, start, page->objects) {
1216 setup_object(s, page, last);
1217 set_freepointer(s, last, p);
1220 setup_object(s, page, last);
1221 set_freepointer(s, last, NULL);
1223 page->freelist = start;
1229 static void __free_slab(struct kmem_cache *s, struct page *page)
1231 int order = compound_order(page);
1232 int pages = 1 << order;
1234 if (kmem_cache_debug(s)) {
1237 slab_pad_check(s, page);
1238 for_each_object(p, s, page_address(page),
1240 check_object(s, page, p, SLUB_RED_INACTIVE);
1243 kmemcheck_free_shadow(page, compound_order(page));
1245 mod_zone_page_state(page_zone(page),
1246 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1247 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1250 __ClearPageSlab(page);
1251 reset_page_mapcount(page);
1252 if (current->reclaim_state)
1253 current->reclaim_state->reclaimed_slab += pages;
1254 __free_pages(page, order);
1257 static void rcu_free_slab(struct rcu_head *h)
1261 page = container_of((struct list_head *)h, struct page, lru);
1262 __free_slab(page->slab, page);
1265 static void free_slab(struct kmem_cache *s, struct page *page)
1267 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1269 * RCU free overloads the RCU head over the LRU
1271 struct rcu_head *head = (void *)&page->lru;
1273 call_rcu(head, rcu_free_slab);
1275 __free_slab(s, page);
1278 static void discard_slab(struct kmem_cache *s, struct page *page)
1280 dec_slabs_node(s, page_to_nid(page), page->objects);
1285 * Per slab locking using the pagelock
1287 static __always_inline void slab_lock(struct page *page)
1289 bit_spin_lock(PG_locked, &page->flags);
1292 static __always_inline void slab_unlock(struct page *page)
1294 __bit_spin_unlock(PG_locked, &page->flags);
1297 static __always_inline int slab_trylock(struct page *page)
1301 rc = bit_spin_trylock(PG_locked, &page->flags);
1306 * Management of partially allocated slabs
1308 static void add_partial(struct kmem_cache_node *n,
1309 struct page *page, int tail)
1311 spin_lock(&n->list_lock);
1314 list_add_tail(&page->lru, &n->partial);
1316 list_add(&page->lru, &n->partial);
1317 spin_unlock(&n->list_lock);
1320 static inline void __remove_partial(struct kmem_cache_node *n,
1323 list_del(&page->lru);
1327 static void remove_partial(struct kmem_cache *s, struct page *page)
1329 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1331 spin_lock(&n->list_lock);
1332 __remove_partial(n, page);
1333 spin_unlock(&n->list_lock);
1337 * Lock slab and remove from the partial list.
1339 * Must hold list_lock.
1341 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1344 if (slab_trylock(page)) {
1345 __remove_partial(n, page);
1346 __SetPageSlubFrozen(page);
1353 * Try to allocate a partial slab from a specific node.
1355 static struct page *get_partial_node(struct kmem_cache_node *n)
1360 * Racy check. If we mistakenly see no partial slabs then we
1361 * just allocate an empty slab. If we mistakenly try to get a
1362 * partial slab and there is none available then get_partials()
1365 if (!n || !n->nr_partial)
1368 spin_lock(&n->list_lock);
1369 list_for_each_entry(page, &n->partial, lru)
1370 if (lock_and_freeze_slab(n, page))
1374 spin_unlock(&n->list_lock);
1379 * Get a page from somewhere. Search in increasing NUMA distances.
1381 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1384 struct zonelist *zonelist;
1387 enum zone_type high_zoneidx = gfp_zone(flags);
1391 * The defrag ratio allows a configuration of the tradeoffs between
1392 * inter node defragmentation and node local allocations. A lower
1393 * defrag_ratio increases the tendency to do local allocations
1394 * instead of attempting to obtain partial slabs from other nodes.
1396 * If the defrag_ratio is set to 0 then kmalloc() always
1397 * returns node local objects. If the ratio is higher then kmalloc()
1398 * may return off node objects because partial slabs are obtained
1399 * from other nodes and filled up.
1401 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1402 * defrag_ratio = 1000) then every (well almost) allocation will
1403 * first attempt to defrag slab caches on other nodes. This means
1404 * scanning over all nodes to look for partial slabs which may be
1405 * expensive if we do it every time we are trying to find a slab
1406 * with available objects.
1408 if (!s->remote_node_defrag_ratio ||
1409 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1413 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1414 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1415 struct kmem_cache_node *n;
1417 n = get_node(s, zone_to_nid(zone));
1419 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1420 n->nr_partial > s->min_partial) {
1421 page = get_partial_node(n);
1434 * Get a partial page, lock it and return it.
1436 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1439 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1441 page = get_partial_node(get_node(s, searchnode));
1442 if (page || node != -1)
1445 return get_any_partial(s, flags);
1449 * Move a page back to the lists.
1451 * Must be called with the slab lock held.
1453 * On exit the slab lock will have been dropped.
1455 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1458 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1460 __ClearPageSlubFrozen(page);
1463 if (page->freelist) {
1464 add_partial(n, page, tail);
1465 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1467 stat(s, DEACTIVATE_FULL);
1468 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1473 stat(s, DEACTIVATE_EMPTY);
1474 if (n->nr_partial < s->min_partial) {
1476 * Adding an empty slab to the partial slabs in order
1477 * to avoid page allocator overhead. This slab needs
1478 * to come after the other slabs with objects in
1479 * so that the others get filled first. That way the
1480 * size of the partial list stays small.
1482 * kmem_cache_shrink can reclaim any empty slabs from
1485 add_partial(n, page, 1);
1490 discard_slab(s, page);
1496 * Remove the cpu slab
1498 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1501 struct page *page = c->page;
1505 stat(s, DEACTIVATE_REMOTE_FREES);
1507 * Merge cpu freelist into slab freelist. Typically we get here
1508 * because both freelists are empty. So this is unlikely
1511 while (unlikely(c->freelist)) {
1514 tail = 0; /* Hot objects. Put the slab first */
1516 /* Retrieve object from cpu_freelist */
1517 object = c->freelist;
1518 c->freelist = get_freepointer(s, c->freelist);
1520 /* And put onto the regular freelist */
1521 set_freepointer(s, object, page->freelist);
1522 page->freelist = object;
1526 unfreeze_slab(s, page, tail);
1529 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1531 stat(s, CPUSLAB_FLUSH);
1533 deactivate_slab(s, c);
1539 * Called from IPI handler with interrupts disabled.
1541 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1543 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1545 if (likely(c && c->page))
1549 static void flush_cpu_slab(void *d)
1551 struct kmem_cache *s = d;
1553 __flush_cpu_slab(s, smp_processor_id());
1556 static void flush_all(struct kmem_cache *s)
1558 on_each_cpu(flush_cpu_slab, s, 1);
1562 * Check if the objects in a per cpu structure fit numa
1563 * locality expectations.
1565 static inline int node_match(struct kmem_cache_cpu *c, int node)
1568 if (node != NUMA_NO_NODE && c->node != node)
1574 static int count_free(struct page *page)
1576 return page->objects - page->inuse;
1579 static unsigned long count_partial(struct kmem_cache_node *n,
1580 int (*get_count)(struct page *))
1582 unsigned long flags;
1583 unsigned long x = 0;
1586 spin_lock_irqsave(&n->list_lock, flags);
1587 list_for_each_entry(page, &n->partial, lru)
1588 x += get_count(page);
1589 spin_unlock_irqrestore(&n->list_lock, flags);
1593 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1595 #ifdef CONFIG_SLUB_DEBUG
1596 return atomic_long_read(&n->total_objects);
1602 static noinline void
1603 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1608 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1610 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1611 "default order: %d, min order: %d\n", s->name, s->objsize,
1612 s->size, oo_order(s->oo), oo_order(s->min));
1614 if (oo_order(s->min) > get_order(s->objsize))
1615 printk(KERN_WARNING " %s debugging increased min order, use "
1616 "slub_debug=O to disable.\n", s->name);
1618 for_each_online_node(node) {
1619 struct kmem_cache_node *n = get_node(s, node);
1620 unsigned long nr_slabs;
1621 unsigned long nr_objs;
1622 unsigned long nr_free;
1627 nr_free = count_partial(n, count_free);
1628 nr_slabs = node_nr_slabs(n);
1629 nr_objs = node_nr_objs(n);
1632 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1633 node, nr_slabs, nr_objs, nr_free);
1638 * Slow path. The lockless freelist is empty or we need to perform
1641 * Interrupts are disabled.
1643 * Processing is still very fast if new objects have been freed to the
1644 * regular freelist. In that case we simply take over the regular freelist
1645 * as the lockless freelist and zap the regular freelist.
1647 * If that is not working then we fall back to the partial lists. We take the
1648 * first element of the freelist as the object to allocate now and move the
1649 * rest of the freelist to the lockless freelist.
1651 * And if we were unable to get a new slab from the partial slab lists then
1652 * we need to allocate a new slab. This is the slowest path since it involves
1653 * a call to the page allocator and the setup of a new slab.
1655 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1656 unsigned long addr, struct kmem_cache_cpu *c)
1661 /* We handle __GFP_ZERO in the caller */
1662 gfpflags &= ~__GFP_ZERO;
1668 if (unlikely(!node_match(c, node)))
1671 stat(s, ALLOC_REFILL);
1674 object = c->page->freelist;
1675 if (unlikely(!object))
1677 if (kmem_cache_debug(s))
1680 c->freelist = get_freepointer(s, object);
1681 c->page->inuse = c->page->objects;
1682 c->page->freelist = NULL;
1683 c->node = page_to_nid(c->page);
1685 slab_unlock(c->page);
1686 stat(s, ALLOC_SLOWPATH);
1690 deactivate_slab(s, c);
1693 new = get_partial(s, gfpflags, node);
1696 stat(s, ALLOC_FROM_PARTIAL);
1700 gfpflags &= gfp_allowed_mask;
1701 if (gfpflags & __GFP_WAIT)
1704 new = new_slab(s, gfpflags, node);
1706 if (gfpflags & __GFP_WAIT)
1707 local_irq_disable();
1710 c = __this_cpu_ptr(s->cpu_slab);
1711 stat(s, ALLOC_SLAB);
1715 __SetPageSlubFrozen(new);
1719 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1720 slab_out_of_memory(s, gfpflags, node);
1723 if (!alloc_debug_processing(s, c->page, object, addr))
1727 c->page->freelist = get_freepointer(s, object);
1728 c->node = NUMA_NO_NODE;
1733 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1734 * have the fastpath folded into their functions. So no function call
1735 * overhead for requests that can be satisfied on the fastpath.
1737 * The fastpath works by first checking if the lockless freelist can be used.
1738 * If not then __slab_alloc is called for slow processing.
1740 * Otherwise we can simply pick the next object from the lockless free list.
1742 static __always_inline void *slab_alloc(struct kmem_cache *s,
1743 gfp_t gfpflags, int node, unsigned long addr)
1746 struct kmem_cache_cpu *c;
1747 unsigned long flags;
1749 if (slab_pre_alloc_hook(s, gfpflags))
1752 local_irq_save(flags);
1753 c = __this_cpu_ptr(s->cpu_slab);
1754 object = c->freelist;
1755 if (unlikely(!object || !node_match(c, node)))
1757 object = __slab_alloc(s, gfpflags, node, addr, c);
1760 c->freelist = get_freepointer(s, object);
1761 stat(s, ALLOC_FASTPATH);
1763 local_irq_restore(flags);
1765 if (unlikely(gfpflags & __GFP_ZERO) && object)
1766 memset(object, 0, s->objsize);
1768 slab_post_alloc_hook(s, gfpflags, object);
1773 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1775 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1777 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1781 EXPORT_SYMBOL(kmem_cache_alloc);
1783 #ifdef CONFIG_TRACING
1784 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1786 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
1787 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
1790 EXPORT_SYMBOL(kmem_cache_alloc_trace);
1792 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1794 void *ret = kmalloc_order(size, flags, order);
1795 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1798 EXPORT_SYMBOL(kmalloc_order_trace);
1802 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1804 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1806 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1807 s->objsize, s->size, gfpflags, node);
1811 EXPORT_SYMBOL(kmem_cache_alloc_node);
1813 #ifdef CONFIG_TRACING
1814 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
1816 int node, size_t size)
1818 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1820 trace_kmalloc_node(_RET_IP_, ret,
1821 size, s->size, gfpflags, node);
1824 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
1829 * Slow patch handling. This may still be called frequently since objects
1830 * have a longer lifetime than the cpu slabs in most processing loads.
1832 * So we still attempt to reduce cache line usage. Just take the slab
1833 * lock and free the item. If there is no additional partial page
1834 * handling required then we can return immediately.
1836 static void __slab_free(struct kmem_cache *s, struct page *page,
1837 void *x, unsigned long addr)
1840 void **object = (void *)x;
1842 stat(s, FREE_SLOWPATH);
1845 if (kmem_cache_debug(s))
1849 prior = page->freelist;
1850 set_freepointer(s, object, prior);
1851 page->freelist = object;
1854 if (unlikely(PageSlubFrozen(page))) {
1855 stat(s, FREE_FROZEN);
1859 if (unlikely(!page->inuse))
1863 * Objects left in the slab. If it was not on the partial list before
1866 if (unlikely(!prior)) {
1867 add_partial(get_node(s, page_to_nid(page)), page, 1);
1868 stat(s, FREE_ADD_PARTIAL);
1878 * Slab still on the partial list.
1880 remove_partial(s, page);
1881 stat(s, FREE_REMOVE_PARTIAL);
1885 discard_slab(s, page);
1889 if (!free_debug_processing(s, page, x, addr))
1895 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1896 * can perform fastpath freeing without additional function calls.
1898 * The fastpath is only possible if we are freeing to the current cpu slab
1899 * of this processor. This typically the case if we have just allocated
1902 * If fastpath is not possible then fall back to __slab_free where we deal
1903 * with all sorts of special processing.
1905 static __always_inline void slab_free(struct kmem_cache *s,
1906 struct page *page, void *x, unsigned long addr)
1908 void **object = (void *)x;
1909 struct kmem_cache_cpu *c;
1910 unsigned long flags;
1912 slab_free_hook(s, x);
1914 local_irq_save(flags);
1915 c = __this_cpu_ptr(s->cpu_slab);
1917 slab_free_hook_irq(s, x);
1919 if (likely(page == c->page && c->node != NUMA_NO_NODE)) {
1920 set_freepointer(s, object, c->freelist);
1921 c->freelist = object;
1922 stat(s, FREE_FASTPATH);
1924 __slab_free(s, page, x, addr);
1926 local_irq_restore(flags);
1929 void kmem_cache_free(struct kmem_cache *s, void *x)
1933 page = virt_to_head_page(x);
1935 slab_free(s, page, x, _RET_IP_);
1937 trace_kmem_cache_free(_RET_IP_, x);
1939 EXPORT_SYMBOL(kmem_cache_free);
1942 * Object placement in a slab is made very easy because we always start at
1943 * offset 0. If we tune the size of the object to the alignment then we can
1944 * get the required alignment by putting one properly sized object after
1947 * Notice that the allocation order determines the sizes of the per cpu
1948 * caches. Each processor has always one slab available for allocations.
1949 * Increasing the allocation order reduces the number of times that slabs
1950 * must be moved on and off the partial lists and is therefore a factor in
1955 * Mininum / Maximum order of slab pages. This influences locking overhead
1956 * and slab fragmentation. A higher order reduces the number of partial slabs
1957 * and increases the number of allocations possible without having to
1958 * take the list_lock.
1960 static int slub_min_order;
1961 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1962 static int slub_min_objects;
1965 * Merge control. If this is set then no merging of slab caches will occur.
1966 * (Could be removed. This was introduced to pacify the merge skeptics.)
1968 static int slub_nomerge;
1971 * Calculate the order of allocation given an slab object size.
1973 * The order of allocation has significant impact on performance and other
1974 * system components. Generally order 0 allocations should be preferred since
1975 * order 0 does not cause fragmentation in the page allocator. Larger objects
1976 * be problematic to put into order 0 slabs because there may be too much
1977 * unused space left. We go to a higher order if more than 1/16th of the slab
1980 * In order to reach satisfactory performance we must ensure that a minimum
1981 * number of objects is in one slab. Otherwise we may generate too much
1982 * activity on the partial lists which requires taking the list_lock. This is
1983 * less a concern for large slabs though which are rarely used.
1985 * slub_max_order specifies the order where we begin to stop considering the
1986 * number of objects in a slab as critical. If we reach slub_max_order then
1987 * we try to keep the page order as low as possible. So we accept more waste
1988 * of space in favor of a small page order.
1990 * Higher order allocations also allow the placement of more objects in a
1991 * slab and thereby reduce object handling overhead. If the user has
1992 * requested a higher mininum order then we start with that one instead of
1993 * the smallest order which will fit the object.
1995 static inline int slab_order(int size, int min_objects,
1996 int max_order, int fract_leftover, int reserved)
2000 int min_order = slub_min_order;
2002 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2003 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2005 for (order = max(min_order,
2006 fls(min_objects * size - 1) - PAGE_SHIFT);
2007 order <= max_order; order++) {
2009 unsigned long slab_size = PAGE_SIZE << order;
2011 if (slab_size < min_objects * size + reserved)
2014 rem = (slab_size - reserved) % size;
2016 if (rem <= slab_size / fract_leftover)
2024 static inline int calculate_order(int size, int reserved)
2032 * Attempt to find best configuration for a slab. This
2033 * works by first attempting to generate a layout with
2034 * the best configuration and backing off gradually.
2036 * First we reduce the acceptable waste in a slab. Then
2037 * we reduce the minimum objects required in a slab.
2039 min_objects = slub_min_objects;
2041 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2042 max_objects = order_objects(slub_max_order, size, reserved);
2043 min_objects = min(min_objects, max_objects);
2045 while (min_objects > 1) {
2047 while (fraction >= 4) {
2048 order = slab_order(size, min_objects,
2049 slub_max_order, fraction, reserved);
2050 if (order <= slub_max_order)
2058 * We were unable to place multiple objects in a slab. Now
2059 * lets see if we can place a single object there.
2061 order = slab_order(size, 1, slub_max_order, 1, reserved);
2062 if (order <= slub_max_order)
2066 * Doh this slab cannot be placed using slub_max_order.
2068 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2069 if (order < MAX_ORDER)
2075 * Figure out what the alignment of the objects will be.
2077 static unsigned long calculate_alignment(unsigned long flags,
2078 unsigned long align, unsigned long size)
2081 * If the user wants hardware cache aligned objects then follow that
2082 * suggestion if the object is sufficiently large.
2084 * The hardware cache alignment cannot override the specified
2085 * alignment though. If that is greater then use it.
2087 if (flags & SLAB_HWCACHE_ALIGN) {
2088 unsigned long ralign = cache_line_size();
2089 while (size <= ralign / 2)
2091 align = max(align, ralign);
2094 if (align < ARCH_SLAB_MINALIGN)
2095 align = ARCH_SLAB_MINALIGN;
2097 return ALIGN(align, sizeof(void *));
2101 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2104 spin_lock_init(&n->list_lock);
2105 INIT_LIST_HEAD(&n->partial);
2106 #ifdef CONFIG_SLUB_DEBUG
2107 atomic_long_set(&n->nr_slabs, 0);
2108 atomic_long_set(&n->total_objects, 0);
2109 INIT_LIST_HEAD(&n->full);
2113 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2115 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2116 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2118 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2120 return s->cpu_slab != NULL;
2123 static struct kmem_cache *kmem_cache_node;
2126 * No kmalloc_node yet so do it by hand. We know that this is the first
2127 * slab on the node for this slabcache. There are no concurrent accesses
2130 * Note that this function only works on the kmalloc_node_cache
2131 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2132 * memory on a fresh node that has no slab structures yet.
2134 static void early_kmem_cache_node_alloc(int node)
2137 struct kmem_cache_node *n;
2138 unsigned long flags;
2140 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2142 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2145 if (page_to_nid(page) != node) {
2146 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2148 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2149 "in order to be able to continue\n");
2154 page->freelist = get_freepointer(kmem_cache_node, n);
2156 kmem_cache_node->node[node] = n;
2157 #ifdef CONFIG_SLUB_DEBUG
2158 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2159 init_tracking(kmem_cache_node, n);
2161 init_kmem_cache_node(n, kmem_cache_node);
2162 inc_slabs_node(kmem_cache_node, node, page->objects);
2165 * lockdep requires consistent irq usage for each lock
2166 * so even though there cannot be a race this early in
2167 * the boot sequence, we still disable irqs.
2169 local_irq_save(flags);
2170 add_partial(n, page, 0);
2171 local_irq_restore(flags);
2174 static void free_kmem_cache_nodes(struct kmem_cache *s)
2178 for_each_node_state(node, N_NORMAL_MEMORY) {
2179 struct kmem_cache_node *n = s->node[node];
2182 kmem_cache_free(kmem_cache_node, n);
2184 s->node[node] = NULL;
2188 static int init_kmem_cache_nodes(struct kmem_cache *s)
2192 for_each_node_state(node, N_NORMAL_MEMORY) {
2193 struct kmem_cache_node *n;
2195 if (slab_state == DOWN) {
2196 early_kmem_cache_node_alloc(node);
2199 n = kmem_cache_alloc_node(kmem_cache_node,
2203 free_kmem_cache_nodes(s);
2208 init_kmem_cache_node(n, s);
2213 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2215 if (min < MIN_PARTIAL)
2217 else if (min > MAX_PARTIAL)
2219 s->min_partial = min;
2223 * calculate_sizes() determines the order and the distribution of data within
2226 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2228 unsigned long flags = s->flags;
2229 unsigned long size = s->objsize;
2230 unsigned long align = s->align;
2234 * Round up object size to the next word boundary. We can only
2235 * place the free pointer at word boundaries and this determines
2236 * the possible location of the free pointer.
2238 size = ALIGN(size, sizeof(void *));
2240 #ifdef CONFIG_SLUB_DEBUG
2242 * Determine if we can poison the object itself. If the user of
2243 * the slab may touch the object after free or before allocation
2244 * then we should never poison the object itself.
2246 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2248 s->flags |= __OBJECT_POISON;
2250 s->flags &= ~__OBJECT_POISON;
2254 * If we are Redzoning then check if there is some space between the
2255 * end of the object and the free pointer. If not then add an
2256 * additional word to have some bytes to store Redzone information.
2258 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2259 size += sizeof(void *);
2263 * With that we have determined the number of bytes in actual use
2264 * by the object. This is the potential offset to the free pointer.
2268 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2271 * Relocate free pointer after the object if it is not
2272 * permitted to overwrite the first word of the object on
2275 * This is the case if we do RCU, have a constructor or
2276 * destructor or are poisoning the objects.
2279 size += sizeof(void *);
2282 #ifdef CONFIG_SLUB_DEBUG
2283 if (flags & SLAB_STORE_USER)
2285 * Need to store information about allocs and frees after
2288 size += 2 * sizeof(struct track);
2290 if (flags & SLAB_RED_ZONE)
2292 * Add some empty padding so that we can catch
2293 * overwrites from earlier objects rather than let
2294 * tracking information or the free pointer be
2295 * corrupted if a user writes before the start
2298 size += sizeof(void *);
2302 * Determine the alignment based on various parameters that the
2303 * user specified and the dynamic determination of cache line size
2306 align = calculate_alignment(flags, align, s->objsize);
2310 * SLUB stores one object immediately after another beginning from
2311 * offset 0. In order to align the objects we have to simply size
2312 * each object to conform to the alignment.
2314 size = ALIGN(size, align);
2316 if (forced_order >= 0)
2317 order = forced_order;
2319 order = calculate_order(size, s->reserved);
2326 s->allocflags |= __GFP_COMP;
2328 if (s->flags & SLAB_CACHE_DMA)
2329 s->allocflags |= SLUB_DMA;
2331 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2332 s->allocflags |= __GFP_RECLAIMABLE;
2335 * Determine the number of objects per slab
2337 s->oo = oo_make(order, size, s->reserved);
2338 s->min = oo_make(get_order(size), size, s->reserved);
2339 if (oo_objects(s->oo) > oo_objects(s->max))
2342 return !!oo_objects(s->oo);
2346 static int kmem_cache_open(struct kmem_cache *s,
2347 const char *name, size_t size,
2348 size_t align, unsigned long flags,
2349 void (*ctor)(void *))
2351 memset(s, 0, kmem_size);
2356 s->flags = kmem_cache_flags(size, flags, name, ctor);
2359 if (!calculate_sizes(s, -1))
2361 if (disable_higher_order_debug) {
2363 * Disable debugging flags that store metadata if the min slab
2366 if (get_order(s->size) > get_order(s->objsize)) {
2367 s->flags &= ~DEBUG_METADATA_FLAGS;
2369 if (!calculate_sizes(s, -1))
2375 * The larger the object size is, the more pages we want on the partial
2376 * list to avoid pounding the page allocator excessively.
2378 set_min_partial(s, ilog2(s->size));
2381 s->remote_node_defrag_ratio = 1000;
2383 if (!init_kmem_cache_nodes(s))
2386 if (alloc_kmem_cache_cpus(s))
2389 free_kmem_cache_nodes(s);
2391 if (flags & SLAB_PANIC)
2392 panic("Cannot create slab %s size=%lu realsize=%u "
2393 "order=%u offset=%u flags=%lx\n",
2394 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2400 * Determine the size of a slab object
2402 unsigned int kmem_cache_size(struct kmem_cache *s)
2406 EXPORT_SYMBOL(kmem_cache_size);
2408 const char *kmem_cache_name(struct kmem_cache *s)
2412 EXPORT_SYMBOL(kmem_cache_name);
2414 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2417 #ifdef CONFIG_SLUB_DEBUG
2418 void *addr = page_address(page);
2420 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2421 sizeof(long), GFP_ATOMIC);
2424 slab_err(s, page, "%s", text);
2426 for_each_free_object(p, s, page->freelist)
2427 set_bit(slab_index(p, s, addr), map);
2429 for_each_object(p, s, addr, page->objects) {
2431 if (!test_bit(slab_index(p, s, addr), map)) {
2432 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2434 print_tracking(s, p);
2443 * Attempt to free all partial slabs on a node.
2445 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2447 unsigned long flags;
2448 struct page *page, *h;
2450 spin_lock_irqsave(&n->list_lock, flags);
2451 list_for_each_entry_safe(page, h, &n->partial, lru) {
2453 __remove_partial(n, page);
2454 discard_slab(s, page);
2456 list_slab_objects(s, page,
2457 "Objects remaining on kmem_cache_close()");
2460 spin_unlock_irqrestore(&n->list_lock, flags);
2464 * Release all resources used by a slab cache.
2466 static inline int kmem_cache_close(struct kmem_cache *s)
2471 free_percpu(s->cpu_slab);
2472 /* Attempt to free all objects */
2473 for_each_node_state(node, N_NORMAL_MEMORY) {
2474 struct kmem_cache_node *n = get_node(s, node);
2477 if (n->nr_partial || slabs_node(s, node))
2480 free_kmem_cache_nodes(s);
2485 * Close a cache and release the kmem_cache structure
2486 * (must be used for caches created using kmem_cache_create)
2488 void kmem_cache_destroy(struct kmem_cache *s)
2490 down_write(&slub_lock);
2494 if (kmem_cache_close(s)) {
2495 printk(KERN_ERR "SLUB %s: %s called for cache that "
2496 "still has objects.\n", s->name, __func__);
2499 if (s->flags & SLAB_DESTROY_BY_RCU)
2501 sysfs_slab_remove(s);
2503 up_write(&slub_lock);
2505 EXPORT_SYMBOL(kmem_cache_destroy);
2507 /********************************************************************
2509 *******************************************************************/
2511 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2512 EXPORT_SYMBOL(kmalloc_caches);
2514 static struct kmem_cache *kmem_cache;
2516 #ifdef CONFIG_ZONE_DMA
2517 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2520 static int __init setup_slub_min_order(char *str)
2522 get_option(&str, &slub_min_order);
2527 __setup("slub_min_order=", setup_slub_min_order);
2529 static int __init setup_slub_max_order(char *str)
2531 get_option(&str, &slub_max_order);
2532 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2537 __setup("slub_max_order=", setup_slub_max_order);
2539 static int __init setup_slub_min_objects(char *str)
2541 get_option(&str, &slub_min_objects);
2546 __setup("slub_min_objects=", setup_slub_min_objects);
2548 static int __init setup_slub_nomerge(char *str)
2554 __setup("slub_nomerge", setup_slub_nomerge);
2556 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2557 int size, unsigned int flags)
2559 struct kmem_cache *s;
2561 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2564 * This function is called with IRQs disabled during early-boot on
2565 * single CPU so there's no need to take slub_lock here.
2567 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2571 list_add(&s->list, &slab_caches);
2575 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2580 * Conversion table for small slabs sizes / 8 to the index in the
2581 * kmalloc array. This is necessary for slabs < 192 since we have non power
2582 * of two cache sizes there. The size of larger slabs can be determined using
2585 static s8 size_index[24] = {
2612 static inline int size_index_elem(size_t bytes)
2614 return (bytes - 1) / 8;
2617 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2623 return ZERO_SIZE_PTR;
2625 index = size_index[size_index_elem(size)];
2627 index = fls(size - 1);
2629 #ifdef CONFIG_ZONE_DMA
2630 if (unlikely((flags & SLUB_DMA)))
2631 return kmalloc_dma_caches[index];
2634 return kmalloc_caches[index];
2637 void *__kmalloc(size_t size, gfp_t flags)
2639 struct kmem_cache *s;
2642 if (unlikely(size > SLUB_MAX_SIZE))
2643 return kmalloc_large(size, flags);
2645 s = get_slab(size, flags);
2647 if (unlikely(ZERO_OR_NULL_PTR(s)))
2650 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2652 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2656 EXPORT_SYMBOL(__kmalloc);
2659 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2664 flags |= __GFP_COMP | __GFP_NOTRACK;
2665 page = alloc_pages_node(node, flags, get_order(size));
2667 ptr = page_address(page);
2669 kmemleak_alloc(ptr, size, 1, flags);
2673 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2675 struct kmem_cache *s;
2678 if (unlikely(size > SLUB_MAX_SIZE)) {
2679 ret = kmalloc_large_node(size, flags, node);
2681 trace_kmalloc_node(_RET_IP_, ret,
2682 size, PAGE_SIZE << get_order(size),
2688 s = get_slab(size, flags);
2690 if (unlikely(ZERO_OR_NULL_PTR(s)))
2693 ret = slab_alloc(s, flags, node, _RET_IP_);
2695 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2699 EXPORT_SYMBOL(__kmalloc_node);
2702 size_t ksize(const void *object)
2705 struct kmem_cache *s;
2707 if (unlikely(object == ZERO_SIZE_PTR))
2710 page = virt_to_head_page(object);
2712 if (unlikely(!PageSlab(page))) {
2713 WARN_ON(!PageCompound(page));
2714 return PAGE_SIZE << compound_order(page);
2718 #ifdef CONFIG_SLUB_DEBUG
2720 * Debugging requires use of the padding between object
2721 * and whatever may come after it.
2723 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2728 * If we have the need to store the freelist pointer
2729 * back there or track user information then we can
2730 * only use the space before that information.
2732 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2735 * Else we can use all the padding etc for the allocation
2739 EXPORT_SYMBOL(ksize);
2741 void kfree(const void *x)
2744 void *object = (void *)x;
2746 trace_kfree(_RET_IP_, x);
2748 if (unlikely(ZERO_OR_NULL_PTR(x)))
2751 page = virt_to_head_page(x);
2752 if (unlikely(!PageSlab(page))) {
2753 BUG_ON(!PageCompound(page));
2758 slab_free(page->slab, page, object, _RET_IP_);
2760 EXPORT_SYMBOL(kfree);
2763 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2764 * the remaining slabs by the number of items in use. The slabs with the
2765 * most items in use come first. New allocations will then fill those up
2766 * and thus they can be removed from the partial lists.
2768 * The slabs with the least items are placed last. This results in them
2769 * being allocated from last increasing the chance that the last objects
2770 * are freed in them.
2772 int kmem_cache_shrink(struct kmem_cache *s)
2776 struct kmem_cache_node *n;
2779 int objects = oo_objects(s->max);
2780 struct list_head *slabs_by_inuse =
2781 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2782 unsigned long flags;
2784 if (!slabs_by_inuse)
2788 for_each_node_state(node, N_NORMAL_MEMORY) {
2789 n = get_node(s, node);
2794 for (i = 0; i < objects; i++)
2795 INIT_LIST_HEAD(slabs_by_inuse + i);
2797 spin_lock_irqsave(&n->list_lock, flags);
2800 * Build lists indexed by the items in use in each slab.
2802 * Note that concurrent frees may occur while we hold the
2803 * list_lock. page->inuse here is the upper limit.
2805 list_for_each_entry_safe(page, t, &n->partial, lru) {
2806 if (!page->inuse && slab_trylock(page)) {
2808 * Must hold slab lock here because slab_free
2809 * may have freed the last object and be
2810 * waiting to release the slab.
2812 __remove_partial(n, page);
2814 discard_slab(s, page);
2816 list_move(&page->lru,
2817 slabs_by_inuse + page->inuse);
2822 * Rebuild the partial list with the slabs filled up most
2823 * first and the least used slabs at the end.
2825 for (i = objects - 1; i >= 0; i--)
2826 list_splice(slabs_by_inuse + i, n->partial.prev);
2828 spin_unlock_irqrestore(&n->list_lock, flags);
2831 kfree(slabs_by_inuse);
2834 EXPORT_SYMBOL(kmem_cache_shrink);
2836 #if defined(CONFIG_MEMORY_HOTPLUG)
2837 static int slab_mem_going_offline_callback(void *arg)
2839 struct kmem_cache *s;
2841 down_read(&slub_lock);
2842 list_for_each_entry(s, &slab_caches, list)
2843 kmem_cache_shrink(s);
2844 up_read(&slub_lock);
2849 static void slab_mem_offline_callback(void *arg)
2851 struct kmem_cache_node *n;
2852 struct kmem_cache *s;
2853 struct memory_notify *marg = arg;
2856 offline_node = marg->status_change_nid;
2859 * If the node still has available memory. we need kmem_cache_node
2862 if (offline_node < 0)
2865 down_read(&slub_lock);
2866 list_for_each_entry(s, &slab_caches, list) {
2867 n = get_node(s, offline_node);
2870 * if n->nr_slabs > 0, slabs still exist on the node
2871 * that is going down. We were unable to free them,
2872 * and offline_pages() function shouldn't call this
2873 * callback. So, we must fail.
2875 BUG_ON(slabs_node(s, offline_node));
2877 s->node[offline_node] = NULL;
2878 kmem_cache_free(kmem_cache_node, n);
2881 up_read(&slub_lock);
2884 static int slab_mem_going_online_callback(void *arg)
2886 struct kmem_cache_node *n;
2887 struct kmem_cache *s;
2888 struct memory_notify *marg = arg;
2889 int nid = marg->status_change_nid;
2893 * If the node's memory is already available, then kmem_cache_node is
2894 * already created. Nothing to do.
2900 * We are bringing a node online. No memory is available yet. We must
2901 * allocate a kmem_cache_node structure in order to bring the node
2904 down_read(&slub_lock);
2905 list_for_each_entry(s, &slab_caches, list) {
2907 * XXX: kmem_cache_alloc_node will fallback to other nodes
2908 * since memory is not yet available from the node that
2911 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
2916 init_kmem_cache_node(n, s);
2920 up_read(&slub_lock);
2924 static int slab_memory_callback(struct notifier_block *self,
2925 unsigned long action, void *arg)
2930 case MEM_GOING_ONLINE:
2931 ret = slab_mem_going_online_callback(arg);
2933 case MEM_GOING_OFFLINE:
2934 ret = slab_mem_going_offline_callback(arg);
2937 case MEM_CANCEL_ONLINE:
2938 slab_mem_offline_callback(arg);
2941 case MEM_CANCEL_OFFLINE:
2945 ret = notifier_from_errno(ret);
2951 #endif /* CONFIG_MEMORY_HOTPLUG */
2953 /********************************************************************
2954 * Basic setup of slabs
2955 *******************************************************************/
2958 * Used for early kmem_cache structures that were allocated using
2959 * the page allocator
2962 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
2966 list_add(&s->list, &slab_caches);
2969 for_each_node_state(node, N_NORMAL_MEMORY) {
2970 struct kmem_cache_node *n = get_node(s, node);
2974 list_for_each_entry(p, &n->partial, lru)
2977 #ifdef CONFIG_SLAB_DEBUG
2978 list_for_each_entry(p, &n->full, lru)
2985 void __init kmem_cache_init(void)
2989 struct kmem_cache *temp_kmem_cache;
2991 struct kmem_cache *temp_kmem_cache_node;
2992 unsigned long kmalloc_size;
2994 kmem_size = offsetof(struct kmem_cache, node) +
2995 nr_node_ids * sizeof(struct kmem_cache_node *);
2997 /* Allocate two kmem_caches from the page allocator */
2998 kmalloc_size = ALIGN(kmem_size, cache_line_size());
2999 order = get_order(2 * kmalloc_size);
3000 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3003 * Must first have the slab cache available for the allocations of the
3004 * struct kmem_cache_node's. There is special bootstrap code in
3005 * kmem_cache_open for slab_state == DOWN.
3007 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3009 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3010 sizeof(struct kmem_cache_node),
3011 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3013 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3015 /* Able to allocate the per node structures */
3016 slab_state = PARTIAL;
3018 temp_kmem_cache = kmem_cache;
3019 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3020 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3021 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3022 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3025 * Allocate kmem_cache_node properly from the kmem_cache slab.
3026 * kmem_cache_node is separately allocated so no need to
3027 * update any list pointers.
3029 temp_kmem_cache_node = kmem_cache_node;
3031 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3032 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3034 kmem_cache_bootstrap_fixup(kmem_cache_node);
3037 kmem_cache_bootstrap_fixup(kmem_cache);
3039 /* Free temporary boot structure */
3040 free_pages((unsigned long)temp_kmem_cache, order);
3042 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3045 * Patch up the size_index table if we have strange large alignment
3046 * requirements for the kmalloc array. This is only the case for
3047 * MIPS it seems. The standard arches will not generate any code here.
3049 * Largest permitted alignment is 256 bytes due to the way we
3050 * handle the index determination for the smaller caches.
3052 * Make sure that nothing crazy happens if someone starts tinkering
3053 * around with ARCH_KMALLOC_MINALIGN
3055 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3056 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3058 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3059 int elem = size_index_elem(i);
3060 if (elem >= ARRAY_SIZE(size_index))
3062 size_index[elem] = KMALLOC_SHIFT_LOW;
3065 if (KMALLOC_MIN_SIZE == 64) {
3067 * The 96 byte size cache is not used if the alignment
3070 for (i = 64 + 8; i <= 96; i += 8)
3071 size_index[size_index_elem(i)] = 7;
3072 } else if (KMALLOC_MIN_SIZE == 128) {
3074 * The 192 byte sized cache is not used if the alignment
3075 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3078 for (i = 128 + 8; i <= 192; i += 8)
3079 size_index[size_index_elem(i)] = 8;
3082 /* Caches that are not of the two-to-the-power-of size */
3083 if (KMALLOC_MIN_SIZE <= 32) {
3084 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3088 if (KMALLOC_MIN_SIZE <= 64) {
3089 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3093 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3094 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3100 /* Provide the correct kmalloc names now that the caches are up */
3101 if (KMALLOC_MIN_SIZE <= 32) {
3102 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3103 BUG_ON(!kmalloc_caches[1]->name);
3106 if (KMALLOC_MIN_SIZE <= 64) {
3107 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3108 BUG_ON(!kmalloc_caches[2]->name);
3111 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3112 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3115 kmalloc_caches[i]->name = s;
3119 register_cpu_notifier(&slab_notifier);
3122 #ifdef CONFIG_ZONE_DMA
3123 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3124 struct kmem_cache *s = kmalloc_caches[i];
3127 char *name = kasprintf(GFP_NOWAIT,
3128 "dma-kmalloc-%d", s->objsize);
3131 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3132 s->objsize, SLAB_CACHE_DMA);
3137 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3138 " CPUs=%d, Nodes=%d\n",
3139 caches, cache_line_size(),
3140 slub_min_order, slub_max_order, slub_min_objects,
3141 nr_cpu_ids, nr_node_ids);
3144 void __init kmem_cache_init_late(void)
3149 * Find a mergeable slab cache
3151 static int slab_unmergeable(struct kmem_cache *s)
3153 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3160 * We may have set a slab to be unmergeable during bootstrap.
3162 if (s->refcount < 0)
3168 static struct kmem_cache *find_mergeable(size_t size,
3169 size_t align, unsigned long flags, const char *name,
3170 void (*ctor)(void *))
3172 struct kmem_cache *s;
3174 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3180 size = ALIGN(size, sizeof(void *));
3181 align = calculate_alignment(flags, align, size);
3182 size = ALIGN(size, align);
3183 flags = kmem_cache_flags(size, flags, name, NULL);
3185 list_for_each_entry(s, &slab_caches, list) {
3186 if (slab_unmergeable(s))
3192 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3195 * Check if alignment is compatible.
3196 * Courtesy of Adrian Drzewiecki
3198 if ((s->size & ~(align - 1)) != s->size)
3201 if (s->size - size >= sizeof(void *))
3209 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3210 size_t align, unsigned long flags, void (*ctor)(void *))
3212 struct kmem_cache *s;
3218 down_write(&slub_lock);
3219 s = find_mergeable(size, align, flags, name, ctor);
3223 * Adjust the object sizes so that we clear
3224 * the complete object on kzalloc.
3226 s->objsize = max(s->objsize, (int)size);
3227 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3229 if (sysfs_slab_alias(s, name)) {
3233 up_write(&slub_lock);
3237 n = kstrdup(name, GFP_KERNEL);
3241 s = kmalloc(kmem_size, GFP_KERNEL);
3243 if (kmem_cache_open(s, n,
3244 size, align, flags, ctor)) {
3245 list_add(&s->list, &slab_caches);
3246 if (sysfs_slab_add(s)) {
3252 up_write(&slub_lock);
3259 up_write(&slub_lock);
3261 if (flags & SLAB_PANIC)
3262 panic("Cannot create slabcache %s\n", name);
3267 EXPORT_SYMBOL(kmem_cache_create);
3271 * Use the cpu notifier to insure that the cpu slabs are flushed when
3274 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3275 unsigned long action, void *hcpu)
3277 long cpu = (long)hcpu;
3278 struct kmem_cache *s;
3279 unsigned long flags;
3282 case CPU_UP_CANCELED:
3283 case CPU_UP_CANCELED_FROZEN:
3285 case CPU_DEAD_FROZEN:
3286 down_read(&slub_lock);
3287 list_for_each_entry(s, &slab_caches, list) {
3288 local_irq_save(flags);
3289 __flush_cpu_slab(s, cpu);
3290 local_irq_restore(flags);
3292 up_read(&slub_lock);
3300 static struct notifier_block __cpuinitdata slab_notifier = {
3301 .notifier_call = slab_cpuup_callback
3306 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3308 struct kmem_cache *s;
3311 if (unlikely(size > SLUB_MAX_SIZE))
3312 return kmalloc_large(size, gfpflags);
3314 s = get_slab(size, gfpflags);
3316 if (unlikely(ZERO_OR_NULL_PTR(s)))
3319 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3321 /* Honor the call site pointer we recieved. */
3322 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3328 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3329 int node, unsigned long caller)
3331 struct kmem_cache *s;
3334 if (unlikely(size > SLUB_MAX_SIZE)) {
3335 ret = kmalloc_large_node(size, gfpflags, node);
3337 trace_kmalloc_node(caller, ret,
3338 size, PAGE_SIZE << get_order(size),
3344 s = get_slab(size, gfpflags);
3346 if (unlikely(ZERO_OR_NULL_PTR(s)))
3349 ret = slab_alloc(s, gfpflags, node, caller);
3351 /* Honor the call site pointer we recieved. */
3352 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3359 static int count_inuse(struct page *page)
3364 static int count_total(struct page *page)
3366 return page->objects;
3370 #ifdef CONFIG_SLUB_DEBUG
3371 static int validate_slab(struct kmem_cache *s, struct page *page,
3375 void *addr = page_address(page);
3377 if (!check_slab(s, page) ||
3378 !on_freelist(s, page, NULL))
3381 /* Now we know that a valid freelist exists */
3382 bitmap_zero(map, page->objects);
3384 for_each_free_object(p, s, page->freelist) {
3385 set_bit(slab_index(p, s, addr), map);
3386 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3390 for_each_object(p, s, addr, page->objects)
3391 if (!test_bit(slab_index(p, s, addr), map))
3392 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3397 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3400 if (slab_trylock(page)) {
3401 validate_slab(s, page, map);
3404 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3408 static int validate_slab_node(struct kmem_cache *s,
3409 struct kmem_cache_node *n, unsigned long *map)
3411 unsigned long count = 0;
3413 unsigned long flags;
3415 spin_lock_irqsave(&n->list_lock, flags);
3417 list_for_each_entry(page, &n->partial, lru) {
3418 validate_slab_slab(s, page, map);
3421 if (count != n->nr_partial)
3422 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3423 "counter=%ld\n", s->name, count, n->nr_partial);
3425 if (!(s->flags & SLAB_STORE_USER))
3428 list_for_each_entry(page, &n->full, lru) {
3429 validate_slab_slab(s, page, map);
3432 if (count != atomic_long_read(&n->nr_slabs))
3433 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3434 "counter=%ld\n", s->name, count,
3435 atomic_long_read(&n->nr_slabs));
3438 spin_unlock_irqrestore(&n->list_lock, flags);
3442 static long validate_slab_cache(struct kmem_cache *s)
3445 unsigned long count = 0;
3446 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3447 sizeof(unsigned long), GFP_KERNEL);
3453 for_each_node_state(node, N_NORMAL_MEMORY) {
3454 struct kmem_cache_node *n = get_node(s, node);
3456 count += validate_slab_node(s, n, map);
3462 * Generate lists of code addresses where slabcache objects are allocated
3467 unsigned long count;
3474 DECLARE_BITMAP(cpus, NR_CPUS);
3480 unsigned long count;
3481 struct location *loc;
3484 static void free_loc_track(struct loc_track *t)
3487 free_pages((unsigned long)t->loc,
3488 get_order(sizeof(struct location) * t->max));
3491 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3496 order = get_order(sizeof(struct location) * max);
3498 l = (void *)__get_free_pages(flags, order);
3503 memcpy(l, t->loc, sizeof(struct location) * t->count);
3511 static int add_location(struct loc_track *t, struct kmem_cache *s,
3512 const struct track *track)
3514 long start, end, pos;
3516 unsigned long caddr;
3517 unsigned long age = jiffies - track->when;
3523 pos = start + (end - start + 1) / 2;
3526 * There is nothing at "end". If we end up there
3527 * we need to add something to before end.
3532 caddr = t->loc[pos].addr;
3533 if (track->addr == caddr) {
3539 if (age < l->min_time)
3541 if (age > l->max_time)
3544 if (track->pid < l->min_pid)
3545 l->min_pid = track->pid;
3546 if (track->pid > l->max_pid)
3547 l->max_pid = track->pid;
3549 cpumask_set_cpu(track->cpu,
3550 to_cpumask(l->cpus));
3552 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3556 if (track->addr < caddr)
3563 * Not found. Insert new tracking element.
3565 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3571 (t->count - pos) * sizeof(struct location));
3574 l->addr = track->addr;
3578 l->min_pid = track->pid;
3579 l->max_pid = track->pid;
3580 cpumask_clear(to_cpumask(l->cpus));
3581 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3582 nodes_clear(l->nodes);
3583 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3587 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3588 struct page *page, enum track_item alloc,
3591 void *addr = page_address(page);
3594 bitmap_zero(map, page->objects);
3595 for_each_free_object(p, s, page->freelist)
3596 set_bit(slab_index(p, s, addr), map);
3598 for_each_object(p, s, addr, page->objects)
3599 if (!test_bit(slab_index(p, s, addr), map))
3600 add_location(t, s, get_track(s, p, alloc));
3603 static int list_locations(struct kmem_cache *s, char *buf,
3604 enum track_item alloc)
3608 struct loc_track t = { 0, 0, NULL };
3610 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3611 sizeof(unsigned long), GFP_KERNEL);
3613 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3616 return sprintf(buf, "Out of memory\n");
3618 /* Push back cpu slabs */
3621 for_each_node_state(node, N_NORMAL_MEMORY) {
3622 struct kmem_cache_node *n = get_node(s, node);
3623 unsigned long flags;
3626 if (!atomic_long_read(&n->nr_slabs))
3629 spin_lock_irqsave(&n->list_lock, flags);
3630 list_for_each_entry(page, &n->partial, lru)
3631 process_slab(&t, s, page, alloc, map);
3632 list_for_each_entry(page, &n->full, lru)
3633 process_slab(&t, s, page, alloc, map);
3634 spin_unlock_irqrestore(&n->list_lock, flags);
3637 for (i = 0; i < t.count; i++) {
3638 struct location *l = &t.loc[i];
3640 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3642 len += sprintf(buf + len, "%7ld ", l->count);
3645 len += sprintf(buf + len, "%pS", (void *)l->addr);
3647 len += sprintf(buf + len, "<not-available>");
3649 if (l->sum_time != l->min_time) {
3650 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3652 (long)div_u64(l->sum_time, l->count),
3655 len += sprintf(buf + len, " age=%ld",
3658 if (l->min_pid != l->max_pid)
3659 len += sprintf(buf + len, " pid=%ld-%ld",
3660 l->min_pid, l->max_pid);
3662 len += sprintf(buf + len, " pid=%ld",
3665 if (num_online_cpus() > 1 &&
3666 !cpumask_empty(to_cpumask(l->cpus)) &&
3667 len < PAGE_SIZE - 60) {
3668 len += sprintf(buf + len, " cpus=");
3669 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3670 to_cpumask(l->cpus));
3673 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3674 len < PAGE_SIZE - 60) {
3675 len += sprintf(buf + len, " nodes=");
3676 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3680 len += sprintf(buf + len, "\n");
3686 len += sprintf(buf, "No data\n");
3691 #ifdef SLUB_RESILIENCY_TEST
3692 static void resiliency_test(void)
3696 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3698 printk(KERN_ERR "SLUB resiliency testing\n");
3699 printk(KERN_ERR "-----------------------\n");
3700 printk(KERN_ERR "A. Corruption after allocation\n");
3702 p = kzalloc(16, GFP_KERNEL);
3704 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3705 " 0x12->0x%p\n\n", p + 16);
3707 validate_slab_cache(kmalloc_caches[4]);
3709 /* Hmmm... The next two are dangerous */
3710 p = kzalloc(32, GFP_KERNEL);
3711 p[32 + sizeof(void *)] = 0x34;
3712 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3713 " 0x34 -> -0x%p\n", p);
3715 "If allocated object is overwritten then not detectable\n\n");
3717 validate_slab_cache(kmalloc_caches[5]);
3718 p = kzalloc(64, GFP_KERNEL);
3719 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3721 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3724 "If allocated object is overwritten then not detectable\n\n");
3725 validate_slab_cache(kmalloc_caches[6]);
3727 printk(KERN_ERR "\nB. Corruption after free\n");
3728 p = kzalloc(128, GFP_KERNEL);
3731 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3732 validate_slab_cache(kmalloc_caches[7]);
3734 p = kzalloc(256, GFP_KERNEL);
3737 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3739 validate_slab_cache(kmalloc_caches[8]);
3741 p = kzalloc(512, GFP_KERNEL);
3744 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3745 validate_slab_cache(kmalloc_caches[9]);
3749 static void resiliency_test(void) {};
3754 enum slab_stat_type {
3755 SL_ALL, /* All slabs */
3756 SL_PARTIAL, /* Only partially allocated slabs */
3757 SL_CPU, /* Only slabs used for cpu caches */
3758 SL_OBJECTS, /* Determine allocated objects not slabs */
3759 SL_TOTAL /* Determine object capacity not slabs */
3762 #define SO_ALL (1 << SL_ALL)
3763 #define SO_PARTIAL (1 << SL_PARTIAL)
3764 #define SO_CPU (1 << SL_CPU)
3765 #define SO_OBJECTS (1 << SL_OBJECTS)
3766 #define SO_TOTAL (1 << SL_TOTAL)
3768 static ssize_t show_slab_objects(struct kmem_cache *s,
3769 char *buf, unsigned long flags)
3771 unsigned long total = 0;
3774 unsigned long *nodes;
3775 unsigned long *per_cpu;
3777 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3780 per_cpu = nodes + nr_node_ids;
3782 if (flags & SO_CPU) {
3785 for_each_possible_cpu(cpu) {
3786 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3788 if (!c || c->node < 0)
3792 if (flags & SO_TOTAL)
3793 x = c->page->objects;
3794 else if (flags & SO_OBJECTS)
3800 nodes[c->node] += x;
3806 lock_memory_hotplug();
3807 #ifdef CONFIG_SLUB_DEBUG
3808 if (flags & SO_ALL) {
3809 for_each_node_state(node, N_NORMAL_MEMORY) {
3810 struct kmem_cache_node *n = get_node(s, node);
3812 if (flags & SO_TOTAL)
3813 x = atomic_long_read(&n->total_objects);
3814 else if (flags & SO_OBJECTS)
3815 x = atomic_long_read(&n->total_objects) -
3816 count_partial(n, count_free);
3819 x = atomic_long_read(&n->nr_slabs);
3826 if (flags & SO_PARTIAL) {
3827 for_each_node_state(node, N_NORMAL_MEMORY) {
3828 struct kmem_cache_node *n = get_node(s, node);
3830 if (flags & SO_TOTAL)
3831 x = count_partial(n, count_total);
3832 else if (flags & SO_OBJECTS)
3833 x = count_partial(n, count_inuse);
3840 x = sprintf(buf, "%lu", total);
3842 for_each_node_state(node, N_NORMAL_MEMORY)
3844 x += sprintf(buf + x, " N%d=%lu",
3847 unlock_memory_hotplug();
3849 return x + sprintf(buf + x, "\n");
3852 #ifdef CONFIG_SLUB_DEBUG
3853 static int any_slab_objects(struct kmem_cache *s)
3857 for_each_online_node(node) {
3858 struct kmem_cache_node *n = get_node(s, node);
3863 if (atomic_long_read(&n->total_objects))
3870 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3871 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3873 struct slab_attribute {
3874 struct attribute attr;
3875 ssize_t (*show)(struct kmem_cache *s, char *buf);
3876 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3879 #define SLAB_ATTR_RO(_name) \
3880 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3882 #define SLAB_ATTR(_name) \
3883 static struct slab_attribute _name##_attr = \
3884 __ATTR(_name, 0644, _name##_show, _name##_store)
3886 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3888 return sprintf(buf, "%d\n", s->size);
3890 SLAB_ATTR_RO(slab_size);
3892 static ssize_t align_show(struct kmem_cache *s, char *buf)
3894 return sprintf(buf, "%d\n", s->align);
3896 SLAB_ATTR_RO(align);
3898 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3900 return sprintf(buf, "%d\n", s->objsize);
3902 SLAB_ATTR_RO(object_size);
3904 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3906 return sprintf(buf, "%d\n", oo_objects(s->oo));
3908 SLAB_ATTR_RO(objs_per_slab);
3910 static ssize_t order_store(struct kmem_cache *s,
3911 const char *buf, size_t length)
3913 unsigned long order;
3916 err = strict_strtoul(buf, 10, &order);
3920 if (order > slub_max_order || order < slub_min_order)
3923 calculate_sizes(s, order);
3927 static ssize_t order_show(struct kmem_cache *s, char *buf)
3929 return sprintf(buf, "%d\n", oo_order(s->oo));
3933 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3935 return sprintf(buf, "%lu\n", s->min_partial);
3938 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3944 err = strict_strtoul(buf, 10, &min);
3948 set_min_partial(s, min);
3951 SLAB_ATTR(min_partial);
3953 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3957 return sprintf(buf, "%pS\n", s->ctor);
3961 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3963 return sprintf(buf, "%d\n", s->refcount - 1);
3965 SLAB_ATTR_RO(aliases);
3967 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3969 return show_slab_objects(s, buf, SO_PARTIAL);
3971 SLAB_ATTR_RO(partial);
3973 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3975 return show_slab_objects(s, buf, SO_CPU);
3977 SLAB_ATTR_RO(cpu_slabs);
3979 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3981 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3983 SLAB_ATTR_RO(objects);
3985 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3987 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3989 SLAB_ATTR_RO(objects_partial);
3991 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3993 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3996 static ssize_t reclaim_account_store(struct kmem_cache *s,
3997 const char *buf, size_t length)
3999 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4001 s->flags |= SLAB_RECLAIM_ACCOUNT;
4004 SLAB_ATTR(reclaim_account);
4006 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4008 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4010 SLAB_ATTR_RO(hwcache_align);
4012 #ifdef CONFIG_ZONE_DMA
4013 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4015 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4017 SLAB_ATTR_RO(cache_dma);
4020 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4022 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4024 SLAB_ATTR_RO(destroy_by_rcu);
4026 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4028 return sprintf(buf, "%d\n", s->reserved);
4030 SLAB_ATTR_RO(reserved);
4032 #ifdef CONFIG_SLUB_DEBUG
4033 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4035 return show_slab_objects(s, buf, SO_ALL);
4037 SLAB_ATTR_RO(slabs);
4039 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4041 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4043 SLAB_ATTR_RO(total_objects);
4045 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4047 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4050 static ssize_t sanity_checks_store(struct kmem_cache *s,
4051 const char *buf, size_t length)
4053 s->flags &= ~SLAB_DEBUG_FREE;
4055 s->flags |= SLAB_DEBUG_FREE;
4058 SLAB_ATTR(sanity_checks);
4060 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4062 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4065 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4068 s->flags &= ~SLAB_TRACE;
4070 s->flags |= SLAB_TRACE;
4075 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4077 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4080 static ssize_t red_zone_store(struct kmem_cache *s,
4081 const char *buf, size_t length)
4083 if (any_slab_objects(s))
4086 s->flags &= ~SLAB_RED_ZONE;
4088 s->flags |= SLAB_RED_ZONE;
4089 calculate_sizes(s, -1);
4092 SLAB_ATTR(red_zone);
4094 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4096 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4099 static ssize_t poison_store(struct kmem_cache *s,
4100 const char *buf, size_t length)
4102 if (any_slab_objects(s))
4105 s->flags &= ~SLAB_POISON;
4107 s->flags |= SLAB_POISON;
4108 calculate_sizes(s, -1);
4113 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4115 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4118 static ssize_t store_user_store(struct kmem_cache *s,
4119 const char *buf, size_t length)
4121 if (any_slab_objects(s))
4124 s->flags &= ~SLAB_STORE_USER;
4126 s->flags |= SLAB_STORE_USER;
4127 calculate_sizes(s, -1);
4130 SLAB_ATTR(store_user);
4132 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4137 static ssize_t validate_store(struct kmem_cache *s,
4138 const char *buf, size_t length)
4142 if (buf[0] == '1') {
4143 ret = validate_slab_cache(s);
4149 SLAB_ATTR(validate);
4151 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4153 if (!(s->flags & SLAB_STORE_USER))
4155 return list_locations(s, buf, TRACK_ALLOC);
4157 SLAB_ATTR_RO(alloc_calls);
4159 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4161 if (!(s->flags & SLAB_STORE_USER))
4163 return list_locations(s, buf, TRACK_FREE);
4165 SLAB_ATTR_RO(free_calls);
4166 #endif /* CONFIG_SLUB_DEBUG */
4168 #ifdef CONFIG_FAILSLAB
4169 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4171 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4174 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4177 s->flags &= ~SLAB_FAILSLAB;
4179 s->flags |= SLAB_FAILSLAB;
4182 SLAB_ATTR(failslab);
4185 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4190 static ssize_t shrink_store(struct kmem_cache *s,
4191 const char *buf, size_t length)
4193 if (buf[0] == '1') {
4194 int rc = kmem_cache_shrink(s);
4205 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4207 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4210 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4211 const char *buf, size_t length)
4213 unsigned long ratio;
4216 err = strict_strtoul(buf, 10, &ratio);
4221 s->remote_node_defrag_ratio = ratio * 10;
4225 SLAB_ATTR(remote_node_defrag_ratio);
4228 #ifdef CONFIG_SLUB_STATS
4229 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4231 unsigned long sum = 0;
4234 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4239 for_each_online_cpu(cpu) {
4240 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4246 len = sprintf(buf, "%lu", sum);
4249 for_each_online_cpu(cpu) {
4250 if (data[cpu] && len < PAGE_SIZE - 20)
4251 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4255 return len + sprintf(buf + len, "\n");
4258 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4262 for_each_online_cpu(cpu)
4263 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4266 #define STAT_ATTR(si, text) \
4267 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4269 return show_stat(s, buf, si); \
4271 static ssize_t text##_store(struct kmem_cache *s, \
4272 const char *buf, size_t length) \
4274 if (buf[0] != '0') \
4276 clear_stat(s, si); \
4281 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4282 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4283 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4284 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4285 STAT_ATTR(FREE_FROZEN, free_frozen);
4286 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4287 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4288 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4289 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4290 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4291 STAT_ATTR(FREE_SLAB, free_slab);
4292 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4293 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4294 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4295 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4296 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4297 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4298 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4301 static struct attribute *slab_attrs[] = {
4302 &slab_size_attr.attr,
4303 &object_size_attr.attr,
4304 &objs_per_slab_attr.attr,
4306 &min_partial_attr.attr,
4308 &objects_partial_attr.attr,
4310 &cpu_slabs_attr.attr,
4314 &hwcache_align_attr.attr,
4315 &reclaim_account_attr.attr,
4316 &destroy_by_rcu_attr.attr,
4318 &reserved_attr.attr,
4319 #ifdef CONFIG_SLUB_DEBUG
4320 &total_objects_attr.attr,
4322 &sanity_checks_attr.attr,
4324 &red_zone_attr.attr,
4326 &store_user_attr.attr,
4327 &validate_attr.attr,
4328 &alloc_calls_attr.attr,
4329 &free_calls_attr.attr,
4331 #ifdef CONFIG_ZONE_DMA
4332 &cache_dma_attr.attr,
4335 &remote_node_defrag_ratio_attr.attr,
4337 #ifdef CONFIG_SLUB_STATS
4338 &alloc_fastpath_attr.attr,
4339 &alloc_slowpath_attr.attr,
4340 &free_fastpath_attr.attr,
4341 &free_slowpath_attr.attr,
4342 &free_frozen_attr.attr,
4343 &free_add_partial_attr.attr,
4344 &free_remove_partial_attr.attr,
4345 &alloc_from_partial_attr.attr,
4346 &alloc_slab_attr.attr,
4347 &alloc_refill_attr.attr,
4348 &free_slab_attr.attr,
4349 &cpuslab_flush_attr.attr,
4350 &deactivate_full_attr.attr,
4351 &deactivate_empty_attr.attr,
4352 &deactivate_to_head_attr.attr,
4353 &deactivate_to_tail_attr.attr,
4354 &deactivate_remote_frees_attr.attr,
4355 &order_fallback_attr.attr,
4357 #ifdef CONFIG_FAILSLAB
4358 &failslab_attr.attr,
4364 static struct attribute_group slab_attr_group = {
4365 .attrs = slab_attrs,
4368 static ssize_t slab_attr_show(struct kobject *kobj,
4369 struct attribute *attr,
4372 struct slab_attribute *attribute;
4373 struct kmem_cache *s;
4376 attribute = to_slab_attr(attr);
4379 if (!attribute->show)
4382 err = attribute->show(s, buf);
4387 static ssize_t slab_attr_store(struct kobject *kobj,
4388 struct attribute *attr,
4389 const char *buf, size_t len)
4391 struct slab_attribute *attribute;
4392 struct kmem_cache *s;
4395 attribute = to_slab_attr(attr);
4398 if (!attribute->store)
4401 err = attribute->store(s, buf, len);
4406 static void kmem_cache_release(struct kobject *kobj)
4408 struct kmem_cache *s = to_slab(kobj);
4414 static const struct sysfs_ops slab_sysfs_ops = {
4415 .show = slab_attr_show,
4416 .store = slab_attr_store,
4419 static struct kobj_type slab_ktype = {
4420 .sysfs_ops = &slab_sysfs_ops,
4421 .release = kmem_cache_release
4424 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4426 struct kobj_type *ktype = get_ktype(kobj);
4428 if (ktype == &slab_ktype)
4433 static const struct kset_uevent_ops slab_uevent_ops = {
4434 .filter = uevent_filter,
4437 static struct kset *slab_kset;
4439 #define ID_STR_LENGTH 64
4441 /* Create a unique string id for a slab cache:
4443 * Format :[flags-]size
4445 static char *create_unique_id(struct kmem_cache *s)
4447 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4454 * First flags affecting slabcache operations. We will only
4455 * get here for aliasable slabs so we do not need to support
4456 * too many flags. The flags here must cover all flags that
4457 * are matched during merging to guarantee that the id is
4460 if (s->flags & SLAB_CACHE_DMA)
4462 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4464 if (s->flags & SLAB_DEBUG_FREE)
4466 if (!(s->flags & SLAB_NOTRACK))
4470 p += sprintf(p, "%07d", s->size);
4471 BUG_ON(p > name + ID_STR_LENGTH - 1);
4475 static int sysfs_slab_add(struct kmem_cache *s)
4481 if (slab_state < SYSFS)
4482 /* Defer until later */
4485 unmergeable = slab_unmergeable(s);
4488 * Slabcache can never be merged so we can use the name proper.
4489 * This is typically the case for debug situations. In that
4490 * case we can catch duplicate names easily.
4492 sysfs_remove_link(&slab_kset->kobj, s->name);
4496 * Create a unique name for the slab as a target
4499 name = create_unique_id(s);
4502 s->kobj.kset = slab_kset;
4503 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4505 kobject_put(&s->kobj);
4509 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4511 kobject_del(&s->kobj);
4512 kobject_put(&s->kobj);
4515 kobject_uevent(&s->kobj, KOBJ_ADD);
4517 /* Setup first alias */
4518 sysfs_slab_alias(s, s->name);
4524 static void sysfs_slab_remove(struct kmem_cache *s)
4526 if (slab_state < SYSFS)
4528 * Sysfs has not been setup yet so no need to remove the
4533 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4534 kobject_del(&s->kobj);
4535 kobject_put(&s->kobj);
4539 * Need to buffer aliases during bootup until sysfs becomes
4540 * available lest we lose that information.
4542 struct saved_alias {
4543 struct kmem_cache *s;
4545 struct saved_alias *next;
4548 static struct saved_alias *alias_list;
4550 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4552 struct saved_alias *al;
4554 if (slab_state == SYSFS) {
4556 * If we have a leftover link then remove it.
4558 sysfs_remove_link(&slab_kset->kobj, name);
4559 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4562 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4568 al->next = alias_list;
4573 static int __init slab_sysfs_init(void)
4575 struct kmem_cache *s;
4578 down_write(&slub_lock);
4580 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4582 up_write(&slub_lock);
4583 printk(KERN_ERR "Cannot register slab subsystem.\n");
4589 list_for_each_entry(s, &slab_caches, list) {
4590 err = sysfs_slab_add(s);
4592 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4593 " to sysfs\n", s->name);
4596 while (alias_list) {
4597 struct saved_alias *al = alias_list;
4599 alias_list = alias_list->next;
4600 err = sysfs_slab_alias(al->s, al->name);
4602 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4603 " %s to sysfs\n", s->name);
4607 up_write(&slub_lock);
4612 __initcall(slab_sysfs_init);
4613 #endif /* CONFIG_SYSFS */
4616 * The /proc/slabinfo ABI
4618 #ifdef CONFIG_SLABINFO
4619 static void print_slabinfo_header(struct seq_file *m)
4621 seq_puts(m, "slabinfo - version: 2.1\n");
4622 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4623 "<objperslab> <pagesperslab>");
4624 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4625 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4629 static void *s_start(struct seq_file *m, loff_t *pos)
4633 down_read(&slub_lock);
4635 print_slabinfo_header(m);
4637 return seq_list_start(&slab_caches, *pos);
4640 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4642 return seq_list_next(p, &slab_caches, pos);
4645 static void s_stop(struct seq_file *m, void *p)
4647 up_read(&slub_lock);
4650 static int s_show(struct seq_file *m, void *p)
4652 unsigned long nr_partials = 0;
4653 unsigned long nr_slabs = 0;
4654 unsigned long nr_inuse = 0;
4655 unsigned long nr_objs = 0;
4656 unsigned long nr_free = 0;
4657 struct kmem_cache *s;
4660 s = list_entry(p, struct kmem_cache, list);
4662 for_each_online_node(node) {
4663 struct kmem_cache_node *n = get_node(s, node);
4668 nr_partials += n->nr_partial;
4669 nr_slabs += atomic_long_read(&n->nr_slabs);
4670 nr_objs += atomic_long_read(&n->total_objects);
4671 nr_free += count_partial(n, count_free);
4674 nr_inuse = nr_objs - nr_free;
4676 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4677 nr_objs, s->size, oo_objects(s->oo),
4678 (1 << oo_order(s->oo)));
4679 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4680 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4686 static const struct seq_operations slabinfo_op = {
4693 static int slabinfo_open(struct inode *inode, struct file *file)
4695 return seq_open(file, &slabinfo_op);
4698 static const struct file_operations proc_slabinfo_operations = {
4699 .open = slabinfo_open,
4701 .llseek = seq_lseek,
4702 .release = seq_release,
4705 static int __init slab_proc_init(void)
4707 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4710 module_init(slab_proc_init);
4711 #endif /* CONFIG_SLABINFO */