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/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/proc_fs.h>
18 #include <linux/seq_file.h>
19 #include <linux/cpu.h>
20 #include <linux/cpuset.h>
21 #include <linux/mempolicy.h>
22 #include <linux/ctype.h>
23 #include <linux/debugobjects.h>
24 #include <linux/kallsyms.h>
25 #include <linux/memory.h>
26 #include <linux/math64.h>
27 #include <linux/kmemtrace.h>
34 * The slab_lock protects operations on the object of a particular
35 * slab and its metadata in the page struct. If the slab lock
36 * has been taken then no allocations nor frees can be performed
37 * on the objects in the slab nor can the slab be added or removed
38 * from the partial or full lists since this would mean modifying
39 * the page_struct of the slab.
41 * The list_lock protects the partial and full list on each node and
42 * the partial slab counter. If taken then no new slabs may be added or
43 * removed from the lists nor make the number of partial slabs be modified.
44 * (Note that the total number of slabs is an atomic value that may be
45 * modified without taking the list lock).
47 * The list_lock is a centralized lock and thus we avoid taking it as
48 * much as possible. As long as SLUB does not have to handle partial
49 * slabs, operations can continue without any centralized lock. F.e.
50 * allocating a long series of objects that fill up slabs does not require
53 * The lock order is sometimes inverted when we are trying to get a slab
54 * off a list. We take the list_lock and then look for a page on the list
55 * to use. While we do that objects in the slabs may be freed. We can
56 * only operate on the slab if we have also taken the slab_lock. So we use
57 * a slab_trylock() on the slab. If trylock was successful then no frees
58 * can occur anymore and we can use the slab for allocations etc. If the
59 * slab_trylock() does not succeed then frees are in progress in the slab and
60 * we must stay away from it for a while since we may cause a bouncing
61 * cacheline if we try to acquire the lock. So go onto the next slab.
62 * If all pages are busy then we may allocate a new slab instead of reusing
63 * a partial slab. A new slab has noone operating on it and thus there is
64 * no danger of cacheline contention.
66 * Interrupts are disabled during allocation and deallocation in order to
67 * make the slab allocator safe to use in the context of an irq. In addition
68 * interrupts are disabled to ensure that the processor does not change
69 * while handling per_cpu slabs, due to kernel preemption.
71 * SLUB assigns one slab for allocation to each processor.
72 * Allocations only occur from these slabs called cpu slabs.
74 * Slabs with free elements are kept on a partial list and during regular
75 * operations no list for full slabs is used. If an object in a full slab is
76 * freed then the slab will show up again on the partial lists.
77 * We track full slabs for debugging purposes though because otherwise we
78 * cannot scan all objects.
80 * Slabs are freed when they become empty. Teardown and setup is
81 * minimal so we rely on the page allocators per cpu caches for
82 * fast frees and allocs.
84 * Overloading of page flags that are otherwise used for LRU management.
86 * PageActive The slab is frozen and exempt from list processing.
87 * This means that the slab is dedicated to a purpose
88 * such as satisfying allocations for a specific
89 * processor. Objects may be freed in the slab while
90 * it is frozen but slab_free will then skip the usual
91 * list operations. It is up to the processor holding
92 * the slab to integrate the slab into the slab lists
93 * when the slab is no longer needed.
95 * One use of this flag is to mark slabs that are
96 * used for allocations. Then such a slab becomes a cpu
97 * slab. The cpu slab may be equipped with an additional
98 * freelist that allows lockless access to
99 * free objects in addition to the regular freelist
100 * that requires the slab lock.
102 * PageError Slab requires special handling due to debug
103 * options set. This moves slab handling out of
104 * the fast path and disables lockless freelists.
107 #ifdef CONFIG_SLUB_DEBUG
114 * Issues still to be resolved:
116 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
118 * - Variable sizing of the per node arrays
121 /* Enable to test recovery from slab corruption on boot */
122 #undef SLUB_RESILIENCY_TEST
125 * Mininum number of partial slabs. These will be left on the partial
126 * lists even if they are empty. kmem_cache_shrink may reclaim them.
128 #define MIN_PARTIAL 5
131 * Maximum number of desirable partial slabs.
132 * The existence of more partial slabs makes kmem_cache_shrink
133 * sort the partial list by the number of objects in the.
135 #define MAX_PARTIAL 10
137 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
138 SLAB_POISON | SLAB_STORE_USER)
141 * Set of flags that will prevent slab merging
143 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
144 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
146 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
149 #ifndef ARCH_KMALLOC_MINALIGN
150 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
153 #ifndef ARCH_SLAB_MINALIGN
154 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
157 /* Internal SLUB flags */
158 #define __OBJECT_POISON 0x80000000 /* Poison object */
159 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
161 static int kmem_size = sizeof(struct kmem_cache);
164 static struct notifier_block slab_notifier;
168 DOWN, /* No slab functionality available */
169 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
170 UP, /* Everything works but does not show up in sysfs */
174 /* A list of all slab caches on the system */
175 static DECLARE_RWSEM(slub_lock);
176 static LIST_HEAD(slab_caches);
179 * Tracking user of a slab.
182 unsigned long addr; /* Called from address */
183 int cpu; /* Was running on cpu */
184 int pid; /* Pid context */
185 unsigned long when; /* When did the operation occur */
188 enum track_item { TRACK_ALLOC, TRACK_FREE };
190 #ifdef CONFIG_SLUB_DEBUG
191 static int sysfs_slab_add(struct kmem_cache *);
192 static int sysfs_slab_alias(struct kmem_cache *, const char *);
193 static void sysfs_slab_remove(struct kmem_cache *);
196 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
197 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
199 static inline void sysfs_slab_remove(struct kmem_cache *s)
206 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
208 #ifdef CONFIG_SLUB_STATS
213 /********************************************************************
214 * Core slab cache functions
215 *******************************************************************/
217 int slab_is_available(void)
219 return slab_state >= UP;
222 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
225 return s->node[node];
227 return &s->local_node;
231 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
234 return s->cpu_slab[cpu];
240 /* Verify that a pointer has an address that is valid within a slab page */
241 static inline int check_valid_pointer(struct kmem_cache *s,
242 struct page *page, const void *object)
249 base = page_address(page);
250 if (object < base || object >= base + page->objects * s->size ||
251 (object - base) % s->size) {
259 * Slow version of get and set free pointer.
261 * This version requires touching the cache lines of kmem_cache which
262 * we avoid to do in the fast alloc free paths. There we obtain the offset
263 * from the page struct.
265 static inline void *get_freepointer(struct kmem_cache *s, void *object)
267 return *(void **)(object + s->offset);
270 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
272 *(void **)(object + s->offset) = fp;
275 /* Loop over all objects in a slab */
276 #define for_each_object(__p, __s, __addr, __objects) \
277 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
281 #define for_each_free_object(__p, __s, __free) \
282 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
284 /* Determine object index from a given position */
285 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
287 return (p - addr) / s->size;
290 static inline struct kmem_cache_order_objects oo_make(int order,
293 struct kmem_cache_order_objects x = {
294 (order << 16) + (PAGE_SIZE << order) / size
300 static inline int oo_order(struct kmem_cache_order_objects x)
305 static inline int oo_objects(struct kmem_cache_order_objects x)
307 return x.x & ((1 << 16) - 1);
310 #ifdef CONFIG_SLUB_DEBUG
314 #ifdef CONFIG_SLUB_DEBUG_ON
315 static int slub_debug = DEBUG_DEFAULT_FLAGS;
317 static int slub_debug;
320 static char *slub_debug_slabs;
325 static void print_section(char *text, u8 *addr, unsigned int length)
333 for (i = 0; i < length; i++) {
335 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
338 printk(KERN_CONT " %02x", addr[i]);
340 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
342 printk(KERN_CONT " %s\n", ascii);
349 printk(KERN_CONT " ");
353 printk(KERN_CONT " %s\n", ascii);
357 static struct track *get_track(struct kmem_cache *s, void *object,
358 enum track_item alloc)
363 p = object + s->offset + sizeof(void *);
365 p = object + s->inuse;
370 static void set_track(struct kmem_cache *s, void *object,
371 enum track_item alloc, unsigned long addr)
376 p = object + s->offset + sizeof(void *);
378 p = object + s->inuse;
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, start, 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 loose 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 > 65535)
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 void inc_slabs_node(struct kmem_cache *s, int node, int objects)
837 struct kmem_cache_node *n = get_node(s, node);
840 * May be called early in order to allocate a slab for the
841 * kmem_cache_node structure. Solve the chicken-egg
842 * dilemma by deferring the increment of the count during
843 * bootstrap (see early_kmem_cache_node_alloc).
845 if (!NUMA_BUILD || n) {
846 atomic_long_inc(&n->nr_slabs);
847 atomic_long_add(objects, &n->total_objects);
850 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
852 struct kmem_cache_node *n = get_node(s, node);
854 atomic_long_dec(&n->nr_slabs);
855 atomic_long_sub(objects, &n->total_objects);
858 /* Object debug checks for alloc/free paths */
859 static void setup_object_debug(struct kmem_cache *s, struct page *page,
862 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
865 init_object(s, object, 0);
866 init_tracking(s, object);
869 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
870 void *object, unsigned long addr)
872 if (!check_slab(s, page))
875 if (!on_freelist(s, page, object)) {
876 object_err(s, page, object, "Object already allocated");
880 if (!check_valid_pointer(s, page, object)) {
881 object_err(s, page, object, "Freelist Pointer check fails");
885 if (!check_object(s, page, object, 0))
888 /* Success perform special debug activities for allocs */
889 if (s->flags & SLAB_STORE_USER)
890 set_track(s, object, TRACK_ALLOC, addr);
891 trace(s, page, object, 1);
892 init_object(s, object, 1);
896 if (PageSlab(page)) {
898 * If this is a slab page then lets do the best we can
899 * to avoid issues in the future. Marking all objects
900 * as used avoids touching the remaining objects.
902 slab_fix(s, "Marking all objects used");
903 page->inuse = page->objects;
904 page->freelist = NULL;
909 static int free_debug_processing(struct kmem_cache *s, struct page *page,
910 void *object, unsigned long addr)
912 if (!check_slab(s, page))
915 if (!check_valid_pointer(s, page, object)) {
916 slab_err(s, page, "Invalid object pointer 0x%p", object);
920 if (on_freelist(s, page, object)) {
921 object_err(s, page, object, "Object already free");
925 if (!check_object(s, page, object, 1))
928 if (unlikely(s != page->slab)) {
929 if (!PageSlab(page)) {
930 slab_err(s, page, "Attempt to free object(0x%p) "
931 "outside of slab", object);
932 } else if (!page->slab) {
934 "SLUB <none>: no slab for object 0x%p.\n",
938 object_err(s, page, object,
939 "page slab pointer corrupt.");
943 /* Special debug activities for freeing objects */
944 if (!PageSlubFrozen(page) && !page->freelist)
945 remove_full(s, page);
946 if (s->flags & SLAB_STORE_USER)
947 set_track(s, object, TRACK_FREE, addr);
948 trace(s, page, object, 0);
949 init_object(s, object, 0);
953 slab_fix(s, "Object at 0x%p not freed", object);
957 static int __init setup_slub_debug(char *str)
959 slub_debug = DEBUG_DEFAULT_FLAGS;
960 if (*str++ != '=' || !*str)
962 * No options specified. Switch on full debugging.
968 * No options but restriction on slabs. This means full
969 * debugging for slabs matching a pattern.
976 * Switch off all debugging measures.
981 * Determine which debug features should be switched on
983 for (; *str && *str != ','; str++) {
984 switch (tolower(*str)) {
986 slub_debug |= SLAB_DEBUG_FREE;
989 slub_debug |= SLAB_RED_ZONE;
992 slub_debug |= SLAB_POISON;
995 slub_debug |= SLAB_STORE_USER;
998 slub_debug |= SLAB_TRACE;
1001 printk(KERN_ERR "slub_debug option '%c' "
1002 "unknown. skipped\n", *str);
1008 slub_debug_slabs = str + 1;
1013 __setup("slub_debug", setup_slub_debug);
1015 static unsigned long kmem_cache_flags(unsigned long objsize,
1016 unsigned long flags, const char *name,
1017 void (*ctor)(void *))
1020 * Enable debugging if selected on the kernel commandline.
1022 if (slub_debug && (!slub_debug_slabs ||
1023 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1024 flags |= slub_debug;
1029 static inline void setup_object_debug(struct kmem_cache *s,
1030 struct page *page, void *object) {}
1032 static inline int alloc_debug_processing(struct kmem_cache *s,
1033 struct page *page, void *object, unsigned long addr) { return 0; }
1035 static inline int free_debug_processing(struct kmem_cache *s,
1036 struct page *page, void *object, unsigned long addr) { return 0; }
1038 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1040 static inline int check_object(struct kmem_cache *s, struct page *page,
1041 void *object, int active) { return 1; }
1042 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1043 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1044 unsigned long flags, const char *name,
1045 void (*ctor)(void *))
1049 #define slub_debug 0
1051 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1053 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1055 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1060 * Slab allocation and freeing
1062 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1063 struct kmem_cache_order_objects oo)
1065 int order = oo_order(oo);
1068 return alloc_pages(flags, order);
1070 return alloc_pages_node(node, flags, order);
1073 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1076 struct kmem_cache_order_objects oo = s->oo;
1078 flags |= s->allocflags;
1080 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1082 if (unlikely(!page)) {
1085 * Allocation may have failed due to fragmentation.
1086 * Try a lower order alloc if possible
1088 page = alloc_slab_page(flags, node, oo);
1092 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1094 page->objects = oo_objects(oo);
1095 mod_zone_page_state(page_zone(page),
1096 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1097 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1103 static void setup_object(struct kmem_cache *s, struct page *page,
1106 setup_object_debug(s, page, object);
1107 if (unlikely(s->ctor))
1111 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1118 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1120 page = allocate_slab(s,
1121 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1125 inc_slabs_node(s, page_to_nid(page), page->objects);
1127 page->flags |= 1 << PG_slab;
1128 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1129 SLAB_STORE_USER | SLAB_TRACE))
1130 __SetPageSlubDebug(page);
1132 start = page_address(page);
1134 if (unlikely(s->flags & SLAB_POISON))
1135 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1138 for_each_object(p, s, start, page->objects) {
1139 setup_object(s, page, last);
1140 set_freepointer(s, last, p);
1143 setup_object(s, page, last);
1144 set_freepointer(s, last, NULL);
1146 page->freelist = start;
1152 static void __free_slab(struct kmem_cache *s, struct page *page)
1154 int order = compound_order(page);
1155 int pages = 1 << order;
1157 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1160 slab_pad_check(s, page);
1161 for_each_object(p, s, page_address(page),
1163 check_object(s, page, p, 0);
1164 __ClearPageSlubDebug(page);
1167 mod_zone_page_state(page_zone(page),
1168 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1169 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1172 __ClearPageSlab(page);
1173 reset_page_mapcount(page);
1174 __free_pages(page, order);
1177 static void rcu_free_slab(struct rcu_head *h)
1181 page = container_of((struct list_head *)h, struct page, lru);
1182 __free_slab(page->slab, page);
1185 static void free_slab(struct kmem_cache *s, struct page *page)
1187 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1189 * RCU free overloads the RCU head over the LRU
1191 struct rcu_head *head = (void *)&page->lru;
1193 call_rcu(head, rcu_free_slab);
1195 __free_slab(s, page);
1198 static void discard_slab(struct kmem_cache *s, struct page *page)
1200 dec_slabs_node(s, page_to_nid(page), page->objects);
1205 * Per slab locking using the pagelock
1207 static __always_inline void slab_lock(struct page *page)
1209 bit_spin_lock(PG_locked, &page->flags);
1212 static __always_inline void slab_unlock(struct page *page)
1214 __bit_spin_unlock(PG_locked, &page->flags);
1217 static __always_inline int slab_trylock(struct page *page)
1221 rc = bit_spin_trylock(PG_locked, &page->flags);
1226 * Management of partially allocated slabs
1228 static void add_partial(struct kmem_cache_node *n,
1229 struct page *page, int tail)
1231 spin_lock(&n->list_lock);
1234 list_add_tail(&page->lru, &n->partial);
1236 list_add(&page->lru, &n->partial);
1237 spin_unlock(&n->list_lock);
1240 static void remove_partial(struct kmem_cache *s, struct page *page)
1242 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1244 spin_lock(&n->list_lock);
1245 list_del(&page->lru);
1247 spin_unlock(&n->list_lock);
1251 * Lock slab and remove from the partial list.
1253 * Must hold list_lock.
1255 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1258 if (slab_trylock(page)) {
1259 list_del(&page->lru);
1261 __SetPageSlubFrozen(page);
1268 * Try to allocate a partial slab from a specific node.
1270 static struct page *get_partial_node(struct kmem_cache_node *n)
1275 * Racy check. If we mistakenly see no partial slabs then we
1276 * just allocate an empty slab. If we mistakenly try to get a
1277 * partial slab and there is none available then get_partials()
1280 if (!n || !n->nr_partial)
1283 spin_lock(&n->list_lock);
1284 list_for_each_entry(page, &n->partial, lru)
1285 if (lock_and_freeze_slab(n, page))
1289 spin_unlock(&n->list_lock);
1294 * Get a page from somewhere. Search in increasing NUMA distances.
1296 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1299 struct zonelist *zonelist;
1302 enum zone_type high_zoneidx = gfp_zone(flags);
1306 * The defrag ratio allows a configuration of the tradeoffs between
1307 * inter node defragmentation and node local allocations. A lower
1308 * defrag_ratio increases the tendency to do local allocations
1309 * instead of attempting to obtain partial slabs from other nodes.
1311 * If the defrag_ratio is set to 0 then kmalloc() always
1312 * returns node local objects. If the ratio is higher then kmalloc()
1313 * may return off node objects because partial slabs are obtained
1314 * from other nodes and filled up.
1316 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1317 * defrag_ratio = 1000) then every (well almost) allocation will
1318 * first attempt to defrag slab caches on other nodes. This means
1319 * scanning over all nodes to look for partial slabs which may be
1320 * expensive if we do it every time we are trying to find a slab
1321 * with available objects.
1323 if (!s->remote_node_defrag_ratio ||
1324 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1327 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1328 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1329 struct kmem_cache_node *n;
1331 n = get_node(s, zone_to_nid(zone));
1333 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1334 n->nr_partial > n->min_partial) {
1335 page = get_partial_node(n);
1345 * Get a partial page, lock it and return it.
1347 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1350 int searchnode = (node == -1) ? numa_node_id() : node;
1352 page = get_partial_node(get_node(s, searchnode));
1353 if (page || (flags & __GFP_THISNODE))
1356 return get_any_partial(s, flags);
1360 * Move a page back to the lists.
1362 * Must be called with the slab lock held.
1364 * On exit the slab lock will have been dropped.
1366 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1368 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1369 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1371 __ClearPageSlubFrozen(page);
1374 if (page->freelist) {
1375 add_partial(n, page, tail);
1376 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1378 stat(c, DEACTIVATE_FULL);
1379 if (SLABDEBUG && PageSlubDebug(page) &&
1380 (s->flags & SLAB_STORE_USER))
1385 stat(c, DEACTIVATE_EMPTY);
1386 if (n->nr_partial < n->min_partial) {
1388 * Adding an empty slab to the partial slabs in order
1389 * to avoid page allocator overhead. This slab needs
1390 * to come after the other slabs with objects in
1391 * so that the others get filled first. That way the
1392 * size of the partial list stays small.
1394 * kmem_cache_shrink can reclaim any empty slabs from
1397 add_partial(n, page, 1);
1401 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1402 discard_slab(s, page);
1408 * Remove the cpu slab
1410 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1412 struct page *page = c->page;
1416 stat(c, DEACTIVATE_REMOTE_FREES);
1418 * Merge cpu freelist into slab freelist. Typically we get here
1419 * because both freelists are empty. So this is unlikely
1422 while (unlikely(c->freelist)) {
1425 tail = 0; /* Hot objects. Put the slab first */
1427 /* Retrieve object from cpu_freelist */
1428 object = c->freelist;
1429 c->freelist = c->freelist[c->offset];
1431 /* And put onto the regular freelist */
1432 object[c->offset] = page->freelist;
1433 page->freelist = object;
1437 unfreeze_slab(s, page, tail);
1440 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1442 stat(c, CPUSLAB_FLUSH);
1444 deactivate_slab(s, c);
1450 * Called from IPI handler with interrupts disabled.
1452 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1454 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1456 if (likely(c && c->page))
1460 static void flush_cpu_slab(void *d)
1462 struct kmem_cache *s = d;
1464 __flush_cpu_slab(s, smp_processor_id());
1467 static void flush_all(struct kmem_cache *s)
1469 on_each_cpu(flush_cpu_slab, s, 1);
1473 * Check if the objects in a per cpu structure fit numa
1474 * locality expectations.
1476 static inline int node_match(struct kmem_cache_cpu *c, int node)
1479 if (node != -1 && c->node != node)
1486 * Slow path. The lockless freelist is empty or we need to perform
1489 * Interrupts are disabled.
1491 * Processing is still very fast if new objects have been freed to the
1492 * regular freelist. In that case we simply take over the regular freelist
1493 * as the lockless freelist and zap the regular freelist.
1495 * If that is not working then we fall back to the partial lists. We take the
1496 * first element of the freelist as the object to allocate now and move the
1497 * rest of the freelist to the lockless freelist.
1499 * And if we were unable to get a new slab from the partial slab lists then
1500 * we need to allocate a new slab. This is the slowest path since it involves
1501 * a call to the page allocator and the setup of a new slab.
1503 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1504 unsigned long addr, struct kmem_cache_cpu *c)
1509 /* We handle __GFP_ZERO in the caller */
1510 gfpflags &= ~__GFP_ZERO;
1516 if (unlikely(!node_match(c, node)))
1519 stat(c, ALLOC_REFILL);
1522 object = c->page->freelist;
1523 if (unlikely(!object))
1525 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1528 c->freelist = object[c->offset];
1529 c->page->inuse = c->page->objects;
1530 c->page->freelist = NULL;
1531 c->node = page_to_nid(c->page);
1533 slab_unlock(c->page);
1534 stat(c, ALLOC_SLOWPATH);
1538 deactivate_slab(s, c);
1541 new = get_partial(s, gfpflags, node);
1544 stat(c, ALLOC_FROM_PARTIAL);
1548 if (gfpflags & __GFP_WAIT)
1551 new = new_slab(s, gfpflags, node);
1553 if (gfpflags & __GFP_WAIT)
1554 local_irq_disable();
1557 c = get_cpu_slab(s, smp_processor_id());
1558 stat(c, ALLOC_SLAB);
1562 __SetPageSlubFrozen(new);
1568 if (!alloc_debug_processing(s, c->page, object, addr))
1572 c->page->freelist = object[c->offset];
1578 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1579 * have the fastpath folded into their functions. So no function call
1580 * overhead for requests that can be satisfied on the fastpath.
1582 * The fastpath works by first checking if the lockless freelist can be used.
1583 * If not then __slab_alloc is called for slow processing.
1585 * Otherwise we can simply pick the next object from the lockless free list.
1587 static __always_inline void *slab_alloc(struct kmem_cache *s,
1588 gfp_t gfpflags, int node, unsigned long addr)
1591 struct kmem_cache_cpu *c;
1592 unsigned long flags;
1593 unsigned int objsize;
1595 local_irq_save(flags);
1596 c = get_cpu_slab(s, smp_processor_id());
1597 objsize = c->objsize;
1598 if (unlikely(!c->freelist || !node_match(c, node)))
1600 object = __slab_alloc(s, gfpflags, node, addr, c);
1603 object = c->freelist;
1604 c->freelist = object[c->offset];
1605 stat(c, ALLOC_FASTPATH);
1607 local_irq_restore(flags);
1609 if (unlikely((gfpflags & __GFP_ZERO) && object))
1610 memset(object, 0, objsize);
1615 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1617 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1619 kmemtrace_mark_alloc(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
1620 s->objsize, s->size, gfpflags);
1624 EXPORT_SYMBOL(kmem_cache_alloc);
1626 #ifdef CONFIG_KMEMTRACE
1627 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1629 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1631 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1635 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1637 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1639 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_CACHE, _RET_IP_, ret,
1640 s->objsize, s->size, gfpflags, node);
1644 EXPORT_SYMBOL(kmem_cache_alloc_node);
1647 #ifdef CONFIG_KMEMTRACE
1648 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1652 return slab_alloc(s, gfpflags, node, _RET_IP_);
1654 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1658 * Slow patch handling. This may still be called frequently since objects
1659 * have a longer lifetime than the cpu slabs in most processing loads.
1661 * So we still attempt to reduce cache line usage. Just take the slab
1662 * lock and free the item. If there is no additional partial page
1663 * handling required then we can return immediately.
1665 static void __slab_free(struct kmem_cache *s, struct page *page,
1666 void *x, unsigned long addr, unsigned int offset)
1669 void **object = (void *)x;
1670 struct kmem_cache_cpu *c;
1672 c = get_cpu_slab(s, raw_smp_processor_id());
1673 stat(c, FREE_SLOWPATH);
1676 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1680 prior = object[offset] = page->freelist;
1681 page->freelist = object;
1684 if (unlikely(PageSlubFrozen(page))) {
1685 stat(c, FREE_FROZEN);
1689 if (unlikely(!page->inuse))
1693 * Objects left in the slab. If it was not on the partial list before
1696 if (unlikely(!prior)) {
1697 add_partial(get_node(s, page_to_nid(page)), page, 1);
1698 stat(c, FREE_ADD_PARTIAL);
1708 * Slab still on the partial list.
1710 remove_partial(s, page);
1711 stat(c, FREE_REMOVE_PARTIAL);
1715 discard_slab(s, page);
1719 if (!free_debug_processing(s, page, x, addr))
1725 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1726 * can perform fastpath freeing without additional function calls.
1728 * The fastpath is only possible if we are freeing to the current cpu slab
1729 * of this processor. This typically the case if we have just allocated
1732 * If fastpath is not possible then fall back to __slab_free where we deal
1733 * with all sorts of special processing.
1735 static __always_inline void slab_free(struct kmem_cache *s,
1736 struct page *page, void *x, unsigned long addr)
1738 void **object = (void *)x;
1739 struct kmem_cache_cpu *c;
1740 unsigned long flags;
1742 local_irq_save(flags);
1743 c = get_cpu_slab(s, smp_processor_id());
1744 debug_check_no_locks_freed(object, c->objsize);
1745 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1746 debug_check_no_obj_freed(object, s->objsize);
1747 if (likely(page == c->page && c->node >= 0)) {
1748 object[c->offset] = c->freelist;
1749 c->freelist = object;
1750 stat(c, FREE_FASTPATH);
1752 __slab_free(s, page, x, addr, c->offset);
1754 local_irq_restore(flags);
1757 void kmem_cache_free(struct kmem_cache *s, void *x)
1761 page = virt_to_head_page(x);
1763 slab_free(s, page, x, _RET_IP_);
1765 kmemtrace_mark_free(KMEMTRACE_TYPE_CACHE, _RET_IP_, x);
1767 EXPORT_SYMBOL(kmem_cache_free);
1769 /* Figure out on which slab object the object resides */
1770 static struct page *get_object_page(const void *x)
1772 struct page *page = virt_to_head_page(x);
1774 if (!PageSlab(page))
1781 * Object placement in a slab is made very easy because we always start at
1782 * offset 0. If we tune the size of the object to the alignment then we can
1783 * get the required alignment by putting one properly sized object after
1786 * Notice that the allocation order determines the sizes of the per cpu
1787 * caches. Each processor has always one slab available for allocations.
1788 * Increasing the allocation order reduces the number of times that slabs
1789 * must be moved on and off the partial lists and is therefore a factor in
1794 * Mininum / Maximum order of slab pages. This influences locking overhead
1795 * and slab fragmentation. A higher order reduces the number of partial slabs
1796 * and increases the number of allocations possible without having to
1797 * take the list_lock.
1799 static int slub_min_order;
1800 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1801 static int slub_min_objects;
1804 * Merge control. If this is set then no merging of slab caches will occur.
1805 * (Could be removed. This was introduced to pacify the merge skeptics.)
1807 static int slub_nomerge;
1810 * Calculate the order of allocation given an slab object size.
1812 * The order of allocation has significant impact on performance and other
1813 * system components. Generally order 0 allocations should be preferred since
1814 * order 0 does not cause fragmentation in the page allocator. Larger objects
1815 * be problematic to put into order 0 slabs because there may be too much
1816 * unused space left. We go to a higher order if more than 1/16th of the slab
1819 * In order to reach satisfactory performance we must ensure that a minimum
1820 * number of objects is in one slab. Otherwise we may generate too much
1821 * activity on the partial lists which requires taking the list_lock. This is
1822 * less a concern for large slabs though which are rarely used.
1824 * slub_max_order specifies the order where we begin to stop considering the
1825 * number of objects in a slab as critical. If we reach slub_max_order then
1826 * we try to keep the page order as low as possible. So we accept more waste
1827 * of space in favor of a small page order.
1829 * Higher order allocations also allow the placement of more objects in a
1830 * slab and thereby reduce object handling overhead. If the user has
1831 * requested a higher mininum order then we start with that one instead of
1832 * the smallest order which will fit the object.
1834 static inline int slab_order(int size, int min_objects,
1835 int max_order, int fract_leftover)
1839 int min_order = slub_min_order;
1841 if ((PAGE_SIZE << min_order) / size > 65535)
1842 return get_order(size * 65535) - 1;
1844 for (order = max(min_order,
1845 fls(min_objects * size - 1) - PAGE_SHIFT);
1846 order <= max_order; order++) {
1848 unsigned long slab_size = PAGE_SIZE << order;
1850 if (slab_size < min_objects * size)
1853 rem = slab_size % size;
1855 if (rem <= slab_size / fract_leftover)
1863 static inline int calculate_order(int size)
1870 * Attempt to find best configuration for a slab. This
1871 * works by first attempting to generate a layout with
1872 * the best configuration and backing off gradually.
1874 * First we reduce the acceptable waste in a slab. Then
1875 * we reduce the minimum objects required in a slab.
1877 min_objects = slub_min_objects;
1879 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1880 while (min_objects > 1) {
1882 while (fraction >= 4) {
1883 order = slab_order(size, min_objects,
1884 slub_max_order, fraction);
1885 if (order <= slub_max_order)
1893 * We were unable to place multiple objects in a slab. Now
1894 * lets see if we can place a single object there.
1896 order = slab_order(size, 1, slub_max_order, 1);
1897 if (order <= slub_max_order)
1901 * Doh this slab cannot be placed using slub_max_order.
1903 order = slab_order(size, 1, MAX_ORDER, 1);
1904 if (order <= MAX_ORDER)
1910 * Figure out what the alignment of the objects will be.
1912 static unsigned long calculate_alignment(unsigned long flags,
1913 unsigned long align, unsigned long size)
1916 * If the user wants hardware cache aligned objects then follow that
1917 * suggestion if the object is sufficiently large.
1919 * The hardware cache alignment cannot override the specified
1920 * alignment though. If that is greater then use it.
1922 if (flags & SLAB_HWCACHE_ALIGN) {
1923 unsigned long ralign = cache_line_size();
1924 while (size <= ralign / 2)
1926 align = max(align, ralign);
1929 if (align < ARCH_SLAB_MINALIGN)
1930 align = ARCH_SLAB_MINALIGN;
1932 return ALIGN(align, sizeof(void *));
1935 static void init_kmem_cache_cpu(struct kmem_cache *s,
1936 struct kmem_cache_cpu *c)
1941 c->offset = s->offset / sizeof(void *);
1942 c->objsize = s->objsize;
1943 #ifdef CONFIG_SLUB_STATS
1944 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1949 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1954 * The larger the object size is, the more pages we want on the partial
1955 * list to avoid pounding the page allocator excessively.
1957 n->min_partial = ilog2(s->size);
1958 if (n->min_partial < MIN_PARTIAL)
1959 n->min_partial = MIN_PARTIAL;
1960 else if (n->min_partial > MAX_PARTIAL)
1961 n->min_partial = MAX_PARTIAL;
1963 spin_lock_init(&n->list_lock);
1964 INIT_LIST_HEAD(&n->partial);
1965 #ifdef CONFIG_SLUB_DEBUG
1966 atomic_long_set(&n->nr_slabs, 0);
1967 atomic_long_set(&n->total_objects, 0);
1968 INIT_LIST_HEAD(&n->full);
1974 * Per cpu array for per cpu structures.
1976 * The per cpu array places all kmem_cache_cpu structures from one processor
1977 * close together meaning that it becomes possible that multiple per cpu
1978 * structures are contained in one cacheline. This may be particularly
1979 * beneficial for the kmalloc caches.
1981 * A desktop system typically has around 60-80 slabs. With 100 here we are
1982 * likely able to get per cpu structures for all caches from the array defined
1983 * here. We must be able to cover all kmalloc caches during bootstrap.
1985 * If the per cpu array is exhausted then fall back to kmalloc
1986 * of individual cachelines. No sharing is possible then.
1988 #define NR_KMEM_CACHE_CPU 100
1990 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1991 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1993 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1994 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1996 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1997 int cpu, gfp_t flags)
1999 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2002 per_cpu(kmem_cache_cpu_free, cpu) =
2003 (void *)c->freelist;
2005 /* Table overflow: So allocate ourselves */
2007 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2008 flags, cpu_to_node(cpu));
2013 init_kmem_cache_cpu(s, c);
2017 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2019 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2020 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2024 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2025 per_cpu(kmem_cache_cpu_free, cpu) = c;
2028 static void free_kmem_cache_cpus(struct kmem_cache *s)
2032 for_each_online_cpu(cpu) {
2033 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2036 s->cpu_slab[cpu] = NULL;
2037 free_kmem_cache_cpu(c, cpu);
2042 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2046 for_each_online_cpu(cpu) {
2047 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2052 c = alloc_kmem_cache_cpu(s, cpu, flags);
2054 free_kmem_cache_cpus(s);
2057 s->cpu_slab[cpu] = c;
2063 * Initialize the per cpu array.
2065 static void init_alloc_cpu_cpu(int cpu)
2069 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2072 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2073 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2075 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2078 static void __init init_alloc_cpu(void)
2082 for_each_online_cpu(cpu)
2083 init_alloc_cpu_cpu(cpu);
2087 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2088 static inline void init_alloc_cpu(void) {}
2090 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2092 init_kmem_cache_cpu(s, &s->cpu_slab);
2099 * No kmalloc_node yet so do it by hand. We know that this is the first
2100 * slab on the node for this slabcache. There are no concurrent accesses
2103 * Note that this function only works on the kmalloc_node_cache
2104 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2105 * memory on a fresh node that has no slab structures yet.
2107 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2111 struct kmem_cache_node *n;
2112 unsigned long flags;
2114 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2116 page = new_slab(kmalloc_caches, gfpflags, node);
2119 if (page_to_nid(page) != node) {
2120 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2122 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2123 "in order to be able to continue\n");
2128 page->freelist = get_freepointer(kmalloc_caches, n);
2130 kmalloc_caches->node[node] = n;
2131 #ifdef CONFIG_SLUB_DEBUG
2132 init_object(kmalloc_caches, n, 1);
2133 init_tracking(kmalloc_caches, n);
2135 init_kmem_cache_node(n, kmalloc_caches);
2136 inc_slabs_node(kmalloc_caches, node, page->objects);
2139 * lockdep requires consistent irq usage for each lock
2140 * so even though there cannot be a race this early in
2141 * the boot sequence, we still disable irqs.
2143 local_irq_save(flags);
2144 add_partial(n, page, 0);
2145 local_irq_restore(flags);
2149 static void free_kmem_cache_nodes(struct kmem_cache *s)
2153 for_each_node_state(node, N_NORMAL_MEMORY) {
2154 struct kmem_cache_node *n = s->node[node];
2155 if (n && n != &s->local_node)
2156 kmem_cache_free(kmalloc_caches, n);
2157 s->node[node] = NULL;
2161 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2166 if (slab_state >= UP)
2167 local_node = page_to_nid(virt_to_page(s));
2171 for_each_node_state(node, N_NORMAL_MEMORY) {
2172 struct kmem_cache_node *n;
2174 if (local_node == node)
2177 if (slab_state == DOWN) {
2178 n = early_kmem_cache_node_alloc(gfpflags,
2182 n = kmem_cache_alloc_node(kmalloc_caches,
2186 free_kmem_cache_nodes(s);
2192 init_kmem_cache_node(n, s);
2197 static void free_kmem_cache_nodes(struct kmem_cache *s)
2201 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2203 init_kmem_cache_node(&s->local_node, s);
2209 * calculate_sizes() determines the order and the distribution of data within
2212 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2214 unsigned long flags = s->flags;
2215 unsigned long size = s->objsize;
2216 unsigned long align = s->align;
2220 * Round up object size to the next word boundary. We can only
2221 * place the free pointer at word boundaries and this determines
2222 * the possible location of the free pointer.
2224 size = ALIGN(size, sizeof(void *));
2226 #ifdef CONFIG_SLUB_DEBUG
2228 * Determine if we can poison the object itself. If the user of
2229 * the slab may touch the object after free or before allocation
2230 * then we should never poison the object itself.
2232 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2234 s->flags |= __OBJECT_POISON;
2236 s->flags &= ~__OBJECT_POISON;
2240 * If we are Redzoning then check if there is some space between the
2241 * end of the object and the free pointer. If not then add an
2242 * additional word to have some bytes to store Redzone information.
2244 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2245 size += sizeof(void *);
2249 * With that we have determined the number of bytes in actual use
2250 * by the object. This is the potential offset to the free pointer.
2254 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2257 * Relocate free pointer after the object if it is not
2258 * permitted to overwrite the first word of the object on
2261 * This is the case if we do RCU, have a constructor or
2262 * destructor or are poisoning the objects.
2265 size += sizeof(void *);
2268 #ifdef CONFIG_SLUB_DEBUG
2269 if (flags & SLAB_STORE_USER)
2271 * Need to store information about allocs and frees after
2274 size += 2 * sizeof(struct track);
2276 if (flags & SLAB_RED_ZONE)
2278 * Add some empty padding so that we can catch
2279 * overwrites from earlier objects rather than let
2280 * tracking information or the free pointer be
2281 * corrupted if an user writes before the start
2284 size += sizeof(void *);
2288 * Determine the alignment based on various parameters that the
2289 * user specified and the dynamic determination of cache line size
2292 align = calculate_alignment(flags, align, s->objsize);
2295 * SLUB stores one object immediately after another beginning from
2296 * offset 0. In order to align the objects we have to simply size
2297 * each object to conform to the alignment.
2299 size = ALIGN(size, align);
2301 if (forced_order >= 0)
2302 order = forced_order;
2304 order = calculate_order(size);
2311 s->allocflags |= __GFP_COMP;
2313 if (s->flags & SLAB_CACHE_DMA)
2314 s->allocflags |= SLUB_DMA;
2316 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2317 s->allocflags |= __GFP_RECLAIMABLE;
2320 * Determine the number of objects per slab
2322 s->oo = oo_make(order, size);
2323 s->min = oo_make(get_order(size), size);
2324 if (oo_objects(s->oo) > oo_objects(s->max))
2327 return !!oo_objects(s->oo);
2331 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2332 const char *name, size_t size,
2333 size_t align, unsigned long flags,
2334 void (*ctor)(void *))
2336 memset(s, 0, kmem_size);
2341 s->flags = kmem_cache_flags(size, flags, name, ctor);
2343 if (!calculate_sizes(s, -1))
2348 s->remote_node_defrag_ratio = 1000;
2350 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2353 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2355 free_kmem_cache_nodes(s);
2357 if (flags & SLAB_PANIC)
2358 panic("Cannot create slab %s size=%lu realsize=%u "
2359 "order=%u offset=%u flags=%lx\n",
2360 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2366 * Check if a given pointer is valid
2368 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2372 page = get_object_page(object);
2374 if (!page || s != page->slab)
2375 /* No slab or wrong slab */
2378 if (!check_valid_pointer(s, page, object))
2382 * We could also check if the object is on the slabs freelist.
2383 * But this would be too expensive and it seems that the main
2384 * purpose of kmem_ptr_valid() is to check if the object belongs
2385 * to a certain slab.
2389 EXPORT_SYMBOL(kmem_ptr_validate);
2392 * Determine the size of a slab object
2394 unsigned int kmem_cache_size(struct kmem_cache *s)
2398 EXPORT_SYMBOL(kmem_cache_size);
2400 const char *kmem_cache_name(struct kmem_cache *s)
2404 EXPORT_SYMBOL(kmem_cache_name);
2406 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2409 #ifdef CONFIG_SLUB_DEBUG
2410 void *addr = page_address(page);
2412 DECLARE_BITMAP(map, page->objects);
2414 bitmap_zero(map, page->objects);
2415 slab_err(s, page, "%s", text);
2417 for_each_free_object(p, s, page->freelist)
2418 set_bit(slab_index(p, s, addr), map);
2420 for_each_object(p, s, addr, page->objects) {
2422 if (!test_bit(slab_index(p, s, addr), map)) {
2423 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2425 print_tracking(s, p);
2433 * Attempt to free all partial slabs on a node.
2435 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2437 unsigned long flags;
2438 struct page *page, *h;
2440 spin_lock_irqsave(&n->list_lock, flags);
2441 list_for_each_entry_safe(page, h, &n->partial, lru) {
2443 list_del(&page->lru);
2444 discard_slab(s, page);
2447 list_slab_objects(s, page,
2448 "Objects remaining on kmem_cache_close()");
2451 spin_unlock_irqrestore(&n->list_lock, flags);
2455 * Release all resources used by a slab cache.
2457 static inline int kmem_cache_close(struct kmem_cache *s)
2463 /* Attempt to free all objects */
2464 free_kmem_cache_cpus(s);
2465 for_each_node_state(node, N_NORMAL_MEMORY) {
2466 struct kmem_cache_node *n = get_node(s, node);
2469 if (n->nr_partial || slabs_node(s, node))
2472 free_kmem_cache_nodes(s);
2477 * Close a cache and release the kmem_cache structure
2478 * (must be used for caches created using kmem_cache_create)
2480 void kmem_cache_destroy(struct kmem_cache *s)
2482 down_write(&slub_lock);
2486 up_write(&slub_lock);
2487 if (kmem_cache_close(s)) {
2488 printk(KERN_ERR "SLUB %s: %s called for cache that "
2489 "still has objects.\n", s->name, __func__);
2492 sysfs_slab_remove(s);
2494 up_write(&slub_lock);
2496 EXPORT_SYMBOL(kmem_cache_destroy);
2498 /********************************************************************
2500 *******************************************************************/
2502 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2503 EXPORT_SYMBOL(kmalloc_caches);
2505 static int __init setup_slub_min_order(char *str)
2507 get_option(&str, &slub_min_order);
2512 __setup("slub_min_order=", setup_slub_min_order);
2514 static int __init setup_slub_max_order(char *str)
2516 get_option(&str, &slub_max_order);
2521 __setup("slub_max_order=", setup_slub_max_order);
2523 static int __init setup_slub_min_objects(char *str)
2525 get_option(&str, &slub_min_objects);
2530 __setup("slub_min_objects=", setup_slub_min_objects);
2532 static int __init setup_slub_nomerge(char *str)
2538 __setup("slub_nomerge", setup_slub_nomerge);
2540 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2541 const char *name, int size, gfp_t gfp_flags)
2543 unsigned int flags = 0;
2545 if (gfp_flags & SLUB_DMA)
2546 flags = SLAB_CACHE_DMA;
2548 down_write(&slub_lock);
2549 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2553 list_add(&s->list, &slab_caches);
2554 up_write(&slub_lock);
2555 if (sysfs_slab_add(s))
2560 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2563 #ifdef CONFIG_ZONE_DMA
2564 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2566 static void sysfs_add_func(struct work_struct *w)
2568 struct kmem_cache *s;
2570 down_write(&slub_lock);
2571 list_for_each_entry(s, &slab_caches, list) {
2572 if (s->flags & __SYSFS_ADD_DEFERRED) {
2573 s->flags &= ~__SYSFS_ADD_DEFERRED;
2577 up_write(&slub_lock);
2580 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2582 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2584 struct kmem_cache *s;
2588 s = kmalloc_caches_dma[index];
2592 /* Dynamically create dma cache */
2593 if (flags & __GFP_WAIT)
2594 down_write(&slub_lock);
2596 if (!down_write_trylock(&slub_lock))
2600 if (kmalloc_caches_dma[index])
2603 realsize = kmalloc_caches[index].objsize;
2604 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2605 (unsigned int)realsize);
2606 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2608 if (!s || !text || !kmem_cache_open(s, flags, text,
2609 realsize, ARCH_KMALLOC_MINALIGN,
2610 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2616 list_add(&s->list, &slab_caches);
2617 kmalloc_caches_dma[index] = s;
2619 schedule_work(&sysfs_add_work);
2622 up_write(&slub_lock);
2624 return kmalloc_caches_dma[index];
2629 * Conversion table for small slabs sizes / 8 to the index in the
2630 * kmalloc array. This is necessary for slabs < 192 since we have non power
2631 * of two cache sizes there. The size of larger slabs can be determined using
2634 static s8 size_index[24] = {
2661 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2667 return ZERO_SIZE_PTR;
2669 index = size_index[(size - 1) / 8];
2671 index = fls(size - 1);
2673 #ifdef CONFIG_ZONE_DMA
2674 if (unlikely((flags & SLUB_DMA)))
2675 return dma_kmalloc_cache(index, flags);
2678 return &kmalloc_caches[index];
2681 void *__kmalloc(size_t size, gfp_t flags)
2683 struct kmem_cache *s;
2686 if (unlikely(size > PAGE_SIZE))
2687 return kmalloc_large(size, flags);
2689 s = get_slab(size, flags);
2691 if (unlikely(ZERO_OR_NULL_PTR(s)))
2694 ret = slab_alloc(s, flags, -1, _RET_IP_);
2696 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, ret,
2697 size, s->size, flags);
2701 EXPORT_SYMBOL(__kmalloc);
2703 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2705 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2709 return page_address(page);
2715 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2717 struct kmem_cache *s;
2720 if (unlikely(size > PAGE_SIZE)) {
2721 ret = kmalloc_large_node(size, flags, node);
2723 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC,
2725 size, PAGE_SIZE << get_order(size),
2731 s = get_slab(size, flags);
2733 if (unlikely(ZERO_OR_NULL_PTR(s)))
2736 ret = slab_alloc(s, flags, node, _RET_IP_);
2738 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, ret,
2739 size, s->size, flags, node);
2743 EXPORT_SYMBOL(__kmalloc_node);
2746 size_t ksize(const void *object)
2749 struct kmem_cache *s;
2751 if (unlikely(object == ZERO_SIZE_PTR))
2754 page = virt_to_head_page(object);
2756 if (unlikely(!PageSlab(page))) {
2757 WARN_ON(!PageCompound(page));
2758 return PAGE_SIZE << compound_order(page);
2762 #ifdef CONFIG_SLUB_DEBUG
2764 * Debugging requires use of the padding between object
2765 * and whatever may come after it.
2767 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2772 * If we have the need to store the freelist pointer
2773 * back there or track user information then we can
2774 * only use the space before that information.
2776 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2779 * Else we can use all the padding etc for the allocation
2784 void kfree(const void *x)
2787 void *object = (void *)x;
2789 if (unlikely(ZERO_OR_NULL_PTR(x)))
2792 page = virt_to_head_page(x);
2793 if (unlikely(!PageSlab(page))) {
2794 BUG_ON(!PageCompound(page));
2798 slab_free(page->slab, page, object, _RET_IP_);
2800 kmemtrace_mark_free(KMEMTRACE_TYPE_KMALLOC, _RET_IP_, x);
2802 EXPORT_SYMBOL(kfree);
2805 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2806 * the remaining slabs by the number of items in use. The slabs with the
2807 * most items in use come first. New allocations will then fill those up
2808 * and thus they can be removed from the partial lists.
2810 * The slabs with the least items are placed last. This results in them
2811 * being allocated from last increasing the chance that the last objects
2812 * are freed in them.
2814 int kmem_cache_shrink(struct kmem_cache *s)
2818 struct kmem_cache_node *n;
2821 int objects = oo_objects(s->max);
2822 struct list_head *slabs_by_inuse =
2823 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2824 unsigned long flags;
2826 if (!slabs_by_inuse)
2830 for_each_node_state(node, N_NORMAL_MEMORY) {
2831 n = get_node(s, node);
2836 for (i = 0; i < objects; i++)
2837 INIT_LIST_HEAD(slabs_by_inuse + i);
2839 spin_lock_irqsave(&n->list_lock, flags);
2842 * Build lists indexed by the items in use in each slab.
2844 * Note that concurrent frees may occur while we hold the
2845 * list_lock. page->inuse here is the upper limit.
2847 list_for_each_entry_safe(page, t, &n->partial, lru) {
2848 if (!page->inuse && slab_trylock(page)) {
2850 * Must hold slab lock here because slab_free
2851 * may have freed the last object and be
2852 * waiting to release the slab.
2854 list_del(&page->lru);
2857 discard_slab(s, page);
2859 list_move(&page->lru,
2860 slabs_by_inuse + page->inuse);
2865 * Rebuild the partial list with the slabs filled up most
2866 * first and the least used slabs at the end.
2868 for (i = objects - 1; i >= 0; i--)
2869 list_splice(slabs_by_inuse + i, n->partial.prev);
2871 spin_unlock_irqrestore(&n->list_lock, flags);
2874 kfree(slabs_by_inuse);
2877 EXPORT_SYMBOL(kmem_cache_shrink);
2879 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2880 static int slab_mem_going_offline_callback(void *arg)
2882 struct kmem_cache *s;
2884 down_read(&slub_lock);
2885 list_for_each_entry(s, &slab_caches, list)
2886 kmem_cache_shrink(s);
2887 up_read(&slub_lock);
2892 static void slab_mem_offline_callback(void *arg)
2894 struct kmem_cache_node *n;
2895 struct kmem_cache *s;
2896 struct memory_notify *marg = arg;
2899 offline_node = marg->status_change_nid;
2902 * If the node still has available memory. we need kmem_cache_node
2905 if (offline_node < 0)
2908 down_read(&slub_lock);
2909 list_for_each_entry(s, &slab_caches, list) {
2910 n = get_node(s, offline_node);
2913 * if n->nr_slabs > 0, slabs still exist on the node
2914 * that is going down. We were unable to free them,
2915 * and offline_pages() function shoudn't call this
2916 * callback. So, we must fail.
2918 BUG_ON(slabs_node(s, offline_node));
2920 s->node[offline_node] = NULL;
2921 kmem_cache_free(kmalloc_caches, n);
2924 up_read(&slub_lock);
2927 static int slab_mem_going_online_callback(void *arg)
2929 struct kmem_cache_node *n;
2930 struct kmem_cache *s;
2931 struct memory_notify *marg = arg;
2932 int nid = marg->status_change_nid;
2936 * If the node's memory is already available, then kmem_cache_node is
2937 * already created. Nothing to do.
2943 * We are bringing a node online. No memory is available yet. We must
2944 * allocate a kmem_cache_node structure in order to bring the node
2947 down_read(&slub_lock);
2948 list_for_each_entry(s, &slab_caches, list) {
2950 * XXX: kmem_cache_alloc_node will fallback to other nodes
2951 * since memory is not yet available from the node that
2954 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2959 init_kmem_cache_node(n, s);
2963 up_read(&slub_lock);
2967 static int slab_memory_callback(struct notifier_block *self,
2968 unsigned long action, void *arg)
2973 case MEM_GOING_ONLINE:
2974 ret = slab_mem_going_online_callback(arg);
2976 case MEM_GOING_OFFLINE:
2977 ret = slab_mem_going_offline_callback(arg);
2980 case MEM_CANCEL_ONLINE:
2981 slab_mem_offline_callback(arg);
2984 case MEM_CANCEL_OFFLINE:
2988 ret = notifier_from_errno(ret);
2994 #endif /* CONFIG_MEMORY_HOTPLUG */
2996 /********************************************************************
2997 * Basic setup of slabs
2998 *******************************************************************/
3000 void __init kmem_cache_init(void)
3009 * Must first have the slab cache available for the allocations of the
3010 * struct kmem_cache_node's. There is special bootstrap code in
3011 * kmem_cache_open for slab_state == DOWN.
3013 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3014 sizeof(struct kmem_cache_node), GFP_KERNEL);
3015 kmalloc_caches[0].refcount = -1;
3018 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3021 /* Able to allocate the per node structures */
3022 slab_state = PARTIAL;
3024 /* Caches that are not of the two-to-the-power-of size */
3025 if (KMALLOC_MIN_SIZE <= 64) {
3026 create_kmalloc_cache(&kmalloc_caches[1],
3027 "kmalloc-96", 96, GFP_KERNEL);
3029 create_kmalloc_cache(&kmalloc_caches[2],
3030 "kmalloc-192", 192, GFP_KERNEL);
3034 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
3035 create_kmalloc_cache(&kmalloc_caches[i],
3036 "kmalloc", 1 << i, GFP_KERNEL);
3042 * Patch up the size_index table if we have strange large alignment
3043 * requirements for the kmalloc array. This is only the case for
3044 * MIPS it seems. The standard arches will not generate any code here.
3046 * Largest permitted alignment is 256 bytes due to the way we
3047 * handle the index determination for the smaller caches.
3049 * Make sure that nothing crazy happens if someone starts tinkering
3050 * around with ARCH_KMALLOC_MINALIGN
3052 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3053 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3055 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3056 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3058 if (KMALLOC_MIN_SIZE == 128) {
3060 * The 192 byte sized cache is not used if the alignment
3061 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3064 for (i = 128 + 8; i <= 192; i += 8)
3065 size_index[(i - 1) / 8] = 8;
3070 /* Provide the correct kmalloc names now that the caches are up */
3071 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3072 kmalloc_caches[i]. name =
3073 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3076 register_cpu_notifier(&slab_notifier);
3077 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3078 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3080 kmem_size = sizeof(struct kmem_cache);
3084 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3085 " CPUs=%d, Nodes=%d\n",
3086 caches, cache_line_size(),
3087 slub_min_order, slub_max_order, slub_min_objects,
3088 nr_cpu_ids, nr_node_ids);
3092 * Find a mergeable slab cache
3094 static int slab_unmergeable(struct kmem_cache *s)
3096 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3103 * We may have set a slab to be unmergeable during bootstrap.
3105 if (s->refcount < 0)
3111 static struct kmem_cache *find_mergeable(size_t size,
3112 size_t align, unsigned long flags, const char *name,
3113 void (*ctor)(void *))
3115 struct kmem_cache *s;
3117 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3123 size = ALIGN(size, sizeof(void *));
3124 align = calculate_alignment(flags, align, size);
3125 size = ALIGN(size, align);
3126 flags = kmem_cache_flags(size, flags, name, NULL);
3128 list_for_each_entry(s, &slab_caches, list) {
3129 if (slab_unmergeable(s))
3135 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3138 * Check if alignment is compatible.
3139 * Courtesy of Adrian Drzewiecki
3141 if ((s->size & ~(align - 1)) != s->size)
3144 if (s->size - size >= sizeof(void *))
3152 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3153 size_t align, unsigned long flags, void (*ctor)(void *))
3155 struct kmem_cache *s;
3157 down_write(&slub_lock);
3158 s = find_mergeable(size, align, flags, name, ctor);
3164 * Adjust the object sizes so that we clear
3165 * the complete object on kzalloc.
3167 s->objsize = max(s->objsize, (int)size);
3170 * And then we need to update the object size in the
3171 * per cpu structures
3173 for_each_online_cpu(cpu)
3174 get_cpu_slab(s, cpu)->objsize = s->objsize;
3176 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3177 up_write(&slub_lock);
3179 if (sysfs_slab_alias(s, name))
3184 s = kmalloc(kmem_size, GFP_KERNEL);
3186 if (kmem_cache_open(s, GFP_KERNEL, name,
3187 size, align, flags, ctor)) {
3188 list_add(&s->list, &slab_caches);
3189 up_write(&slub_lock);
3190 if (sysfs_slab_add(s))
3196 up_write(&slub_lock);
3199 if (flags & SLAB_PANIC)
3200 panic("Cannot create slabcache %s\n", name);
3205 EXPORT_SYMBOL(kmem_cache_create);
3209 * Use the cpu notifier to insure that the cpu slabs are flushed when
3212 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3213 unsigned long action, void *hcpu)
3215 long cpu = (long)hcpu;
3216 struct kmem_cache *s;
3217 unsigned long flags;
3220 case CPU_UP_PREPARE:
3221 case CPU_UP_PREPARE_FROZEN:
3222 init_alloc_cpu_cpu(cpu);
3223 down_read(&slub_lock);
3224 list_for_each_entry(s, &slab_caches, list)
3225 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3227 up_read(&slub_lock);
3230 case CPU_UP_CANCELED:
3231 case CPU_UP_CANCELED_FROZEN:
3233 case CPU_DEAD_FROZEN:
3234 down_read(&slub_lock);
3235 list_for_each_entry(s, &slab_caches, list) {
3236 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3238 local_irq_save(flags);
3239 __flush_cpu_slab(s, cpu);
3240 local_irq_restore(flags);
3241 free_kmem_cache_cpu(c, cpu);
3242 s->cpu_slab[cpu] = NULL;
3244 up_read(&slub_lock);
3252 static struct notifier_block __cpuinitdata slab_notifier = {
3253 .notifier_call = slab_cpuup_callback
3258 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3260 struct kmem_cache *s;
3263 if (unlikely(size > PAGE_SIZE))
3264 return kmalloc_large(size, gfpflags);
3266 s = get_slab(size, gfpflags);
3268 if (unlikely(ZERO_OR_NULL_PTR(s)))
3271 ret = slab_alloc(s, gfpflags, -1, caller);
3273 /* Honor the call site pointer we recieved. */
3274 kmemtrace_mark_alloc(KMEMTRACE_TYPE_KMALLOC, caller, ret, size,
3280 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3281 int node, unsigned long caller)
3283 struct kmem_cache *s;
3286 if (unlikely(size > PAGE_SIZE))
3287 return kmalloc_large_node(size, gfpflags, node);
3289 s = get_slab(size, gfpflags);
3291 if (unlikely(ZERO_OR_NULL_PTR(s)))
3294 ret = slab_alloc(s, gfpflags, node, caller);
3296 /* Honor the call site pointer we recieved. */
3297 kmemtrace_mark_alloc_node(KMEMTRACE_TYPE_KMALLOC, caller, ret,
3298 size, s->size, gfpflags, node);
3303 #ifdef CONFIG_SLUB_DEBUG
3304 static unsigned long count_partial(struct kmem_cache_node *n,
3305 int (*get_count)(struct page *))
3307 unsigned long flags;
3308 unsigned long x = 0;
3311 spin_lock_irqsave(&n->list_lock, flags);
3312 list_for_each_entry(page, &n->partial, lru)
3313 x += get_count(page);
3314 spin_unlock_irqrestore(&n->list_lock, flags);
3318 static int count_inuse(struct page *page)
3323 static int count_total(struct page *page)
3325 return page->objects;
3328 static int count_free(struct page *page)
3330 return page->objects - page->inuse;
3333 static int validate_slab(struct kmem_cache *s, struct page *page,
3337 void *addr = page_address(page);
3339 if (!check_slab(s, page) ||
3340 !on_freelist(s, page, NULL))
3343 /* Now we know that a valid freelist exists */
3344 bitmap_zero(map, page->objects);
3346 for_each_free_object(p, s, page->freelist) {
3347 set_bit(slab_index(p, s, addr), map);
3348 if (!check_object(s, page, p, 0))
3352 for_each_object(p, s, addr, page->objects)
3353 if (!test_bit(slab_index(p, s, addr), map))
3354 if (!check_object(s, page, p, 1))
3359 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3362 if (slab_trylock(page)) {
3363 validate_slab(s, page, map);
3366 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3369 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3370 if (!PageSlubDebug(page))
3371 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3372 "on slab 0x%p\n", s->name, page);
3374 if (PageSlubDebug(page))
3375 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3376 "slab 0x%p\n", s->name, page);
3380 static int validate_slab_node(struct kmem_cache *s,
3381 struct kmem_cache_node *n, unsigned long *map)
3383 unsigned long count = 0;
3385 unsigned long flags;
3387 spin_lock_irqsave(&n->list_lock, flags);
3389 list_for_each_entry(page, &n->partial, lru) {
3390 validate_slab_slab(s, page, map);
3393 if (count != n->nr_partial)
3394 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3395 "counter=%ld\n", s->name, count, n->nr_partial);
3397 if (!(s->flags & SLAB_STORE_USER))
3400 list_for_each_entry(page, &n->full, lru) {
3401 validate_slab_slab(s, page, map);
3404 if (count != atomic_long_read(&n->nr_slabs))
3405 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3406 "counter=%ld\n", s->name, count,
3407 atomic_long_read(&n->nr_slabs));
3410 spin_unlock_irqrestore(&n->list_lock, flags);
3414 static long validate_slab_cache(struct kmem_cache *s)
3417 unsigned long count = 0;
3418 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3419 sizeof(unsigned long), GFP_KERNEL);
3425 for_each_node_state(node, N_NORMAL_MEMORY) {
3426 struct kmem_cache_node *n = get_node(s, node);
3428 count += validate_slab_node(s, n, map);
3434 #ifdef SLUB_RESILIENCY_TEST
3435 static void resiliency_test(void)
3439 printk(KERN_ERR "SLUB resiliency testing\n");
3440 printk(KERN_ERR "-----------------------\n");
3441 printk(KERN_ERR "A. Corruption after allocation\n");
3443 p = kzalloc(16, GFP_KERNEL);
3445 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3446 " 0x12->0x%p\n\n", p + 16);
3448 validate_slab_cache(kmalloc_caches + 4);
3450 /* Hmmm... The next two are dangerous */
3451 p = kzalloc(32, GFP_KERNEL);
3452 p[32 + sizeof(void *)] = 0x34;
3453 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3454 " 0x34 -> -0x%p\n", p);
3456 "If allocated object is overwritten then not detectable\n\n");
3458 validate_slab_cache(kmalloc_caches + 5);
3459 p = kzalloc(64, GFP_KERNEL);
3460 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3462 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3465 "If allocated object is overwritten then not detectable\n\n");
3466 validate_slab_cache(kmalloc_caches + 6);
3468 printk(KERN_ERR "\nB. Corruption after free\n");
3469 p = kzalloc(128, GFP_KERNEL);
3472 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3473 validate_slab_cache(kmalloc_caches + 7);
3475 p = kzalloc(256, GFP_KERNEL);
3478 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3480 validate_slab_cache(kmalloc_caches + 8);
3482 p = kzalloc(512, GFP_KERNEL);
3485 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3486 validate_slab_cache(kmalloc_caches + 9);
3489 static void resiliency_test(void) {};
3493 * Generate lists of code addresses where slabcache objects are allocated
3498 unsigned long count;
3511 unsigned long count;
3512 struct location *loc;
3515 static void free_loc_track(struct loc_track *t)
3518 free_pages((unsigned long)t->loc,
3519 get_order(sizeof(struct location) * t->max));
3522 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3527 order = get_order(sizeof(struct location) * max);
3529 l = (void *)__get_free_pages(flags, order);
3534 memcpy(l, t->loc, sizeof(struct location) * t->count);
3542 static int add_location(struct loc_track *t, struct kmem_cache *s,
3543 const struct track *track)
3545 long start, end, pos;
3547 unsigned long caddr;
3548 unsigned long age = jiffies - track->when;
3554 pos = start + (end - start + 1) / 2;
3557 * There is nothing at "end". If we end up there
3558 * we need to add something to before end.
3563 caddr = t->loc[pos].addr;
3564 if (track->addr == caddr) {
3570 if (age < l->min_time)
3572 if (age > l->max_time)
3575 if (track->pid < l->min_pid)
3576 l->min_pid = track->pid;
3577 if (track->pid > l->max_pid)
3578 l->max_pid = track->pid;
3580 cpu_set(track->cpu, l->cpus);
3582 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3586 if (track->addr < caddr)
3593 * Not found. Insert new tracking element.
3595 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3601 (t->count - pos) * sizeof(struct location));
3604 l->addr = track->addr;
3608 l->min_pid = track->pid;
3609 l->max_pid = track->pid;
3610 cpus_clear(l->cpus);
3611 cpu_set(track->cpu, l->cpus);
3612 nodes_clear(l->nodes);
3613 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3617 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3618 struct page *page, enum track_item alloc)
3620 void *addr = page_address(page);
3621 DECLARE_BITMAP(map, page->objects);
3624 bitmap_zero(map, page->objects);
3625 for_each_free_object(p, s, page->freelist)
3626 set_bit(slab_index(p, s, addr), map);
3628 for_each_object(p, s, addr, page->objects)
3629 if (!test_bit(slab_index(p, s, addr), map))
3630 add_location(t, s, get_track(s, p, alloc));
3633 static int list_locations(struct kmem_cache *s, char *buf,
3634 enum track_item alloc)
3638 struct loc_track t = { 0, 0, NULL };
3641 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3643 return sprintf(buf, "Out of memory\n");
3645 /* Push back cpu slabs */
3648 for_each_node_state(node, N_NORMAL_MEMORY) {
3649 struct kmem_cache_node *n = get_node(s, node);
3650 unsigned long flags;
3653 if (!atomic_long_read(&n->nr_slabs))
3656 spin_lock_irqsave(&n->list_lock, flags);
3657 list_for_each_entry(page, &n->partial, lru)
3658 process_slab(&t, s, page, alloc);
3659 list_for_each_entry(page, &n->full, lru)
3660 process_slab(&t, s, page, alloc);
3661 spin_unlock_irqrestore(&n->list_lock, flags);
3664 for (i = 0; i < t.count; i++) {
3665 struct location *l = &t.loc[i];
3667 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3669 len += sprintf(buf + len, "%7ld ", l->count);
3672 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3674 len += sprintf(buf + len, "<not-available>");
3676 if (l->sum_time != l->min_time) {
3677 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3679 (long)div_u64(l->sum_time, l->count),
3682 len += sprintf(buf + len, " age=%ld",
3685 if (l->min_pid != l->max_pid)
3686 len += sprintf(buf + len, " pid=%ld-%ld",
3687 l->min_pid, l->max_pid);
3689 len += sprintf(buf + len, " pid=%ld",
3692 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3693 len < PAGE_SIZE - 60) {
3694 len += sprintf(buf + len, " cpus=");
3695 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3699 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3700 len < PAGE_SIZE - 60) {
3701 len += sprintf(buf + len, " nodes=");
3702 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3706 len += sprintf(buf + len, "\n");
3711 len += sprintf(buf, "No data\n");
3715 enum slab_stat_type {
3716 SL_ALL, /* All slabs */
3717 SL_PARTIAL, /* Only partially allocated slabs */
3718 SL_CPU, /* Only slabs used for cpu caches */
3719 SL_OBJECTS, /* Determine allocated objects not slabs */
3720 SL_TOTAL /* Determine object capacity not slabs */
3723 #define SO_ALL (1 << SL_ALL)
3724 #define SO_PARTIAL (1 << SL_PARTIAL)
3725 #define SO_CPU (1 << SL_CPU)
3726 #define SO_OBJECTS (1 << SL_OBJECTS)
3727 #define SO_TOTAL (1 << SL_TOTAL)
3729 static ssize_t show_slab_objects(struct kmem_cache *s,
3730 char *buf, unsigned long flags)
3732 unsigned long total = 0;
3735 unsigned long *nodes;
3736 unsigned long *per_cpu;
3738 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3741 per_cpu = nodes + nr_node_ids;
3743 if (flags & SO_CPU) {
3746 for_each_possible_cpu(cpu) {
3747 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3749 if (!c || c->node < 0)
3753 if (flags & SO_TOTAL)
3754 x = c->page->objects;
3755 else if (flags & SO_OBJECTS)
3761 nodes[c->node] += x;
3767 if (flags & SO_ALL) {
3768 for_each_node_state(node, N_NORMAL_MEMORY) {
3769 struct kmem_cache_node *n = get_node(s, node);
3771 if (flags & SO_TOTAL)
3772 x = atomic_long_read(&n->total_objects);
3773 else if (flags & SO_OBJECTS)
3774 x = atomic_long_read(&n->total_objects) -
3775 count_partial(n, count_free);
3778 x = atomic_long_read(&n->nr_slabs);
3783 } else if (flags & SO_PARTIAL) {
3784 for_each_node_state(node, N_NORMAL_MEMORY) {
3785 struct kmem_cache_node *n = get_node(s, node);
3787 if (flags & SO_TOTAL)
3788 x = count_partial(n, count_total);
3789 else if (flags & SO_OBJECTS)
3790 x = count_partial(n, count_inuse);
3797 x = sprintf(buf, "%lu", total);
3799 for_each_node_state(node, N_NORMAL_MEMORY)
3801 x += sprintf(buf + x, " N%d=%lu",
3805 return x + sprintf(buf + x, "\n");
3808 static int any_slab_objects(struct kmem_cache *s)
3812 for_each_online_node(node) {
3813 struct kmem_cache_node *n = get_node(s, node);
3818 if (atomic_long_read(&n->total_objects))
3824 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3825 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3827 struct slab_attribute {
3828 struct attribute attr;
3829 ssize_t (*show)(struct kmem_cache *s, char *buf);
3830 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3833 #define SLAB_ATTR_RO(_name) \
3834 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3836 #define SLAB_ATTR(_name) \
3837 static struct slab_attribute _name##_attr = \
3838 __ATTR(_name, 0644, _name##_show, _name##_store)
3840 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3842 return sprintf(buf, "%d\n", s->size);
3844 SLAB_ATTR_RO(slab_size);
3846 static ssize_t align_show(struct kmem_cache *s, char *buf)
3848 return sprintf(buf, "%d\n", s->align);
3850 SLAB_ATTR_RO(align);
3852 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3854 return sprintf(buf, "%d\n", s->objsize);
3856 SLAB_ATTR_RO(object_size);
3858 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3860 return sprintf(buf, "%d\n", oo_objects(s->oo));
3862 SLAB_ATTR_RO(objs_per_slab);
3864 static ssize_t order_store(struct kmem_cache *s,
3865 const char *buf, size_t length)
3867 unsigned long order;
3870 err = strict_strtoul(buf, 10, &order);
3874 if (order > slub_max_order || order < slub_min_order)
3877 calculate_sizes(s, order);
3881 static ssize_t order_show(struct kmem_cache *s, char *buf)
3883 return sprintf(buf, "%d\n", oo_order(s->oo));
3887 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3890 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3892 return n + sprintf(buf + n, "\n");
3898 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3900 return sprintf(buf, "%d\n", s->refcount - 1);
3902 SLAB_ATTR_RO(aliases);
3904 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3906 return show_slab_objects(s, buf, SO_ALL);
3908 SLAB_ATTR_RO(slabs);
3910 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3912 return show_slab_objects(s, buf, SO_PARTIAL);
3914 SLAB_ATTR_RO(partial);
3916 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3918 return show_slab_objects(s, buf, SO_CPU);
3920 SLAB_ATTR_RO(cpu_slabs);
3922 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3924 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3926 SLAB_ATTR_RO(objects);
3928 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3930 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3932 SLAB_ATTR_RO(objects_partial);
3934 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3936 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3938 SLAB_ATTR_RO(total_objects);
3940 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3942 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3945 static ssize_t sanity_checks_store(struct kmem_cache *s,
3946 const char *buf, size_t length)
3948 s->flags &= ~SLAB_DEBUG_FREE;
3950 s->flags |= SLAB_DEBUG_FREE;
3953 SLAB_ATTR(sanity_checks);
3955 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3957 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3960 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3963 s->flags &= ~SLAB_TRACE;
3965 s->flags |= SLAB_TRACE;
3970 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3972 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3975 static ssize_t reclaim_account_store(struct kmem_cache *s,
3976 const char *buf, size_t length)
3978 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3980 s->flags |= SLAB_RECLAIM_ACCOUNT;
3983 SLAB_ATTR(reclaim_account);
3985 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3987 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3989 SLAB_ATTR_RO(hwcache_align);
3991 #ifdef CONFIG_ZONE_DMA
3992 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3994 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3996 SLAB_ATTR_RO(cache_dma);
3999 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4001 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4003 SLAB_ATTR_RO(destroy_by_rcu);
4005 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4007 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4010 static ssize_t red_zone_store(struct kmem_cache *s,
4011 const char *buf, size_t length)
4013 if (any_slab_objects(s))
4016 s->flags &= ~SLAB_RED_ZONE;
4018 s->flags |= SLAB_RED_ZONE;
4019 calculate_sizes(s, -1);
4022 SLAB_ATTR(red_zone);
4024 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4026 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4029 static ssize_t poison_store(struct kmem_cache *s,
4030 const char *buf, size_t length)
4032 if (any_slab_objects(s))
4035 s->flags &= ~SLAB_POISON;
4037 s->flags |= SLAB_POISON;
4038 calculate_sizes(s, -1);
4043 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4045 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4048 static ssize_t store_user_store(struct kmem_cache *s,
4049 const char *buf, size_t length)
4051 if (any_slab_objects(s))
4054 s->flags &= ~SLAB_STORE_USER;
4056 s->flags |= SLAB_STORE_USER;
4057 calculate_sizes(s, -1);
4060 SLAB_ATTR(store_user);
4062 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4067 static ssize_t validate_store(struct kmem_cache *s,
4068 const char *buf, size_t length)
4072 if (buf[0] == '1') {
4073 ret = validate_slab_cache(s);
4079 SLAB_ATTR(validate);
4081 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4086 static ssize_t shrink_store(struct kmem_cache *s,
4087 const char *buf, size_t length)
4089 if (buf[0] == '1') {
4090 int rc = kmem_cache_shrink(s);
4100 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4102 if (!(s->flags & SLAB_STORE_USER))
4104 return list_locations(s, buf, TRACK_ALLOC);
4106 SLAB_ATTR_RO(alloc_calls);
4108 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4110 if (!(s->flags & SLAB_STORE_USER))
4112 return list_locations(s, buf, TRACK_FREE);
4114 SLAB_ATTR_RO(free_calls);
4117 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4119 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4122 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4123 const char *buf, size_t length)
4125 unsigned long ratio;
4128 err = strict_strtoul(buf, 10, &ratio);
4133 s->remote_node_defrag_ratio = ratio * 10;
4137 SLAB_ATTR(remote_node_defrag_ratio);
4140 #ifdef CONFIG_SLUB_STATS
4141 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4143 unsigned long sum = 0;
4146 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4151 for_each_online_cpu(cpu) {
4152 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4158 len = sprintf(buf, "%lu", sum);
4161 for_each_online_cpu(cpu) {
4162 if (data[cpu] && len < PAGE_SIZE - 20)
4163 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4167 return len + sprintf(buf + len, "\n");
4170 #define STAT_ATTR(si, text) \
4171 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4173 return show_stat(s, buf, si); \
4175 SLAB_ATTR_RO(text); \
4177 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4178 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4179 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4180 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4181 STAT_ATTR(FREE_FROZEN, free_frozen);
4182 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4183 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4184 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4185 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4186 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4187 STAT_ATTR(FREE_SLAB, free_slab);
4188 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4189 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4190 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4191 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4192 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4193 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4194 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4197 static struct attribute *slab_attrs[] = {
4198 &slab_size_attr.attr,
4199 &object_size_attr.attr,
4200 &objs_per_slab_attr.attr,
4203 &objects_partial_attr.attr,
4204 &total_objects_attr.attr,
4207 &cpu_slabs_attr.attr,
4211 &sanity_checks_attr.attr,
4213 &hwcache_align_attr.attr,
4214 &reclaim_account_attr.attr,
4215 &destroy_by_rcu_attr.attr,
4216 &red_zone_attr.attr,
4218 &store_user_attr.attr,
4219 &validate_attr.attr,
4221 &alloc_calls_attr.attr,
4222 &free_calls_attr.attr,
4223 #ifdef CONFIG_ZONE_DMA
4224 &cache_dma_attr.attr,
4227 &remote_node_defrag_ratio_attr.attr,
4229 #ifdef CONFIG_SLUB_STATS
4230 &alloc_fastpath_attr.attr,
4231 &alloc_slowpath_attr.attr,
4232 &free_fastpath_attr.attr,
4233 &free_slowpath_attr.attr,
4234 &free_frozen_attr.attr,
4235 &free_add_partial_attr.attr,
4236 &free_remove_partial_attr.attr,
4237 &alloc_from_partial_attr.attr,
4238 &alloc_slab_attr.attr,
4239 &alloc_refill_attr.attr,
4240 &free_slab_attr.attr,
4241 &cpuslab_flush_attr.attr,
4242 &deactivate_full_attr.attr,
4243 &deactivate_empty_attr.attr,
4244 &deactivate_to_head_attr.attr,
4245 &deactivate_to_tail_attr.attr,
4246 &deactivate_remote_frees_attr.attr,
4247 &order_fallback_attr.attr,
4252 static struct attribute_group slab_attr_group = {
4253 .attrs = slab_attrs,
4256 static ssize_t slab_attr_show(struct kobject *kobj,
4257 struct attribute *attr,
4260 struct slab_attribute *attribute;
4261 struct kmem_cache *s;
4264 attribute = to_slab_attr(attr);
4267 if (!attribute->show)
4270 err = attribute->show(s, buf);
4275 static ssize_t slab_attr_store(struct kobject *kobj,
4276 struct attribute *attr,
4277 const char *buf, size_t len)
4279 struct slab_attribute *attribute;
4280 struct kmem_cache *s;
4283 attribute = to_slab_attr(attr);
4286 if (!attribute->store)
4289 err = attribute->store(s, buf, len);
4294 static void kmem_cache_release(struct kobject *kobj)
4296 struct kmem_cache *s = to_slab(kobj);
4301 static struct sysfs_ops slab_sysfs_ops = {
4302 .show = slab_attr_show,
4303 .store = slab_attr_store,
4306 static struct kobj_type slab_ktype = {
4307 .sysfs_ops = &slab_sysfs_ops,
4308 .release = kmem_cache_release
4311 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4313 struct kobj_type *ktype = get_ktype(kobj);
4315 if (ktype == &slab_ktype)
4320 static struct kset_uevent_ops slab_uevent_ops = {
4321 .filter = uevent_filter,
4324 static struct kset *slab_kset;
4326 #define ID_STR_LENGTH 64
4328 /* Create a unique string id for a slab cache:
4330 * Format :[flags-]size
4332 static char *create_unique_id(struct kmem_cache *s)
4334 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4341 * First flags affecting slabcache operations. We will only
4342 * get here for aliasable slabs so we do not need to support
4343 * too many flags. The flags here must cover all flags that
4344 * are matched during merging to guarantee that the id is
4347 if (s->flags & SLAB_CACHE_DMA)
4349 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4351 if (s->flags & SLAB_DEBUG_FREE)
4355 p += sprintf(p, "%07d", s->size);
4356 BUG_ON(p > name + ID_STR_LENGTH - 1);
4360 static int sysfs_slab_add(struct kmem_cache *s)
4366 if (slab_state < SYSFS)
4367 /* Defer until later */
4370 unmergeable = slab_unmergeable(s);
4373 * Slabcache can never be merged so we can use the name proper.
4374 * This is typically the case for debug situations. In that
4375 * case we can catch duplicate names easily.
4377 sysfs_remove_link(&slab_kset->kobj, s->name);
4381 * Create a unique name for the slab as a target
4384 name = create_unique_id(s);
4387 s->kobj.kset = slab_kset;
4388 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4390 kobject_put(&s->kobj);
4394 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4397 kobject_uevent(&s->kobj, KOBJ_ADD);
4399 /* Setup first alias */
4400 sysfs_slab_alias(s, s->name);
4406 static void sysfs_slab_remove(struct kmem_cache *s)
4408 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4409 kobject_del(&s->kobj);
4410 kobject_put(&s->kobj);
4414 * Need to buffer aliases during bootup until sysfs becomes
4415 * available lest we loose that information.
4417 struct saved_alias {
4418 struct kmem_cache *s;
4420 struct saved_alias *next;
4423 static struct saved_alias *alias_list;
4425 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4427 struct saved_alias *al;
4429 if (slab_state == SYSFS) {
4431 * If we have a leftover link then remove it.
4433 sysfs_remove_link(&slab_kset->kobj, name);
4434 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4437 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4443 al->next = alias_list;
4448 static int __init slab_sysfs_init(void)
4450 struct kmem_cache *s;
4453 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4455 printk(KERN_ERR "Cannot register slab subsystem.\n");
4461 list_for_each_entry(s, &slab_caches, list) {
4462 err = sysfs_slab_add(s);
4464 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4465 " to sysfs\n", s->name);
4468 while (alias_list) {
4469 struct saved_alias *al = alias_list;
4471 alias_list = alias_list->next;
4472 err = sysfs_slab_alias(al->s, al->name);
4474 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4475 " %s to sysfs\n", s->name);
4483 __initcall(slab_sysfs_init);
4487 * The /proc/slabinfo ABI
4489 #ifdef CONFIG_SLABINFO
4490 static void print_slabinfo_header(struct seq_file *m)
4492 seq_puts(m, "slabinfo - version: 2.1\n");
4493 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4494 "<objperslab> <pagesperslab>");
4495 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4496 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4500 static void *s_start(struct seq_file *m, loff_t *pos)
4504 down_read(&slub_lock);
4506 print_slabinfo_header(m);
4508 return seq_list_start(&slab_caches, *pos);
4511 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4513 return seq_list_next(p, &slab_caches, pos);
4516 static void s_stop(struct seq_file *m, void *p)
4518 up_read(&slub_lock);
4521 static int s_show(struct seq_file *m, void *p)
4523 unsigned long nr_partials = 0;
4524 unsigned long nr_slabs = 0;
4525 unsigned long nr_inuse = 0;
4526 unsigned long nr_objs = 0;
4527 unsigned long nr_free = 0;
4528 struct kmem_cache *s;
4531 s = list_entry(p, struct kmem_cache, list);
4533 for_each_online_node(node) {
4534 struct kmem_cache_node *n = get_node(s, node);
4539 nr_partials += n->nr_partial;
4540 nr_slabs += atomic_long_read(&n->nr_slabs);
4541 nr_objs += atomic_long_read(&n->total_objects);
4542 nr_free += count_partial(n, count_free);
4545 nr_inuse = nr_objs - nr_free;
4547 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4548 nr_objs, s->size, oo_objects(s->oo),
4549 (1 << oo_order(s->oo)));
4550 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4551 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4557 static const struct seq_operations slabinfo_op = {
4564 static int slabinfo_open(struct inode *inode, struct file *file)
4566 return seq_open(file, &slabinfo_op);
4569 static const struct file_operations proc_slabinfo_operations = {
4570 .open = slabinfo_open,
4572 .llseek = seq_lseek,
4573 .release = seq_release,
4576 static int __init slab_proc_init(void)
4578 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4581 module_init(slab_proc_init);
4582 #endif /* CONFIG_SLABINFO */