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)
156 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
157 SLAB_CACHE_DMA | SLAB_NOTRACK)
159 #ifndef ARCH_KMALLOC_MINALIGN
160 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
163 #ifndef ARCH_SLAB_MINALIGN
164 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
168 #define OO_MASK ((1 << OO_SHIFT) - 1)
169 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
171 /* Internal SLUB flags */
172 #define __OBJECT_POISON 0x80000000 /* Poison object */
173 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
175 static int kmem_size = sizeof(struct kmem_cache);
178 static struct notifier_block slab_notifier;
182 DOWN, /* No slab functionality available */
183 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
184 UP, /* Everything works but does not show up in sysfs */
188 /* A list of all slab caches on the system */
189 static DECLARE_RWSEM(slub_lock);
190 static LIST_HEAD(slab_caches);
193 * Tracking user of a slab.
196 unsigned long addr; /* Called from address */
197 int cpu; /* Was running on cpu */
198 int pid; /* Pid context */
199 unsigned long when; /* When did the operation occur */
202 enum track_item { TRACK_ALLOC, TRACK_FREE };
204 #ifdef CONFIG_SLUB_DEBUG
205 static int sysfs_slab_add(struct kmem_cache *);
206 static int sysfs_slab_alias(struct kmem_cache *, const char *);
207 static void sysfs_slab_remove(struct kmem_cache *);
210 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
211 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
213 static inline void sysfs_slab_remove(struct kmem_cache *s)
220 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
222 #ifdef CONFIG_SLUB_STATS
227 /********************************************************************
228 * Core slab cache functions
229 *******************************************************************/
231 int slab_is_available(void)
233 return slab_state >= UP;
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
239 return s->node[node];
241 return &s->local_node;
245 /* Verify that a pointer has an address that is valid within a slab page */
246 static inline int check_valid_pointer(struct kmem_cache *s,
247 struct page *page, const void *object)
254 base = page_address(page);
255 if (object < base || object >= base + page->objects * s->size ||
256 (object - base) % s->size) {
264 * Slow version of get and set free pointer.
266 * This version requires touching the cache lines of kmem_cache which
267 * we avoid to do in the fast alloc free paths. There we obtain the offset
268 * from the page struct.
270 static inline void *get_freepointer(struct kmem_cache *s, void *object)
272 return *(void **)(object + s->offset);
275 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
277 *(void **)(object + s->offset) = fp;
280 /* Loop over all objects in a slab */
281 #define for_each_object(__p, __s, __addr, __objects) \
282 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
286 #define for_each_free_object(__p, __s, __free) \
287 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 static inline struct kmem_cache_order_objects oo_make(int order,
298 struct kmem_cache_order_objects x = {
299 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
305 static inline int oo_order(struct kmem_cache_order_objects x)
307 return x.x >> OO_SHIFT;
310 static inline int oo_objects(struct kmem_cache_order_objects x)
312 return x.x & OO_MASK;
315 #ifdef CONFIG_SLUB_DEBUG
319 #ifdef CONFIG_SLUB_DEBUG_ON
320 static int slub_debug = DEBUG_DEFAULT_FLAGS;
322 static int slub_debug;
325 static char *slub_debug_slabs;
326 static int disable_higher_order_debug;
331 static void print_section(char *text, u8 *addr, unsigned int length)
339 for (i = 0; i < length; i++) {
341 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
344 printk(KERN_CONT " %02x", addr[i]);
346 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
348 printk(KERN_CONT " %s\n", ascii);
355 printk(KERN_CONT " ");
359 printk(KERN_CONT " %s\n", ascii);
363 static struct track *get_track(struct kmem_cache *s, void *object,
364 enum track_item alloc)
369 p = object + s->offset + sizeof(void *);
371 p = object + s->inuse;
376 static void set_track(struct kmem_cache *s, void *object,
377 enum track_item alloc, unsigned long addr)
379 struct track *p = get_track(s, object, alloc);
383 p->cpu = smp_processor_id();
384 p->pid = current->pid;
387 memset(p, 0, sizeof(struct track));
390 static void init_tracking(struct kmem_cache *s, void *object)
392 if (!(s->flags & SLAB_STORE_USER))
395 set_track(s, object, TRACK_FREE, 0UL);
396 set_track(s, object, TRACK_ALLOC, 0UL);
399 static void print_track(const char *s, struct track *t)
404 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
405 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
408 static void print_tracking(struct kmem_cache *s, void *object)
410 if (!(s->flags & SLAB_STORE_USER))
413 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
414 print_track("Freed", get_track(s, object, TRACK_FREE));
417 static void print_page_info(struct page *page)
419 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
420 page, page->objects, page->inuse, page->freelist, page->flags);
424 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
430 vsnprintf(buf, sizeof(buf), fmt, args);
432 printk(KERN_ERR "========================================"
433 "=====================================\n");
434 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
435 printk(KERN_ERR "----------------------------------------"
436 "-------------------------------------\n\n");
439 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
445 vsnprintf(buf, sizeof(buf), fmt, args);
447 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
450 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
452 unsigned int off; /* Offset of last byte */
453 u8 *addr = page_address(page);
455 print_tracking(s, p);
457 print_page_info(page);
459 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
460 p, p - addr, get_freepointer(s, p));
463 print_section("Bytes b4", p - 16, 16);
465 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
467 if (s->flags & SLAB_RED_ZONE)
468 print_section("Redzone", p + s->objsize,
469 s->inuse - s->objsize);
472 off = s->offset + sizeof(void *);
476 if (s->flags & SLAB_STORE_USER)
477 off += 2 * sizeof(struct track);
480 /* Beginning of the filler is the free pointer */
481 print_section("Padding", p + off, s->size - off);
486 static void object_err(struct kmem_cache *s, struct page *page,
487 u8 *object, char *reason)
489 slab_bug(s, "%s", reason);
490 print_trailer(s, page, object);
493 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
499 vsnprintf(buf, sizeof(buf), fmt, args);
501 slab_bug(s, "%s", buf);
502 print_page_info(page);
506 static void init_object(struct kmem_cache *s, void *object, int active)
510 if (s->flags & __OBJECT_POISON) {
511 memset(p, POISON_FREE, s->objsize - 1);
512 p[s->objsize - 1] = POISON_END;
515 if (s->flags & SLAB_RED_ZONE)
516 memset(p + s->objsize,
517 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
518 s->inuse - s->objsize);
521 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
524 if (*start != (u8)value)
532 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
533 void *from, void *to)
535 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
536 memset(from, data, to - from);
539 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
540 u8 *object, char *what,
541 u8 *start, unsigned int value, unsigned int bytes)
546 fault = check_bytes(start, value, bytes);
551 while (end > fault && end[-1] == value)
554 slab_bug(s, "%s overwritten", what);
555 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
556 fault, end - 1, fault[0], value);
557 print_trailer(s, page, object);
559 restore_bytes(s, what, value, fault, end);
567 * Bytes of the object to be managed.
568 * If the freepointer may overlay the object then the free
569 * pointer is the first word of the object.
571 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
574 * object + s->objsize
575 * Padding to reach word boundary. This is also used for Redzoning.
576 * Padding is extended by another word if Redzoning is enabled and
579 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
580 * 0xcc (RED_ACTIVE) for objects in use.
583 * Meta data starts here.
585 * A. Free pointer (if we cannot overwrite object on free)
586 * B. Tracking data for SLAB_STORE_USER
587 * C. Padding to reach required alignment boundary or at mininum
588 * one word if debugging is on to be able to detect writes
589 * before the word boundary.
591 * Padding is done using 0x5a (POISON_INUSE)
594 * Nothing is used beyond s->size.
596 * If slabcaches are merged then the objsize and inuse boundaries are mostly
597 * ignored. And therefore no slab options that rely on these boundaries
598 * may be used with merged slabcaches.
601 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
603 unsigned long off = s->inuse; /* The end of info */
606 /* Freepointer is placed after the object. */
607 off += sizeof(void *);
609 if (s->flags & SLAB_STORE_USER)
610 /* We also have user information there */
611 off += 2 * sizeof(struct track);
616 return check_bytes_and_report(s, page, p, "Object padding",
617 p + off, POISON_INUSE, s->size - off);
620 /* Check the pad bytes at the end of a slab page */
621 static int slab_pad_check(struct kmem_cache *s, struct page *page)
629 if (!(s->flags & SLAB_POISON))
632 start = page_address(page);
633 length = (PAGE_SIZE << compound_order(page));
634 end = start + length;
635 remainder = length % s->size;
639 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
642 while (end > fault && end[-1] == POISON_INUSE)
645 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
646 print_section("Padding", end - remainder, remainder);
648 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
652 static int check_object(struct kmem_cache *s, struct page *page,
653 void *object, int active)
656 u8 *endobject = object + s->objsize;
658 if (s->flags & SLAB_RED_ZONE) {
660 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
662 if (!check_bytes_and_report(s, page, object, "Redzone",
663 endobject, red, s->inuse - s->objsize))
666 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
667 check_bytes_and_report(s, page, p, "Alignment padding",
668 endobject, POISON_INUSE, s->inuse - s->objsize);
672 if (s->flags & SLAB_POISON) {
673 if (!active && (s->flags & __OBJECT_POISON) &&
674 (!check_bytes_and_report(s, page, p, "Poison", p,
675 POISON_FREE, s->objsize - 1) ||
676 !check_bytes_and_report(s, page, p, "Poison",
677 p + s->objsize - 1, POISON_END, 1)))
680 * check_pad_bytes cleans up on its own.
682 check_pad_bytes(s, page, p);
685 if (!s->offset && active)
687 * Object and freepointer overlap. Cannot check
688 * freepointer while object is allocated.
692 /* Check free pointer validity */
693 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
694 object_err(s, page, p, "Freepointer corrupt");
696 * No choice but to zap it and thus lose the remainder
697 * of the free objects in this slab. May cause
698 * another error because the object count is now wrong.
700 set_freepointer(s, p, NULL);
706 static int check_slab(struct kmem_cache *s, struct page *page)
710 VM_BUG_ON(!irqs_disabled());
712 if (!PageSlab(page)) {
713 slab_err(s, page, "Not a valid slab page");
717 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
718 if (page->objects > maxobj) {
719 slab_err(s, page, "objects %u > max %u",
720 s->name, page->objects, maxobj);
723 if (page->inuse > page->objects) {
724 slab_err(s, page, "inuse %u > max %u",
725 s->name, page->inuse, page->objects);
728 /* Slab_pad_check fixes things up after itself */
729 slab_pad_check(s, page);
734 * Determine if a certain object on a page is on the freelist. Must hold the
735 * slab lock to guarantee that the chains are in a consistent state.
737 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
740 void *fp = page->freelist;
742 unsigned long max_objects;
744 while (fp && nr <= page->objects) {
747 if (!check_valid_pointer(s, page, fp)) {
749 object_err(s, page, object,
750 "Freechain corrupt");
751 set_freepointer(s, object, NULL);
754 slab_err(s, page, "Freepointer corrupt");
755 page->freelist = NULL;
756 page->inuse = page->objects;
757 slab_fix(s, "Freelist cleared");
763 fp = get_freepointer(s, object);
767 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
768 if (max_objects > MAX_OBJS_PER_PAGE)
769 max_objects = MAX_OBJS_PER_PAGE;
771 if (page->objects != max_objects) {
772 slab_err(s, page, "Wrong number of objects. Found %d but "
773 "should be %d", page->objects, max_objects);
774 page->objects = max_objects;
775 slab_fix(s, "Number of objects adjusted.");
777 if (page->inuse != page->objects - nr) {
778 slab_err(s, page, "Wrong object count. Counter is %d but "
779 "counted were %d", page->inuse, page->objects - nr);
780 page->inuse = page->objects - nr;
781 slab_fix(s, "Object count adjusted.");
783 return search == NULL;
786 static void trace(struct kmem_cache *s, struct page *page, void *object,
789 if (s->flags & SLAB_TRACE) {
790 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
792 alloc ? "alloc" : "free",
797 print_section("Object", (void *)object, s->objsize);
804 * Tracking of fully allocated slabs for debugging purposes.
806 static void add_full(struct kmem_cache_node *n, struct page *page)
808 spin_lock(&n->list_lock);
809 list_add(&page->lru, &n->full);
810 spin_unlock(&n->list_lock);
813 static void remove_full(struct kmem_cache *s, struct page *page)
815 struct kmem_cache_node *n;
817 if (!(s->flags & SLAB_STORE_USER))
820 n = get_node(s, page_to_nid(page));
822 spin_lock(&n->list_lock);
823 list_del(&page->lru);
824 spin_unlock(&n->list_lock);
827 /* Tracking of the number of slabs for debugging purposes */
828 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
830 struct kmem_cache_node *n = get_node(s, node);
832 return atomic_long_read(&n->nr_slabs);
835 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
837 return atomic_long_read(&n->nr_slabs);
840 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
842 struct kmem_cache_node *n = get_node(s, node);
845 * May be called early in order to allocate a slab for the
846 * kmem_cache_node structure. Solve the chicken-egg
847 * dilemma by deferring the increment of the count during
848 * bootstrap (see early_kmem_cache_node_alloc).
850 if (!NUMA_BUILD || n) {
851 atomic_long_inc(&n->nr_slabs);
852 atomic_long_add(objects, &n->total_objects);
855 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
857 struct kmem_cache_node *n = get_node(s, node);
859 atomic_long_dec(&n->nr_slabs);
860 atomic_long_sub(objects, &n->total_objects);
863 /* Object debug checks for alloc/free paths */
864 static void setup_object_debug(struct kmem_cache *s, struct page *page,
867 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
870 init_object(s, object, 0);
871 init_tracking(s, object);
874 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
875 void *object, unsigned long addr)
877 if (!check_slab(s, page))
880 if (!on_freelist(s, page, object)) {
881 object_err(s, page, object, "Object already allocated");
885 if (!check_valid_pointer(s, page, object)) {
886 object_err(s, page, object, "Freelist Pointer check fails");
890 if (!check_object(s, page, object, 0))
893 /* Success perform special debug activities for allocs */
894 if (s->flags & SLAB_STORE_USER)
895 set_track(s, object, TRACK_ALLOC, addr);
896 trace(s, page, object, 1);
897 init_object(s, object, 1);
901 if (PageSlab(page)) {
903 * If this is a slab page then lets do the best we can
904 * to avoid issues in the future. Marking all objects
905 * as used avoids touching the remaining objects.
907 slab_fix(s, "Marking all objects used");
908 page->inuse = page->objects;
909 page->freelist = NULL;
914 static int free_debug_processing(struct kmem_cache *s, struct page *page,
915 void *object, unsigned long addr)
917 if (!check_slab(s, page))
920 if (!check_valid_pointer(s, page, object)) {
921 slab_err(s, page, "Invalid object pointer 0x%p", object);
925 if (on_freelist(s, page, object)) {
926 object_err(s, page, object, "Object already free");
930 if (!check_object(s, page, object, 1))
933 if (unlikely(s != page->slab)) {
934 if (!PageSlab(page)) {
935 slab_err(s, page, "Attempt to free object(0x%p) "
936 "outside of slab", object);
937 } else if (!page->slab) {
939 "SLUB <none>: no slab for object 0x%p.\n",
943 object_err(s, page, object,
944 "page slab pointer corrupt.");
948 /* Special debug activities for freeing objects */
949 if (!PageSlubFrozen(page) && !page->freelist)
950 remove_full(s, page);
951 if (s->flags & SLAB_STORE_USER)
952 set_track(s, object, TRACK_FREE, addr);
953 trace(s, page, object, 0);
954 init_object(s, object, 0);
958 slab_fix(s, "Object at 0x%p not freed", object);
962 static int __init setup_slub_debug(char *str)
964 slub_debug = DEBUG_DEFAULT_FLAGS;
965 if (*str++ != '=' || !*str)
967 * No options specified. Switch on full debugging.
973 * No options but restriction on slabs. This means full
974 * debugging for slabs matching a pattern.
978 if (tolower(*str) == 'o') {
980 * Avoid enabling debugging on caches if its minimum order
981 * would increase as a result.
983 disable_higher_order_debug = 1;
990 * Switch off all debugging measures.
995 * Determine which debug features should be switched on
997 for (; *str && *str != ','; str++) {
998 switch (tolower(*str)) {
1000 slub_debug |= SLAB_DEBUG_FREE;
1003 slub_debug |= SLAB_RED_ZONE;
1006 slub_debug |= SLAB_POISON;
1009 slub_debug |= SLAB_STORE_USER;
1012 slub_debug |= SLAB_TRACE;
1015 printk(KERN_ERR "slub_debug option '%c' "
1016 "unknown. skipped\n", *str);
1022 slub_debug_slabs = str + 1;
1027 __setup("slub_debug", setup_slub_debug);
1029 static unsigned long kmem_cache_flags(unsigned long objsize,
1030 unsigned long flags, const char *name,
1031 void (*ctor)(void *))
1034 * Enable debugging if selected on the kernel commandline.
1036 if (slub_debug && (!slub_debug_slabs ||
1037 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1038 flags |= slub_debug;
1043 static inline void setup_object_debug(struct kmem_cache *s,
1044 struct page *page, void *object) {}
1046 static inline int alloc_debug_processing(struct kmem_cache *s,
1047 struct page *page, void *object, unsigned long addr) { return 0; }
1049 static inline int free_debug_processing(struct kmem_cache *s,
1050 struct page *page, void *object, unsigned long addr) { return 0; }
1052 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1054 static inline int check_object(struct kmem_cache *s, struct page *page,
1055 void *object, int active) { return 1; }
1056 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1057 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1058 unsigned long flags, const char *name,
1059 void (*ctor)(void *))
1063 #define slub_debug 0
1065 #define disable_higher_order_debug 0
1067 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1069 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1071 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1073 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1078 * Slab allocation and freeing
1080 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1081 struct kmem_cache_order_objects oo)
1083 int order = oo_order(oo);
1085 flags |= __GFP_NOTRACK;
1088 return alloc_pages(flags, order);
1090 return alloc_pages_node(node, flags, order);
1093 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1096 struct kmem_cache_order_objects oo = s->oo;
1099 flags |= s->allocflags;
1102 * Let the initial higher-order allocation fail under memory pressure
1103 * so we fall-back to the minimum order allocation.
1105 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1107 page = alloc_slab_page(alloc_gfp, node, oo);
1108 if (unlikely(!page)) {
1111 * Allocation may have failed due to fragmentation.
1112 * Try a lower order alloc if possible
1114 page = alloc_slab_page(flags, node, oo);
1118 stat(this_cpu_ptr(s->cpu_slab), ORDER_FALLBACK);
1121 if (kmemcheck_enabled
1122 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1123 int pages = 1 << oo_order(oo);
1125 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1128 * Objects from caches that have a constructor don't get
1129 * cleared when they're allocated, so we need to do it here.
1132 kmemcheck_mark_uninitialized_pages(page, pages);
1134 kmemcheck_mark_unallocated_pages(page, pages);
1137 page->objects = oo_objects(oo);
1138 mod_zone_page_state(page_zone(page),
1139 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1140 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1146 static void setup_object(struct kmem_cache *s, struct page *page,
1149 setup_object_debug(s, page, object);
1150 if (unlikely(s->ctor))
1154 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1161 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1163 page = allocate_slab(s,
1164 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1168 inc_slabs_node(s, page_to_nid(page), page->objects);
1170 page->flags |= 1 << PG_slab;
1171 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1172 SLAB_STORE_USER | SLAB_TRACE))
1173 __SetPageSlubDebug(page);
1175 start = page_address(page);
1177 if (unlikely(s->flags & SLAB_POISON))
1178 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1181 for_each_object(p, s, start, page->objects) {
1182 setup_object(s, page, last);
1183 set_freepointer(s, last, p);
1186 setup_object(s, page, last);
1187 set_freepointer(s, last, NULL);
1189 page->freelist = start;
1195 static void __free_slab(struct kmem_cache *s, struct page *page)
1197 int order = compound_order(page);
1198 int pages = 1 << order;
1200 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1203 slab_pad_check(s, page);
1204 for_each_object(p, s, page_address(page),
1206 check_object(s, page, p, 0);
1207 __ClearPageSlubDebug(page);
1210 kmemcheck_free_shadow(page, compound_order(page));
1212 mod_zone_page_state(page_zone(page),
1213 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1214 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1217 __ClearPageSlab(page);
1218 reset_page_mapcount(page);
1219 if (current->reclaim_state)
1220 current->reclaim_state->reclaimed_slab += pages;
1221 __free_pages(page, order);
1224 static void rcu_free_slab(struct rcu_head *h)
1228 page = container_of((struct list_head *)h, struct page, lru);
1229 __free_slab(page->slab, page);
1232 static void free_slab(struct kmem_cache *s, struct page *page)
1234 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1236 * RCU free overloads the RCU head over the LRU
1238 struct rcu_head *head = (void *)&page->lru;
1240 call_rcu(head, rcu_free_slab);
1242 __free_slab(s, page);
1245 static void discard_slab(struct kmem_cache *s, struct page *page)
1247 dec_slabs_node(s, page_to_nid(page), page->objects);
1252 * Per slab locking using the pagelock
1254 static __always_inline void slab_lock(struct page *page)
1256 bit_spin_lock(PG_locked, &page->flags);
1259 static __always_inline void slab_unlock(struct page *page)
1261 __bit_spin_unlock(PG_locked, &page->flags);
1264 static __always_inline int slab_trylock(struct page *page)
1268 rc = bit_spin_trylock(PG_locked, &page->flags);
1273 * Management of partially allocated slabs
1275 static void add_partial(struct kmem_cache_node *n,
1276 struct page *page, int tail)
1278 spin_lock(&n->list_lock);
1281 list_add_tail(&page->lru, &n->partial);
1283 list_add(&page->lru, &n->partial);
1284 spin_unlock(&n->list_lock);
1287 static void remove_partial(struct kmem_cache *s, struct page *page)
1289 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1291 spin_lock(&n->list_lock);
1292 list_del(&page->lru);
1294 spin_unlock(&n->list_lock);
1298 * Lock slab and remove from the partial list.
1300 * Must hold list_lock.
1302 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1305 if (slab_trylock(page)) {
1306 list_del(&page->lru);
1308 __SetPageSlubFrozen(page);
1315 * Try to allocate a partial slab from a specific node.
1317 static struct page *get_partial_node(struct kmem_cache_node *n)
1322 * Racy check. If we mistakenly see no partial slabs then we
1323 * just allocate an empty slab. If we mistakenly try to get a
1324 * partial slab and there is none available then get_partials()
1327 if (!n || !n->nr_partial)
1330 spin_lock(&n->list_lock);
1331 list_for_each_entry(page, &n->partial, lru)
1332 if (lock_and_freeze_slab(n, page))
1336 spin_unlock(&n->list_lock);
1341 * Get a page from somewhere. Search in increasing NUMA distances.
1343 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1346 struct zonelist *zonelist;
1349 enum zone_type high_zoneidx = gfp_zone(flags);
1353 * The defrag ratio allows a configuration of the tradeoffs between
1354 * inter node defragmentation and node local allocations. A lower
1355 * defrag_ratio increases the tendency to do local allocations
1356 * instead of attempting to obtain partial slabs from other nodes.
1358 * If the defrag_ratio is set to 0 then kmalloc() always
1359 * returns node local objects. If the ratio is higher then kmalloc()
1360 * may return off node objects because partial slabs are obtained
1361 * from other nodes and filled up.
1363 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1364 * defrag_ratio = 1000) then every (well almost) allocation will
1365 * first attempt to defrag slab caches on other nodes. This means
1366 * scanning over all nodes to look for partial slabs which may be
1367 * expensive if we do it every time we are trying to find a slab
1368 * with available objects.
1370 if (!s->remote_node_defrag_ratio ||
1371 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1374 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1375 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1376 struct kmem_cache_node *n;
1378 n = get_node(s, zone_to_nid(zone));
1380 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1381 n->nr_partial > s->min_partial) {
1382 page = get_partial_node(n);
1392 * Get a partial page, lock it and return it.
1394 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1397 int searchnode = (node == -1) ? numa_node_id() : node;
1399 page = get_partial_node(get_node(s, searchnode));
1400 if (page || (flags & __GFP_THISNODE))
1403 return get_any_partial(s, flags);
1407 * Move a page back to the lists.
1409 * Must be called with the slab lock held.
1411 * On exit the slab lock will have been dropped.
1413 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1415 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1416 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1418 __ClearPageSlubFrozen(page);
1421 if (page->freelist) {
1422 add_partial(n, page, tail);
1423 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1425 stat(c, DEACTIVATE_FULL);
1426 if (SLABDEBUG && PageSlubDebug(page) &&
1427 (s->flags & SLAB_STORE_USER))
1432 stat(c, DEACTIVATE_EMPTY);
1433 if (n->nr_partial < s->min_partial) {
1435 * Adding an empty slab to the partial slabs in order
1436 * to avoid page allocator overhead. This slab needs
1437 * to come after the other slabs with objects in
1438 * so that the others get filled first. That way the
1439 * size of the partial list stays small.
1441 * kmem_cache_shrink can reclaim any empty slabs from
1444 add_partial(n, page, 1);
1448 stat(__this_cpu_ptr(s->cpu_slab), FREE_SLAB);
1449 discard_slab(s, page);
1455 * Remove the cpu slab
1457 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1459 struct page *page = c->page;
1463 stat(c, DEACTIVATE_REMOTE_FREES);
1465 * Merge cpu freelist into slab freelist. Typically we get here
1466 * because both freelists are empty. So this is unlikely
1469 while (unlikely(c->freelist)) {
1472 tail = 0; /* Hot objects. Put the slab first */
1474 /* Retrieve object from cpu_freelist */
1475 object = c->freelist;
1476 c->freelist = c->freelist[c->offset];
1478 /* And put onto the regular freelist */
1479 object[c->offset] = page->freelist;
1480 page->freelist = object;
1484 unfreeze_slab(s, page, tail);
1487 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1489 stat(c, CPUSLAB_FLUSH);
1491 deactivate_slab(s, c);
1497 * Called from IPI handler with interrupts disabled.
1499 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1501 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1503 if (likely(c && c->page))
1507 static void flush_cpu_slab(void *d)
1509 struct kmem_cache *s = d;
1511 __flush_cpu_slab(s, smp_processor_id());
1514 static void flush_all(struct kmem_cache *s)
1516 on_each_cpu(flush_cpu_slab, s, 1);
1520 * Check if the objects in a per cpu structure fit numa
1521 * locality expectations.
1523 static inline int node_match(struct kmem_cache_cpu *c, int node)
1526 if (node != -1 && c->node != node)
1532 static int count_free(struct page *page)
1534 return page->objects - page->inuse;
1537 static unsigned long count_partial(struct kmem_cache_node *n,
1538 int (*get_count)(struct page *))
1540 unsigned long flags;
1541 unsigned long x = 0;
1544 spin_lock_irqsave(&n->list_lock, flags);
1545 list_for_each_entry(page, &n->partial, lru)
1546 x += get_count(page);
1547 spin_unlock_irqrestore(&n->list_lock, flags);
1551 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1553 #ifdef CONFIG_SLUB_DEBUG
1554 return atomic_long_read(&n->total_objects);
1560 static noinline void
1561 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1566 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1568 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1569 "default order: %d, min order: %d\n", s->name, s->objsize,
1570 s->size, oo_order(s->oo), oo_order(s->min));
1572 if (oo_order(s->min) > get_order(s->objsize))
1573 printk(KERN_WARNING " %s debugging increased min order, use "
1574 "slub_debug=O to disable.\n", s->name);
1576 for_each_online_node(node) {
1577 struct kmem_cache_node *n = get_node(s, node);
1578 unsigned long nr_slabs;
1579 unsigned long nr_objs;
1580 unsigned long nr_free;
1585 nr_free = count_partial(n, count_free);
1586 nr_slabs = node_nr_slabs(n);
1587 nr_objs = node_nr_objs(n);
1590 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1591 node, nr_slabs, nr_objs, nr_free);
1596 * Slow path. The lockless freelist is empty or we need to perform
1599 * Interrupts are disabled.
1601 * Processing is still very fast if new objects have been freed to the
1602 * regular freelist. In that case we simply take over the regular freelist
1603 * as the lockless freelist and zap the regular freelist.
1605 * If that is not working then we fall back to the partial lists. We take the
1606 * first element of the freelist as the object to allocate now and move the
1607 * rest of the freelist to the lockless freelist.
1609 * And if we were unable to get a new slab from the partial slab lists then
1610 * we need to allocate a new slab. This is the slowest path since it involves
1611 * a call to the page allocator and the setup of a new slab.
1613 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1614 unsigned long addr, struct kmem_cache_cpu *c)
1619 /* We handle __GFP_ZERO in the caller */
1620 gfpflags &= ~__GFP_ZERO;
1626 if (unlikely(!node_match(c, node)))
1629 stat(c, ALLOC_REFILL);
1632 object = c->page->freelist;
1633 if (unlikely(!object))
1635 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1638 c->freelist = object[c->offset];
1639 c->page->inuse = c->page->objects;
1640 c->page->freelist = NULL;
1641 c->node = page_to_nid(c->page);
1643 slab_unlock(c->page);
1644 stat(c, ALLOC_SLOWPATH);
1648 deactivate_slab(s, c);
1651 new = get_partial(s, gfpflags, node);
1654 stat(c, ALLOC_FROM_PARTIAL);
1658 if (gfpflags & __GFP_WAIT)
1661 new = new_slab(s, gfpflags, node);
1663 if (gfpflags & __GFP_WAIT)
1664 local_irq_disable();
1667 c = __this_cpu_ptr(s->cpu_slab);
1668 stat(c, ALLOC_SLAB);
1672 __SetPageSlubFrozen(new);
1676 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1677 slab_out_of_memory(s, gfpflags, node);
1680 if (!alloc_debug_processing(s, c->page, object, addr))
1684 c->page->freelist = object[c->offset];
1690 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1691 * have the fastpath folded into their functions. So no function call
1692 * overhead for requests that can be satisfied on the fastpath.
1694 * The fastpath works by first checking if the lockless freelist can be used.
1695 * If not then __slab_alloc is called for slow processing.
1697 * Otherwise we can simply pick the next object from the lockless free list.
1699 static __always_inline void *slab_alloc(struct kmem_cache *s,
1700 gfp_t gfpflags, int node, unsigned long addr)
1703 struct kmem_cache_cpu *c;
1704 unsigned long flags;
1705 unsigned long objsize;
1707 gfpflags &= gfp_allowed_mask;
1709 lockdep_trace_alloc(gfpflags);
1710 might_sleep_if(gfpflags & __GFP_WAIT);
1712 if (should_failslab(s->objsize, gfpflags))
1715 local_irq_save(flags);
1716 c = __this_cpu_ptr(s->cpu_slab);
1717 object = c->freelist;
1718 objsize = c->objsize;
1719 if (unlikely(!object || !node_match(c, node)))
1721 object = __slab_alloc(s, gfpflags, node, addr, c);
1724 c->freelist = object[c->offset];
1725 stat(c, ALLOC_FASTPATH);
1727 local_irq_restore(flags);
1729 if (unlikely(gfpflags & __GFP_ZERO) && object)
1730 memset(object, 0, objsize);
1732 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize);
1733 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
1738 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1740 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1742 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1746 EXPORT_SYMBOL(kmem_cache_alloc);
1748 #ifdef CONFIG_TRACING
1749 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1751 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1753 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1757 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1759 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1761 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1762 s->objsize, s->size, gfpflags, node);
1766 EXPORT_SYMBOL(kmem_cache_alloc_node);
1769 #ifdef CONFIG_TRACING
1770 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1774 return slab_alloc(s, gfpflags, node, _RET_IP_);
1776 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1780 * Slow patch handling. This may still be called frequently since objects
1781 * have a longer lifetime than the cpu slabs in most processing loads.
1783 * So we still attempt to reduce cache line usage. Just take the slab
1784 * lock and free the item. If there is no additional partial page
1785 * handling required then we can return immediately.
1787 static void __slab_free(struct kmem_cache *s, struct page *page,
1788 void *x, unsigned long addr, unsigned int offset)
1791 void **object = (void *)x;
1792 struct kmem_cache_cpu *c;
1794 c = __this_cpu_ptr(s->cpu_slab);
1795 stat(c, FREE_SLOWPATH);
1798 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1802 prior = object[offset] = page->freelist;
1803 page->freelist = object;
1806 if (unlikely(PageSlubFrozen(page))) {
1807 stat(c, FREE_FROZEN);
1811 if (unlikely(!page->inuse))
1815 * Objects left in the slab. If it was not on the partial list before
1818 if (unlikely(!prior)) {
1819 add_partial(get_node(s, page_to_nid(page)), page, 1);
1820 stat(c, FREE_ADD_PARTIAL);
1830 * Slab still on the partial list.
1832 remove_partial(s, page);
1833 stat(c, FREE_REMOVE_PARTIAL);
1837 discard_slab(s, page);
1841 if (!free_debug_processing(s, page, x, addr))
1847 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1848 * can perform fastpath freeing without additional function calls.
1850 * The fastpath is only possible if we are freeing to the current cpu slab
1851 * of this processor. This typically the case if we have just allocated
1854 * If fastpath is not possible then fall back to __slab_free where we deal
1855 * with all sorts of special processing.
1857 static __always_inline void slab_free(struct kmem_cache *s,
1858 struct page *page, void *x, unsigned long addr)
1860 void **object = (void *)x;
1861 struct kmem_cache_cpu *c;
1862 unsigned long flags;
1864 kmemleak_free_recursive(x, s->flags);
1865 local_irq_save(flags);
1866 c = __this_cpu_ptr(s->cpu_slab);
1867 kmemcheck_slab_free(s, object, c->objsize);
1868 debug_check_no_locks_freed(object, c->objsize);
1869 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1870 debug_check_no_obj_freed(object, c->objsize);
1871 if (likely(page == c->page && c->node >= 0)) {
1872 object[c->offset] = c->freelist;
1873 c->freelist = object;
1874 stat(c, FREE_FASTPATH);
1876 __slab_free(s, page, x, addr, c->offset);
1878 local_irq_restore(flags);
1881 void kmem_cache_free(struct kmem_cache *s, void *x)
1885 page = virt_to_head_page(x);
1887 slab_free(s, page, x, _RET_IP_);
1889 trace_kmem_cache_free(_RET_IP_, x);
1891 EXPORT_SYMBOL(kmem_cache_free);
1893 /* Figure out on which slab page the object resides */
1894 static struct page *get_object_page(const void *x)
1896 struct page *page = virt_to_head_page(x);
1898 if (!PageSlab(page))
1905 * Object placement in a slab is made very easy because we always start at
1906 * offset 0. If we tune the size of the object to the alignment then we can
1907 * get the required alignment by putting one properly sized object after
1910 * Notice that the allocation order determines the sizes of the per cpu
1911 * caches. Each processor has always one slab available for allocations.
1912 * Increasing the allocation order reduces the number of times that slabs
1913 * must be moved on and off the partial lists and is therefore a factor in
1918 * Mininum / Maximum order of slab pages. This influences locking overhead
1919 * and slab fragmentation. A higher order reduces the number of partial slabs
1920 * and increases the number of allocations possible without having to
1921 * take the list_lock.
1923 static int slub_min_order;
1924 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1925 static int slub_min_objects;
1928 * Merge control. If this is set then no merging of slab caches will occur.
1929 * (Could be removed. This was introduced to pacify the merge skeptics.)
1931 static int slub_nomerge;
1934 * Calculate the order of allocation given an slab object size.
1936 * The order of allocation has significant impact on performance and other
1937 * system components. Generally order 0 allocations should be preferred since
1938 * order 0 does not cause fragmentation in the page allocator. Larger objects
1939 * be problematic to put into order 0 slabs because there may be too much
1940 * unused space left. We go to a higher order if more than 1/16th of the slab
1943 * In order to reach satisfactory performance we must ensure that a minimum
1944 * number of objects is in one slab. Otherwise we may generate too much
1945 * activity on the partial lists which requires taking the list_lock. This is
1946 * less a concern for large slabs though which are rarely used.
1948 * slub_max_order specifies the order where we begin to stop considering the
1949 * number of objects in a slab as critical. If we reach slub_max_order then
1950 * we try to keep the page order as low as possible. So we accept more waste
1951 * of space in favor of a small page order.
1953 * Higher order allocations also allow the placement of more objects in a
1954 * slab and thereby reduce object handling overhead. If the user has
1955 * requested a higher mininum order then we start with that one instead of
1956 * the smallest order which will fit the object.
1958 static inline int slab_order(int size, int min_objects,
1959 int max_order, int fract_leftover)
1963 int min_order = slub_min_order;
1965 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1966 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1968 for (order = max(min_order,
1969 fls(min_objects * size - 1) - PAGE_SHIFT);
1970 order <= max_order; order++) {
1972 unsigned long slab_size = PAGE_SIZE << order;
1974 if (slab_size < min_objects * size)
1977 rem = slab_size % size;
1979 if (rem <= slab_size / fract_leftover)
1987 static inline int calculate_order(int size)
1995 * Attempt to find best configuration for a slab. This
1996 * works by first attempting to generate a layout with
1997 * the best configuration and backing off gradually.
1999 * First we reduce the acceptable waste in a slab. Then
2000 * we reduce the minimum objects required in a slab.
2002 min_objects = slub_min_objects;
2004 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2005 max_objects = (PAGE_SIZE << slub_max_order)/size;
2006 min_objects = min(min_objects, max_objects);
2008 while (min_objects > 1) {
2010 while (fraction >= 4) {
2011 order = slab_order(size, min_objects,
2012 slub_max_order, fraction);
2013 if (order <= slub_max_order)
2021 * We were unable to place multiple objects in a slab. Now
2022 * lets see if we can place a single object there.
2024 order = slab_order(size, 1, slub_max_order, 1);
2025 if (order <= slub_max_order)
2029 * Doh this slab cannot be placed using slub_max_order.
2031 order = slab_order(size, 1, MAX_ORDER, 1);
2032 if (order < MAX_ORDER)
2038 * Figure out what the alignment of the objects will be.
2040 static unsigned long calculate_alignment(unsigned long flags,
2041 unsigned long align, unsigned long size)
2044 * If the user wants hardware cache aligned objects then follow that
2045 * suggestion if the object is sufficiently large.
2047 * The hardware cache alignment cannot override the specified
2048 * alignment though. If that is greater then use it.
2050 if (flags & SLAB_HWCACHE_ALIGN) {
2051 unsigned long ralign = cache_line_size();
2052 while (size <= ralign / 2)
2054 align = max(align, ralign);
2057 if (align < ARCH_SLAB_MINALIGN)
2058 align = ARCH_SLAB_MINALIGN;
2060 return ALIGN(align, sizeof(void *));
2063 static void init_kmem_cache_cpu(struct kmem_cache *s,
2064 struct kmem_cache_cpu *c)
2069 c->offset = s->offset / sizeof(void *);
2070 c->objsize = s->objsize;
2071 #ifdef CONFIG_SLUB_STATS
2072 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
2077 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2080 spin_lock_init(&n->list_lock);
2081 INIT_LIST_HEAD(&n->partial);
2082 #ifdef CONFIG_SLUB_DEBUG
2083 atomic_long_set(&n->nr_slabs, 0);
2084 atomic_long_set(&n->total_objects, 0);
2085 INIT_LIST_HEAD(&n->full);
2089 static DEFINE_PER_CPU(struct kmem_cache_cpu, kmalloc_percpu[SLUB_PAGE_SHIFT]);
2091 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2095 if (s < kmalloc_caches + KMALLOC_CACHES && s >= kmalloc_caches)
2097 * Boot time creation of the kmalloc array. Use static per cpu data
2098 * since the per cpu allocator is not available yet.
2100 s->cpu_slab = per_cpu_var(kmalloc_percpu) + (s - kmalloc_caches);
2102 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2107 for_each_possible_cpu(cpu)
2108 init_kmem_cache_cpu(s, per_cpu_ptr(s->cpu_slab, cpu));
2114 * No kmalloc_node yet so do it by hand. We know that this is the first
2115 * slab on the node for this slabcache. There are no concurrent accesses
2118 * Note that this function only works on the kmalloc_node_cache
2119 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2120 * memory on a fresh node that has no slab structures yet.
2122 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2125 struct kmem_cache_node *n;
2126 unsigned long flags;
2128 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2130 page = new_slab(kmalloc_caches, gfpflags, node);
2133 if (page_to_nid(page) != node) {
2134 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2136 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2137 "in order to be able to continue\n");
2142 page->freelist = get_freepointer(kmalloc_caches, n);
2144 kmalloc_caches->node[node] = n;
2145 #ifdef CONFIG_SLUB_DEBUG
2146 init_object(kmalloc_caches, n, 1);
2147 init_tracking(kmalloc_caches, n);
2149 init_kmem_cache_node(n, kmalloc_caches);
2150 inc_slabs_node(kmalloc_caches, node, page->objects);
2153 * lockdep requires consistent irq usage for each lock
2154 * so even though there cannot be a race this early in
2155 * the boot sequence, we still disable irqs.
2157 local_irq_save(flags);
2158 add_partial(n, page, 0);
2159 local_irq_restore(flags);
2162 static void free_kmem_cache_nodes(struct kmem_cache *s)
2166 for_each_node_state(node, N_NORMAL_MEMORY) {
2167 struct kmem_cache_node *n = s->node[node];
2168 if (n && n != &s->local_node)
2169 kmem_cache_free(kmalloc_caches, n);
2170 s->node[node] = NULL;
2174 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2179 if (slab_state >= UP)
2180 local_node = page_to_nid(virt_to_page(s));
2184 for_each_node_state(node, N_NORMAL_MEMORY) {
2185 struct kmem_cache_node *n;
2187 if (local_node == node)
2190 if (slab_state == DOWN) {
2191 early_kmem_cache_node_alloc(gfpflags, node);
2194 n = kmem_cache_alloc_node(kmalloc_caches,
2198 free_kmem_cache_nodes(s);
2204 init_kmem_cache_node(n, s);
2209 static void free_kmem_cache_nodes(struct kmem_cache *s)
2213 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2215 init_kmem_cache_node(&s->local_node, s);
2220 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2222 if (min < MIN_PARTIAL)
2224 else if (min > MAX_PARTIAL)
2226 s->min_partial = min;
2230 * calculate_sizes() determines the order and the distribution of data within
2233 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2235 unsigned long flags = s->flags;
2236 unsigned long size = s->objsize;
2237 unsigned long align = s->align;
2241 * Round up object size to the next word boundary. We can only
2242 * place the free pointer at word boundaries and this determines
2243 * the possible location of the free pointer.
2245 size = ALIGN(size, sizeof(void *));
2247 #ifdef CONFIG_SLUB_DEBUG
2249 * Determine if we can poison the object itself. If the user of
2250 * the slab may touch the object after free or before allocation
2251 * then we should never poison the object itself.
2253 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2255 s->flags |= __OBJECT_POISON;
2257 s->flags &= ~__OBJECT_POISON;
2261 * If we are Redzoning then check if there is some space between the
2262 * end of the object and the free pointer. If not then add an
2263 * additional word to have some bytes to store Redzone information.
2265 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2266 size += sizeof(void *);
2270 * With that we have determined the number of bytes in actual use
2271 * by the object. This is the potential offset to the free pointer.
2275 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2278 * Relocate free pointer after the object if it is not
2279 * permitted to overwrite the first word of the object on
2282 * This is the case if we do RCU, have a constructor or
2283 * destructor or are poisoning the objects.
2286 size += sizeof(void *);
2289 #ifdef CONFIG_SLUB_DEBUG
2290 if (flags & SLAB_STORE_USER)
2292 * Need to store information about allocs and frees after
2295 size += 2 * sizeof(struct track);
2297 if (flags & SLAB_RED_ZONE)
2299 * Add some empty padding so that we can catch
2300 * overwrites from earlier objects rather than let
2301 * tracking information or the free pointer be
2302 * corrupted if a user writes before the start
2305 size += sizeof(void *);
2309 * Determine the alignment based on various parameters that the
2310 * user specified and the dynamic determination of cache line size
2313 align = calculate_alignment(flags, align, s->objsize);
2317 * SLUB stores one object immediately after another beginning from
2318 * offset 0. In order to align the objects we have to simply size
2319 * each object to conform to the alignment.
2321 size = ALIGN(size, align);
2323 if (forced_order >= 0)
2324 order = forced_order;
2326 order = calculate_order(size);
2333 s->allocflags |= __GFP_COMP;
2335 if (s->flags & SLAB_CACHE_DMA)
2336 s->allocflags |= SLUB_DMA;
2338 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2339 s->allocflags |= __GFP_RECLAIMABLE;
2342 * Determine the number of objects per slab
2344 s->oo = oo_make(order, size);
2345 s->min = oo_make(get_order(size), size);
2346 if (oo_objects(s->oo) > oo_objects(s->max))
2349 return !!oo_objects(s->oo);
2353 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2354 const char *name, size_t size,
2355 size_t align, unsigned long flags,
2356 void (*ctor)(void *))
2358 memset(s, 0, kmem_size);
2363 s->flags = kmem_cache_flags(size, flags, name, ctor);
2365 if (!calculate_sizes(s, -1))
2367 if (disable_higher_order_debug) {
2369 * Disable debugging flags that store metadata if the min slab
2372 if (get_order(s->size) > get_order(s->objsize)) {
2373 s->flags &= ~DEBUG_METADATA_FLAGS;
2375 if (!calculate_sizes(s, -1))
2381 * The larger the object size is, the more pages we want on the partial
2382 * list to avoid pounding the page allocator excessively.
2384 set_min_partial(s, ilog2(s->size));
2387 s->remote_node_defrag_ratio = 1000;
2389 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2392 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2394 free_kmem_cache_nodes(s);
2396 if (flags & SLAB_PANIC)
2397 panic("Cannot create slab %s size=%lu realsize=%u "
2398 "order=%u offset=%u flags=%lx\n",
2399 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2405 * Check if a given pointer is valid
2407 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2411 page = get_object_page(object);
2413 if (!page || s != page->slab)
2414 /* No slab or wrong slab */
2417 if (!check_valid_pointer(s, page, object))
2421 * We could also check if the object is on the slabs freelist.
2422 * But this would be too expensive and it seems that the main
2423 * purpose of kmem_ptr_valid() is to check if the object belongs
2424 * to a certain slab.
2428 EXPORT_SYMBOL(kmem_ptr_validate);
2431 * Determine the size of a slab object
2433 unsigned int kmem_cache_size(struct kmem_cache *s)
2437 EXPORT_SYMBOL(kmem_cache_size);
2439 const char *kmem_cache_name(struct kmem_cache *s)
2443 EXPORT_SYMBOL(kmem_cache_name);
2445 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2448 #ifdef CONFIG_SLUB_DEBUG
2449 void *addr = page_address(page);
2451 DECLARE_BITMAP(map, page->objects);
2453 bitmap_zero(map, page->objects);
2454 slab_err(s, page, "%s", text);
2456 for_each_free_object(p, s, page->freelist)
2457 set_bit(slab_index(p, s, addr), map);
2459 for_each_object(p, s, addr, page->objects) {
2461 if (!test_bit(slab_index(p, s, addr), map)) {
2462 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2464 print_tracking(s, p);
2472 * Attempt to free all partial slabs on a node.
2474 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2476 unsigned long flags;
2477 struct page *page, *h;
2479 spin_lock_irqsave(&n->list_lock, flags);
2480 list_for_each_entry_safe(page, h, &n->partial, lru) {
2482 list_del(&page->lru);
2483 discard_slab(s, page);
2486 list_slab_objects(s, page,
2487 "Objects remaining on kmem_cache_close()");
2490 spin_unlock_irqrestore(&n->list_lock, flags);
2494 * Release all resources used by a slab cache.
2496 static inline int kmem_cache_close(struct kmem_cache *s)
2501 free_percpu(s->cpu_slab);
2502 /* Attempt to free all objects */
2503 for_each_node_state(node, N_NORMAL_MEMORY) {
2504 struct kmem_cache_node *n = get_node(s, node);
2507 if (n->nr_partial || slabs_node(s, node))
2510 free_kmem_cache_nodes(s);
2515 * Close a cache and release the kmem_cache structure
2516 * (must be used for caches created using kmem_cache_create)
2518 void kmem_cache_destroy(struct kmem_cache *s)
2520 down_write(&slub_lock);
2524 up_write(&slub_lock);
2525 if (kmem_cache_close(s)) {
2526 printk(KERN_ERR "SLUB %s: %s called for cache that "
2527 "still has objects.\n", s->name, __func__);
2530 if (s->flags & SLAB_DESTROY_BY_RCU)
2532 sysfs_slab_remove(s);
2534 up_write(&slub_lock);
2536 EXPORT_SYMBOL(kmem_cache_destroy);
2538 /********************************************************************
2540 *******************************************************************/
2542 struct kmem_cache kmalloc_caches[KMALLOC_CACHES] __cacheline_aligned;
2543 EXPORT_SYMBOL(kmalloc_caches);
2545 static int __init setup_slub_min_order(char *str)
2547 get_option(&str, &slub_min_order);
2552 __setup("slub_min_order=", setup_slub_min_order);
2554 static int __init setup_slub_max_order(char *str)
2556 get_option(&str, &slub_max_order);
2557 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2562 __setup("slub_max_order=", setup_slub_max_order);
2564 static int __init setup_slub_min_objects(char *str)
2566 get_option(&str, &slub_min_objects);
2571 __setup("slub_min_objects=", setup_slub_min_objects);
2573 static int __init setup_slub_nomerge(char *str)
2579 __setup("slub_nomerge", setup_slub_nomerge);
2581 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2582 const char *name, int size, gfp_t gfp_flags)
2584 unsigned int flags = 0;
2586 if (gfp_flags & SLUB_DMA)
2587 flags = SLAB_CACHE_DMA;
2590 * This function is called with IRQs disabled during early-boot on
2591 * single CPU so there's no need to take slub_lock here.
2593 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2597 list_add(&s->list, &slab_caches);
2599 if (sysfs_slab_add(s))
2604 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2607 #ifdef CONFIG_ZONE_DMA
2608 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2610 static void sysfs_add_func(struct work_struct *w)
2612 struct kmem_cache *s;
2614 down_write(&slub_lock);
2615 list_for_each_entry(s, &slab_caches, list) {
2616 if (s->flags & __SYSFS_ADD_DEFERRED) {
2617 s->flags &= ~__SYSFS_ADD_DEFERRED;
2621 up_write(&slub_lock);
2624 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2626 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2628 struct kmem_cache *s;
2631 unsigned long slabflags;
2634 s = kmalloc_caches_dma[index];
2638 /* Dynamically create dma cache */
2639 if (flags & __GFP_WAIT)
2640 down_write(&slub_lock);
2642 if (!down_write_trylock(&slub_lock))
2646 if (kmalloc_caches_dma[index])
2649 realsize = kmalloc_caches[index].objsize;
2650 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2651 (unsigned int)realsize);
2654 for (i = 0; i < KMALLOC_CACHES; i++)
2655 if (!kmalloc_caches[i].size)
2658 BUG_ON(i >= KMALLOC_CACHES);
2659 s = kmalloc_caches + i;
2662 * Must defer sysfs creation to a workqueue because we don't know
2663 * what context we are called from. Before sysfs comes up, we don't
2664 * need to do anything because our sysfs initcall will start by
2665 * adding all existing slabs to sysfs.
2667 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
2668 if (slab_state >= SYSFS)
2669 slabflags |= __SYSFS_ADD_DEFERRED;
2671 if (!s || !text || !kmem_cache_open(s, flags, text,
2672 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
2678 list_add(&s->list, &slab_caches);
2679 kmalloc_caches_dma[index] = s;
2681 if (slab_state >= SYSFS)
2682 schedule_work(&sysfs_add_work);
2685 up_write(&slub_lock);
2687 return kmalloc_caches_dma[index];
2692 * Conversion table for small slabs sizes / 8 to the index in the
2693 * kmalloc array. This is necessary for slabs < 192 since we have non power
2694 * of two cache sizes there. The size of larger slabs can be determined using
2697 static s8 size_index[24] = {
2724 static inline int size_index_elem(size_t bytes)
2726 return (bytes - 1) / 8;
2729 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2735 return ZERO_SIZE_PTR;
2737 index = size_index[size_index_elem(size)];
2739 index = fls(size - 1);
2741 #ifdef CONFIG_ZONE_DMA
2742 if (unlikely((flags & SLUB_DMA)))
2743 return dma_kmalloc_cache(index, flags);
2746 return &kmalloc_caches[index];
2749 void *__kmalloc(size_t size, gfp_t flags)
2751 struct kmem_cache *s;
2754 if (unlikely(size > SLUB_MAX_SIZE))
2755 return kmalloc_large(size, flags);
2757 s = get_slab(size, flags);
2759 if (unlikely(ZERO_OR_NULL_PTR(s)))
2762 ret = slab_alloc(s, flags, -1, _RET_IP_);
2764 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2768 EXPORT_SYMBOL(__kmalloc);
2770 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2775 flags |= __GFP_COMP | __GFP_NOTRACK;
2776 page = alloc_pages_node(node, flags, get_order(size));
2778 ptr = page_address(page);
2780 kmemleak_alloc(ptr, size, 1, flags);
2785 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2787 struct kmem_cache *s;
2790 if (unlikely(size > SLUB_MAX_SIZE)) {
2791 ret = kmalloc_large_node(size, flags, node);
2793 trace_kmalloc_node(_RET_IP_, ret,
2794 size, PAGE_SIZE << get_order(size),
2800 s = get_slab(size, flags);
2802 if (unlikely(ZERO_OR_NULL_PTR(s)))
2805 ret = slab_alloc(s, flags, node, _RET_IP_);
2807 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2811 EXPORT_SYMBOL(__kmalloc_node);
2814 size_t ksize(const void *object)
2817 struct kmem_cache *s;
2819 if (unlikely(object == ZERO_SIZE_PTR))
2822 page = virt_to_head_page(object);
2824 if (unlikely(!PageSlab(page))) {
2825 WARN_ON(!PageCompound(page));
2826 return PAGE_SIZE << compound_order(page);
2830 #ifdef CONFIG_SLUB_DEBUG
2832 * Debugging requires use of the padding between object
2833 * and whatever may come after it.
2835 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2840 * If we have the need to store the freelist pointer
2841 * back there or track user information then we can
2842 * only use the space before that information.
2844 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2847 * Else we can use all the padding etc for the allocation
2851 EXPORT_SYMBOL(ksize);
2853 void kfree(const void *x)
2856 void *object = (void *)x;
2858 trace_kfree(_RET_IP_, x);
2860 if (unlikely(ZERO_OR_NULL_PTR(x)))
2863 page = virt_to_head_page(x);
2864 if (unlikely(!PageSlab(page))) {
2865 BUG_ON(!PageCompound(page));
2870 slab_free(page->slab, page, object, _RET_IP_);
2872 EXPORT_SYMBOL(kfree);
2875 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2876 * the remaining slabs by the number of items in use. The slabs with the
2877 * most items in use come first. New allocations will then fill those up
2878 * and thus they can be removed from the partial lists.
2880 * The slabs with the least items are placed last. This results in them
2881 * being allocated from last increasing the chance that the last objects
2882 * are freed in them.
2884 int kmem_cache_shrink(struct kmem_cache *s)
2888 struct kmem_cache_node *n;
2891 int objects = oo_objects(s->max);
2892 struct list_head *slabs_by_inuse =
2893 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2894 unsigned long flags;
2896 if (!slabs_by_inuse)
2900 for_each_node_state(node, N_NORMAL_MEMORY) {
2901 n = get_node(s, node);
2906 for (i = 0; i < objects; i++)
2907 INIT_LIST_HEAD(slabs_by_inuse + i);
2909 spin_lock_irqsave(&n->list_lock, flags);
2912 * Build lists indexed by the items in use in each slab.
2914 * Note that concurrent frees may occur while we hold the
2915 * list_lock. page->inuse here is the upper limit.
2917 list_for_each_entry_safe(page, t, &n->partial, lru) {
2918 if (!page->inuse && slab_trylock(page)) {
2920 * Must hold slab lock here because slab_free
2921 * may have freed the last object and be
2922 * waiting to release the slab.
2924 list_del(&page->lru);
2927 discard_slab(s, page);
2929 list_move(&page->lru,
2930 slabs_by_inuse + page->inuse);
2935 * Rebuild the partial list with the slabs filled up most
2936 * first and the least used slabs at the end.
2938 for (i = objects - 1; i >= 0; i--)
2939 list_splice(slabs_by_inuse + i, n->partial.prev);
2941 spin_unlock_irqrestore(&n->list_lock, flags);
2944 kfree(slabs_by_inuse);
2947 EXPORT_SYMBOL(kmem_cache_shrink);
2949 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2950 static int slab_mem_going_offline_callback(void *arg)
2952 struct kmem_cache *s;
2954 down_read(&slub_lock);
2955 list_for_each_entry(s, &slab_caches, list)
2956 kmem_cache_shrink(s);
2957 up_read(&slub_lock);
2962 static void slab_mem_offline_callback(void *arg)
2964 struct kmem_cache_node *n;
2965 struct kmem_cache *s;
2966 struct memory_notify *marg = arg;
2969 offline_node = marg->status_change_nid;
2972 * If the node still has available memory. we need kmem_cache_node
2975 if (offline_node < 0)
2978 down_read(&slub_lock);
2979 list_for_each_entry(s, &slab_caches, list) {
2980 n = get_node(s, offline_node);
2983 * if n->nr_slabs > 0, slabs still exist on the node
2984 * that is going down. We were unable to free them,
2985 * and offline_pages() function shoudn't call this
2986 * callback. So, we must fail.
2988 BUG_ON(slabs_node(s, offline_node));
2990 s->node[offline_node] = NULL;
2991 kmem_cache_free(kmalloc_caches, n);
2994 up_read(&slub_lock);
2997 static int slab_mem_going_online_callback(void *arg)
2999 struct kmem_cache_node *n;
3000 struct kmem_cache *s;
3001 struct memory_notify *marg = arg;
3002 int nid = marg->status_change_nid;
3006 * If the node's memory is already available, then kmem_cache_node is
3007 * already created. Nothing to do.
3013 * We are bringing a node online. No memory is available yet. We must
3014 * allocate a kmem_cache_node structure in order to bring the node
3017 down_read(&slub_lock);
3018 list_for_each_entry(s, &slab_caches, list) {
3020 * XXX: kmem_cache_alloc_node will fallback to other nodes
3021 * since memory is not yet available from the node that
3024 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3029 init_kmem_cache_node(n, s);
3033 up_read(&slub_lock);
3037 static int slab_memory_callback(struct notifier_block *self,
3038 unsigned long action, void *arg)
3043 case MEM_GOING_ONLINE:
3044 ret = slab_mem_going_online_callback(arg);
3046 case MEM_GOING_OFFLINE:
3047 ret = slab_mem_going_offline_callback(arg);
3050 case MEM_CANCEL_ONLINE:
3051 slab_mem_offline_callback(arg);
3054 case MEM_CANCEL_OFFLINE:
3058 ret = notifier_from_errno(ret);
3064 #endif /* CONFIG_MEMORY_HOTPLUG */
3066 /********************************************************************
3067 * Basic setup of slabs
3068 *******************************************************************/
3070 void __init kmem_cache_init(void)
3077 * Must first have the slab cache available for the allocations of the
3078 * struct kmem_cache_node's. There is special bootstrap code in
3079 * kmem_cache_open for slab_state == DOWN.
3081 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3082 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3083 kmalloc_caches[0].refcount = -1;
3086 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3089 /* Able to allocate the per node structures */
3090 slab_state = PARTIAL;
3092 /* Caches that are not of the two-to-the-power-of size */
3093 if (KMALLOC_MIN_SIZE <= 32) {
3094 create_kmalloc_cache(&kmalloc_caches[1],
3095 "kmalloc-96", 96, GFP_NOWAIT);
3098 if (KMALLOC_MIN_SIZE <= 64) {
3099 create_kmalloc_cache(&kmalloc_caches[2],
3100 "kmalloc-192", 192, GFP_NOWAIT);
3104 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3105 create_kmalloc_cache(&kmalloc_caches[i],
3106 "kmalloc", 1 << i, GFP_NOWAIT);
3112 * Patch up the size_index table if we have strange large alignment
3113 * requirements for the kmalloc array. This is only the case for
3114 * MIPS it seems. The standard arches will not generate any code here.
3116 * Largest permitted alignment is 256 bytes due to the way we
3117 * handle the index determination for the smaller caches.
3119 * Make sure that nothing crazy happens if someone starts tinkering
3120 * around with ARCH_KMALLOC_MINALIGN
3122 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3123 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3125 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3126 int elem = size_index_elem(i);
3127 if (elem >= ARRAY_SIZE(size_index))
3129 size_index[elem] = KMALLOC_SHIFT_LOW;
3132 if (KMALLOC_MIN_SIZE == 64) {
3134 * The 96 byte size cache is not used if the alignment
3137 for (i = 64 + 8; i <= 96; i += 8)
3138 size_index[size_index_elem(i)] = 7;
3139 } else if (KMALLOC_MIN_SIZE == 128) {
3141 * The 192 byte sized cache is not used if the alignment
3142 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3145 for (i = 128 + 8; i <= 192; i += 8)
3146 size_index[size_index_elem(i)] = 8;
3151 /* Provide the correct kmalloc names now that the caches are up */
3152 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3153 kmalloc_caches[i]. name =
3154 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3157 register_cpu_notifier(&slab_notifier);
3160 kmem_size = offsetof(struct kmem_cache, node) +
3161 nr_node_ids * sizeof(struct kmem_cache_node *);
3163 kmem_size = sizeof(struct kmem_cache);
3167 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3168 " CPUs=%d, Nodes=%d\n",
3169 caches, cache_line_size(),
3170 slub_min_order, slub_max_order, slub_min_objects,
3171 nr_cpu_ids, nr_node_ids);
3174 void __init kmem_cache_init_late(void)
3179 * Find a mergeable slab cache
3181 static int slab_unmergeable(struct kmem_cache *s)
3183 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3190 * We may have set a slab to be unmergeable during bootstrap.
3192 if (s->refcount < 0)
3198 static struct kmem_cache *find_mergeable(size_t size,
3199 size_t align, unsigned long flags, const char *name,
3200 void (*ctor)(void *))
3202 struct kmem_cache *s;
3204 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3210 size = ALIGN(size, sizeof(void *));
3211 align = calculate_alignment(flags, align, size);
3212 size = ALIGN(size, align);
3213 flags = kmem_cache_flags(size, flags, name, NULL);
3215 list_for_each_entry(s, &slab_caches, list) {
3216 if (slab_unmergeable(s))
3222 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3225 * Check if alignment is compatible.
3226 * Courtesy of Adrian Drzewiecki
3228 if ((s->size & ~(align - 1)) != s->size)
3231 if (s->size - size >= sizeof(void *))
3239 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3240 size_t align, unsigned long flags, void (*ctor)(void *))
3242 struct kmem_cache *s;
3247 down_write(&slub_lock);
3248 s = find_mergeable(size, align, flags, name, ctor);
3254 * Adjust the object sizes so that we clear
3255 * the complete object on kzalloc.
3257 s->objsize = max(s->objsize, (int)size);
3260 * And then we need to update the object size in the
3261 * per cpu structures
3263 for_each_online_cpu(cpu)
3264 per_cpu_ptr(s->cpu_slab, cpu)->objsize = s->objsize;
3266 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3267 up_write(&slub_lock);
3269 if (sysfs_slab_alias(s, name)) {
3270 down_write(&slub_lock);
3272 up_write(&slub_lock);
3278 s = kmalloc(kmem_size, GFP_KERNEL);
3280 if (kmem_cache_open(s, GFP_KERNEL, name,
3281 size, align, flags, ctor)) {
3282 list_add(&s->list, &slab_caches);
3283 up_write(&slub_lock);
3284 if (sysfs_slab_add(s)) {
3285 down_write(&slub_lock);
3287 up_write(&slub_lock);
3295 up_write(&slub_lock);
3298 if (flags & SLAB_PANIC)
3299 panic("Cannot create slabcache %s\n", name);
3304 EXPORT_SYMBOL(kmem_cache_create);
3308 * Use the cpu notifier to insure that the cpu slabs are flushed when
3311 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3312 unsigned long action, void *hcpu)
3314 long cpu = (long)hcpu;
3315 struct kmem_cache *s;
3316 unsigned long flags;
3319 case CPU_UP_PREPARE:
3320 case CPU_UP_PREPARE_FROZEN:
3321 down_read(&slub_lock);
3322 list_for_each_entry(s, &slab_caches, list)
3323 init_kmem_cache_cpu(s, per_cpu_ptr(s->cpu_slab, cpu));
3324 up_read(&slub_lock);
3327 case CPU_UP_CANCELED:
3328 case CPU_UP_CANCELED_FROZEN:
3330 case CPU_DEAD_FROZEN:
3331 down_read(&slub_lock);
3332 list_for_each_entry(s, &slab_caches, list) {
3333 local_irq_save(flags);
3334 __flush_cpu_slab(s, cpu);
3335 local_irq_restore(flags);
3337 up_read(&slub_lock);
3345 static struct notifier_block __cpuinitdata slab_notifier = {
3346 .notifier_call = slab_cpuup_callback
3351 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3353 struct kmem_cache *s;
3356 if (unlikely(size > SLUB_MAX_SIZE))
3357 return kmalloc_large(size, gfpflags);
3359 s = get_slab(size, gfpflags);
3361 if (unlikely(ZERO_OR_NULL_PTR(s)))
3364 ret = slab_alloc(s, gfpflags, -1, caller);
3366 /* Honor the call site pointer we recieved. */
3367 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3372 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3373 int node, unsigned long caller)
3375 struct kmem_cache *s;
3378 if (unlikely(size > SLUB_MAX_SIZE))
3379 return kmalloc_large_node(size, gfpflags, node);
3381 s = get_slab(size, gfpflags);
3383 if (unlikely(ZERO_OR_NULL_PTR(s)))
3386 ret = slab_alloc(s, gfpflags, node, caller);
3388 /* Honor the call site pointer we recieved. */
3389 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3394 #ifdef CONFIG_SLUB_DEBUG
3395 static int count_inuse(struct page *page)
3400 static int count_total(struct page *page)
3402 return page->objects;
3405 static int validate_slab(struct kmem_cache *s, struct page *page,
3409 void *addr = page_address(page);
3411 if (!check_slab(s, page) ||
3412 !on_freelist(s, page, NULL))
3415 /* Now we know that a valid freelist exists */
3416 bitmap_zero(map, page->objects);
3418 for_each_free_object(p, s, page->freelist) {
3419 set_bit(slab_index(p, s, addr), map);
3420 if (!check_object(s, page, p, 0))
3424 for_each_object(p, s, addr, page->objects)
3425 if (!test_bit(slab_index(p, s, addr), map))
3426 if (!check_object(s, page, p, 1))
3431 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3434 if (slab_trylock(page)) {
3435 validate_slab(s, page, map);
3438 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3441 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3442 if (!PageSlubDebug(page))
3443 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3444 "on slab 0x%p\n", s->name, page);
3446 if (PageSlubDebug(page))
3447 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3448 "slab 0x%p\n", s->name, page);
3452 static int validate_slab_node(struct kmem_cache *s,
3453 struct kmem_cache_node *n, unsigned long *map)
3455 unsigned long count = 0;
3457 unsigned long flags;
3459 spin_lock_irqsave(&n->list_lock, flags);
3461 list_for_each_entry(page, &n->partial, lru) {
3462 validate_slab_slab(s, page, map);
3465 if (count != n->nr_partial)
3466 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3467 "counter=%ld\n", s->name, count, n->nr_partial);
3469 if (!(s->flags & SLAB_STORE_USER))
3472 list_for_each_entry(page, &n->full, lru) {
3473 validate_slab_slab(s, page, map);
3476 if (count != atomic_long_read(&n->nr_slabs))
3477 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3478 "counter=%ld\n", s->name, count,
3479 atomic_long_read(&n->nr_slabs));
3482 spin_unlock_irqrestore(&n->list_lock, flags);
3486 static long validate_slab_cache(struct kmem_cache *s)
3489 unsigned long count = 0;
3490 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3491 sizeof(unsigned long), GFP_KERNEL);
3497 for_each_node_state(node, N_NORMAL_MEMORY) {
3498 struct kmem_cache_node *n = get_node(s, node);
3500 count += validate_slab_node(s, n, map);
3506 #ifdef SLUB_RESILIENCY_TEST
3507 static void resiliency_test(void)
3511 printk(KERN_ERR "SLUB resiliency testing\n");
3512 printk(KERN_ERR "-----------------------\n");
3513 printk(KERN_ERR "A. Corruption after allocation\n");
3515 p = kzalloc(16, GFP_KERNEL);
3517 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3518 " 0x12->0x%p\n\n", p + 16);
3520 validate_slab_cache(kmalloc_caches + 4);
3522 /* Hmmm... The next two are dangerous */
3523 p = kzalloc(32, GFP_KERNEL);
3524 p[32 + sizeof(void *)] = 0x34;
3525 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3526 " 0x34 -> -0x%p\n", p);
3528 "If allocated object is overwritten then not detectable\n\n");
3530 validate_slab_cache(kmalloc_caches + 5);
3531 p = kzalloc(64, GFP_KERNEL);
3532 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3534 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3537 "If allocated object is overwritten then not detectable\n\n");
3538 validate_slab_cache(kmalloc_caches + 6);
3540 printk(KERN_ERR "\nB. Corruption after free\n");
3541 p = kzalloc(128, GFP_KERNEL);
3544 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3545 validate_slab_cache(kmalloc_caches + 7);
3547 p = kzalloc(256, GFP_KERNEL);
3550 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3552 validate_slab_cache(kmalloc_caches + 8);
3554 p = kzalloc(512, GFP_KERNEL);
3557 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3558 validate_slab_cache(kmalloc_caches + 9);
3561 static void resiliency_test(void) {};
3565 * Generate lists of code addresses where slabcache objects are allocated
3570 unsigned long count;
3577 DECLARE_BITMAP(cpus, NR_CPUS);
3583 unsigned long count;
3584 struct location *loc;
3587 static void free_loc_track(struct loc_track *t)
3590 free_pages((unsigned long)t->loc,
3591 get_order(sizeof(struct location) * t->max));
3594 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3599 order = get_order(sizeof(struct location) * max);
3601 l = (void *)__get_free_pages(flags, order);
3606 memcpy(l, t->loc, sizeof(struct location) * t->count);
3614 static int add_location(struct loc_track *t, struct kmem_cache *s,
3615 const struct track *track)
3617 long start, end, pos;
3619 unsigned long caddr;
3620 unsigned long age = jiffies - track->when;
3626 pos = start + (end - start + 1) / 2;
3629 * There is nothing at "end". If we end up there
3630 * we need to add something to before end.
3635 caddr = t->loc[pos].addr;
3636 if (track->addr == caddr) {
3642 if (age < l->min_time)
3644 if (age > l->max_time)
3647 if (track->pid < l->min_pid)
3648 l->min_pid = track->pid;
3649 if (track->pid > l->max_pid)
3650 l->max_pid = track->pid;
3652 cpumask_set_cpu(track->cpu,
3653 to_cpumask(l->cpus));
3655 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3659 if (track->addr < caddr)
3666 * Not found. Insert new tracking element.
3668 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3674 (t->count - pos) * sizeof(struct location));
3677 l->addr = track->addr;
3681 l->min_pid = track->pid;
3682 l->max_pid = track->pid;
3683 cpumask_clear(to_cpumask(l->cpus));
3684 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3685 nodes_clear(l->nodes);
3686 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3690 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3691 struct page *page, enum track_item alloc)
3693 void *addr = page_address(page);
3694 DECLARE_BITMAP(map, page->objects);
3697 bitmap_zero(map, page->objects);
3698 for_each_free_object(p, s, page->freelist)
3699 set_bit(slab_index(p, s, addr), map);
3701 for_each_object(p, s, addr, page->objects)
3702 if (!test_bit(slab_index(p, s, addr), map))
3703 add_location(t, s, get_track(s, p, alloc));
3706 static int list_locations(struct kmem_cache *s, char *buf,
3707 enum track_item alloc)
3711 struct loc_track t = { 0, 0, NULL };
3714 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3716 return sprintf(buf, "Out of memory\n");
3718 /* Push back cpu slabs */
3721 for_each_node_state(node, N_NORMAL_MEMORY) {
3722 struct kmem_cache_node *n = get_node(s, node);
3723 unsigned long flags;
3726 if (!atomic_long_read(&n->nr_slabs))
3729 spin_lock_irqsave(&n->list_lock, flags);
3730 list_for_each_entry(page, &n->partial, lru)
3731 process_slab(&t, s, page, alloc);
3732 list_for_each_entry(page, &n->full, lru)
3733 process_slab(&t, s, page, alloc);
3734 spin_unlock_irqrestore(&n->list_lock, flags);
3737 for (i = 0; i < t.count; i++) {
3738 struct location *l = &t.loc[i];
3740 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3742 len += sprintf(buf + len, "%7ld ", l->count);
3745 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3747 len += sprintf(buf + len, "<not-available>");
3749 if (l->sum_time != l->min_time) {
3750 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3752 (long)div_u64(l->sum_time, l->count),
3755 len += sprintf(buf + len, " age=%ld",
3758 if (l->min_pid != l->max_pid)
3759 len += sprintf(buf + len, " pid=%ld-%ld",
3760 l->min_pid, l->max_pid);
3762 len += sprintf(buf + len, " pid=%ld",
3765 if (num_online_cpus() > 1 &&
3766 !cpumask_empty(to_cpumask(l->cpus)) &&
3767 len < PAGE_SIZE - 60) {
3768 len += sprintf(buf + len, " cpus=");
3769 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3770 to_cpumask(l->cpus));
3773 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3774 len < PAGE_SIZE - 60) {
3775 len += sprintf(buf + len, " nodes=");
3776 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3780 len += sprintf(buf + len, "\n");
3785 len += sprintf(buf, "No data\n");
3789 enum slab_stat_type {
3790 SL_ALL, /* All slabs */
3791 SL_PARTIAL, /* Only partially allocated slabs */
3792 SL_CPU, /* Only slabs used for cpu caches */
3793 SL_OBJECTS, /* Determine allocated objects not slabs */
3794 SL_TOTAL /* Determine object capacity not slabs */
3797 #define SO_ALL (1 << SL_ALL)
3798 #define SO_PARTIAL (1 << SL_PARTIAL)
3799 #define SO_CPU (1 << SL_CPU)
3800 #define SO_OBJECTS (1 << SL_OBJECTS)
3801 #define SO_TOTAL (1 << SL_TOTAL)
3803 static ssize_t show_slab_objects(struct kmem_cache *s,
3804 char *buf, unsigned long flags)
3806 unsigned long total = 0;
3809 unsigned long *nodes;
3810 unsigned long *per_cpu;
3812 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3815 per_cpu = nodes + nr_node_ids;
3817 if (flags & SO_CPU) {
3820 for_each_possible_cpu(cpu) {
3821 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3823 if (!c || c->node < 0)
3827 if (flags & SO_TOTAL)
3828 x = c->page->objects;
3829 else if (flags & SO_OBJECTS)
3835 nodes[c->node] += x;
3841 if (flags & SO_ALL) {
3842 for_each_node_state(node, N_NORMAL_MEMORY) {
3843 struct kmem_cache_node *n = get_node(s, node);
3845 if (flags & SO_TOTAL)
3846 x = atomic_long_read(&n->total_objects);
3847 else if (flags & SO_OBJECTS)
3848 x = atomic_long_read(&n->total_objects) -
3849 count_partial(n, count_free);
3852 x = atomic_long_read(&n->nr_slabs);
3857 } else if (flags & SO_PARTIAL) {
3858 for_each_node_state(node, N_NORMAL_MEMORY) {
3859 struct kmem_cache_node *n = get_node(s, node);
3861 if (flags & SO_TOTAL)
3862 x = count_partial(n, count_total);
3863 else if (flags & SO_OBJECTS)
3864 x = count_partial(n, count_inuse);
3871 x = sprintf(buf, "%lu", total);
3873 for_each_node_state(node, N_NORMAL_MEMORY)
3875 x += sprintf(buf + x, " N%d=%lu",
3879 return x + sprintf(buf + x, "\n");
3882 static int any_slab_objects(struct kmem_cache *s)
3886 for_each_online_node(node) {
3887 struct kmem_cache_node *n = get_node(s, node);
3892 if (atomic_long_read(&n->total_objects))
3898 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3899 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3901 struct slab_attribute {
3902 struct attribute attr;
3903 ssize_t (*show)(struct kmem_cache *s, char *buf);
3904 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3907 #define SLAB_ATTR_RO(_name) \
3908 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3910 #define SLAB_ATTR(_name) \
3911 static struct slab_attribute _name##_attr = \
3912 __ATTR(_name, 0644, _name##_show, _name##_store)
3914 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3916 return sprintf(buf, "%d\n", s->size);
3918 SLAB_ATTR_RO(slab_size);
3920 static ssize_t align_show(struct kmem_cache *s, char *buf)
3922 return sprintf(buf, "%d\n", s->align);
3924 SLAB_ATTR_RO(align);
3926 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3928 return sprintf(buf, "%d\n", s->objsize);
3930 SLAB_ATTR_RO(object_size);
3932 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3934 return sprintf(buf, "%d\n", oo_objects(s->oo));
3936 SLAB_ATTR_RO(objs_per_slab);
3938 static ssize_t order_store(struct kmem_cache *s,
3939 const char *buf, size_t length)
3941 unsigned long order;
3944 err = strict_strtoul(buf, 10, &order);
3948 if (order > slub_max_order || order < slub_min_order)
3951 calculate_sizes(s, order);
3955 static ssize_t order_show(struct kmem_cache *s, char *buf)
3957 return sprintf(buf, "%d\n", oo_order(s->oo));
3961 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3963 return sprintf(buf, "%lu\n", s->min_partial);
3966 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3972 err = strict_strtoul(buf, 10, &min);
3976 set_min_partial(s, min);
3979 SLAB_ATTR(min_partial);
3981 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3984 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3986 return n + sprintf(buf + n, "\n");
3992 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3994 return sprintf(buf, "%d\n", s->refcount - 1);
3996 SLAB_ATTR_RO(aliases);
3998 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4000 return show_slab_objects(s, buf, SO_ALL);
4002 SLAB_ATTR_RO(slabs);
4004 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4006 return show_slab_objects(s, buf, SO_PARTIAL);
4008 SLAB_ATTR_RO(partial);
4010 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4012 return show_slab_objects(s, buf, SO_CPU);
4014 SLAB_ATTR_RO(cpu_slabs);
4016 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4018 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4020 SLAB_ATTR_RO(objects);
4022 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4024 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4026 SLAB_ATTR_RO(objects_partial);
4028 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4030 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4032 SLAB_ATTR_RO(total_objects);
4034 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4036 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4039 static ssize_t sanity_checks_store(struct kmem_cache *s,
4040 const char *buf, size_t length)
4042 s->flags &= ~SLAB_DEBUG_FREE;
4044 s->flags |= SLAB_DEBUG_FREE;
4047 SLAB_ATTR(sanity_checks);
4049 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4051 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4054 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4057 s->flags &= ~SLAB_TRACE;
4059 s->flags |= SLAB_TRACE;
4064 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4066 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4069 static ssize_t reclaim_account_store(struct kmem_cache *s,
4070 const char *buf, size_t length)
4072 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4074 s->flags |= SLAB_RECLAIM_ACCOUNT;
4077 SLAB_ATTR(reclaim_account);
4079 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4081 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4083 SLAB_ATTR_RO(hwcache_align);
4085 #ifdef CONFIG_ZONE_DMA
4086 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4088 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4090 SLAB_ATTR_RO(cache_dma);
4093 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4095 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4097 SLAB_ATTR_RO(destroy_by_rcu);
4099 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4101 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4104 static ssize_t red_zone_store(struct kmem_cache *s,
4105 const char *buf, size_t length)
4107 if (any_slab_objects(s))
4110 s->flags &= ~SLAB_RED_ZONE;
4112 s->flags |= SLAB_RED_ZONE;
4113 calculate_sizes(s, -1);
4116 SLAB_ATTR(red_zone);
4118 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4123 static ssize_t poison_store(struct kmem_cache *s,
4124 const char *buf, size_t length)
4126 if (any_slab_objects(s))
4129 s->flags &= ~SLAB_POISON;
4131 s->flags |= SLAB_POISON;
4132 calculate_sizes(s, -1);
4137 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4139 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4142 static ssize_t store_user_store(struct kmem_cache *s,
4143 const char *buf, size_t length)
4145 if (any_slab_objects(s))
4148 s->flags &= ~SLAB_STORE_USER;
4150 s->flags |= SLAB_STORE_USER;
4151 calculate_sizes(s, -1);
4154 SLAB_ATTR(store_user);
4156 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4161 static ssize_t validate_store(struct kmem_cache *s,
4162 const char *buf, size_t length)
4166 if (buf[0] == '1') {
4167 ret = validate_slab_cache(s);
4173 SLAB_ATTR(validate);
4175 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4180 static ssize_t shrink_store(struct kmem_cache *s,
4181 const char *buf, size_t length)
4183 if (buf[0] == '1') {
4184 int rc = kmem_cache_shrink(s);
4194 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4196 if (!(s->flags & SLAB_STORE_USER))
4198 return list_locations(s, buf, TRACK_ALLOC);
4200 SLAB_ATTR_RO(alloc_calls);
4202 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4204 if (!(s->flags & SLAB_STORE_USER))
4206 return list_locations(s, buf, TRACK_FREE);
4208 SLAB_ATTR_RO(free_calls);
4211 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4213 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4216 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4217 const char *buf, size_t length)
4219 unsigned long ratio;
4222 err = strict_strtoul(buf, 10, &ratio);
4227 s->remote_node_defrag_ratio = ratio * 10;
4231 SLAB_ATTR(remote_node_defrag_ratio);
4234 #ifdef CONFIG_SLUB_STATS
4235 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4237 unsigned long sum = 0;
4240 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4245 for_each_online_cpu(cpu) {
4246 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4252 len = sprintf(buf, "%lu", sum);
4255 for_each_online_cpu(cpu) {
4256 if (data[cpu] && len < PAGE_SIZE - 20)
4257 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4261 return len + sprintf(buf + len, "\n");
4264 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4268 for_each_online_cpu(cpu)
4269 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4272 #define STAT_ATTR(si, text) \
4273 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4275 return show_stat(s, buf, si); \
4277 static ssize_t text##_store(struct kmem_cache *s, \
4278 const char *buf, size_t length) \
4280 if (buf[0] != '0') \
4282 clear_stat(s, si); \
4287 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4288 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4289 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4290 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4291 STAT_ATTR(FREE_FROZEN, free_frozen);
4292 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4293 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4294 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4295 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4296 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4297 STAT_ATTR(FREE_SLAB, free_slab);
4298 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4299 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4300 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4301 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4302 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4303 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4304 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4307 static struct attribute *slab_attrs[] = {
4308 &slab_size_attr.attr,
4309 &object_size_attr.attr,
4310 &objs_per_slab_attr.attr,
4312 &min_partial_attr.attr,
4314 &objects_partial_attr.attr,
4315 &total_objects_attr.attr,
4318 &cpu_slabs_attr.attr,
4322 &sanity_checks_attr.attr,
4324 &hwcache_align_attr.attr,
4325 &reclaim_account_attr.attr,
4326 &destroy_by_rcu_attr.attr,
4327 &red_zone_attr.attr,
4329 &store_user_attr.attr,
4330 &validate_attr.attr,
4332 &alloc_calls_attr.attr,
4333 &free_calls_attr.attr,
4334 #ifdef CONFIG_ZONE_DMA
4335 &cache_dma_attr.attr,
4338 &remote_node_defrag_ratio_attr.attr,
4340 #ifdef CONFIG_SLUB_STATS
4341 &alloc_fastpath_attr.attr,
4342 &alloc_slowpath_attr.attr,
4343 &free_fastpath_attr.attr,
4344 &free_slowpath_attr.attr,
4345 &free_frozen_attr.attr,
4346 &free_add_partial_attr.attr,
4347 &free_remove_partial_attr.attr,
4348 &alloc_from_partial_attr.attr,
4349 &alloc_slab_attr.attr,
4350 &alloc_refill_attr.attr,
4351 &free_slab_attr.attr,
4352 &cpuslab_flush_attr.attr,
4353 &deactivate_full_attr.attr,
4354 &deactivate_empty_attr.attr,
4355 &deactivate_to_head_attr.attr,
4356 &deactivate_to_tail_attr.attr,
4357 &deactivate_remote_frees_attr.attr,
4358 &order_fallback_attr.attr,
4363 static struct attribute_group slab_attr_group = {
4364 .attrs = slab_attrs,
4367 static ssize_t slab_attr_show(struct kobject *kobj,
4368 struct attribute *attr,
4371 struct slab_attribute *attribute;
4372 struct kmem_cache *s;
4375 attribute = to_slab_attr(attr);
4378 if (!attribute->show)
4381 err = attribute->show(s, buf);
4386 static ssize_t slab_attr_store(struct kobject *kobj,
4387 struct attribute *attr,
4388 const char *buf, size_t len)
4390 struct slab_attribute *attribute;
4391 struct kmem_cache *s;
4394 attribute = to_slab_attr(attr);
4397 if (!attribute->store)
4400 err = attribute->store(s, buf, len);
4405 static void kmem_cache_release(struct kobject *kobj)
4407 struct kmem_cache *s = to_slab(kobj);
4412 static struct sysfs_ops slab_sysfs_ops = {
4413 .show = slab_attr_show,
4414 .store = slab_attr_store,
4417 static struct kobj_type slab_ktype = {
4418 .sysfs_ops = &slab_sysfs_ops,
4419 .release = kmem_cache_release
4422 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4424 struct kobj_type *ktype = get_ktype(kobj);
4426 if (ktype == &slab_ktype)
4431 static struct kset_uevent_ops slab_uevent_ops = {
4432 .filter = uevent_filter,
4435 static struct kset *slab_kset;
4437 #define ID_STR_LENGTH 64
4439 /* Create a unique string id for a slab cache:
4441 * Format :[flags-]size
4443 static char *create_unique_id(struct kmem_cache *s)
4445 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4452 * First flags affecting slabcache operations. We will only
4453 * get here for aliasable slabs so we do not need to support
4454 * too many flags. The flags here must cover all flags that
4455 * are matched during merging to guarantee that the id is
4458 if (s->flags & SLAB_CACHE_DMA)
4460 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4462 if (s->flags & SLAB_DEBUG_FREE)
4464 if (!(s->flags & SLAB_NOTRACK))
4468 p += sprintf(p, "%07d", s->size);
4469 BUG_ON(p > name + ID_STR_LENGTH - 1);
4473 static int sysfs_slab_add(struct kmem_cache *s)
4479 if (slab_state < SYSFS)
4480 /* Defer until later */
4483 unmergeable = slab_unmergeable(s);
4486 * Slabcache can never be merged so we can use the name proper.
4487 * This is typically the case for debug situations. In that
4488 * case we can catch duplicate names easily.
4490 sysfs_remove_link(&slab_kset->kobj, s->name);
4494 * Create a unique name for the slab as a target
4497 name = create_unique_id(s);
4500 s->kobj.kset = slab_kset;
4501 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4503 kobject_put(&s->kobj);
4507 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4509 kobject_del(&s->kobj);
4510 kobject_put(&s->kobj);
4513 kobject_uevent(&s->kobj, KOBJ_ADD);
4515 /* Setup first alias */
4516 sysfs_slab_alias(s, s->name);
4522 static void sysfs_slab_remove(struct kmem_cache *s)
4524 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4525 kobject_del(&s->kobj);
4526 kobject_put(&s->kobj);
4530 * Need to buffer aliases during bootup until sysfs becomes
4531 * available lest we lose that information.
4533 struct saved_alias {
4534 struct kmem_cache *s;
4536 struct saved_alias *next;
4539 static struct saved_alias *alias_list;
4541 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4543 struct saved_alias *al;
4545 if (slab_state == SYSFS) {
4547 * If we have a leftover link then remove it.
4549 sysfs_remove_link(&slab_kset->kobj, name);
4550 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4553 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4559 al->next = alias_list;
4564 static int __init slab_sysfs_init(void)
4566 struct kmem_cache *s;
4569 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4571 printk(KERN_ERR "Cannot register slab subsystem.\n");
4577 list_for_each_entry(s, &slab_caches, list) {
4578 err = sysfs_slab_add(s);
4580 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4581 " to sysfs\n", s->name);
4584 while (alias_list) {
4585 struct saved_alias *al = alias_list;
4587 alias_list = alias_list->next;
4588 err = sysfs_slab_alias(al->s, al->name);
4590 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4591 " %s to sysfs\n", s->name);
4599 __initcall(slab_sysfs_init);
4603 * The /proc/slabinfo ABI
4605 #ifdef CONFIG_SLABINFO
4606 static void print_slabinfo_header(struct seq_file *m)
4608 seq_puts(m, "slabinfo - version: 2.1\n");
4609 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4610 "<objperslab> <pagesperslab>");
4611 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4612 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4616 static void *s_start(struct seq_file *m, loff_t *pos)
4620 down_read(&slub_lock);
4622 print_slabinfo_header(m);
4624 return seq_list_start(&slab_caches, *pos);
4627 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4629 return seq_list_next(p, &slab_caches, pos);
4632 static void s_stop(struct seq_file *m, void *p)
4634 up_read(&slub_lock);
4637 static int s_show(struct seq_file *m, void *p)
4639 unsigned long nr_partials = 0;
4640 unsigned long nr_slabs = 0;
4641 unsigned long nr_inuse = 0;
4642 unsigned long nr_objs = 0;
4643 unsigned long nr_free = 0;
4644 struct kmem_cache *s;
4647 s = list_entry(p, struct kmem_cache, list);
4649 for_each_online_node(node) {
4650 struct kmem_cache_node *n = get_node(s, node);
4655 nr_partials += n->nr_partial;
4656 nr_slabs += atomic_long_read(&n->nr_slabs);
4657 nr_objs += atomic_long_read(&n->total_objects);
4658 nr_free += count_partial(n, count_free);
4661 nr_inuse = nr_objs - nr_free;
4663 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4664 nr_objs, s->size, oo_objects(s->oo),
4665 (1 << oo_order(s->oo)));
4666 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4667 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4673 static const struct seq_operations slabinfo_op = {
4680 static int slabinfo_open(struct inode *inode, struct file *file)
4682 return seq_open(file, &slabinfo_op);
4685 static const struct file_operations proc_slabinfo_operations = {
4686 .open = slabinfo_open,
4688 .llseek = seq_lseek,
4689 .release = seq_release,
4692 static int __init slab_proc_init(void)
4694 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
4697 module_init(slab_proc_init);
4698 #endif /* CONFIG_SLABINFO */