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/fault-inject.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 void *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, void *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, NULL);
396 set_track(s, object, TRACK_ALLOC, NULL);
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, 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, void *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, void *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, void *addr) { return 0; }
1035 static inline int free_debug_processing(struct kmem_cache *s,
1036 struct page *page, void *object, void *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,
1504 gfp_t gfpflags, int node, void *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, void *addr)
1591 struct kmem_cache_cpu *c;
1592 unsigned long flags;
1593 unsigned int objsize;
1595 if (should_failslab(s->objsize, gfpflags))
1598 local_irq_save(flags);
1599 c = get_cpu_slab(s, smp_processor_id());
1600 objsize = c->objsize;
1601 if (unlikely(!c->freelist || !node_match(c, node)))
1603 object = __slab_alloc(s, gfpflags, node, addr, c);
1606 object = c->freelist;
1607 c->freelist = object[c->offset];
1608 stat(c, ALLOC_FASTPATH);
1610 local_irq_restore(flags);
1612 if (unlikely((gfpflags & __GFP_ZERO) && object))
1613 memset(object, 0, objsize);
1618 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1620 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1622 EXPORT_SYMBOL(kmem_cache_alloc);
1625 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1627 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1629 EXPORT_SYMBOL(kmem_cache_alloc_node);
1633 * Slow patch handling. This may still be called frequently since objects
1634 * have a longer lifetime than the cpu slabs in most processing loads.
1636 * So we still attempt to reduce cache line usage. Just take the slab
1637 * lock and free the item. If there is no additional partial page
1638 * handling required then we can return immediately.
1640 static void __slab_free(struct kmem_cache *s, struct page *page,
1641 void *x, void *addr, unsigned int offset)
1644 void **object = (void *)x;
1645 struct kmem_cache_cpu *c;
1647 c = get_cpu_slab(s, raw_smp_processor_id());
1648 stat(c, FREE_SLOWPATH);
1651 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1655 prior = object[offset] = page->freelist;
1656 page->freelist = object;
1659 if (unlikely(PageSlubFrozen(page))) {
1660 stat(c, FREE_FROZEN);
1664 if (unlikely(!page->inuse))
1668 * Objects left in the slab. If it was not on the partial list before
1671 if (unlikely(!prior)) {
1672 add_partial(get_node(s, page_to_nid(page)), page, 1);
1673 stat(c, FREE_ADD_PARTIAL);
1683 * Slab still on the partial list.
1685 remove_partial(s, page);
1686 stat(c, FREE_REMOVE_PARTIAL);
1690 discard_slab(s, page);
1694 if (!free_debug_processing(s, page, x, addr))
1700 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1701 * can perform fastpath freeing without additional function calls.
1703 * The fastpath is only possible if we are freeing to the current cpu slab
1704 * of this processor. This typically the case if we have just allocated
1707 * If fastpath is not possible then fall back to __slab_free where we deal
1708 * with all sorts of special processing.
1710 static __always_inline void slab_free(struct kmem_cache *s,
1711 struct page *page, void *x, void *addr)
1713 void **object = (void *)x;
1714 struct kmem_cache_cpu *c;
1715 unsigned long flags;
1717 local_irq_save(flags);
1718 c = get_cpu_slab(s, smp_processor_id());
1719 debug_check_no_locks_freed(object, c->objsize);
1720 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1721 debug_check_no_obj_freed(object, s->objsize);
1722 if (likely(page == c->page && c->node >= 0)) {
1723 object[c->offset] = c->freelist;
1724 c->freelist = object;
1725 stat(c, FREE_FASTPATH);
1727 __slab_free(s, page, x, addr, c->offset);
1729 local_irq_restore(flags);
1732 void kmem_cache_free(struct kmem_cache *s, void *x)
1736 page = virt_to_head_page(x);
1738 slab_free(s, page, x, __builtin_return_address(0));
1740 EXPORT_SYMBOL(kmem_cache_free);
1742 /* Figure out on which slab object the object resides */
1743 static struct page *get_object_page(const void *x)
1745 struct page *page = virt_to_head_page(x);
1747 if (!PageSlab(page))
1754 * Object placement in a slab is made very easy because we always start at
1755 * offset 0. If we tune the size of the object to the alignment then we can
1756 * get the required alignment by putting one properly sized object after
1759 * Notice that the allocation order determines the sizes of the per cpu
1760 * caches. Each processor has always one slab available for allocations.
1761 * Increasing the allocation order reduces the number of times that slabs
1762 * must be moved on and off the partial lists and is therefore a factor in
1767 * Mininum / Maximum order of slab pages. This influences locking overhead
1768 * and slab fragmentation. A higher order reduces the number of partial slabs
1769 * and increases the number of allocations possible without having to
1770 * take the list_lock.
1772 static int slub_min_order;
1773 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1774 static int slub_min_objects;
1777 * Merge control. If this is set then no merging of slab caches will occur.
1778 * (Could be removed. This was introduced to pacify the merge skeptics.)
1780 static int slub_nomerge;
1783 * Calculate the order of allocation given an slab object size.
1785 * The order of allocation has significant impact on performance and other
1786 * system components. Generally order 0 allocations should be preferred since
1787 * order 0 does not cause fragmentation in the page allocator. Larger objects
1788 * be problematic to put into order 0 slabs because there may be too much
1789 * unused space left. We go to a higher order if more than 1/16th of the slab
1792 * In order to reach satisfactory performance we must ensure that a minimum
1793 * number of objects is in one slab. Otherwise we may generate too much
1794 * activity on the partial lists which requires taking the list_lock. This is
1795 * less a concern for large slabs though which are rarely used.
1797 * slub_max_order specifies the order where we begin to stop considering the
1798 * number of objects in a slab as critical. If we reach slub_max_order then
1799 * we try to keep the page order as low as possible. So we accept more waste
1800 * of space in favor of a small page order.
1802 * Higher order allocations also allow the placement of more objects in a
1803 * slab and thereby reduce object handling overhead. If the user has
1804 * requested a higher mininum order then we start with that one instead of
1805 * the smallest order which will fit the object.
1807 static inline int slab_order(int size, int min_objects,
1808 int max_order, int fract_leftover)
1812 int min_order = slub_min_order;
1814 if ((PAGE_SIZE << min_order) / size > 65535)
1815 return get_order(size * 65535) - 1;
1817 for (order = max(min_order,
1818 fls(min_objects * size - 1) - PAGE_SHIFT);
1819 order <= max_order; order++) {
1821 unsigned long slab_size = PAGE_SIZE << order;
1823 if (slab_size < min_objects * size)
1826 rem = slab_size % size;
1828 if (rem <= slab_size / fract_leftover)
1836 static inline int calculate_order(int size)
1843 * Attempt to find best configuration for a slab. This
1844 * works by first attempting to generate a layout with
1845 * the best configuration and backing off gradually.
1847 * First we reduce the acceptable waste in a slab. Then
1848 * we reduce the minimum objects required in a slab.
1850 min_objects = slub_min_objects;
1852 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1853 while (min_objects > 1) {
1855 while (fraction >= 4) {
1856 order = slab_order(size, min_objects,
1857 slub_max_order, fraction);
1858 if (order <= slub_max_order)
1866 * We were unable to place multiple objects in a slab. Now
1867 * lets see if we can place a single object there.
1869 order = slab_order(size, 1, slub_max_order, 1);
1870 if (order <= slub_max_order)
1874 * Doh this slab cannot be placed using slub_max_order.
1876 order = slab_order(size, 1, MAX_ORDER, 1);
1877 if (order <= MAX_ORDER)
1883 * Figure out what the alignment of the objects will be.
1885 static unsigned long calculate_alignment(unsigned long flags,
1886 unsigned long align, unsigned long size)
1889 * If the user wants hardware cache aligned objects then follow that
1890 * suggestion if the object is sufficiently large.
1892 * The hardware cache alignment cannot override the specified
1893 * alignment though. If that is greater then use it.
1895 if (flags & SLAB_HWCACHE_ALIGN) {
1896 unsigned long ralign = cache_line_size();
1897 while (size <= ralign / 2)
1899 align = max(align, ralign);
1902 if (align < ARCH_SLAB_MINALIGN)
1903 align = ARCH_SLAB_MINALIGN;
1905 return ALIGN(align, sizeof(void *));
1908 static void init_kmem_cache_cpu(struct kmem_cache *s,
1909 struct kmem_cache_cpu *c)
1914 c->offset = s->offset / sizeof(void *);
1915 c->objsize = s->objsize;
1916 #ifdef CONFIG_SLUB_STATS
1917 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1922 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1927 * The larger the object size is, the more pages we want on the partial
1928 * list to avoid pounding the page allocator excessively.
1930 n->min_partial = ilog2(s->size);
1931 if (n->min_partial < MIN_PARTIAL)
1932 n->min_partial = MIN_PARTIAL;
1933 else if (n->min_partial > MAX_PARTIAL)
1934 n->min_partial = MAX_PARTIAL;
1936 spin_lock_init(&n->list_lock);
1937 INIT_LIST_HEAD(&n->partial);
1938 #ifdef CONFIG_SLUB_DEBUG
1939 atomic_long_set(&n->nr_slabs, 0);
1940 atomic_long_set(&n->total_objects, 0);
1941 INIT_LIST_HEAD(&n->full);
1947 * Per cpu array for per cpu structures.
1949 * The per cpu array places all kmem_cache_cpu structures from one processor
1950 * close together meaning that it becomes possible that multiple per cpu
1951 * structures are contained in one cacheline. This may be particularly
1952 * beneficial for the kmalloc caches.
1954 * A desktop system typically has around 60-80 slabs. With 100 here we are
1955 * likely able to get per cpu structures for all caches from the array defined
1956 * here. We must be able to cover all kmalloc caches during bootstrap.
1958 * If the per cpu array is exhausted then fall back to kmalloc
1959 * of individual cachelines. No sharing is possible then.
1961 #define NR_KMEM_CACHE_CPU 100
1963 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1964 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1966 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1967 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1969 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1970 int cpu, gfp_t flags)
1972 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1975 per_cpu(kmem_cache_cpu_free, cpu) =
1976 (void *)c->freelist;
1978 /* Table overflow: So allocate ourselves */
1980 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1981 flags, cpu_to_node(cpu));
1986 init_kmem_cache_cpu(s, c);
1990 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1992 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1993 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1997 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1998 per_cpu(kmem_cache_cpu_free, cpu) = c;
2001 static void free_kmem_cache_cpus(struct kmem_cache *s)
2005 for_each_online_cpu(cpu) {
2006 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2009 s->cpu_slab[cpu] = NULL;
2010 free_kmem_cache_cpu(c, cpu);
2015 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2019 for_each_online_cpu(cpu) {
2020 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2025 c = alloc_kmem_cache_cpu(s, cpu, flags);
2027 free_kmem_cache_cpus(s);
2030 s->cpu_slab[cpu] = c;
2036 * Initialize the per cpu array.
2038 static void init_alloc_cpu_cpu(int cpu)
2042 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2045 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2046 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2048 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2051 static void __init init_alloc_cpu(void)
2055 for_each_online_cpu(cpu)
2056 init_alloc_cpu_cpu(cpu);
2060 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2061 static inline void init_alloc_cpu(void) {}
2063 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2065 init_kmem_cache_cpu(s, &s->cpu_slab);
2072 * No kmalloc_node yet so do it by hand. We know that this is the first
2073 * slab on the node for this slabcache. There are no concurrent accesses
2076 * Note that this function only works on the kmalloc_node_cache
2077 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2078 * memory on a fresh node that has no slab structures yet.
2080 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2084 struct kmem_cache_node *n;
2085 unsigned long flags;
2087 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2089 page = new_slab(kmalloc_caches, gfpflags, node);
2092 if (page_to_nid(page) != node) {
2093 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2095 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2096 "in order to be able to continue\n");
2101 page->freelist = get_freepointer(kmalloc_caches, n);
2103 kmalloc_caches->node[node] = n;
2104 #ifdef CONFIG_SLUB_DEBUG
2105 init_object(kmalloc_caches, n, 1);
2106 init_tracking(kmalloc_caches, n);
2108 init_kmem_cache_node(n, kmalloc_caches);
2109 inc_slabs_node(kmalloc_caches, node, page->objects);
2112 * lockdep requires consistent irq usage for each lock
2113 * so even though there cannot be a race this early in
2114 * the boot sequence, we still disable irqs.
2116 local_irq_save(flags);
2117 add_partial(n, page, 0);
2118 local_irq_restore(flags);
2122 static void free_kmem_cache_nodes(struct kmem_cache *s)
2126 for_each_node_state(node, N_NORMAL_MEMORY) {
2127 struct kmem_cache_node *n = s->node[node];
2128 if (n && n != &s->local_node)
2129 kmem_cache_free(kmalloc_caches, n);
2130 s->node[node] = NULL;
2134 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2139 if (slab_state >= UP)
2140 local_node = page_to_nid(virt_to_page(s));
2144 for_each_node_state(node, N_NORMAL_MEMORY) {
2145 struct kmem_cache_node *n;
2147 if (local_node == node)
2150 if (slab_state == DOWN) {
2151 n = early_kmem_cache_node_alloc(gfpflags,
2155 n = kmem_cache_alloc_node(kmalloc_caches,
2159 free_kmem_cache_nodes(s);
2165 init_kmem_cache_node(n, s);
2170 static void free_kmem_cache_nodes(struct kmem_cache *s)
2174 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2176 init_kmem_cache_node(&s->local_node, s);
2182 * calculate_sizes() determines the order and the distribution of data within
2185 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2187 unsigned long flags = s->flags;
2188 unsigned long size = s->objsize;
2189 unsigned long align = s->align;
2193 * Round up object size to the next word boundary. We can only
2194 * place the free pointer at word boundaries and this determines
2195 * the possible location of the free pointer.
2197 size = ALIGN(size, sizeof(void *));
2199 #ifdef CONFIG_SLUB_DEBUG
2201 * Determine if we can poison the object itself. If the user of
2202 * the slab may touch the object after free or before allocation
2203 * then we should never poison the object itself.
2205 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2207 s->flags |= __OBJECT_POISON;
2209 s->flags &= ~__OBJECT_POISON;
2213 * If we are Redzoning then check if there is some space between the
2214 * end of the object and the free pointer. If not then add an
2215 * additional word to have some bytes to store Redzone information.
2217 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2218 size += sizeof(void *);
2222 * With that we have determined the number of bytes in actual use
2223 * by the object. This is the potential offset to the free pointer.
2227 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2230 * Relocate free pointer after the object if it is not
2231 * permitted to overwrite the first word of the object on
2234 * This is the case if we do RCU, have a constructor or
2235 * destructor or are poisoning the objects.
2238 size += sizeof(void *);
2241 #ifdef CONFIG_SLUB_DEBUG
2242 if (flags & SLAB_STORE_USER)
2244 * Need to store information about allocs and frees after
2247 size += 2 * sizeof(struct track);
2249 if (flags & SLAB_RED_ZONE)
2251 * Add some empty padding so that we can catch
2252 * overwrites from earlier objects rather than let
2253 * tracking information or the free pointer be
2254 * corrupted if an user writes before the start
2257 size += sizeof(void *);
2261 * Determine the alignment based on various parameters that the
2262 * user specified and the dynamic determination of cache line size
2265 align = calculate_alignment(flags, align, s->objsize);
2268 * SLUB stores one object immediately after another beginning from
2269 * offset 0. In order to align the objects we have to simply size
2270 * each object to conform to the alignment.
2272 size = ALIGN(size, align);
2274 if (forced_order >= 0)
2275 order = forced_order;
2277 order = calculate_order(size);
2284 s->allocflags |= __GFP_COMP;
2286 if (s->flags & SLAB_CACHE_DMA)
2287 s->allocflags |= SLUB_DMA;
2289 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2290 s->allocflags |= __GFP_RECLAIMABLE;
2293 * Determine the number of objects per slab
2295 s->oo = oo_make(order, size);
2296 s->min = oo_make(get_order(size), size);
2297 if (oo_objects(s->oo) > oo_objects(s->max))
2300 return !!oo_objects(s->oo);
2304 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2305 const char *name, size_t size,
2306 size_t align, unsigned long flags,
2307 void (*ctor)(void *))
2309 memset(s, 0, kmem_size);
2314 s->flags = kmem_cache_flags(size, flags, name, ctor);
2316 if (!calculate_sizes(s, -1))
2321 s->remote_node_defrag_ratio = 1000;
2323 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2326 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2328 free_kmem_cache_nodes(s);
2330 if (flags & SLAB_PANIC)
2331 panic("Cannot create slab %s size=%lu realsize=%u "
2332 "order=%u offset=%u flags=%lx\n",
2333 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2339 * Check if a given pointer is valid
2341 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2345 page = get_object_page(object);
2347 if (!page || s != page->slab)
2348 /* No slab or wrong slab */
2351 if (!check_valid_pointer(s, page, object))
2355 * We could also check if the object is on the slabs freelist.
2356 * But this would be too expensive and it seems that the main
2357 * purpose of kmem_ptr_valid() is to check if the object belongs
2358 * to a certain slab.
2362 EXPORT_SYMBOL(kmem_ptr_validate);
2365 * Determine the size of a slab object
2367 unsigned int kmem_cache_size(struct kmem_cache *s)
2371 EXPORT_SYMBOL(kmem_cache_size);
2373 const char *kmem_cache_name(struct kmem_cache *s)
2377 EXPORT_SYMBOL(kmem_cache_name);
2379 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2382 #ifdef CONFIG_SLUB_DEBUG
2383 void *addr = page_address(page);
2385 DECLARE_BITMAP(map, page->objects);
2387 bitmap_zero(map, page->objects);
2388 slab_err(s, page, "%s", text);
2390 for_each_free_object(p, s, page->freelist)
2391 set_bit(slab_index(p, s, addr), map);
2393 for_each_object(p, s, addr, page->objects) {
2395 if (!test_bit(slab_index(p, s, addr), map)) {
2396 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2398 print_tracking(s, p);
2406 * Attempt to free all partial slabs on a node.
2408 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2410 unsigned long flags;
2411 struct page *page, *h;
2413 spin_lock_irqsave(&n->list_lock, flags);
2414 list_for_each_entry_safe(page, h, &n->partial, lru) {
2416 list_del(&page->lru);
2417 discard_slab(s, page);
2420 list_slab_objects(s, page,
2421 "Objects remaining on kmem_cache_close()");
2424 spin_unlock_irqrestore(&n->list_lock, flags);
2428 * Release all resources used by a slab cache.
2430 static inline int kmem_cache_close(struct kmem_cache *s)
2436 /* Attempt to free all objects */
2437 free_kmem_cache_cpus(s);
2438 for_each_node_state(node, N_NORMAL_MEMORY) {
2439 struct kmem_cache_node *n = get_node(s, node);
2442 if (n->nr_partial || slabs_node(s, node))
2445 free_kmem_cache_nodes(s);
2450 * Close a cache and release the kmem_cache structure
2451 * (must be used for caches created using kmem_cache_create)
2453 void kmem_cache_destroy(struct kmem_cache *s)
2455 down_write(&slub_lock);
2459 up_write(&slub_lock);
2460 if (kmem_cache_close(s)) {
2461 printk(KERN_ERR "SLUB %s: %s called for cache that "
2462 "still has objects.\n", s->name, __func__);
2465 sysfs_slab_remove(s);
2467 up_write(&slub_lock);
2469 EXPORT_SYMBOL(kmem_cache_destroy);
2471 /********************************************************************
2473 *******************************************************************/
2475 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2476 EXPORT_SYMBOL(kmalloc_caches);
2478 static int __init setup_slub_min_order(char *str)
2480 get_option(&str, &slub_min_order);
2485 __setup("slub_min_order=", setup_slub_min_order);
2487 static int __init setup_slub_max_order(char *str)
2489 get_option(&str, &slub_max_order);
2494 __setup("slub_max_order=", setup_slub_max_order);
2496 static int __init setup_slub_min_objects(char *str)
2498 get_option(&str, &slub_min_objects);
2503 __setup("slub_min_objects=", setup_slub_min_objects);
2505 static int __init setup_slub_nomerge(char *str)
2511 __setup("slub_nomerge", setup_slub_nomerge);
2513 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2514 const char *name, int size, gfp_t gfp_flags)
2516 unsigned int flags = 0;
2518 if (gfp_flags & SLUB_DMA)
2519 flags = SLAB_CACHE_DMA;
2521 down_write(&slub_lock);
2522 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2526 list_add(&s->list, &slab_caches);
2527 up_write(&slub_lock);
2528 if (sysfs_slab_add(s))
2533 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2536 #ifdef CONFIG_ZONE_DMA
2537 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2539 static void sysfs_add_func(struct work_struct *w)
2541 struct kmem_cache *s;
2543 down_write(&slub_lock);
2544 list_for_each_entry(s, &slab_caches, list) {
2545 if (s->flags & __SYSFS_ADD_DEFERRED) {
2546 s->flags &= ~__SYSFS_ADD_DEFERRED;
2550 up_write(&slub_lock);
2553 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2555 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2557 struct kmem_cache *s;
2561 s = kmalloc_caches_dma[index];
2565 /* Dynamically create dma cache */
2566 if (flags & __GFP_WAIT)
2567 down_write(&slub_lock);
2569 if (!down_write_trylock(&slub_lock))
2573 if (kmalloc_caches_dma[index])
2576 realsize = kmalloc_caches[index].objsize;
2577 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2578 (unsigned int)realsize);
2579 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2581 if (!s || !text || !kmem_cache_open(s, flags, text,
2582 realsize, ARCH_KMALLOC_MINALIGN,
2583 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2589 list_add(&s->list, &slab_caches);
2590 kmalloc_caches_dma[index] = s;
2592 schedule_work(&sysfs_add_work);
2595 up_write(&slub_lock);
2597 return kmalloc_caches_dma[index];
2602 * Conversion table for small slabs sizes / 8 to the index in the
2603 * kmalloc array. This is necessary for slabs < 192 since we have non power
2604 * of two cache sizes there. The size of larger slabs can be determined using
2607 static s8 size_index[24] = {
2634 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2640 return ZERO_SIZE_PTR;
2642 index = size_index[(size - 1) / 8];
2644 index = fls(size - 1);
2646 #ifdef CONFIG_ZONE_DMA
2647 if (unlikely((flags & SLUB_DMA)))
2648 return dma_kmalloc_cache(index, flags);
2651 return &kmalloc_caches[index];
2654 void *__kmalloc(size_t size, gfp_t flags)
2656 struct kmem_cache *s;
2658 if (unlikely(size > PAGE_SIZE))
2659 return kmalloc_large(size, flags);
2661 s = get_slab(size, flags);
2663 if (unlikely(ZERO_OR_NULL_PTR(s)))
2666 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2668 EXPORT_SYMBOL(__kmalloc);
2670 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2672 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2676 return page_address(page);
2682 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2684 struct kmem_cache *s;
2686 if (unlikely(size > PAGE_SIZE))
2687 return kmalloc_large_node(size, flags, node);
2689 s = get_slab(size, flags);
2691 if (unlikely(ZERO_OR_NULL_PTR(s)))
2694 return slab_alloc(s, flags, node, __builtin_return_address(0));
2696 EXPORT_SYMBOL(__kmalloc_node);
2699 size_t ksize(const void *object)
2702 struct kmem_cache *s;
2704 if (unlikely(object == ZERO_SIZE_PTR))
2707 page = virt_to_head_page(object);
2709 if (unlikely(!PageSlab(page))) {
2710 WARN_ON(!PageCompound(page));
2711 return PAGE_SIZE << compound_order(page);
2715 #ifdef CONFIG_SLUB_DEBUG
2717 * Debugging requires use of the padding between object
2718 * and whatever may come after it.
2720 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2725 * If we have the need to store the freelist pointer
2726 * back there or track user information then we can
2727 * only use the space before that information.
2729 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2732 * Else we can use all the padding etc for the allocation
2737 void kfree(const void *x)
2740 void *object = (void *)x;
2742 if (unlikely(ZERO_OR_NULL_PTR(x)))
2745 page = virt_to_head_page(x);
2746 if (unlikely(!PageSlab(page))) {
2747 BUG_ON(!PageCompound(page));
2751 slab_free(page->slab, page, object, __builtin_return_address(0));
2753 EXPORT_SYMBOL(kfree);
2756 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2757 * the remaining slabs by the number of items in use. The slabs with the
2758 * most items in use come first. New allocations will then fill those up
2759 * and thus they can be removed from the partial lists.
2761 * The slabs with the least items are placed last. This results in them
2762 * being allocated from last increasing the chance that the last objects
2763 * are freed in them.
2765 int kmem_cache_shrink(struct kmem_cache *s)
2769 struct kmem_cache_node *n;
2772 int objects = oo_objects(s->max);
2773 struct list_head *slabs_by_inuse =
2774 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2775 unsigned long flags;
2777 if (!slabs_by_inuse)
2781 for_each_node_state(node, N_NORMAL_MEMORY) {
2782 n = get_node(s, node);
2787 for (i = 0; i < objects; i++)
2788 INIT_LIST_HEAD(slabs_by_inuse + i);
2790 spin_lock_irqsave(&n->list_lock, flags);
2793 * Build lists indexed by the items in use in each slab.
2795 * Note that concurrent frees may occur while we hold the
2796 * list_lock. page->inuse here is the upper limit.
2798 list_for_each_entry_safe(page, t, &n->partial, lru) {
2799 if (!page->inuse && slab_trylock(page)) {
2801 * Must hold slab lock here because slab_free
2802 * may have freed the last object and be
2803 * waiting to release the slab.
2805 list_del(&page->lru);
2808 discard_slab(s, page);
2810 list_move(&page->lru,
2811 slabs_by_inuse + page->inuse);
2816 * Rebuild the partial list with the slabs filled up most
2817 * first and the least used slabs at the end.
2819 for (i = objects - 1; i >= 0; i--)
2820 list_splice(slabs_by_inuse + i, n->partial.prev);
2822 spin_unlock_irqrestore(&n->list_lock, flags);
2825 kfree(slabs_by_inuse);
2828 EXPORT_SYMBOL(kmem_cache_shrink);
2830 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2831 static int slab_mem_going_offline_callback(void *arg)
2833 struct kmem_cache *s;
2835 down_read(&slub_lock);
2836 list_for_each_entry(s, &slab_caches, list)
2837 kmem_cache_shrink(s);
2838 up_read(&slub_lock);
2843 static void slab_mem_offline_callback(void *arg)
2845 struct kmem_cache_node *n;
2846 struct kmem_cache *s;
2847 struct memory_notify *marg = arg;
2850 offline_node = marg->status_change_nid;
2853 * If the node still has available memory. we need kmem_cache_node
2856 if (offline_node < 0)
2859 down_read(&slub_lock);
2860 list_for_each_entry(s, &slab_caches, list) {
2861 n = get_node(s, offline_node);
2864 * if n->nr_slabs > 0, slabs still exist on the node
2865 * that is going down. We were unable to free them,
2866 * and offline_pages() function shoudn't call this
2867 * callback. So, we must fail.
2869 BUG_ON(slabs_node(s, offline_node));
2871 s->node[offline_node] = NULL;
2872 kmem_cache_free(kmalloc_caches, n);
2875 up_read(&slub_lock);
2878 static int slab_mem_going_online_callback(void *arg)
2880 struct kmem_cache_node *n;
2881 struct kmem_cache *s;
2882 struct memory_notify *marg = arg;
2883 int nid = marg->status_change_nid;
2887 * If the node's memory is already available, then kmem_cache_node is
2888 * already created. Nothing to do.
2894 * We are bringing a node online. No memory is available yet. We must
2895 * allocate a kmem_cache_node structure in order to bring the node
2898 down_read(&slub_lock);
2899 list_for_each_entry(s, &slab_caches, list) {
2901 * XXX: kmem_cache_alloc_node will fallback to other nodes
2902 * since memory is not yet available from the node that
2905 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2910 init_kmem_cache_node(n, s);
2914 up_read(&slub_lock);
2918 static int slab_memory_callback(struct notifier_block *self,
2919 unsigned long action, void *arg)
2924 case MEM_GOING_ONLINE:
2925 ret = slab_mem_going_online_callback(arg);
2927 case MEM_GOING_OFFLINE:
2928 ret = slab_mem_going_offline_callback(arg);
2931 case MEM_CANCEL_ONLINE:
2932 slab_mem_offline_callback(arg);
2935 case MEM_CANCEL_OFFLINE:
2939 ret = notifier_from_errno(ret);
2945 #endif /* CONFIG_MEMORY_HOTPLUG */
2947 /********************************************************************
2948 * Basic setup of slabs
2949 *******************************************************************/
2951 void __init kmem_cache_init(void)
2960 * Must first have the slab cache available for the allocations of the
2961 * struct kmem_cache_node's. There is special bootstrap code in
2962 * kmem_cache_open for slab_state == DOWN.
2964 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2965 sizeof(struct kmem_cache_node), GFP_KERNEL);
2966 kmalloc_caches[0].refcount = -1;
2969 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
2972 /* Able to allocate the per node structures */
2973 slab_state = PARTIAL;
2975 /* Caches that are not of the two-to-the-power-of size */
2976 if (KMALLOC_MIN_SIZE <= 64) {
2977 create_kmalloc_cache(&kmalloc_caches[1],
2978 "kmalloc-96", 96, GFP_KERNEL);
2980 create_kmalloc_cache(&kmalloc_caches[2],
2981 "kmalloc-192", 192, GFP_KERNEL);
2985 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
2986 create_kmalloc_cache(&kmalloc_caches[i],
2987 "kmalloc", 1 << i, GFP_KERNEL);
2993 * Patch up the size_index table if we have strange large alignment
2994 * requirements for the kmalloc array. This is only the case for
2995 * MIPS it seems. The standard arches will not generate any code here.
2997 * Largest permitted alignment is 256 bytes due to the way we
2998 * handle the index determination for the smaller caches.
3000 * Make sure that nothing crazy happens if someone starts tinkering
3001 * around with ARCH_KMALLOC_MINALIGN
3003 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3004 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3006 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3007 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3009 if (KMALLOC_MIN_SIZE == 128) {
3011 * The 192 byte sized cache is not used if the alignment
3012 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3015 for (i = 128 + 8; i <= 192; i += 8)
3016 size_index[(i - 1) / 8] = 8;
3021 /* Provide the correct kmalloc names now that the caches are up */
3022 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3023 kmalloc_caches[i]. name =
3024 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3027 register_cpu_notifier(&slab_notifier);
3028 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3029 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3031 kmem_size = sizeof(struct kmem_cache);
3035 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3036 " CPUs=%d, Nodes=%d\n",
3037 caches, cache_line_size(),
3038 slub_min_order, slub_max_order, slub_min_objects,
3039 nr_cpu_ids, nr_node_ids);
3043 * Find a mergeable slab cache
3045 static int slab_unmergeable(struct kmem_cache *s)
3047 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3054 * We may have set a slab to be unmergeable during bootstrap.
3056 if (s->refcount < 0)
3062 static struct kmem_cache *find_mergeable(size_t size,
3063 size_t align, unsigned long flags, const char *name,
3064 void (*ctor)(void *))
3066 struct kmem_cache *s;
3068 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3074 size = ALIGN(size, sizeof(void *));
3075 align = calculate_alignment(flags, align, size);
3076 size = ALIGN(size, align);
3077 flags = kmem_cache_flags(size, flags, name, NULL);
3079 list_for_each_entry(s, &slab_caches, list) {
3080 if (slab_unmergeable(s))
3086 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3089 * Check if alignment is compatible.
3090 * Courtesy of Adrian Drzewiecki
3092 if ((s->size & ~(align - 1)) != s->size)
3095 if (s->size - size >= sizeof(void *))
3103 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3104 size_t align, unsigned long flags, void (*ctor)(void *))
3106 struct kmem_cache *s;
3108 down_write(&slub_lock);
3109 s = find_mergeable(size, align, flags, name, ctor);
3115 * Adjust the object sizes so that we clear
3116 * the complete object on kzalloc.
3118 s->objsize = max(s->objsize, (int)size);
3121 * And then we need to update the object size in the
3122 * per cpu structures
3124 for_each_online_cpu(cpu)
3125 get_cpu_slab(s, cpu)->objsize = s->objsize;
3127 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3128 up_write(&slub_lock);
3130 if (sysfs_slab_alias(s, name))
3135 s = kmalloc(kmem_size, GFP_KERNEL);
3137 if (kmem_cache_open(s, GFP_KERNEL, name,
3138 size, align, flags, ctor)) {
3139 list_add(&s->list, &slab_caches);
3140 up_write(&slub_lock);
3141 if (sysfs_slab_add(s))
3147 up_write(&slub_lock);
3150 if (flags & SLAB_PANIC)
3151 panic("Cannot create slabcache %s\n", name);
3156 EXPORT_SYMBOL(kmem_cache_create);
3160 * Use the cpu notifier to insure that the cpu slabs are flushed when
3163 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3164 unsigned long action, void *hcpu)
3166 long cpu = (long)hcpu;
3167 struct kmem_cache *s;
3168 unsigned long flags;
3171 case CPU_UP_PREPARE:
3172 case CPU_UP_PREPARE_FROZEN:
3173 init_alloc_cpu_cpu(cpu);
3174 down_read(&slub_lock);
3175 list_for_each_entry(s, &slab_caches, list)
3176 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3178 up_read(&slub_lock);
3181 case CPU_UP_CANCELED:
3182 case CPU_UP_CANCELED_FROZEN:
3184 case CPU_DEAD_FROZEN:
3185 down_read(&slub_lock);
3186 list_for_each_entry(s, &slab_caches, list) {
3187 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3189 local_irq_save(flags);
3190 __flush_cpu_slab(s, cpu);
3191 local_irq_restore(flags);
3192 free_kmem_cache_cpu(c, cpu);
3193 s->cpu_slab[cpu] = NULL;
3195 up_read(&slub_lock);
3203 static struct notifier_block __cpuinitdata slab_notifier = {
3204 .notifier_call = slab_cpuup_callback
3209 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3211 struct kmem_cache *s;
3213 if (unlikely(size > PAGE_SIZE))
3214 return kmalloc_large(size, gfpflags);
3216 s = get_slab(size, gfpflags);
3218 if (unlikely(ZERO_OR_NULL_PTR(s)))
3221 return slab_alloc(s, gfpflags, -1, caller);
3224 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3225 int node, void *caller)
3227 struct kmem_cache *s;
3229 if (unlikely(size > PAGE_SIZE))
3230 return kmalloc_large_node(size, gfpflags, node);
3232 s = get_slab(size, gfpflags);
3234 if (unlikely(ZERO_OR_NULL_PTR(s)))
3237 return slab_alloc(s, gfpflags, node, caller);
3240 #ifdef CONFIG_SLUB_DEBUG
3241 static unsigned long count_partial(struct kmem_cache_node *n,
3242 int (*get_count)(struct page *))
3244 unsigned long flags;
3245 unsigned long x = 0;
3248 spin_lock_irqsave(&n->list_lock, flags);
3249 list_for_each_entry(page, &n->partial, lru)
3250 x += get_count(page);
3251 spin_unlock_irqrestore(&n->list_lock, flags);
3255 static int count_inuse(struct page *page)
3260 static int count_total(struct page *page)
3262 return page->objects;
3265 static int count_free(struct page *page)
3267 return page->objects - page->inuse;
3270 static int validate_slab(struct kmem_cache *s, struct page *page,
3274 void *addr = page_address(page);
3276 if (!check_slab(s, page) ||
3277 !on_freelist(s, page, NULL))
3280 /* Now we know that a valid freelist exists */
3281 bitmap_zero(map, page->objects);
3283 for_each_free_object(p, s, page->freelist) {
3284 set_bit(slab_index(p, s, addr), map);
3285 if (!check_object(s, page, p, 0))
3289 for_each_object(p, s, addr, page->objects)
3290 if (!test_bit(slab_index(p, s, addr), map))
3291 if (!check_object(s, page, p, 1))
3296 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3299 if (slab_trylock(page)) {
3300 validate_slab(s, page, map);
3303 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3306 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3307 if (!PageSlubDebug(page))
3308 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3309 "on slab 0x%p\n", s->name, page);
3311 if (PageSlubDebug(page))
3312 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3313 "slab 0x%p\n", s->name, page);
3317 static int validate_slab_node(struct kmem_cache *s,
3318 struct kmem_cache_node *n, unsigned long *map)
3320 unsigned long count = 0;
3322 unsigned long flags;
3324 spin_lock_irqsave(&n->list_lock, flags);
3326 list_for_each_entry(page, &n->partial, lru) {
3327 validate_slab_slab(s, page, map);
3330 if (count != n->nr_partial)
3331 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3332 "counter=%ld\n", s->name, count, n->nr_partial);
3334 if (!(s->flags & SLAB_STORE_USER))
3337 list_for_each_entry(page, &n->full, lru) {
3338 validate_slab_slab(s, page, map);
3341 if (count != atomic_long_read(&n->nr_slabs))
3342 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3343 "counter=%ld\n", s->name, count,
3344 atomic_long_read(&n->nr_slabs));
3347 spin_unlock_irqrestore(&n->list_lock, flags);
3351 static long validate_slab_cache(struct kmem_cache *s)
3354 unsigned long count = 0;
3355 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3356 sizeof(unsigned long), GFP_KERNEL);
3362 for_each_node_state(node, N_NORMAL_MEMORY) {
3363 struct kmem_cache_node *n = get_node(s, node);
3365 count += validate_slab_node(s, n, map);
3371 #ifdef SLUB_RESILIENCY_TEST
3372 static void resiliency_test(void)
3376 printk(KERN_ERR "SLUB resiliency testing\n");
3377 printk(KERN_ERR "-----------------------\n");
3378 printk(KERN_ERR "A. Corruption after allocation\n");
3380 p = kzalloc(16, GFP_KERNEL);
3382 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3383 " 0x12->0x%p\n\n", p + 16);
3385 validate_slab_cache(kmalloc_caches + 4);
3387 /* Hmmm... The next two are dangerous */
3388 p = kzalloc(32, GFP_KERNEL);
3389 p[32 + sizeof(void *)] = 0x34;
3390 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3391 " 0x34 -> -0x%p\n", p);
3393 "If allocated object is overwritten then not detectable\n\n");
3395 validate_slab_cache(kmalloc_caches + 5);
3396 p = kzalloc(64, GFP_KERNEL);
3397 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3399 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3402 "If allocated object is overwritten then not detectable\n\n");
3403 validate_slab_cache(kmalloc_caches + 6);
3405 printk(KERN_ERR "\nB. Corruption after free\n");
3406 p = kzalloc(128, GFP_KERNEL);
3409 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3410 validate_slab_cache(kmalloc_caches + 7);
3412 p = kzalloc(256, GFP_KERNEL);
3415 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3417 validate_slab_cache(kmalloc_caches + 8);
3419 p = kzalloc(512, GFP_KERNEL);
3422 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3423 validate_slab_cache(kmalloc_caches + 9);
3426 static void resiliency_test(void) {};
3430 * Generate lists of code addresses where slabcache objects are allocated
3435 unsigned long count;
3448 unsigned long count;
3449 struct location *loc;
3452 static void free_loc_track(struct loc_track *t)
3455 free_pages((unsigned long)t->loc,
3456 get_order(sizeof(struct location) * t->max));
3459 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3464 order = get_order(sizeof(struct location) * max);
3466 l = (void *)__get_free_pages(flags, order);
3471 memcpy(l, t->loc, sizeof(struct location) * t->count);
3479 static int add_location(struct loc_track *t, struct kmem_cache *s,
3480 const struct track *track)
3482 long start, end, pos;
3485 unsigned long age = jiffies - track->when;
3491 pos = start + (end - start + 1) / 2;
3494 * There is nothing at "end". If we end up there
3495 * we need to add something to before end.
3500 caddr = t->loc[pos].addr;
3501 if (track->addr == caddr) {
3507 if (age < l->min_time)
3509 if (age > l->max_time)
3512 if (track->pid < l->min_pid)
3513 l->min_pid = track->pid;
3514 if (track->pid > l->max_pid)
3515 l->max_pid = track->pid;
3517 cpu_set(track->cpu, l->cpus);
3519 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3523 if (track->addr < caddr)
3530 * Not found. Insert new tracking element.
3532 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3538 (t->count - pos) * sizeof(struct location));
3541 l->addr = track->addr;
3545 l->min_pid = track->pid;
3546 l->max_pid = track->pid;
3547 cpus_clear(l->cpus);
3548 cpu_set(track->cpu, l->cpus);
3549 nodes_clear(l->nodes);
3550 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3554 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3555 struct page *page, enum track_item alloc)
3557 void *addr = page_address(page);
3558 DECLARE_BITMAP(map, page->objects);
3561 bitmap_zero(map, page->objects);
3562 for_each_free_object(p, s, page->freelist)
3563 set_bit(slab_index(p, s, addr), map);
3565 for_each_object(p, s, addr, page->objects)
3566 if (!test_bit(slab_index(p, s, addr), map))
3567 add_location(t, s, get_track(s, p, alloc));
3570 static int list_locations(struct kmem_cache *s, char *buf,
3571 enum track_item alloc)
3575 struct loc_track t = { 0, 0, NULL };
3578 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3580 return sprintf(buf, "Out of memory\n");
3582 /* Push back cpu slabs */
3585 for_each_node_state(node, N_NORMAL_MEMORY) {
3586 struct kmem_cache_node *n = get_node(s, node);
3587 unsigned long flags;
3590 if (!atomic_long_read(&n->nr_slabs))
3593 spin_lock_irqsave(&n->list_lock, flags);
3594 list_for_each_entry(page, &n->partial, lru)
3595 process_slab(&t, s, page, alloc);
3596 list_for_each_entry(page, &n->full, lru)
3597 process_slab(&t, s, page, alloc);
3598 spin_unlock_irqrestore(&n->list_lock, flags);
3601 for (i = 0; i < t.count; i++) {
3602 struct location *l = &t.loc[i];
3604 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3606 len += sprintf(buf + len, "%7ld ", l->count);
3609 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3611 len += sprintf(buf + len, "<not-available>");
3613 if (l->sum_time != l->min_time) {
3614 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3616 (long)div_u64(l->sum_time, l->count),
3619 len += sprintf(buf + len, " age=%ld",
3622 if (l->min_pid != l->max_pid)
3623 len += sprintf(buf + len, " pid=%ld-%ld",
3624 l->min_pid, l->max_pid);
3626 len += sprintf(buf + len, " pid=%ld",
3629 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3630 len < PAGE_SIZE - 60) {
3631 len += sprintf(buf + len, " cpus=");
3632 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3636 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3637 len < PAGE_SIZE - 60) {
3638 len += sprintf(buf + len, " nodes=");
3639 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3643 len += sprintf(buf + len, "\n");
3648 len += sprintf(buf, "No data\n");
3652 enum slab_stat_type {
3653 SL_ALL, /* All slabs */
3654 SL_PARTIAL, /* Only partially allocated slabs */
3655 SL_CPU, /* Only slabs used for cpu caches */
3656 SL_OBJECTS, /* Determine allocated objects not slabs */
3657 SL_TOTAL /* Determine object capacity not slabs */
3660 #define SO_ALL (1 << SL_ALL)
3661 #define SO_PARTIAL (1 << SL_PARTIAL)
3662 #define SO_CPU (1 << SL_CPU)
3663 #define SO_OBJECTS (1 << SL_OBJECTS)
3664 #define SO_TOTAL (1 << SL_TOTAL)
3666 static ssize_t show_slab_objects(struct kmem_cache *s,
3667 char *buf, unsigned long flags)
3669 unsigned long total = 0;
3672 unsigned long *nodes;
3673 unsigned long *per_cpu;
3675 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3678 per_cpu = nodes + nr_node_ids;
3680 if (flags & SO_CPU) {
3683 for_each_possible_cpu(cpu) {
3684 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3686 if (!c || c->node < 0)
3690 if (flags & SO_TOTAL)
3691 x = c->page->objects;
3692 else if (flags & SO_OBJECTS)
3698 nodes[c->node] += x;
3704 if (flags & SO_ALL) {
3705 for_each_node_state(node, N_NORMAL_MEMORY) {
3706 struct kmem_cache_node *n = get_node(s, node);
3708 if (flags & SO_TOTAL)
3709 x = atomic_long_read(&n->total_objects);
3710 else if (flags & SO_OBJECTS)
3711 x = atomic_long_read(&n->total_objects) -
3712 count_partial(n, count_free);
3715 x = atomic_long_read(&n->nr_slabs);
3720 } else if (flags & SO_PARTIAL) {
3721 for_each_node_state(node, N_NORMAL_MEMORY) {
3722 struct kmem_cache_node *n = get_node(s, node);
3724 if (flags & SO_TOTAL)
3725 x = count_partial(n, count_total);
3726 else if (flags & SO_OBJECTS)
3727 x = count_partial(n, count_inuse);
3734 x = sprintf(buf, "%lu", total);
3736 for_each_node_state(node, N_NORMAL_MEMORY)
3738 x += sprintf(buf + x, " N%d=%lu",
3742 return x + sprintf(buf + x, "\n");
3745 static int any_slab_objects(struct kmem_cache *s)
3749 for_each_online_node(node) {
3750 struct kmem_cache_node *n = get_node(s, node);
3755 if (atomic_long_read(&n->total_objects))
3761 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3762 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3764 struct slab_attribute {
3765 struct attribute attr;
3766 ssize_t (*show)(struct kmem_cache *s, char *buf);
3767 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3770 #define SLAB_ATTR_RO(_name) \
3771 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3773 #define SLAB_ATTR(_name) \
3774 static struct slab_attribute _name##_attr = \
3775 __ATTR(_name, 0644, _name##_show, _name##_store)
3777 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3779 return sprintf(buf, "%d\n", s->size);
3781 SLAB_ATTR_RO(slab_size);
3783 static ssize_t align_show(struct kmem_cache *s, char *buf)
3785 return sprintf(buf, "%d\n", s->align);
3787 SLAB_ATTR_RO(align);
3789 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3791 return sprintf(buf, "%d\n", s->objsize);
3793 SLAB_ATTR_RO(object_size);
3795 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3797 return sprintf(buf, "%d\n", oo_objects(s->oo));
3799 SLAB_ATTR_RO(objs_per_slab);
3801 static ssize_t order_store(struct kmem_cache *s,
3802 const char *buf, size_t length)
3804 unsigned long order;
3807 err = strict_strtoul(buf, 10, &order);
3811 if (order > slub_max_order || order < slub_min_order)
3814 calculate_sizes(s, order);
3818 static ssize_t order_show(struct kmem_cache *s, char *buf)
3820 return sprintf(buf, "%d\n", oo_order(s->oo));
3824 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3827 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3829 return n + sprintf(buf + n, "\n");
3835 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3837 return sprintf(buf, "%d\n", s->refcount - 1);
3839 SLAB_ATTR_RO(aliases);
3841 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3843 return show_slab_objects(s, buf, SO_ALL);
3845 SLAB_ATTR_RO(slabs);
3847 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3849 return show_slab_objects(s, buf, SO_PARTIAL);
3851 SLAB_ATTR_RO(partial);
3853 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3855 return show_slab_objects(s, buf, SO_CPU);
3857 SLAB_ATTR_RO(cpu_slabs);
3859 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3861 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3863 SLAB_ATTR_RO(objects);
3865 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3867 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3869 SLAB_ATTR_RO(objects_partial);
3871 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3873 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3875 SLAB_ATTR_RO(total_objects);
3877 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3879 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3882 static ssize_t sanity_checks_store(struct kmem_cache *s,
3883 const char *buf, size_t length)
3885 s->flags &= ~SLAB_DEBUG_FREE;
3887 s->flags |= SLAB_DEBUG_FREE;
3890 SLAB_ATTR(sanity_checks);
3892 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3894 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3897 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3900 s->flags &= ~SLAB_TRACE;
3902 s->flags |= SLAB_TRACE;
3907 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3909 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3912 static ssize_t reclaim_account_store(struct kmem_cache *s,
3913 const char *buf, size_t length)
3915 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3917 s->flags |= SLAB_RECLAIM_ACCOUNT;
3920 SLAB_ATTR(reclaim_account);
3922 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3924 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3926 SLAB_ATTR_RO(hwcache_align);
3928 #ifdef CONFIG_ZONE_DMA
3929 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3931 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3933 SLAB_ATTR_RO(cache_dma);
3936 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3938 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3940 SLAB_ATTR_RO(destroy_by_rcu);
3942 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3944 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3947 static ssize_t red_zone_store(struct kmem_cache *s,
3948 const char *buf, size_t length)
3950 if (any_slab_objects(s))
3953 s->flags &= ~SLAB_RED_ZONE;
3955 s->flags |= SLAB_RED_ZONE;
3956 calculate_sizes(s, -1);
3959 SLAB_ATTR(red_zone);
3961 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3963 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3966 static ssize_t poison_store(struct kmem_cache *s,
3967 const char *buf, size_t length)
3969 if (any_slab_objects(s))
3972 s->flags &= ~SLAB_POISON;
3974 s->flags |= SLAB_POISON;
3975 calculate_sizes(s, -1);
3980 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3982 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3985 static ssize_t store_user_store(struct kmem_cache *s,
3986 const char *buf, size_t length)
3988 if (any_slab_objects(s))
3991 s->flags &= ~SLAB_STORE_USER;
3993 s->flags |= SLAB_STORE_USER;
3994 calculate_sizes(s, -1);
3997 SLAB_ATTR(store_user);
3999 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4004 static ssize_t validate_store(struct kmem_cache *s,
4005 const char *buf, size_t length)
4009 if (buf[0] == '1') {
4010 ret = validate_slab_cache(s);
4016 SLAB_ATTR(validate);
4018 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4023 static ssize_t shrink_store(struct kmem_cache *s,
4024 const char *buf, size_t length)
4026 if (buf[0] == '1') {
4027 int rc = kmem_cache_shrink(s);
4037 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4039 if (!(s->flags & SLAB_STORE_USER))
4041 return list_locations(s, buf, TRACK_ALLOC);
4043 SLAB_ATTR_RO(alloc_calls);
4045 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4047 if (!(s->flags & SLAB_STORE_USER))
4049 return list_locations(s, buf, TRACK_FREE);
4051 SLAB_ATTR_RO(free_calls);
4054 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4056 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4059 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4060 const char *buf, size_t length)
4062 unsigned long ratio;
4065 err = strict_strtoul(buf, 10, &ratio);
4070 s->remote_node_defrag_ratio = ratio * 10;
4074 SLAB_ATTR(remote_node_defrag_ratio);
4077 #ifdef CONFIG_SLUB_STATS
4078 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4080 unsigned long sum = 0;
4083 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4088 for_each_online_cpu(cpu) {
4089 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4095 len = sprintf(buf, "%lu", sum);
4098 for_each_online_cpu(cpu) {
4099 if (data[cpu] && len < PAGE_SIZE - 20)
4100 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4104 return len + sprintf(buf + len, "\n");
4107 #define STAT_ATTR(si, text) \
4108 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4110 return show_stat(s, buf, si); \
4112 SLAB_ATTR_RO(text); \
4114 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4115 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4116 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4117 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4118 STAT_ATTR(FREE_FROZEN, free_frozen);
4119 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4120 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4121 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4122 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4123 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4124 STAT_ATTR(FREE_SLAB, free_slab);
4125 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4126 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4127 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4128 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4129 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4130 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4131 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4134 static struct attribute *slab_attrs[] = {
4135 &slab_size_attr.attr,
4136 &object_size_attr.attr,
4137 &objs_per_slab_attr.attr,
4140 &objects_partial_attr.attr,
4141 &total_objects_attr.attr,
4144 &cpu_slabs_attr.attr,
4148 &sanity_checks_attr.attr,
4150 &hwcache_align_attr.attr,
4151 &reclaim_account_attr.attr,
4152 &destroy_by_rcu_attr.attr,
4153 &red_zone_attr.attr,
4155 &store_user_attr.attr,
4156 &validate_attr.attr,
4158 &alloc_calls_attr.attr,
4159 &free_calls_attr.attr,
4160 #ifdef CONFIG_ZONE_DMA
4161 &cache_dma_attr.attr,
4164 &remote_node_defrag_ratio_attr.attr,
4166 #ifdef CONFIG_SLUB_STATS
4167 &alloc_fastpath_attr.attr,
4168 &alloc_slowpath_attr.attr,
4169 &free_fastpath_attr.attr,
4170 &free_slowpath_attr.attr,
4171 &free_frozen_attr.attr,
4172 &free_add_partial_attr.attr,
4173 &free_remove_partial_attr.attr,
4174 &alloc_from_partial_attr.attr,
4175 &alloc_slab_attr.attr,
4176 &alloc_refill_attr.attr,
4177 &free_slab_attr.attr,
4178 &cpuslab_flush_attr.attr,
4179 &deactivate_full_attr.attr,
4180 &deactivate_empty_attr.attr,
4181 &deactivate_to_head_attr.attr,
4182 &deactivate_to_tail_attr.attr,
4183 &deactivate_remote_frees_attr.attr,
4184 &order_fallback_attr.attr,
4189 static struct attribute_group slab_attr_group = {
4190 .attrs = slab_attrs,
4193 static ssize_t slab_attr_show(struct kobject *kobj,
4194 struct attribute *attr,
4197 struct slab_attribute *attribute;
4198 struct kmem_cache *s;
4201 attribute = to_slab_attr(attr);
4204 if (!attribute->show)
4207 err = attribute->show(s, buf);
4212 static ssize_t slab_attr_store(struct kobject *kobj,
4213 struct attribute *attr,
4214 const char *buf, size_t len)
4216 struct slab_attribute *attribute;
4217 struct kmem_cache *s;
4220 attribute = to_slab_attr(attr);
4223 if (!attribute->store)
4226 err = attribute->store(s, buf, len);
4231 static void kmem_cache_release(struct kobject *kobj)
4233 struct kmem_cache *s = to_slab(kobj);
4238 static struct sysfs_ops slab_sysfs_ops = {
4239 .show = slab_attr_show,
4240 .store = slab_attr_store,
4243 static struct kobj_type slab_ktype = {
4244 .sysfs_ops = &slab_sysfs_ops,
4245 .release = kmem_cache_release
4248 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4250 struct kobj_type *ktype = get_ktype(kobj);
4252 if (ktype == &slab_ktype)
4257 static struct kset_uevent_ops slab_uevent_ops = {
4258 .filter = uevent_filter,
4261 static struct kset *slab_kset;
4263 #define ID_STR_LENGTH 64
4265 /* Create a unique string id for a slab cache:
4267 * Format :[flags-]size
4269 static char *create_unique_id(struct kmem_cache *s)
4271 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4278 * First flags affecting slabcache operations. We will only
4279 * get here for aliasable slabs so we do not need to support
4280 * too many flags. The flags here must cover all flags that
4281 * are matched during merging to guarantee that the id is
4284 if (s->flags & SLAB_CACHE_DMA)
4286 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4288 if (s->flags & SLAB_DEBUG_FREE)
4292 p += sprintf(p, "%07d", s->size);
4293 BUG_ON(p > name + ID_STR_LENGTH - 1);
4297 static int sysfs_slab_add(struct kmem_cache *s)
4303 if (slab_state < SYSFS)
4304 /* Defer until later */
4307 unmergeable = slab_unmergeable(s);
4310 * Slabcache can never be merged so we can use the name proper.
4311 * This is typically the case for debug situations. In that
4312 * case we can catch duplicate names easily.
4314 sysfs_remove_link(&slab_kset->kobj, s->name);
4318 * Create a unique name for the slab as a target
4321 name = create_unique_id(s);
4324 s->kobj.kset = slab_kset;
4325 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4327 kobject_put(&s->kobj);
4331 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4334 kobject_uevent(&s->kobj, KOBJ_ADD);
4336 /* Setup first alias */
4337 sysfs_slab_alias(s, s->name);
4343 static void sysfs_slab_remove(struct kmem_cache *s)
4345 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4346 kobject_del(&s->kobj);
4347 kobject_put(&s->kobj);
4351 * Need to buffer aliases during bootup until sysfs becomes
4352 * available lest we loose that information.
4354 struct saved_alias {
4355 struct kmem_cache *s;
4357 struct saved_alias *next;
4360 static struct saved_alias *alias_list;
4362 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4364 struct saved_alias *al;
4366 if (slab_state == SYSFS) {
4368 * If we have a leftover link then remove it.
4370 sysfs_remove_link(&slab_kset->kobj, name);
4371 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4374 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4380 al->next = alias_list;
4385 static int __init slab_sysfs_init(void)
4387 struct kmem_cache *s;
4390 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4392 printk(KERN_ERR "Cannot register slab subsystem.\n");
4398 list_for_each_entry(s, &slab_caches, list) {
4399 err = sysfs_slab_add(s);
4401 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4402 " to sysfs\n", s->name);
4405 while (alias_list) {
4406 struct saved_alias *al = alias_list;
4408 alias_list = alias_list->next;
4409 err = sysfs_slab_alias(al->s, al->name);
4411 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4412 " %s to sysfs\n", s->name);
4420 __initcall(slab_sysfs_init);
4424 * The /proc/slabinfo ABI
4426 #ifdef CONFIG_SLABINFO
4427 static void print_slabinfo_header(struct seq_file *m)
4429 seq_puts(m, "slabinfo - version: 2.1\n");
4430 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4431 "<objperslab> <pagesperslab>");
4432 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4433 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4437 static void *s_start(struct seq_file *m, loff_t *pos)
4441 down_read(&slub_lock);
4443 print_slabinfo_header(m);
4445 return seq_list_start(&slab_caches, *pos);
4448 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4450 return seq_list_next(p, &slab_caches, pos);
4453 static void s_stop(struct seq_file *m, void *p)
4455 up_read(&slub_lock);
4458 static int s_show(struct seq_file *m, void *p)
4460 unsigned long nr_partials = 0;
4461 unsigned long nr_slabs = 0;
4462 unsigned long nr_inuse = 0;
4463 unsigned long nr_objs = 0;
4464 unsigned long nr_free = 0;
4465 struct kmem_cache *s;
4468 s = list_entry(p, struct kmem_cache, list);
4470 for_each_online_node(node) {
4471 struct kmem_cache_node *n = get_node(s, node);
4476 nr_partials += n->nr_partial;
4477 nr_slabs += atomic_long_read(&n->nr_slabs);
4478 nr_objs += atomic_long_read(&n->total_objects);
4479 nr_free += count_partial(n, count_free);
4482 nr_inuse = nr_objs - nr_free;
4484 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4485 nr_objs, s->size, oo_objects(s->oo),
4486 (1 << oo_order(s->oo)));
4487 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4488 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4494 static const struct seq_operations slabinfo_op = {
4501 static int slabinfo_open(struct inode *inode, struct file *file)
4503 return seq_open(file, &slabinfo_op);
4506 static const struct file_operations proc_slabinfo_operations = {
4507 .open = slabinfo_open,
4509 .llseek = seq_lseek,
4510 .release = seq_release,
4513 static int __init slab_proc_init(void)
4515 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4518 module_init(slab_proc_init);
4519 #endif /* CONFIG_SLABINFO */