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/kmemtrace.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
37 * The slab_lock protects operations on the object of a particular
38 * slab and its metadata in the page struct. If the slab lock
39 * has been taken then no allocations nor frees can be performed
40 * on the objects in the slab nor can the slab be added or removed
41 * from the partial or full lists since this would mean modifying
42 * the page_struct of the slab.
44 * The list_lock protects the partial and full list on each node and
45 * the partial slab counter. If taken then no new slabs may be added or
46 * removed from the lists nor make the number of partial slabs be modified.
47 * (Note that the total number of slabs is an atomic value that may be
48 * modified without taking the list lock).
50 * The list_lock is a centralized lock and thus we avoid taking it as
51 * much as possible. As long as SLUB does not have to handle partial
52 * slabs, operations can continue without any centralized lock. F.e.
53 * allocating a long series of objects that fill up slabs does not require
56 * The lock order is sometimes inverted when we are trying to get a slab
57 * off a list. We take the list_lock and then look for a page on the list
58 * to use. While we do that objects in the slabs may be freed. We can
59 * only operate on the slab if we have also taken the slab_lock. So we use
60 * a slab_trylock() on the slab. If trylock was successful then no frees
61 * can occur anymore and we can use the slab for allocations etc. If the
62 * slab_trylock() does not succeed then frees are in progress in the slab and
63 * we must stay away from it for a while since we may cause a bouncing
64 * cacheline if we try to acquire the lock. So go onto the next slab.
65 * If all pages are busy then we may allocate a new slab instead of reusing
66 * a partial slab. A new slab has noone operating on it and thus there is
67 * no danger of cacheline contention.
69 * Interrupts are disabled during allocation and deallocation in order to
70 * make the slab allocator safe to use in the context of an irq. In addition
71 * interrupts are disabled to ensure that the processor does not change
72 * while handling per_cpu slabs, due to kernel preemption.
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
77 * Slabs with free elements are kept on a partial list and during regular
78 * operations no list for full slabs is used. If an object in a full slab is
79 * freed then the slab will show up again on the partial lists.
80 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
83 * Slabs are freed when they become empty. Teardown and setup is
84 * minimal so we rely on the page allocators per cpu caches for
85 * fast frees and allocs.
87 * Overloading of page flags that are otherwise used for LRU management.
89 * PageActive The slab is frozen and exempt from list processing.
90 * This means that the slab is dedicated to a purpose
91 * such as satisfying allocations for a specific
92 * processor. Objects may be freed in the slab while
93 * it is frozen but slab_free will then skip the usual
94 * list operations. It is up to the processor holding
95 * the slab to integrate the slab into the slab lists
96 * when the slab is no longer needed.
98 * One use of this flag is to mark slabs that are
99 * used for allocations. Then such a slab becomes a cpu
100 * slab. The cpu slab may be equipped with an additional
101 * freelist that allows lockless access to
102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
107 * the fast path and disables lockless freelists.
110 #ifdef CONFIG_SLUB_DEBUG
117 * Issues still to be resolved:
119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
121 * - Variable sizing of the per node arrays
124 /* Enable to test recovery from slab corruption on boot */
125 #undef SLUB_RESILIENCY_TEST
128 * Mininum number of partial slabs. These will be left on the partial
129 * lists even if they are empty. kmem_cache_shrink may reclaim them.
131 #define MIN_PARTIAL 5
134 * Maximum number of desirable partial slabs.
135 * The existence of more partial slabs makes kmem_cache_shrink
136 * sort the partial list by the number of objects in the.
138 #define MAX_PARTIAL 10
140 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
141 SLAB_POISON | SLAB_STORE_USER)
144 * Debugging flags that require metadata to be stored in the slab. These get
145 * disabled when slub_debug=O is used and a cache's min order increases with
148 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
151 * Set of flags that will prevent slab merging
153 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
154 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
157 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
158 SLAB_CACHE_DMA | SLAB_NOTRACK)
160 #ifndef ARCH_KMALLOC_MINALIGN
161 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
164 #ifndef ARCH_SLAB_MINALIGN
165 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
169 #define OO_MASK ((1 << OO_SHIFT) - 1)
170 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
172 /* Internal SLUB flags */
173 #define __OBJECT_POISON 0x80000000 /* Poison object */
174 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
176 static int kmem_size = sizeof(struct kmem_cache);
179 static struct notifier_block slab_notifier;
183 DOWN, /* No slab functionality available */
184 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
185 UP, /* Everything works but does not show up in sysfs */
189 /* A list of all slab caches on the system */
190 static DECLARE_RWSEM(slub_lock);
191 static LIST_HEAD(slab_caches);
194 * Tracking user of a slab.
197 unsigned long addr; /* Called from address */
198 int cpu; /* Was running on cpu */
199 int pid; /* Pid context */
200 unsigned long when; /* When did the operation occur */
203 enum track_item { TRACK_ALLOC, TRACK_FREE };
205 #ifdef CONFIG_SLUB_DEBUG
206 static int sysfs_slab_add(struct kmem_cache *);
207 static int sysfs_slab_alias(struct kmem_cache *, const char *);
208 static void sysfs_slab_remove(struct kmem_cache *);
211 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
212 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
214 static inline void sysfs_slab_remove(struct kmem_cache *s)
221 static inline void stat(struct kmem_cache *s, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
224 __this_cpu_inc(s->cpu_slab->stat[si]);
228 /********************************************************************
229 * Core slab cache functions
230 *******************************************************************/
232 int slab_is_available(void)
234 return slab_state >= UP;
237 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
240 return s->node[node];
242 return &s->local_node;
246 /* Verify that a pointer has an address that is valid within a slab page */
247 static inline int check_valid_pointer(struct kmem_cache *s,
248 struct page *page, const void *object)
255 base = page_address(page);
256 if (object < base || object >= base + page->objects * s->size ||
257 (object - base) % s->size) {
264 static inline void *get_freepointer(struct kmem_cache *s, void *object)
266 return *(void **)(object + s->offset);
269 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
271 *(void **)(object + s->offset) = fp;
274 /* Loop over all objects in a slab */
275 #define for_each_object(__p, __s, __addr, __objects) \
276 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
280 #define for_each_free_object(__p, __s, __free) \
281 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
283 /* Determine object index from a given position */
284 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
286 return (p - addr) / s->size;
289 static inline struct kmem_cache_order_objects oo_make(int order,
292 struct kmem_cache_order_objects x = {
293 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
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, int active)
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,
511 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
512 s->inuse - s->objsize);
515 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
518 if (*start != (u8)value)
526 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
527 void *from, void *to)
529 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
530 memset(from, data, to - from);
533 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
534 u8 *object, char *what,
535 u8 *start, unsigned int value, unsigned int bytes)
540 fault = check_bytes(start, value, bytes);
545 while (end > fault && end[-1] == value)
548 slab_bug(s, "%s overwritten", what);
549 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
550 fault, end - 1, fault[0], value);
551 print_trailer(s, page, object);
553 restore_bytes(s, what, value, fault, end);
561 * Bytes of the object to be managed.
562 * If the freepointer may overlay the object then the free
563 * pointer is the first word of the object.
565 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
568 * object + s->objsize
569 * Padding to reach word boundary. This is also used for Redzoning.
570 * Padding is extended by another word if Redzoning is enabled and
573 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
574 * 0xcc (RED_ACTIVE) for objects in use.
577 * Meta data starts here.
579 * A. Free pointer (if we cannot overwrite object on free)
580 * B. Tracking data for SLAB_STORE_USER
581 * C. Padding to reach required alignment boundary or at mininum
582 * one word if debugging is on to be able to detect writes
583 * before the word boundary.
585 * Padding is done using 0x5a (POISON_INUSE)
588 * Nothing is used beyond s->size.
590 * If slabcaches are merged then the objsize and inuse boundaries are mostly
591 * ignored. And therefore no slab options that rely on these boundaries
592 * may be used with merged slabcaches.
595 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
597 unsigned long off = s->inuse; /* The end of info */
600 /* Freepointer is placed after the object. */
601 off += sizeof(void *);
603 if (s->flags & SLAB_STORE_USER)
604 /* We also have user information there */
605 off += 2 * sizeof(struct track);
610 return check_bytes_and_report(s, page, p, "Object padding",
611 p + off, POISON_INUSE, s->size - off);
614 /* Check the pad bytes at the end of a slab page */
615 static int slab_pad_check(struct kmem_cache *s, struct page *page)
623 if (!(s->flags & SLAB_POISON))
626 start = page_address(page);
627 length = (PAGE_SIZE << compound_order(page));
628 end = start + length;
629 remainder = length % s->size;
633 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
636 while (end > fault && end[-1] == POISON_INUSE)
639 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
640 print_section("Padding", end - remainder, remainder);
642 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
646 static int check_object(struct kmem_cache *s, struct page *page,
647 void *object, int active)
650 u8 *endobject = object + s->objsize;
652 if (s->flags & SLAB_RED_ZONE) {
654 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
656 if (!check_bytes_and_report(s, page, object, "Redzone",
657 endobject, red, s->inuse - s->objsize))
660 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
661 check_bytes_and_report(s, page, p, "Alignment padding",
662 endobject, POISON_INUSE, s->inuse - s->objsize);
666 if (s->flags & SLAB_POISON) {
667 if (!active && (s->flags & __OBJECT_POISON) &&
668 (!check_bytes_and_report(s, page, p, "Poison", p,
669 POISON_FREE, s->objsize - 1) ||
670 !check_bytes_and_report(s, page, p, "Poison",
671 p + s->objsize - 1, POISON_END, 1)))
674 * check_pad_bytes cleans up on its own.
676 check_pad_bytes(s, page, p);
679 if (!s->offset && active)
681 * Object and freepointer overlap. Cannot check
682 * freepointer while object is allocated.
686 /* Check free pointer validity */
687 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
688 object_err(s, page, p, "Freepointer corrupt");
690 * No choice but to zap it and thus lose the remainder
691 * of the free objects in this slab. May cause
692 * another error because the object count is now wrong.
694 set_freepointer(s, p, NULL);
700 static int check_slab(struct kmem_cache *s, struct page *page)
704 VM_BUG_ON(!irqs_disabled());
706 if (!PageSlab(page)) {
707 slab_err(s, page, "Not a valid slab page");
711 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
712 if (page->objects > maxobj) {
713 slab_err(s, page, "objects %u > max %u",
714 s->name, page->objects, maxobj);
717 if (page->inuse > page->objects) {
718 slab_err(s, page, "inuse %u > max %u",
719 s->name, page->inuse, page->objects);
722 /* Slab_pad_check fixes things up after itself */
723 slab_pad_check(s, page);
728 * Determine if a certain object on a page is on the freelist. Must hold the
729 * slab lock to guarantee that the chains are in a consistent state.
731 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
734 void *fp = page->freelist;
736 unsigned long max_objects;
738 while (fp && nr <= page->objects) {
741 if (!check_valid_pointer(s, page, fp)) {
743 object_err(s, page, object,
744 "Freechain corrupt");
745 set_freepointer(s, object, NULL);
748 slab_err(s, page, "Freepointer corrupt");
749 page->freelist = NULL;
750 page->inuse = page->objects;
751 slab_fix(s, "Freelist cleared");
757 fp = get_freepointer(s, object);
761 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
762 if (max_objects > MAX_OBJS_PER_PAGE)
763 max_objects = MAX_OBJS_PER_PAGE;
765 if (page->objects != max_objects) {
766 slab_err(s, page, "Wrong number of objects. Found %d but "
767 "should be %d", page->objects, max_objects);
768 page->objects = max_objects;
769 slab_fix(s, "Number of objects adjusted.");
771 if (page->inuse != page->objects - nr) {
772 slab_err(s, page, "Wrong object count. Counter is %d but "
773 "counted were %d", page->inuse, page->objects - nr);
774 page->inuse = page->objects - nr;
775 slab_fix(s, "Object count adjusted.");
777 return search == NULL;
780 static void trace(struct kmem_cache *s, struct page *page, void *object,
783 if (s->flags & SLAB_TRACE) {
784 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
786 alloc ? "alloc" : "free",
791 print_section("Object", (void *)object, s->objsize);
798 * Tracking of fully allocated slabs for debugging purposes.
800 static void add_full(struct kmem_cache_node *n, struct page *page)
802 spin_lock(&n->list_lock);
803 list_add(&page->lru, &n->full);
804 spin_unlock(&n->list_lock);
807 static void remove_full(struct kmem_cache *s, struct page *page)
809 struct kmem_cache_node *n;
811 if (!(s->flags & SLAB_STORE_USER))
814 n = get_node(s, page_to_nid(page));
816 spin_lock(&n->list_lock);
817 list_del(&page->lru);
818 spin_unlock(&n->list_lock);
821 /* Tracking of the number of slabs for debugging purposes */
822 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
824 struct kmem_cache_node *n = get_node(s, node);
826 return atomic_long_read(&n->nr_slabs);
829 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
831 return atomic_long_read(&n->nr_slabs);
834 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
836 struct kmem_cache_node *n = get_node(s, node);
839 * May be called early in order to allocate a slab for the
840 * kmem_cache_node structure. Solve the chicken-egg
841 * dilemma by deferring the increment of the count during
842 * bootstrap (see early_kmem_cache_node_alloc).
844 if (!NUMA_BUILD || n) {
845 atomic_long_inc(&n->nr_slabs);
846 atomic_long_add(objects, &n->total_objects);
849 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
851 struct kmem_cache_node *n = get_node(s, node);
853 atomic_long_dec(&n->nr_slabs);
854 atomic_long_sub(objects, &n->total_objects);
857 /* Object debug checks for alloc/free paths */
858 static void setup_object_debug(struct kmem_cache *s, struct page *page,
861 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
864 init_object(s, object, 0);
865 init_tracking(s, object);
868 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
869 void *object, unsigned long addr)
871 if (!check_slab(s, page))
874 if (!on_freelist(s, page, object)) {
875 object_err(s, page, object, "Object already allocated");
879 if (!check_valid_pointer(s, page, object)) {
880 object_err(s, page, object, "Freelist Pointer check fails");
884 if (!check_object(s, page, object, 0))
887 /* Success perform special debug activities for allocs */
888 if (s->flags & SLAB_STORE_USER)
889 set_track(s, object, TRACK_ALLOC, addr);
890 trace(s, page, object, 1);
891 init_object(s, object, 1);
895 if (PageSlab(page)) {
897 * If this is a slab page then lets do the best we can
898 * to avoid issues in the future. Marking all objects
899 * as used avoids touching the remaining objects.
901 slab_fix(s, "Marking all objects used");
902 page->inuse = page->objects;
903 page->freelist = NULL;
908 static int free_debug_processing(struct kmem_cache *s, struct page *page,
909 void *object, unsigned long addr)
911 if (!check_slab(s, page))
914 if (!check_valid_pointer(s, page, object)) {
915 slab_err(s, page, "Invalid object pointer 0x%p", object);
919 if (on_freelist(s, page, object)) {
920 object_err(s, page, object, "Object already free");
924 if (!check_object(s, page, object, 1))
927 if (unlikely(s != page->slab)) {
928 if (!PageSlab(page)) {
929 slab_err(s, page, "Attempt to free object(0x%p) "
930 "outside of slab", object);
931 } else if (!page->slab) {
933 "SLUB <none>: no slab for object 0x%p.\n",
937 object_err(s, page, object,
938 "page slab pointer corrupt.");
942 /* Special debug activities for freeing objects */
943 if (!PageSlubFrozen(page) && !page->freelist)
944 remove_full(s, page);
945 if (s->flags & SLAB_STORE_USER)
946 set_track(s, object, TRACK_FREE, addr);
947 trace(s, page, object, 0);
948 init_object(s, object, 0);
952 slab_fix(s, "Object at 0x%p not freed", object);
956 static int __init setup_slub_debug(char *str)
958 slub_debug = DEBUG_DEFAULT_FLAGS;
959 if (*str++ != '=' || !*str)
961 * No options specified. Switch on full debugging.
967 * No options but restriction on slabs. This means full
968 * debugging for slabs matching a pattern.
972 if (tolower(*str) == 'o') {
974 * Avoid enabling debugging on caches if its minimum order
975 * would increase as a result.
977 disable_higher_order_debug = 1;
984 * Switch off all debugging measures.
989 * Determine which debug features should be switched on
991 for (; *str && *str != ','; str++) {
992 switch (tolower(*str)) {
994 slub_debug |= SLAB_DEBUG_FREE;
997 slub_debug |= SLAB_RED_ZONE;
1000 slub_debug |= SLAB_POISON;
1003 slub_debug |= SLAB_STORE_USER;
1006 slub_debug |= SLAB_TRACE;
1009 slub_debug |= SLAB_FAILSLAB;
1012 printk(KERN_ERR "slub_debug option '%c' "
1013 "unknown. skipped\n", *str);
1019 slub_debug_slabs = str + 1;
1024 __setup("slub_debug", setup_slub_debug);
1026 static unsigned long kmem_cache_flags(unsigned long objsize,
1027 unsigned long flags, const char *name,
1028 void (*ctor)(void *))
1031 * Enable debugging if selected on the kernel commandline.
1033 if (slub_debug && (!slub_debug_slabs ||
1034 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1035 flags |= slub_debug;
1040 static inline void setup_object_debug(struct kmem_cache *s,
1041 struct page *page, void *object) {}
1043 static inline int alloc_debug_processing(struct kmem_cache *s,
1044 struct page *page, void *object, unsigned long addr) { return 0; }
1046 static inline int free_debug_processing(struct kmem_cache *s,
1047 struct page *page, void *object, unsigned long addr) { return 0; }
1049 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1051 static inline int check_object(struct kmem_cache *s, struct page *page,
1052 void *object, int active) { return 1; }
1053 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1054 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1055 unsigned long flags, const char *name,
1056 void (*ctor)(void *))
1060 #define slub_debug 0
1062 #define disable_higher_order_debug 0
1064 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1066 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1068 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1070 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1075 * Slab allocation and freeing
1077 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1078 struct kmem_cache_order_objects oo)
1080 int order = oo_order(oo);
1082 flags |= __GFP_NOTRACK;
1085 return alloc_pages(flags, order);
1087 return alloc_pages_node(node, flags, order);
1090 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1093 struct kmem_cache_order_objects oo = s->oo;
1096 flags |= s->allocflags;
1099 * Let the initial higher-order allocation fail under memory pressure
1100 * so we fall-back to the minimum order allocation.
1102 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1104 page = alloc_slab_page(alloc_gfp, node, oo);
1105 if (unlikely(!page)) {
1108 * Allocation may have failed due to fragmentation.
1109 * Try a lower order alloc if possible
1111 page = alloc_slab_page(flags, node, oo);
1115 stat(s, ORDER_FALLBACK);
1118 if (kmemcheck_enabled
1119 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1120 int pages = 1 << oo_order(oo);
1122 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1125 * Objects from caches that have a constructor don't get
1126 * cleared when they're allocated, so we need to do it here.
1129 kmemcheck_mark_uninitialized_pages(page, pages);
1131 kmemcheck_mark_unallocated_pages(page, pages);
1134 page->objects = oo_objects(oo);
1135 mod_zone_page_state(page_zone(page),
1136 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1137 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1143 static void setup_object(struct kmem_cache *s, struct page *page,
1146 setup_object_debug(s, page, object);
1147 if (unlikely(s->ctor))
1151 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1158 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1160 page = allocate_slab(s,
1161 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1165 inc_slabs_node(s, page_to_nid(page), page->objects);
1167 page->flags |= 1 << PG_slab;
1168 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1169 SLAB_STORE_USER | SLAB_TRACE))
1170 __SetPageSlubDebug(page);
1172 start = page_address(page);
1174 if (unlikely(s->flags & SLAB_POISON))
1175 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1178 for_each_object(p, s, start, page->objects) {
1179 setup_object(s, page, last);
1180 set_freepointer(s, last, p);
1183 setup_object(s, page, last);
1184 set_freepointer(s, last, NULL);
1186 page->freelist = start;
1192 static void __free_slab(struct kmem_cache *s, struct page *page)
1194 int order = compound_order(page);
1195 int pages = 1 << order;
1197 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1200 slab_pad_check(s, page);
1201 for_each_object(p, s, page_address(page),
1203 check_object(s, page, p, 0);
1204 __ClearPageSlubDebug(page);
1207 kmemcheck_free_shadow(page, compound_order(page));
1209 mod_zone_page_state(page_zone(page),
1210 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1211 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1214 __ClearPageSlab(page);
1215 reset_page_mapcount(page);
1216 if (current->reclaim_state)
1217 current->reclaim_state->reclaimed_slab += pages;
1218 __free_pages(page, order);
1221 static void rcu_free_slab(struct rcu_head *h)
1225 page = container_of((struct list_head *)h, struct page, lru);
1226 __free_slab(page->slab, page);
1229 static void free_slab(struct kmem_cache *s, struct page *page)
1231 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1233 * RCU free overloads the RCU head over the LRU
1235 struct rcu_head *head = (void *)&page->lru;
1237 call_rcu(head, rcu_free_slab);
1239 __free_slab(s, page);
1242 static void discard_slab(struct kmem_cache *s, struct page *page)
1244 dec_slabs_node(s, page_to_nid(page), page->objects);
1249 * Per slab locking using the pagelock
1251 static __always_inline void slab_lock(struct page *page)
1253 bit_spin_lock(PG_locked, &page->flags);
1256 static __always_inline void slab_unlock(struct page *page)
1258 __bit_spin_unlock(PG_locked, &page->flags);
1261 static __always_inline int slab_trylock(struct page *page)
1265 rc = bit_spin_trylock(PG_locked, &page->flags);
1270 * Management of partially allocated slabs
1272 static void add_partial(struct kmem_cache_node *n,
1273 struct page *page, int tail)
1275 spin_lock(&n->list_lock);
1278 list_add_tail(&page->lru, &n->partial);
1280 list_add(&page->lru, &n->partial);
1281 spin_unlock(&n->list_lock);
1284 static void remove_partial(struct kmem_cache *s, struct page *page)
1286 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1288 spin_lock(&n->list_lock);
1289 list_del(&page->lru);
1291 spin_unlock(&n->list_lock);
1295 * Lock slab and remove from the partial list.
1297 * Must hold list_lock.
1299 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1302 if (slab_trylock(page)) {
1303 list_del(&page->lru);
1305 __SetPageSlubFrozen(page);
1312 * Try to allocate a partial slab from a specific node.
1314 static struct page *get_partial_node(struct kmem_cache_node *n)
1319 * Racy check. If we mistakenly see no partial slabs then we
1320 * just allocate an empty slab. If we mistakenly try to get a
1321 * partial slab and there is none available then get_partials()
1324 if (!n || !n->nr_partial)
1327 spin_lock(&n->list_lock);
1328 list_for_each_entry(page, &n->partial, lru)
1329 if (lock_and_freeze_slab(n, page))
1333 spin_unlock(&n->list_lock);
1338 * Get a page from somewhere. Search in increasing NUMA distances.
1340 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1343 struct zonelist *zonelist;
1346 enum zone_type high_zoneidx = gfp_zone(flags);
1350 * The defrag ratio allows a configuration of the tradeoffs between
1351 * inter node defragmentation and node local allocations. A lower
1352 * defrag_ratio increases the tendency to do local allocations
1353 * instead of attempting to obtain partial slabs from other nodes.
1355 * If the defrag_ratio is set to 0 then kmalloc() always
1356 * returns node local objects. If the ratio is higher then kmalloc()
1357 * may return off node objects because partial slabs are obtained
1358 * from other nodes and filled up.
1360 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1361 * defrag_ratio = 1000) then every (well almost) allocation will
1362 * first attempt to defrag slab caches on other nodes. This means
1363 * scanning over all nodes to look for partial slabs which may be
1364 * expensive if we do it every time we are trying to find a slab
1365 * with available objects.
1367 if (!s->remote_node_defrag_ratio ||
1368 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1371 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1372 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1373 struct kmem_cache_node *n;
1375 n = get_node(s, zone_to_nid(zone));
1377 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1378 n->nr_partial > s->min_partial) {
1379 page = get_partial_node(n);
1389 * Get a partial page, lock it and return it.
1391 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1394 int searchnode = (node == -1) ? numa_node_id() : node;
1396 page = get_partial_node(get_node(s, searchnode));
1397 if (page || (flags & __GFP_THISNODE))
1400 return get_any_partial(s, flags);
1404 * Move a page back to the lists.
1406 * Must be called with the slab lock held.
1408 * On exit the slab lock will have been dropped.
1410 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1412 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1414 __ClearPageSlubFrozen(page);
1417 if (page->freelist) {
1418 add_partial(n, page, tail);
1419 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1421 stat(s, DEACTIVATE_FULL);
1422 if (SLABDEBUG && PageSlubDebug(page) &&
1423 (s->flags & SLAB_STORE_USER))
1428 stat(s, DEACTIVATE_EMPTY);
1429 if (n->nr_partial < s->min_partial) {
1431 * Adding an empty slab to the partial slabs in order
1432 * to avoid page allocator overhead. This slab needs
1433 * to come after the other slabs with objects in
1434 * so that the others get filled first. That way the
1435 * size of the partial list stays small.
1437 * kmem_cache_shrink can reclaim any empty slabs from
1440 add_partial(n, page, 1);
1445 discard_slab(s, page);
1451 * Remove the cpu slab
1453 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1455 struct page *page = c->page;
1459 stat(s, DEACTIVATE_REMOTE_FREES);
1461 * Merge cpu freelist into slab freelist. Typically we get here
1462 * because both freelists are empty. So this is unlikely
1465 while (unlikely(c->freelist)) {
1468 tail = 0; /* Hot objects. Put the slab first */
1470 /* Retrieve object from cpu_freelist */
1471 object = c->freelist;
1472 c->freelist = get_freepointer(s, c->freelist);
1474 /* And put onto the regular freelist */
1475 set_freepointer(s, object, page->freelist);
1476 page->freelist = object;
1480 unfreeze_slab(s, page, tail);
1483 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1485 stat(s, CPUSLAB_FLUSH);
1487 deactivate_slab(s, c);
1493 * Called from IPI handler with interrupts disabled.
1495 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1497 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1499 if (likely(c && c->page))
1503 static void flush_cpu_slab(void *d)
1505 struct kmem_cache *s = d;
1507 __flush_cpu_slab(s, smp_processor_id());
1510 static void flush_all(struct kmem_cache *s)
1512 on_each_cpu(flush_cpu_slab, s, 1);
1516 * Check if the objects in a per cpu structure fit numa
1517 * locality expectations.
1519 static inline int node_match(struct kmem_cache_cpu *c, int node)
1522 if (node != -1 && c->node != node)
1528 static int count_free(struct page *page)
1530 return page->objects - page->inuse;
1533 static unsigned long count_partial(struct kmem_cache_node *n,
1534 int (*get_count)(struct page *))
1536 unsigned long flags;
1537 unsigned long x = 0;
1540 spin_lock_irqsave(&n->list_lock, flags);
1541 list_for_each_entry(page, &n->partial, lru)
1542 x += get_count(page);
1543 spin_unlock_irqrestore(&n->list_lock, flags);
1547 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1549 #ifdef CONFIG_SLUB_DEBUG
1550 return atomic_long_read(&n->total_objects);
1556 static noinline void
1557 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1562 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1564 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1565 "default order: %d, min order: %d\n", s->name, s->objsize,
1566 s->size, oo_order(s->oo), oo_order(s->min));
1568 if (oo_order(s->min) > get_order(s->objsize))
1569 printk(KERN_WARNING " %s debugging increased min order, use "
1570 "slub_debug=O to disable.\n", s->name);
1572 for_each_online_node(node) {
1573 struct kmem_cache_node *n = get_node(s, node);
1574 unsigned long nr_slabs;
1575 unsigned long nr_objs;
1576 unsigned long nr_free;
1581 nr_free = count_partial(n, count_free);
1582 nr_slabs = node_nr_slabs(n);
1583 nr_objs = node_nr_objs(n);
1586 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1587 node, nr_slabs, nr_objs, nr_free);
1592 * Slow path. The lockless freelist is empty or we need to perform
1595 * Interrupts are disabled.
1597 * Processing is still very fast if new objects have been freed to the
1598 * regular freelist. In that case we simply take over the regular freelist
1599 * as the lockless freelist and zap the regular freelist.
1601 * If that is not working then we fall back to the partial lists. We take the
1602 * first element of the freelist as the object to allocate now and move the
1603 * rest of the freelist to the lockless freelist.
1605 * And if we were unable to get a new slab from the partial slab lists then
1606 * we need to allocate a new slab. This is the slowest path since it involves
1607 * a call to the page allocator and the setup of a new slab.
1609 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1610 unsigned long addr, struct kmem_cache_cpu *c)
1615 /* We handle __GFP_ZERO in the caller */
1616 gfpflags &= ~__GFP_ZERO;
1622 if (unlikely(!node_match(c, node)))
1625 stat(s, ALLOC_REFILL);
1628 object = c->page->freelist;
1629 if (unlikely(!object))
1631 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1634 c->freelist = get_freepointer(s, object);
1635 c->page->inuse = c->page->objects;
1636 c->page->freelist = NULL;
1637 c->node = page_to_nid(c->page);
1639 slab_unlock(c->page);
1640 stat(s, ALLOC_SLOWPATH);
1644 deactivate_slab(s, c);
1647 new = get_partial(s, gfpflags, node);
1650 stat(s, ALLOC_FROM_PARTIAL);
1654 if (gfpflags & __GFP_WAIT)
1657 new = new_slab(s, gfpflags, node);
1659 if (gfpflags & __GFP_WAIT)
1660 local_irq_disable();
1663 c = __this_cpu_ptr(s->cpu_slab);
1664 stat(s, ALLOC_SLAB);
1668 __SetPageSlubFrozen(new);
1672 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1673 slab_out_of_memory(s, gfpflags, node);
1676 if (!alloc_debug_processing(s, c->page, object, addr))
1680 c->page->freelist = get_freepointer(s, object);
1686 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1687 * have the fastpath folded into their functions. So no function call
1688 * overhead for requests that can be satisfied on the fastpath.
1690 * The fastpath works by first checking if the lockless freelist can be used.
1691 * If not then __slab_alloc is called for slow processing.
1693 * Otherwise we can simply pick the next object from the lockless free list.
1695 static __always_inline void *slab_alloc(struct kmem_cache *s,
1696 gfp_t gfpflags, int node, unsigned long addr)
1699 struct kmem_cache_cpu *c;
1700 unsigned long flags;
1702 gfpflags &= gfp_allowed_mask;
1704 lockdep_trace_alloc(gfpflags);
1705 might_sleep_if(gfpflags & __GFP_WAIT);
1707 if (should_failslab(s->objsize, gfpflags, s->flags))
1710 local_irq_save(flags);
1711 c = __this_cpu_ptr(s->cpu_slab);
1712 object = c->freelist;
1713 if (unlikely(!object || !node_match(c, node)))
1715 object = __slab_alloc(s, gfpflags, node, addr, c);
1718 c->freelist = get_freepointer(s, object);
1719 stat(s, ALLOC_FASTPATH);
1721 local_irq_restore(flags);
1723 if (unlikely(gfpflags & __GFP_ZERO) && object)
1724 memset(object, 0, s->objsize);
1726 kmemcheck_slab_alloc(s, gfpflags, object, s->objsize);
1727 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, gfpflags);
1732 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1734 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1736 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1740 EXPORT_SYMBOL(kmem_cache_alloc);
1742 #ifdef CONFIG_TRACING
1743 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1745 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1747 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1751 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1753 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1755 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1756 s->objsize, s->size, gfpflags, node);
1760 EXPORT_SYMBOL(kmem_cache_alloc_node);
1763 #ifdef CONFIG_TRACING
1764 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1768 return slab_alloc(s, gfpflags, node, _RET_IP_);
1770 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1774 * Slow patch handling. This may still be called frequently since objects
1775 * have a longer lifetime than the cpu slabs in most processing loads.
1777 * So we still attempt to reduce cache line usage. Just take the slab
1778 * lock and free the item. If there is no additional partial page
1779 * handling required then we can return immediately.
1781 static void __slab_free(struct kmem_cache *s, struct page *page,
1782 void *x, unsigned long addr)
1785 void **object = (void *)x;
1787 stat(s, FREE_SLOWPATH);
1790 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1794 prior = page->freelist;
1795 set_freepointer(s, object, prior);
1796 page->freelist = object;
1799 if (unlikely(PageSlubFrozen(page))) {
1800 stat(s, FREE_FROZEN);
1804 if (unlikely(!page->inuse))
1808 * Objects left in the slab. If it was not on the partial list before
1811 if (unlikely(!prior)) {
1812 add_partial(get_node(s, page_to_nid(page)), page, 1);
1813 stat(s, FREE_ADD_PARTIAL);
1823 * Slab still on the partial list.
1825 remove_partial(s, page);
1826 stat(s, FREE_REMOVE_PARTIAL);
1830 discard_slab(s, page);
1834 if (!free_debug_processing(s, page, x, addr))
1840 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1841 * can perform fastpath freeing without additional function calls.
1843 * The fastpath is only possible if we are freeing to the current cpu slab
1844 * of this processor. This typically the case if we have just allocated
1847 * If fastpath is not possible then fall back to __slab_free where we deal
1848 * with all sorts of special processing.
1850 static __always_inline void slab_free(struct kmem_cache *s,
1851 struct page *page, void *x, unsigned long addr)
1853 void **object = (void *)x;
1854 struct kmem_cache_cpu *c;
1855 unsigned long flags;
1857 kmemleak_free_recursive(x, s->flags);
1858 local_irq_save(flags);
1859 c = __this_cpu_ptr(s->cpu_slab);
1860 kmemcheck_slab_free(s, object, s->objsize);
1861 debug_check_no_locks_freed(object, s->objsize);
1862 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1863 debug_check_no_obj_freed(object, s->objsize);
1864 if (likely(page == c->page && c->node >= 0)) {
1865 set_freepointer(s, object, c->freelist);
1866 c->freelist = object;
1867 stat(s, FREE_FASTPATH);
1869 __slab_free(s, page, x, addr);
1871 local_irq_restore(flags);
1874 void kmem_cache_free(struct kmem_cache *s, void *x)
1878 page = virt_to_head_page(x);
1880 slab_free(s, page, x, _RET_IP_);
1882 trace_kmem_cache_free(_RET_IP_, x);
1884 EXPORT_SYMBOL(kmem_cache_free);
1886 /* Figure out on which slab page the object resides */
1887 static struct page *get_object_page(const void *x)
1889 struct page *page = virt_to_head_page(x);
1891 if (!PageSlab(page))
1898 * Object placement in a slab is made very easy because we always start at
1899 * offset 0. If we tune the size of the object to the alignment then we can
1900 * get the required alignment by putting one properly sized object after
1903 * Notice that the allocation order determines the sizes of the per cpu
1904 * caches. Each processor has always one slab available for allocations.
1905 * Increasing the allocation order reduces the number of times that slabs
1906 * must be moved on and off the partial lists and is therefore a factor in
1911 * Mininum / Maximum order of slab pages. This influences locking overhead
1912 * and slab fragmentation. A higher order reduces the number of partial slabs
1913 * and increases the number of allocations possible without having to
1914 * take the list_lock.
1916 static int slub_min_order;
1917 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1918 static int slub_min_objects;
1921 * Merge control. If this is set then no merging of slab caches will occur.
1922 * (Could be removed. This was introduced to pacify the merge skeptics.)
1924 static int slub_nomerge;
1927 * Calculate the order of allocation given an slab object size.
1929 * The order of allocation has significant impact on performance and other
1930 * system components. Generally order 0 allocations should be preferred since
1931 * order 0 does not cause fragmentation in the page allocator. Larger objects
1932 * be problematic to put into order 0 slabs because there may be too much
1933 * unused space left. We go to a higher order if more than 1/16th of the slab
1936 * In order to reach satisfactory performance we must ensure that a minimum
1937 * number of objects is in one slab. Otherwise we may generate too much
1938 * activity on the partial lists which requires taking the list_lock. This is
1939 * less a concern for large slabs though which are rarely used.
1941 * slub_max_order specifies the order where we begin to stop considering the
1942 * number of objects in a slab as critical. If we reach slub_max_order then
1943 * we try to keep the page order as low as possible. So we accept more waste
1944 * of space in favor of a small page order.
1946 * Higher order allocations also allow the placement of more objects in a
1947 * slab and thereby reduce object handling overhead. If the user has
1948 * requested a higher mininum order then we start with that one instead of
1949 * the smallest order which will fit the object.
1951 static inline int slab_order(int size, int min_objects,
1952 int max_order, int fract_leftover)
1956 int min_order = slub_min_order;
1958 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1959 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1961 for (order = max(min_order,
1962 fls(min_objects * size - 1) - PAGE_SHIFT);
1963 order <= max_order; order++) {
1965 unsigned long slab_size = PAGE_SIZE << order;
1967 if (slab_size < min_objects * size)
1970 rem = slab_size % size;
1972 if (rem <= slab_size / fract_leftover)
1980 static inline int calculate_order(int size)
1988 * Attempt to find best configuration for a slab. This
1989 * works by first attempting to generate a layout with
1990 * the best configuration and backing off gradually.
1992 * First we reduce the acceptable waste in a slab. Then
1993 * we reduce the minimum objects required in a slab.
1995 min_objects = slub_min_objects;
1997 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1998 max_objects = (PAGE_SIZE << slub_max_order)/size;
1999 min_objects = min(min_objects, max_objects);
2001 while (min_objects > 1) {
2003 while (fraction >= 4) {
2004 order = slab_order(size, min_objects,
2005 slub_max_order, fraction);
2006 if (order <= slub_max_order)
2014 * We were unable to place multiple objects in a slab. Now
2015 * lets see if we can place a single object there.
2017 order = slab_order(size, 1, slub_max_order, 1);
2018 if (order <= slub_max_order)
2022 * Doh this slab cannot be placed using slub_max_order.
2024 order = slab_order(size, 1, MAX_ORDER, 1);
2025 if (order < MAX_ORDER)
2031 * Figure out what the alignment of the objects will be.
2033 static unsigned long calculate_alignment(unsigned long flags,
2034 unsigned long align, unsigned long size)
2037 * If the user wants hardware cache aligned objects then follow that
2038 * suggestion if the object is sufficiently large.
2040 * The hardware cache alignment cannot override the specified
2041 * alignment though. If that is greater then use it.
2043 if (flags & SLAB_HWCACHE_ALIGN) {
2044 unsigned long ralign = cache_line_size();
2045 while (size <= ralign / 2)
2047 align = max(align, ralign);
2050 if (align < ARCH_SLAB_MINALIGN)
2051 align = ARCH_SLAB_MINALIGN;
2053 return ALIGN(align, sizeof(void *));
2057 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2060 spin_lock_init(&n->list_lock);
2061 INIT_LIST_HEAD(&n->partial);
2062 #ifdef CONFIG_SLUB_DEBUG
2063 atomic_long_set(&n->nr_slabs, 0);
2064 atomic_long_set(&n->total_objects, 0);
2065 INIT_LIST_HEAD(&n->full);
2069 static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[KMALLOC_CACHES]);
2071 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2073 if (s < kmalloc_caches + KMALLOC_CACHES && s >= kmalloc_caches)
2075 * Boot time creation of the kmalloc array. Use static per cpu data
2076 * since the per cpu allocator is not available yet.
2078 s->cpu_slab = kmalloc_percpu + (s - kmalloc_caches);
2080 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2090 * No kmalloc_node yet so do it by hand. We know that this is the first
2091 * slab on the node for this slabcache. There are no concurrent accesses
2094 * Note that this function only works on the kmalloc_node_cache
2095 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2096 * memory on a fresh node that has no slab structures yet.
2098 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2101 struct kmem_cache_node *n;
2102 unsigned long flags;
2104 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2106 page = new_slab(kmalloc_caches, gfpflags, node);
2109 if (page_to_nid(page) != node) {
2110 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2112 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2113 "in order to be able to continue\n");
2118 page->freelist = get_freepointer(kmalloc_caches, n);
2120 kmalloc_caches->node[node] = n;
2121 #ifdef CONFIG_SLUB_DEBUG
2122 init_object(kmalloc_caches, n, 1);
2123 init_tracking(kmalloc_caches, n);
2125 init_kmem_cache_node(n, kmalloc_caches);
2126 inc_slabs_node(kmalloc_caches, node, page->objects);
2129 * lockdep requires consistent irq usage for each lock
2130 * so even though there cannot be a race this early in
2131 * the boot sequence, we still disable irqs.
2133 local_irq_save(flags);
2134 add_partial(n, page, 0);
2135 local_irq_restore(flags);
2138 static void free_kmem_cache_nodes(struct kmem_cache *s)
2142 for_each_node_state(node, N_NORMAL_MEMORY) {
2143 struct kmem_cache_node *n = s->node[node];
2144 if (n && n != &s->local_node)
2145 kmem_cache_free(kmalloc_caches, n);
2146 s->node[node] = NULL;
2150 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2155 if (slab_state >= UP && (s < kmalloc_caches ||
2156 s >= kmalloc_caches + KMALLOC_CACHES))
2157 local_node = page_to_nid(virt_to_page(s));
2161 for_each_node_state(node, N_NORMAL_MEMORY) {
2162 struct kmem_cache_node *n;
2164 if (local_node == node)
2167 if (slab_state == DOWN) {
2168 early_kmem_cache_node_alloc(gfpflags, node);
2171 n = kmem_cache_alloc_node(kmalloc_caches,
2175 free_kmem_cache_nodes(s);
2181 init_kmem_cache_node(n, s);
2186 static void free_kmem_cache_nodes(struct kmem_cache *s)
2190 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2192 init_kmem_cache_node(&s->local_node, s);
2197 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2199 if (min < MIN_PARTIAL)
2201 else if (min > MAX_PARTIAL)
2203 s->min_partial = min;
2207 * calculate_sizes() determines the order and the distribution of data within
2210 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2212 unsigned long flags = s->flags;
2213 unsigned long size = s->objsize;
2214 unsigned long align = s->align;
2218 * Round up object size to the next word boundary. We can only
2219 * place the free pointer at word boundaries and this determines
2220 * the possible location of the free pointer.
2222 size = ALIGN(size, sizeof(void *));
2224 #ifdef CONFIG_SLUB_DEBUG
2226 * Determine if we can poison the object itself. If the user of
2227 * the slab may touch the object after free or before allocation
2228 * then we should never poison the object itself.
2230 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2232 s->flags |= __OBJECT_POISON;
2234 s->flags &= ~__OBJECT_POISON;
2238 * If we are Redzoning then check if there is some space between the
2239 * end of the object and the free pointer. If not then add an
2240 * additional word to have some bytes to store Redzone information.
2242 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2243 size += sizeof(void *);
2247 * With that we have determined the number of bytes in actual use
2248 * by the object. This is the potential offset to the free pointer.
2252 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2255 * Relocate free pointer after the object if it is not
2256 * permitted to overwrite the first word of the object on
2259 * This is the case if we do RCU, have a constructor or
2260 * destructor or are poisoning the objects.
2263 size += sizeof(void *);
2266 #ifdef CONFIG_SLUB_DEBUG
2267 if (flags & SLAB_STORE_USER)
2269 * Need to store information about allocs and frees after
2272 size += 2 * sizeof(struct track);
2274 if (flags & SLAB_RED_ZONE)
2276 * Add some empty padding so that we can catch
2277 * overwrites from earlier objects rather than let
2278 * tracking information or the free pointer be
2279 * corrupted if a user writes before the start
2282 size += sizeof(void *);
2286 * Determine the alignment based on various parameters that the
2287 * user specified and the dynamic determination of cache line size
2290 align = calculate_alignment(flags, align, s->objsize);
2294 * SLUB stores one object immediately after another beginning from
2295 * offset 0. In order to align the objects we have to simply size
2296 * each object to conform to the alignment.
2298 size = ALIGN(size, align);
2300 if (forced_order >= 0)
2301 order = forced_order;
2303 order = calculate_order(size);
2310 s->allocflags |= __GFP_COMP;
2312 if (s->flags & SLAB_CACHE_DMA)
2313 s->allocflags |= SLUB_DMA;
2315 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2316 s->allocflags |= __GFP_RECLAIMABLE;
2319 * Determine the number of objects per slab
2321 s->oo = oo_make(order, size);
2322 s->min = oo_make(get_order(size), size);
2323 if (oo_objects(s->oo) > oo_objects(s->max))
2326 return !!oo_objects(s->oo);
2330 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2331 const char *name, size_t size,
2332 size_t align, unsigned long flags,
2333 void (*ctor)(void *))
2335 memset(s, 0, kmem_size);
2340 s->flags = kmem_cache_flags(size, flags, name, ctor);
2342 if (!calculate_sizes(s, -1))
2344 if (disable_higher_order_debug) {
2346 * Disable debugging flags that store metadata if the min slab
2349 if (get_order(s->size) > get_order(s->objsize)) {
2350 s->flags &= ~DEBUG_METADATA_FLAGS;
2352 if (!calculate_sizes(s, -1))
2358 * The larger the object size is, the more pages we want on the partial
2359 * list to avoid pounding the page allocator excessively.
2361 set_min_partial(s, ilog2(s->size));
2364 s->remote_node_defrag_ratio = 1000;
2366 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2369 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2372 free_kmem_cache_nodes(s);
2374 if (flags & SLAB_PANIC)
2375 panic("Cannot create slab %s size=%lu realsize=%u "
2376 "order=%u offset=%u flags=%lx\n",
2377 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2383 * Check if a given pointer is valid
2385 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2389 if (!kern_ptr_validate(object, s->size))
2392 page = get_object_page(object);
2394 if (!page || s != page->slab)
2395 /* No slab or wrong slab */
2398 if (!check_valid_pointer(s, page, object))
2402 * We could also check if the object is on the slabs freelist.
2403 * But this would be too expensive and it seems that the main
2404 * purpose of kmem_ptr_valid() is to check if the object belongs
2405 * to a certain slab.
2409 EXPORT_SYMBOL(kmem_ptr_validate);
2412 * Determine the size of a slab object
2414 unsigned int kmem_cache_size(struct kmem_cache *s)
2418 EXPORT_SYMBOL(kmem_cache_size);
2420 const char *kmem_cache_name(struct kmem_cache *s)
2424 EXPORT_SYMBOL(kmem_cache_name);
2426 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2429 #ifdef CONFIG_SLUB_DEBUG
2430 void *addr = page_address(page);
2432 long *map = kzalloc(BITS_TO_LONGS(page->objects) * sizeof(long),
2437 slab_err(s, page, "%s", text);
2439 for_each_free_object(p, s, page->freelist)
2440 set_bit(slab_index(p, s, addr), map);
2442 for_each_object(p, s, addr, page->objects) {
2444 if (!test_bit(slab_index(p, s, addr), map)) {
2445 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2447 print_tracking(s, p);
2456 * Attempt to free all partial slabs on a node.
2458 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2460 unsigned long flags;
2461 struct page *page, *h;
2463 spin_lock_irqsave(&n->list_lock, flags);
2464 list_for_each_entry_safe(page, h, &n->partial, lru) {
2466 list_del(&page->lru);
2467 discard_slab(s, page);
2470 list_slab_objects(s, page,
2471 "Objects remaining on kmem_cache_close()");
2474 spin_unlock_irqrestore(&n->list_lock, flags);
2478 * Release all resources used by a slab cache.
2480 static inline int kmem_cache_close(struct kmem_cache *s)
2485 free_percpu(s->cpu_slab);
2486 /* Attempt to free all objects */
2487 for_each_node_state(node, N_NORMAL_MEMORY) {
2488 struct kmem_cache_node *n = get_node(s, node);
2491 if (n->nr_partial || slabs_node(s, node))
2494 free_kmem_cache_nodes(s);
2499 * Close a cache and release the kmem_cache structure
2500 * (must be used for caches created using kmem_cache_create)
2502 void kmem_cache_destroy(struct kmem_cache *s)
2504 down_write(&slub_lock);
2508 up_write(&slub_lock);
2509 if (kmem_cache_close(s)) {
2510 printk(KERN_ERR "SLUB %s: %s called for cache that "
2511 "still has objects.\n", s->name, __func__);
2514 if (s->flags & SLAB_DESTROY_BY_RCU)
2516 sysfs_slab_remove(s);
2518 up_write(&slub_lock);
2520 EXPORT_SYMBOL(kmem_cache_destroy);
2522 /********************************************************************
2524 *******************************************************************/
2526 struct kmem_cache kmalloc_caches[KMALLOC_CACHES] __cacheline_aligned;
2527 EXPORT_SYMBOL(kmalloc_caches);
2529 static int __init setup_slub_min_order(char *str)
2531 get_option(&str, &slub_min_order);
2536 __setup("slub_min_order=", setup_slub_min_order);
2538 static int __init setup_slub_max_order(char *str)
2540 get_option(&str, &slub_max_order);
2541 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2546 __setup("slub_max_order=", setup_slub_max_order);
2548 static int __init setup_slub_min_objects(char *str)
2550 get_option(&str, &slub_min_objects);
2555 __setup("slub_min_objects=", setup_slub_min_objects);
2557 static int __init setup_slub_nomerge(char *str)
2563 __setup("slub_nomerge", setup_slub_nomerge);
2565 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2566 const char *name, int size, gfp_t gfp_flags)
2568 unsigned int flags = 0;
2570 if (gfp_flags & SLUB_DMA)
2571 flags = SLAB_CACHE_DMA;
2574 * This function is called with IRQs disabled during early-boot on
2575 * single CPU so there's no need to take slub_lock here.
2577 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2581 list_add(&s->list, &slab_caches);
2583 if (sysfs_slab_add(s))
2588 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2591 #ifdef CONFIG_ZONE_DMA
2592 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2594 static void sysfs_add_func(struct work_struct *w)
2596 struct kmem_cache *s;
2598 down_write(&slub_lock);
2599 list_for_each_entry(s, &slab_caches, list) {
2600 if (s->flags & __SYSFS_ADD_DEFERRED) {
2601 s->flags &= ~__SYSFS_ADD_DEFERRED;
2605 up_write(&slub_lock);
2608 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2610 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2612 struct kmem_cache *s;
2615 unsigned long slabflags;
2618 s = kmalloc_caches_dma[index];
2622 /* Dynamically create dma cache */
2623 if (flags & __GFP_WAIT)
2624 down_write(&slub_lock);
2626 if (!down_write_trylock(&slub_lock))
2630 if (kmalloc_caches_dma[index])
2633 realsize = kmalloc_caches[index].objsize;
2634 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2635 (unsigned int)realsize);
2638 for (i = 0; i < KMALLOC_CACHES; i++)
2639 if (!kmalloc_caches[i].size)
2642 BUG_ON(i >= KMALLOC_CACHES);
2643 s = kmalloc_caches + i;
2646 * Must defer sysfs creation to a workqueue because we don't know
2647 * what context we are called from. Before sysfs comes up, we don't
2648 * need to do anything because our sysfs initcall will start by
2649 * adding all existing slabs to sysfs.
2651 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
2652 if (slab_state >= SYSFS)
2653 slabflags |= __SYSFS_ADD_DEFERRED;
2655 if (!text || !kmem_cache_open(s, flags, text,
2656 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
2662 list_add(&s->list, &slab_caches);
2663 kmalloc_caches_dma[index] = s;
2665 if (slab_state >= SYSFS)
2666 schedule_work(&sysfs_add_work);
2669 up_write(&slub_lock);
2671 return kmalloc_caches_dma[index];
2676 * Conversion table for small slabs sizes / 8 to the index in the
2677 * kmalloc array. This is necessary for slabs < 192 since we have non power
2678 * of two cache sizes there. The size of larger slabs can be determined using
2681 static s8 size_index[24] = {
2708 static inline int size_index_elem(size_t bytes)
2710 return (bytes - 1) / 8;
2713 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2719 return ZERO_SIZE_PTR;
2721 index = size_index[size_index_elem(size)];
2723 index = fls(size - 1);
2725 #ifdef CONFIG_ZONE_DMA
2726 if (unlikely((flags & SLUB_DMA)))
2727 return dma_kmalloc_cache(index, flags);
2730 return &kmalloc_caches[index];
2733 void *__kmalloc(size_t size, gfp_t flags)
2735 struct kmem_cache *s;
2738 if (unlikely(size > SLUB_MAX_SIZE))
2739 return kmalloc_large(size, flags);
2741 s = get_slab(size, flags);
2743 if (unlikely(ZERO_OR_NULL_PTR(s)))
2746 ret = slab_alloc(s, flags, -1, _RET_IP_);
2748 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2752 EXPORT_SYMBOL(__kmalloc);
2754 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2759 flags |= __GFP_COMP | __GFP_NOTRACK;
2760 page = alloc_pages_node(node, flags, get_order(size));
2762 ptr = page_address(page);
2764 kmemleak_alloc(ptr, size, 1, flags);
2769 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2771 struct kmem_cache *s;
2774 if (unlikely(size > SLUB_MAX_SIZE)) {
2775 ret = kmalloc_large_node(size, flags, node);
2777 trace_kmalloc_node(_RET_IP_, ret,
2778 size, PAGE_SIZE << get_order(size),
2784 s = get_slab(size, flags);
2786 if (unlikely(ZERO_OR_NULL_PTR(s)))
2789 ret = slab_alloc(s, flags, node, _RET_IP_);
2791 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2795 EXPORT_SYMBOL(__kmalloc_node);
2798 size_t ksize(const void *object)
2801 struct kmem_cache *s;
2803 if (unlikely(object == ZERO_SIZE_PTR))
2806 page = virt_to_head_page(object);
2808 if (unlikely(!PageSlab(page))) {
2809 WARN_ON(!PageCompound(page));
2810 return PAGE_SIZE << compound_order(page);
2814 #ifdef CONFIG_SLUB_DEBUG
2816 * Debugging requires use of the padding between object
2817 * and whatever may come after it.
2819 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2824 * If we have the need to store the freelist pointer
2825 * back there or track user information then we can
2826 * only use the space before that information.
2828 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2831 * Else we can use all the padding etc for the allocation
2835 EXPORT_SYMBOL(ksize);
2837 void kfree(const void *x)
2840 void *object = (void *)x;
2842 trace_kfree(_RET_IP_, x);
2844 if (unlikely(ZERO_OR_NULL_PTR(x)))
2847 page = virt_to_head_page(x);
2848 if (unlikely(!PageSlab(page))) {
2849 BUG_ON(!PageCompound(page));
2854 slab_free(page->slab, page, object, _RET_IP_);
2856 EXPORT_SYMBOL(kfree);
2859 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2860 * the remaining slabs by the number of items in use. The slabs with the
2861 * most items in use come first. New allocations will then fill those up
2862 * and thus they can be removed from the partial lists.
2864 * The slabs with the least items are placed last. This results in them
2865 * being allocated from last increasing the chance that the last objects
2866 * are freed in them.
2868 int kmem_cache_shrink(struct kmem_cache *s)
2872 struct kmem_cache_node *n;
2875 int objects = oo_objects(s->max);
2876 struct list_head *slabs_by_inuse =
2877 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2878 unsigned long flags;
2880 if (!slabs_by_inuse)
2884 for_each_node_state(node, N_NORMAL_MEMORY) {
2885 n = get_node(s, node);
2890 for (i = 0; i < objects; i++)
2891 INIT_LIST_HEAD(slabs_by_inuse + i);
2893 spin_lock_irqsave(&n->list_lock, flags);
2896 * Build lists indexed by the items in use in each slab.
2898 * Note that concurrent frees may occur while we hold the
2899 * list_lock. page->inuse here is the upper limit.
2901 list_for_each_entry_safe(page, t, &n->partial, lru) {
2902 if (!page->inuse && slab_trylock(page)) {
2904 * Must hold slab lock here because slab_free
2905 * may have freed the last object and be
2906 * waiting to release the slab.
2908 list_del(&page->lru);
2911 discard_slab(s, page);
2913 list_move(&page->lru,
2914 slabs_by_inuse + page->inuse);
2919 * Rebuild the partial list with the slabs filled up most
2920 * first and the least used slabs at the end.
2922 for (i = objects - 1; i >= 0; i--)
2923 list_splice(slabs_by_inuse + i, n->partial.prev);
2925 spin_unlock_irqrestore(&n->list_lock, flags);
2928 kfree(slabs_by_inuse);
2931 EXPORT_SYMBOL(kmem_cache_shrink);
2933 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2934 static int slab_mem_going_offline_callback(void *arg)
2936 struct kmem_cache *s;
2938 down_read(&slub_lock);
2939 list_for_each_entry(s, &slab_caches, list)
2940 kmem_cache_shrink(s);
2941 up_read(&slub_lock);
2946 static void slab_mem_offline_callback(void *arg)
2948 struct kmem_cache_node *n;
2949 struct kmem_cache *s;
2950 struct memory_notify *marg = arg;
2953 offline_node = marg->status_change_nid;
2956 * If the node still has available memory. we need kmem_cache_node
2959 if (offline_node < 0)
2962 down_read(&slub_lock);
2963 list_for_each_entry(s, &slab_caches, list) {
2964 n = get_node(s, offline_node);
2967 * if n->nr_slabs > 0, slabs still exist on the node
2968 * that is going down. We were unable to free them,
2969 * and offline_pages() function shouldn't call this
2970 * callback. So, we must fail.
2972 BUG_ON(slabs_node(s, offline_node));
2974 s->node[offline_node] = NULL;
2975 kmem_cache_free(kmalloc_caches, n);
2978 up_read(&slub_lock);
2981 static int slab_mem_going_online_callback(void *arg)
2983 struct kmem_cache_node *n;
2984 struct kmem_cache *s;
2985 struct memory_notify *marg = arg;
2986 int nid = marg->status_change_nid;
2990 * If the node's memory is already available, then kmem_cache_node is
2991 * already created. Nothing to do.
2997 * We are bringing a node online. No memory is available yet. We must
2998 * allocate a kmem_cache_node structure in order to bring the node
3001 down_read(&slub_lock);
3002 list_for_each_entry(s, &slab_caches, list) {
3004 * XXX: kmem_cache_alloc_node will fallback to other nodes
3005 * since memory is not yet available from the node that
3008 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3013 init_kmem_cache_node(n, s);
3017 up_read(&slub_lock);
3021 static int slab_memory_callback(struct notifier_block *self,
3022 unsigned long action, void *arg)
3027 case MEM_GOING_ONLINE:
3028 ret = slab_mem_going_online_callback(arg);
3030 case MEM_GOING_OFFLINE:
3031 ret = slab_mem_going_offline_callback(arg);
3034 case MEM_CANCEL_ONLINE:
3035 slab_mem_offline_callback(arg);
3038 case MEM_CANCEL_OFFLINE:
3042 ret = notifier_from_errno(ret);
3048 #endif /* CONFIG_MEMORY_HOTPLUG */
3050 /********************************************************************
3051 * Basic setup of slabs
3052 *******************************************************************/
3054 void __init kmem_cache_init(void)
3061 * Must first have the slab cache available for the allocations of the
3062 * struct kmem_cache_node's. There is special bootstrap code in
3063 * kmem_cache_open for slab_state == DOWN.
3065 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3066 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3067 kmalloc_caches[0].refcount = -1;
3070 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3073 /* Able to allocate the per node structures */
3074 slab_state = PARTIAL;
3076 /* Caches that are not of the two-to-the-power-of size */
3077 if (KMALLOC_MIN_SIZE <= 32) {
3078 create_kmalloc_cache(&kmalloc_caches[1],
3079 "kmalloc-96", 96, GFP_NOWAIT);
3082 if (KMALLOC_MIN_SIZE <= 64) {
3083 create_kmalloc_cache(&kmalloc_caches[2],
3084 "kmalloc-192", 192, GFP_NOWAIT);
3088 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3089 create_kmalloc_cache(&kmalloc_caches[i],
3090 "kmalloc", 1 << i, GFP_NOWAIT);
3096 * Patch up the size_index table if we have strange large alignment
3097 * requirements for the kmalloc array. This is only the case for
3098 * MIPS it seems. The standard arches will not generate any code here.
3100 * Largest permitted alignment is 256 bytes due to the way we
3101 * handle the index determination for the smaller caches.
3103 * Make sure that nothing crazy happens if someone starts tinkering
3104 * around with ARCH_KMALLOC_MINALIGN
3106 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3107 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3109 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3110 int elem = size_index_elem(i);
3111 if (elem >= ARRAY_SIZE(size_index))
3113 size_index[elem] = KMALLOC_SHIFT_LOW;
3116 if (KMALLOC_MIN_SIZE == 64) {
3118 * The 96 byte size cache is not used if the alignment
3121 for (i = 64 + 8; i <= 96; i += 8)
3122 size_index[size_index_elem(i)] = 7;
3123 } else if (KMALLOC_MIN_SIZE == 128) {
3125 * The 192 byte sized cache is not used if the alignment
3126 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3129 for (i = 128 + 8; i <= 192; i += 8)
3130 size_index[size_index_elem(i)] = 8;
3135 /* Provide the correct kmalloc names now that the caches are up */
3136 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3137 kmalloc_caches[i]. name =
3138 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3141 register_cpu_notifier(&slab_notifier);
3144 kmem_size = offsetof(struct kmem_cache, node) +
3145 nr_node_ids * sizeof(struct kmem_cache_node *);
3147 kmem_size = sizeof(struct kmem_cache);
3151 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3152 " CPUs=%d, Nodes=%d\n",
3153 caches, cache_line_size(),
3154 slub_min_order, slub_max_order, slub_min_objects,
3155 nr_cpu_ids, nr_node_ids);
3158 void __init kmem_cache_init_late(void)
3163 * Find a mergeable slab cache
3165 static int slab_unmergeable(struct kmem_cache *s)
3167 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3174 * We may have set a slab to be unmergeable during bootstrap.
3176 if (s->refcount < 0)
3182 static struct kmem_cache *find_mergeable(size_t size,
3183 size_t align, unsigned long flags, const char *name,
3184 void (*ctor)(void *))
3186 struct kmem_cache *s;
3188 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3194 size = ALIGN(size, sizeof(void *));
3195 align = calculate_alignment(flags, align, size);
3196 size = ALIGN(size, align);
3197 flags = kmem_cache_flags(size, flags, name, NULL);
3199 list_for_each_entry(s, &slab_caches, list) {
3200 if (slab_unmergeable(s))
3206 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3209 * Check if alignment is compatible.
3210 * Courtesy of Adrian Drzewiecki
3212 if ((s->size & ~(align - 1)) != s->size)
3215 if (s->size - size >= sizeof(void *))
3223 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3224 size_t align, unsigned long flags, void (*ctor)(void *))
3226 struct kmem_cache *s;
3231 down_write(&slub_lock);
3232 s = find_mergeable(size, align, flags, name, ctor);
3236 * Adjust the object sizes so that we clear
3237 * the complete object on kzalloc.
3239 s->objsize = max(s->objsize, (int)size);
3240 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3241 up_write(&slub_lock);
3243 if (sysfs_slab_alias(s, name)) {
3244 down_write(&slub_lock);
3246 up_write(&slub_lock);
3252 s = kmalloc(kmem_size, GFP_KERNEL);
3254 if (kmem_cache_open(s, GFP_KERNEL, name,
3255 size, align, flags, ctor)) {
3256 list_add(&s->list, &slab_caches);
3257 up_write(&slub_lock);
3258 if (sysfs_slab_add(s)) {
3259 down_write(&slub_lock);
3261 up_write(&slub_lock);
3269 up_write(&slub_lock);
3272 if (flags & SLAB_PANIC)
3273 panic("Cannot create slabcache %s\n", name);
3278 EXPORT_SYMBOL(kmem_cache_create);
3282 * Use the cpu notifier to insure that the cpu slabs are flushed when
3285 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3286 unsigned long action, void *hcpu)
3288 long cpu = (long)hcpu;
3289 struct kmem_cache *s;
3290 unsigned long flags;
3293 case CPU_UP_CANCELED:
3294 case CPU_UP_CANCELED_FROZEN:
3296 case CPU_DEAD_FROZEN:
3297 down_read(&slub_lock);
3298 list_for_each_entry(s, &slab_caches, list) {
3299 local_irq_save(flags);
3300 __flush_cpu_slab(s, cpu);
3301 local_irq_restore(flags);
3303 up_read(&slub_lock);
3311 static struct notifier_block __cpuinitdata slab_notifier = {
3312 .notifier_call = slab_cpuup_callback
3317 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3319 struct kmem_cache *s;
3322 if (unlikely(size > SLUB_MAX_SIZE))
3323 return kmalloc_large(size, gfpflags);
3325 s = get_slab(size, gfpflags);
3327 if (unlikely(ZERO_OR_NULL_PTR(s)))
3330 ret = slab_alloc(s, gfpflags, -1, caller);
3332 /* Honor the call site pointer we recieved. */
3333 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3338 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3339 int node, unsigned long caller)
3341 struct kmem_cache *s;
3344 if (unlikely(size > SLUB_MAX_SIZE)) {
3345 ret = kmalloc_large_node(size, gfpflags, node);
3347 trace_kmalloc_node(caller, ret,
3348 size, PAGE_SIZE << get_order(size),
3354 s = get_slab(size, gfpflags);
3356 if (unlikely(ZERO_OR_NULL_PTR(s)))
3359 ret = slab_alloc(s, gfpflags, node, caller);
3361 /* Honor the call site pointer we recieved. */
3362 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3367 #ifdef CONFIG_SLUB_DEBUG
3368 static int count_inuse(struct page *page)
3373 static int count_total(struct page *page)
3375 return page->objects;
3378 static int validate_slab(struct kmem_cache *s, struct page *page,
3382 void *addr = page_address(page);
3384 if (!check_slab(s, page) ||
3385 !on_freelist(s, page, NULL))
3388 /* Now we know that a valid freelist exists */
3389 bitmap_zero(map, page->objects);
3391 for_each_free_object(p, s, page->freelist) {
3392 set_bit(slab_index(p, s, addr), map);
3393 if (!check_object(s, page, p, 0))
3397 for_each_object(p, s, addr, page->objects)
3398 if (!test_bit(slab_index(p, s, addr), map))
3399 if (!check_object(s, page, p, 1))
3404 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3407 if (slab_trylock(page)) {
3408 validate_slab(s, page, map);
3411 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3414 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3415 if (!PageSlubDebug(page))
3416 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3417 "on slab 0x%p\n", s->name, page);
3419 if (PageSlubDebug(page))
3420 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3421 "slab 0x%p\n", s->name, page);
3425 static int validate_slab_node(struct kmem_cache *s,
3426 struct kmem_cache_node *n, unsigned long *map)
3428 unsigned long count = 0;
3430 unsigned long flags;
3432 spin_lock_irqsave(&n->list_lock, flags);
3434 list_for_each_entry(page, &n->partial, lru) {
3435 validate_slab_slab(s, page, map);
3438 if (count != n->nr_partial)
3439 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3440 "counter=%ld\n", s->name, count, n->nr_partial);
3442 if (!(s->flags & SLAB_STORE_USER))
3445 list_for_each_entry(page, &n->full, lru) {
3446 validate_slab_slab(s, page, map);
3449 if (count != atomic_long_read(&n->nr_slabs))
3450 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3451 "counter=%ld\n", s->name, count,
3452 atomic_long_read(&n->nr_slabs));
3455 spin_unlock_irqrestore(&n->list_lock, flags);
3459 static long validate_slab_cache(struct kmem_cache *s)
3462 unsigned long count = 0;
3463 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3464 sizeof(unsigned long), GFP_KERNEL);
3470 for_each_node_state(node, N_NORMAL_MEMORY) {
3471 struct kmem_cache_node *n = get_node(s, node);
3473 count += validate_slab_node(s, n, map);
3479 #ifdef SLUB_RESILIENCY_TEST
3480 static void resiliency_test(void)
3484 printk(KERN_ERR "SLUB resiliency testing\n");
3485 printk(KERN_ERR "-----------------------\n");
3486 printk(KERN_ERR "A. Corruption after allocation\n");
3488 p = kzalloc(16, GFP_KERNEL);
3490 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3491 " 0x12->0x%p\n\n", p + 16);
3493 validate_slab_cache(kmalloc_caches + 4);
3495 /* Hmmm... The next two are dangerous */
3496 p = kzalloc(32, GFP_KERNEL);
3497 p[32 + sizeof(void *)] = 0x34;
3498 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3499 " 0x34 -> -0x%p\n", p);
3501 "If allocated object is overwritten then not detectable\n\n");
3503 validate_slab_cache(kmalloc_caches + 5);
3504 p = kzalloc(64, GFP_KERNEL);
3505 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3507 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3510 "If allocated object is overwritten then not detectable\n\n");
3511 validate_slab_cache(kmalloc_caches + 6);
3513 printk(KERN_ERR "\nB. Corruption after free\n");
3514 p = kzalloc(128, GFP_KERNEL);
3517 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3518 validate_slab_cache(kmalloc_caches + 7);
3520 p = kzalloc(256, GFP_KERNEL);
3523 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3525 validate_slab_cache(kmalloc_caches + 8);
3527 p = kzalloc(512, GFP_KERNEL);
3530 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3531 validate_slab_cache(kmalloc_caches + 9);
3534 static void resiliency_test(void) {};
3538 * Generate lists of code addresses where slabcache objects are allocated
3543 unsigned long count;
3550 DECLARE_BITMAP(cpus, NR_CPUS);
3556 unsigned long count;
3557 struct location *loc;
3560 static void free_loc_track(struct loc_track *t)
3563 free_pages((unsigned long)t->loc,
3564 get_order(sizeof(struct location) * t->max));
3567 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3572 order = get_order(sizeof(struct location) * max);
3574 l = (void *)__get_free_pages(flags, order);
3579 memcpy(l, t->loc, sizeof(struct location) * t->count);
3587 static int add_location(struct loc_track *t, struct kmem_cache *s,
3588 const struct track *track)
3590 long start, end, pos;
3592 unsigned long caddr;
3593 unsigned long age = jiffies - track->when;
3599 pos = start + (end - start + 1) / 2;
3602 * There is nothing at "end". If we end up there
3603 * we need to add something to before end.
3608 caddr = t->loc[pos].addr;
3609 if (track->addr == caddr) {
3615 if (age < l->min_time)
3617 if (age > l->max_time)
3620 if (track->pid < l->min_pid)
3621 l->min_pid = track->pid;
3622 if (track->pid > l->max_pid)
3623 l->max_pid = track->pid;
3625 cpumask_set_cpu(track->cpu,
3626 to_cpumask(l->cpus));
3628 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3632 if (track->addr < caddr)
3639 * Not found. Insert new tracking element.
3641 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3647 (t->count - pos) * sizeof(struct location));
3650 l->addr = track->addr;
3654 l->min_pid = track->pid;
3655 l->max_pid = track->pid;
3656 cpumask_clear(to_cpumask(l->cpus));
3657 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3658 nodes_clear(l->nodes);
3659 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3663 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3664 struct page *page, enum track_item alloc,
3667 void *addr = page_address(page);
3670 bitmap_zero(map, page->objects);
3671 for_each_free_object(p, s, page->freelist)
3672 set_bit(slab_index(p, s, addr), map);
3674 for_each_object(p, s, addr, page->objects)
3675 if (!test_bit(slab_index(p, s, addr), map))
3676 add_location(t, s, get_track(s, p, alloc));
3679 static int list_locations(struct kmem_cache *s, char *buf,
3680 enum track_item alloc)
3684 struct loc_track t = { 0, 0, NULL };
3686 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3687 sizeof(unsigned long), GFP_KERNEL);
3689 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3692 return sprintf(buf, "Out of memory\n");
3694 /* Push back cpu slabs */
3697 for_each_node_state(node, N_NORMAL_MEMORY) {
3698 struct kmem_cache_node *n = get_node(s, node);
3699 unsigned long flags;
3702 if (!atomic_long_read(&n->nr_slabs))
3705 spin_lock_irqsave(&n->list_lock, flags);
3706 list_for_each_entry(page, &n->partial, lru)
3707 process_slab(&t, s, page, alloc, map);
3708 list_for_each_entry(page, &n->full, lru)
3709 process_slab(&t, s, page, alloc, map);
3710 spin_unlock_irqrestore(&n->list_lock, flags);
3713 for (i = 0; i < t.count; i++) {
3714 struct location *l = &t.loc[i];
3716 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3718 len += sprintf(buf + len, "%7ld ", l->count);
3721 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3723 len += sprintf(buf + len, "<not-available>");
3725 if (l->sum_time != l->min_time) {
3726 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3728 (long)div_u64(l->sum_time, l->count),
3731 len += sprintf(buf + len, " age=%ld",
3734 if (l->min_pid != l->max_pid)
3735 len += sprintf(buf + len, " pid=%ld-%ld",
3736 l->min_pid, l->max_pid);
3738 len += sprintf(buf + len, " pid=%ld",
3741 if (num_online_cpus() > 1 &&
3742 !cpumask_empty(to_cpumask(l->cpus)) &&
3743 len < PAGE_SIZE - 60) {
3744 len += sprintf(buf + len, " cpus=");
3745 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3746 to_cpumask(l->cpus));
3749 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3750 len < PAGE_SIZE - 60) {
3751 len += sprintf(buf + len, " nodes=");
3752 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3756 len += sprintf(buf + len, "\n");
3762 len += sprintf(buf, "No data\n");
3766 enum slab_stat_type {
3767 SL_ALL, /* All slabs */
3768 SL_PARTIAL, /* Only partially allocated slabs */
3769 SL_CPU, /* Only slabs used for cpu caches */
3770 SL_OBJECTS, /* Determine allocated objects not slabs */
3771 SL_TOTAL /* Determine object capacity not slabs */
3774 #define SO_ALL (1 << SL_ALL)
3775 #define SO_PARTIAL (1 << SL_PARTIAL)
3776 #define SO_CPU (1 << SL_CPU)
3777 #define SO_OBJECTS (1 << SL_OBJECTS)
3778 #define SO_TOTAL (1 << SL_TOTAL)
3780 static ssize_t show_slab_objects(struct kmem_cache *s,
3781 char *buf, unsigned long flags)
3783 unsigned long total = 0;
3786 unsigned long *nodes;
3787 unsigned long *per_cpu;
3789 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3792 per_cpu = nodes + nr_node_ids;
3794 if (flags & SO_CPU) {
3797 for_each_possible_cpu(cpu) {
3798 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3800 if (!c || c->node < 0)
3804 if (flags & SO_TOTAL)
3805 x = c->page->objects;
3806 else if (flags & SO_OBJECTS)
3812 nodes[c->node] += x;
3818 if (flags & SO_ALL) {
3819 for_each_node_state(node, N_NORMAL_MEMORY) {
3820 struct kmem_cache_node *n = get_node(s, node);
3822 if (flags & SO_TOTAL)
3823 x = atomic_long_read(&n->total_objects);
3824 else if (flags & SO_OBJECTS)
3825 x = atomic_long_read(&n->total_objects) -
3826 count_partial(n, count_free);
3829 x = atomic_long_read(&n->nr_slabs);
3834 } else if (flags & SO_PARTIAL) {
3835 for_each_node_state(node, N_NORMAL_MEMORY) {
3836 struct kmem_cache_node *n = get_node(s, node);
3838 if (flags & SO_TOTAL)
3839 x = count_partial(n, count_total);
3840 else if (flags & SO_OBJECTS)
3841 x = count_partial(n, count_inuse);
3848 x = sprintf(buf, "%lu", total);
3850 for_each_node_state(node, N_NORMAL_MEMORY)
3852 x += sprintf(buf + x, " N%d=%lu",
3856 return x + sprintf(buf + x, "\n");
3859 static int any_slab_objects(struct kmem_cache *s)
3863 for_each_online_node(node) {
3864 struct kmem_cache_node *n = get_node(s, node);
3869 if (atomic_long_read(&n->total_objects))
3875 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3876 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3878 struct slab_attribute {
3879 struct attribute attr;
3880 ssize_t (*show)(struct kmem_cache *s, char *buf);
3881 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3884 #define SLAB_ATTR_RO(_name) \
3885 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3887 #define SLAB_ATTR(_name) \
3888 static struct slab_attribute _name##_attr = \
3889 __ATTR(_name, 0644, _name##_show, _name##_store)
3891 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3893 return sprintf(buf, "%d\n", s->size);
3895 SLAB_ATTR_RO(slab_size);
3897 static ssize_t align_show(struct kmem_cache *s, char *buf)
3899 return sprintf(buf, "%d\n", s->align);
3901 SLAB_ATTR_RO(align);
3903 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3905 return sprintf(buf, "%d\n", s->objsize);
3907 SLAB_ATTR_RO(object_size);
3909 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3911 return sprintf(buf, "%d\n", oo_objects(s->oo));
3913 SLAB_ATTR_RO(objs_per_slab);
3915 static ssize_t order_store(struct kmem_cache *s,
3916 const char *buf, size_t length)
3918 unsigned long order;
3921 err = strict_strtoul(buf, 10, &order);
3925 if (order > slub_max_order || order < slub_min_order)
3928 calculate_sizes(s, order);
3932 static ssize_t order_show(struct kmem_cache *s, char *buf)
3934 return sprintf(buf, "%d\n", oo_order(s->oo));
3938 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3940 return sprintf(buf, "%lu\n", s->min_partial);
3943 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3949 err = strict_strtoul(buf, 10, &min);
3953 set_min_partial(s, min);
3956 SLAB_ATTR(min_partial);
3958 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3961 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3963 return n + sprintf(buf + n, "\n");
3969 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3971 return sprintf(buf, "%d\n", s->refcount - 1);
3973 SLAB_ATTR_RO(aliases);
3975 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3977 return show_slab_objects(s, buf, SO_ALL);
3979 SLAB_ATTR_RO(slabs);
3981 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3983 return show_slab_objects(s, buf, SO_PARTIAL);
3985 SLAB_ATTR_RO(partial);
3987 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3989 return show_slab_objects(s, buf, SO_CPU);
3991 SLAB_ATTR_RO(cpu_slabs);
3993 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3995 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3997 SLAB_ATTR_RO(objects);
3999 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4001 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4003 SLAB_ATTR_RO(objects_partial);
4005 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4007 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4009 SLAB_ATTR_RO(total_objects);
4011 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4013 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4016 static ssize_t sanity_checks_store(struct kmem_cache *s,
4017 const char *buf, size_t length)
4019 s->flags &= ~SLAB_DEBUG_FREE;
4021 s->flags |= SLAB_DEBUG_FREE;
4024 SLAB_ATTR(sanity_checks);
4026 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4028 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4031 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4034 s->flags &= ~SLAB_TRACE;
4036 s->flags |= SLAB_TRACE;
4041 #ifdef CONFIG_FAILSLAB
4042 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4044 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4047 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4050 s->flags &= ~SLAB_FAILSLAB;
4052 s->flags |= SLAB_FAILSLAB;
4055 SLAB_ATTR(failslab);
4058 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4060 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4063 static ssize_t reclaim_account_store(struct kmem_cache *s,
4064 const char *buf, size_t length)
4066 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4068 s->flags |= SLAB_RECLAIM_ACCOUNT;
4071 SLAB_ATTR(reclaim_account);
4073 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4075 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4077 SLAB_ATTR_RO(hwcache_align);
4079 #ifdef CONFIG_ZONE_DMA
4080 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4082 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4084 SLAB_ATTR_RO(cache_dma);
4087 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4089 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4091 SLAB_ATTR_RO(destroy_by_rcu);
4093 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4095 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4098 static ssize_t red_zone_store(struct kmem_cache *s,
4099 const char *buf, size_t length)
4101 if (any_slab_objects(s))
4104 s->flags &= ~SLAB_RED_ZONE;
4106 s->flags |= SLAB_RED_ZONE;
4107 calculate_sizes(s, -1);
4110 SLAB_ATTR(red_zone);
4112 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4114 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4117 static ssize_t poison_store(struct kmem_cache *s,
4118 const char *buf, size_t length)
4120 if (any_slab_objects(s))
4123 s->flags &= ~SLAB_POISON;
4125 s->flags |= SLAB_POISON;
4126 calculate_sizes(s, -1);
4131 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4133 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4136 static ssize_t store_user_store(struct kmem_cache *s,
4137 const char *buf, size_t length)
4139 if (any_slab_objects(s))
4142 s->flags &= ~SLAB_STORE_USER;
4144 s->flags |= SLAB_STORE_USER;
4145 calculate_sizes(s, -1);
4148 SLAB_ATTR(store_user);
4150 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4155 static ssize_t validate_store(struct kmem_cache *s,
4156 const char *buf, size_t length)
4160 if (buf[0] == '1') {
4161 ret = validate_slab_cache(s);
4167 SLAB_ATTR(validate);
4169 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4174 static ssize_t shrink_store(struct kmem_cache *s,
4175 const char *buf, size_t length)
4177 if (buf[0] == '1') {
4178 int rc = kmem_cache_shrink(s);
4188 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4190 if (!(s->flags & SLAB_STORE_USER))
4192 return list_locations(s, buf, TRACK_ALLOC);
4194 SLAB_ATTR_RO(alloc_calls);
4196 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4198 if (!(s->flags & SLAB_STORE_USER))
4200 return list_locations(s, buf, TRACK_FREE);
4202 SLAB_ATTR_RO(free_calls);
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,
4309 &total_objects_attr.attr,
4312 &cpu_slabs_attr.attr,
4316 &sanity_checks_attr.attr,
4318 &hwcache_align_attr.attr,
4319 &reclaim_account_attr.attr,
4320 &destroy_by_rcu_attr.attr,
4321 &red_zone_attr.attr,
4323 &store_user_attr.attr,
4324 &validate_attr.attr,
4326 &alloc_calls_attr.attr,
4327 &free_calls_attr.attr,
4328 #ifdef CONFIG_ZONE_DMA
4329 &cache_dma_attr.attr,
4332 &remote_node_defrag_ratio_attr.attr,
4334 #ifdef CONFIG_SLUB_STATS
4335 &alloc_fastpath_attr.attr,
4336 &alloc_slowpath_attr.attr,
4337 &free_fastpath_attr.attr,
4338 &free_slowpath_attr.attr,
4339 &free_frozen_attr.attr,
4340 &free_add_partial_attr.attr,
4341 &free_remove_partial_attr.attr,
4342 &alloc_from_partial_attr.attr,
4343 &alloc_slab_attr.attr,
4344 &alloc_refill_attr.attr,
4345 &free_slab_attr.attr,
4346 &cpuslab_flush_attr.attr,
4347 &deactivate_full_attr.attr,
4348 &deactivate_empty_attr.attr,
4349 &deactivate_to_head_attr.attr,
4350 &deactivate_to_tail_attr.attr,
4351 &deactivate_remote_frees_attr.attr,
4352 &order_fallback_attr.attr,
4354 #ifdef CONFIG_FAILSLAB
4355 &failslab_attr.attr,
4361 static struct attribute_group slab_attr_group = {
4362 .attrs = slab_attrs,
4365 static ssize_t slab_attr_show(struct kobject *kobj,
4366 struct attribute *attr,
4369 struct slab_attribute *attribute;
4370 struct kmem_cache *s;
4373 attribute = to_slab_attr(attr);
4376 if (!attribute->show)
4379 err = attribute->show(s, buf);
4384 static ssize_t slab_attr_store(struct kobject *kobj,
4385 struct attribute *attr,
4386 const char *buf, size_t len)
4388 struct slab_attribute *attribute;
4389 struct kmem_cache *s;
4392 attribute = to_slab_attr(attr);
4395 if (!attribute->store)
4398 err = attribute->store(s, buf, len);
4403 static void kmem_cache_release(struct kobject *kobj)
4405 struct kmem_cache *s = to_slab(kobj);
4410 static const struct sysfs_ops slab_sysfs_ops = {
4411 .show = slab_attr_show,
4412 .store = slab_attr_store,
4415 static struct kobj_type slab_ktype = {
4416 .sysfs_ops = &slab_sysfs_ops,
4417 .release = kmem_cache_release
4420 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4422 struct kobj_type *ktype = get_ktype(kobj);
4424 if (ktype == &slab_ktype)
4429 static const struct kset_uevent_ops slab_uevent_ops = {
4430 .filter = uevent_filter,
4433 static struct kset *slab_kset;
4435 #define ID_STR_LENGTH 64
4437 /* Create a unique string id for a slab cache:
4439 * Format :[flags-]size
4441 static char *create_unique_id(struct kmem_cache *s)
4443 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4450 * First flags affecting slabcache operations. We will only
4451 * get here for aliasable slabs so we do not need to support
4452 * too many flags. The flags here must cover all flags that
4453 * are matched during merging to guarantee that the id is
4456 if (s->flags & SLAB_CACHE_DMA)
4458 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4460 if (s->flags & SLAB_DEBUG_FREE)
4462 if (!(s->flags & SLAB_NOTRACK))
4466 p += sprintf(p, "%07d", s->size);
4467 BUG_ON(p > name + ID_STR_LENGTH - 1);
4471 static int sysfs_slab_add(struct kmem_cache *s)
4477 if (slab_state < SYSFS)
4478 /* Defer until later */
4481 unmergeable = slab_unmergeable(s);
4484 * Slabcache can never be merged so we can use the name proper.
4485 * This is typically the case for debug situations. In that
4486 * case we can catch duplicate names easily.
4488 sysfs_remove_link(&slab_kset->kobj, s->name);
4492 * Create a unique name for the slab as a target
4495 name = create_unique_id(s);
4498 s->kobj.kset = slab_kset;
4499 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4501 kobject_put(&s->kobj);
4505 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4507 kobject_del(&s->kobj);
4508 kobject_put(&s->kobj);
4511 kobject_uevent(&s->kobj, KOBJ_ADD);
4513 /* Setup first alias */
4514 sysfs_slab_alias(s, s->name);
4520 static void sysfs_slab_remove(struct kmem_cache *s)
4522 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4523 kobject_del(&s->kobj);
4524 kobject_put(&s->kobj);
4528 * Need to buffer aliases during bootup until sysfs becomes
4529 * available lest we lose that information.
4531 struct saved_alias {
4532 struct kmem_cache *s;
4534 struct saved_alias *next;
4537 static struct saved_alias *alias_list;
4539 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4541 struct saved_alias *al;
4543 if (slab_state == SYSFS) {
4545 * If we have a leftover link then remove it.
4547 sysfs_remove_link(&slab_kset->kobj, name);
4548 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4551 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4557 al->next = alias_list;
4562 static int __init slab_sysfs_init(void)
4564 struct kmem_cache *s;
4567 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4569 printk(KERN_ERR "Cannot register slab subsystem.\n");
4575 list_for_each_entry(s, &slab_caches, list) {
4576 err = sysfs_slab_add(s);
4578 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4579 " to sysfs\n", s->name);
4582 while (alias_list) {
4583 struct saved_alias *al = alias_list;
4585 alias_list = alias_list->next;
4586 err = sysfs_slab_alias(al->s, al->name);
4588 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4589 " %s to sysfs\n", s->name);
4597 __initcall(slab_sysfs_init);
4601 * The /proc/slabinfo ABI
4603 #ifdef CONFIG_SLABINFO
4604 static void print_slabinfo_header(struct seq_file *m)
4606 seq_puts(m, "slabinfo - version: 2.1\n");
4607 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4608 "<objperslab> <pagesperslab>");
4609 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4610 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4614 static void *s_start(struct seq_file *m, loff_t *pos)
4618 down_read(&slub_lock);
4620 print_slabinfo_header(m);
4622 return seq_list_start(&slab_caches, *pos);
4625 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4627 return seq_list_next(p, &slab_caches, pos);
4630 static void s_stop(struct seq_file *m, void *p)
4632 up_read(&slub_lock);
4635 static int s_show(struct seq_file *m, void *p)
4637 unsigned long nr_partials = 0;
4638 unsigned long nr_slabs = 0;
4639 unsigned long nr_inuse = 0;
4640 unsigned long nr_objs = 0;
4641 unsigned long nr_free = 0;
4642 struct kmem_cache *s;
4645 s = list_entry(p, struct kmem_cache, list);
4647 for_each_online_node(node) {
4648 struct kmem_cache_node *n = get_node(s, node);
4653 nr_partials += n->nr_partial;
4654 nr_slabs += atomic_long_read(&n->nr_slabs);
4655 nr_objs += atomic_long_read(&n->total_objects);
4656 nr_free += count_partial(n, count_free);
4659 nr_inuse = nr_objs - nr_free;
4661 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4662 nr_objs, s->size, oo_objects(s->oo),
4663 (1 << oo_order(s->oo)));
4664 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4665 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4671 static const struct seq_operations slabinfo_op = {
4678 static int slabinfo_open(struct inode *inode, struct file *file)
4680 return seq_open(file, &slabinfo_op);
4683 static const struct file_operations proc_slabinfo_operations = {
4684 .open = slabinfo_open,
4686 .llseek = seq_lseek,
4687 .release = seq_release,
4690 static int __init slab_proc_init(void)
4692 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4695 module_init(slab_proc_init);
4696 #endif /* CONFIG_SLABINFO */