2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemtrace.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
37 * The slab_lock protects operations on the object of a particular
38 * slab and its metadata in the page struct. If the slab lock
39 * has been taken then no allocations nor frees can be performed
40 * on the objects in the slab nor can the slab be added or removed
41 * from the partial or full lists since this would mean modifying
42 * the page_struct of the slab.
44 * The list_lock protects the partial and full list on each node and
45 * the partial slab counter. If taken then no new slabs may be added or
46 * removed from the lists nor make the number of partial slabs be modified.
47 * (Note that the total number of slabs is an atomic value that may be
48 * modified without taking the list lock).
50 * The list_lock is a centralized lock and thus we avoid taking it as
51 * much as possible. As long as SLUB does not have to handle partial
52 * slabs, operations can continue without any centralized lock. F.e.
53 * allocating a long series of objects that fill up slabs does not require
56 * The lock order is sometimes inverted when we are trying to get a slab
57 * off a list. We take the list_lock and then look for a page on the list
58 * to use. While we do that objects in the slabs may be freed. We can
59 * only operate on the slab if we have also taken the slab_lock. So we use
60 * a slab_trylock() on the slab. If trylock was successful then no frees
61 * can occur anymore and we can use the slab for allocations etc. If the
62 * slab_trylock() does not succeed then frees are in progress in the slab and
63 * we must stay away from it for a while since we may cause a bouncing
64 * cacheline if we try to acquire the lock. So go onto the next slab.
65 * If all pages are busy then we may allocate a new slab instead of reusing
66 * a partial slab. A new slab has noone operating on it and thus there is
67 * no danger of cacheline contention.
69 * Interrupts are disabled during allocation and deallocation in order to
70 * make the slab allocator safe to use in the context of an irq. In addition
71 * interrupts are disabled to ensure that the processor does not change
72 * while handling per_cpu slabs, due to kernel preemption.
74 * SLUB assigns one slab for allocation to each processor.
75 * Allocations only occur from these slabs called cpu slabs.
77 * Slabs with free elements are kept on a partial list and during regular
78 * operations no list for full slabs is used. If an object in a full slab is
79 * freed then the slab will show up again on the partial lists.
80 * We track full slabs for debugging purposes though because otherwise we
81 * cannot scan all objects.
83 * Slabs are freed when they become empty. Teardown and setup is
84 * minimal so we rely on the page allocators per cpu caches for
85 * fast frees and allocs.
87 * Overloading of page flags that are otherwise used for LRU management.
89 * PageActive The slab is frozen and exempt from list processing.
90 * This means that the slab is dedicated to a purpose
91 * such as satisfying allocations for a specific
92 * processor. Objects may be freed in the slab while
93 * it is frozen but slab_free will then skip the usual
94 * list operations. It is up to the processor holding
95 * the slab to integrate the slab into the slab lists
96 * when the slab is no longer needed.
98 * One use of this flag is to mark slabs that are
99 * used for allocations. Then such a slab becomes a cpu
100 * slab. The cpu slab may be equipped with an additional
101 * freelist that allows lockless access to
102 * free objects in addition to the regular freelist
103 * that requires the slab lock.
105 * PageError Slab requires special handling due to debug
106 * options set. This moves slab handling out of
107 * the fast path and disables lockless freelists.
110 #ifdef CONFIG_SLUB_DEBUG
117 * Issues still to be resolved:
119 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
121 * - Variable sizing of the per node arrays
124 /* Enable to test recovery from slab corruption on boot */
125 #undef SLUB_RESILIENCY_TEST
128 * Mininum number of partial slabs. These will be left on the partial
129 * lists even if they are empty. kmem_cache_shrink may reclaim them.
131 #define MIN_PARTIAL 5
134 * Maximum number of desirable partial slabs.
135 * The existence of more partial slabs makes kmem_cache_shrink
136 * sort the partial list by the number of objects in the.
138 #define MAX_PARTIAL 10
140 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
141 SLAB_POISON | SLAB_STORE_USER)
144 * Set of flags that will prevent slab merging
146 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
147 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
149 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
150 SLAB_CACHE_DMA | SLAB_NOTRACK)
152 #ifndef ARCH_KMALLOC_MINALIGN
153 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
156 #ifndef ARCH_SLAB_MINALIGN
157 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
161 #define OO_MASK ((1 << OO_SHIFT) - 1)
162 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
164 /* Internal SLUB flags */
165 #define __OBJECT_POISON 0x80000000 /* Poison object */
166 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
168 static int kmem_size = sizeof(struct kmem_cache);
171 static struct notifier_block slab_notifier;
175 DOWN, /* No slab functionality available */
176 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
177 UP, /* Everything works but does not show up in sysfs */
181 /* A list of all slab caches on the system */
182 static DECLARE_RWSEM(slub_lock);
183 static LIST_HEAD(slab_caches);
186 * Tracking user of a slab.
189 unsigned long addr; /* Called from address */
190 int cpu; /* Was running on cpu */
191 int pid; /* Pid context */
192 unsigned long when; /* When did the operation occur */
195 enum track_item { TRACK_ALLOC, TRACK_FREE };
197 #ifdef CONFIG_SLUB_DEBUG
198 static int sysfs_slab_add(struct kmem_cache *);
199 static int sysfs_slab_alias(struct kmem_cache *, const char *);
200 static void sysfs_slab_remove(struct kmem_cache *);
203 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
204 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
206 static inline void sysfs_slab_remove(struct kmem_cache *s)
213 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
215 #ifdef CONFIG_SLUB_STATS
220 /********************************************************************
221 * Core slab cache functions
222 *******************************************************************/
224 int slab_is_available(void)
226 return slab_state >= UP;
229 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
232 return s->node[node];
234 return &s->local_node;
238 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
241 return s->cpu_slab[cpu];
247 /* Verify that a pointer has an address that is valid within a slab page */
248 static inline int check_valid_pointer(struct kmem_cache *s,
249 struct page *page, const void *object)
256 base = page_address(page);
257 if (object < base || object >= base + page->objects * s->size ||
258 (object - base) % s->size) {
266 * Slow version of get and set free pointer.
268 * This version requires touching the cache lines of kmem_cache which
269 * we avoid to do in the fast alloc free paths. There we obtain the offset
270 * from the page struct.
272 static inline void *get_freepointer(struct kmem_cache *s, void *object)
274 return *(void **)(object + s->offset);
277 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
279 *(void **)(object + s->offset) = fp;
282 /* Loop over all objects in a slab */
283 #define for_each_object(__p, __s, __addr, __objects) \
284 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
288 #define for_each_free_object(__p, __s, __free) \
289 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
294 return (p - addr) / s->size;
297 static inline struct kmem_cache_order_objects oo_make(int order,
300 struct kmem_cache_order_objects x = {
301 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
307 static inline int oo_order(struct kmem_cache_order_objects x)
309 return x.x >> OO_SHIFT;
312 static inline int oo_objects(struct kmem_cache_order_objects x)
314 return x.x & OO_MASK;
317 #ifdef CONFIG_SLUB_DEBUG
321 #ifdef CONFIG_SLUB_DEBUG_ON
322 static int slub_debug = DEBUG_DEFAULT_FLAGS;
324 static int slub_debug;
327 static char *slub_debug_slabs;
332 static void print_section(char *text, u8 *addr, unsigned int length)
340 for (i = 0; i < length; i++) {
342 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
345 printk(KERN_CONT " %02x", addr[i]);
347 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
349 printk(KERN_CONT " %s\n", ascii);
356 printk(KERN_CONT " ");
360 printk(KERN_CONT " %s\n", ascii);
364 static struct track *get_track(struct kmem_cache *s, void *object,
365 enum track_item alloc)
370 p = object + s->offset + sizeof(void *);
372 p = object + s->inuse;
377 static void set_track(struct kmem_cache *s, void *object,
378 enum track_item alloc, unsigned long addr)
380 struct track *p = get_track(s, object, alloc);
384 p->cpu = smp_processor_id();
385 p->pid = current->pid;
388 memset(p, 0, sizeof(struct track));
391 static void init_tracking(struct kmem_cache *s, void *object)
393 if (!(s->flags & SLAB_STORE_USER))
396 set_track(s, object, TRACK_FREE, 0UL);
397 set_track(s, object, TRACK_ALLOC, 0UL);
400 static void print_track(const char *s, struct track *t)
405 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
406 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
409 static void print_tracking(struct kmem_cache *s, void *object)
411 if (!(s->flags & SLAB_STORE_USER))
414 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
415 print_track("Freed", get_track(s, object, TRACK_FREE));
418 static void print_page_info(struct page *page)
420 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
421 page, page->objects, page->inuse, page->freelist, page->flags);
425 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
431 vsnprintf(buf, sizeof(buf), fmt, args);
433 printk(KERN_ERR "========================================"
434 "=====================================\n");
435 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
436 printk(KERN_ERR "----------------------------------------"
437 "-------------------------------------\n\n");
440 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
446 vsnprintf(buf, sizeof(buf), fmt, args);
448 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
451 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
453 unsigned int off; /* Offset of last byte */
454 u8 *addr = page_address(page);
456 print_tracking(s, p);
458 print_page_info(page);
460 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
461 p, p - addr, get_freepointer(s, p));
464 print_section("Bytes b4", p - 16, 16);
466 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
468 if (s->flags & SLAB_RED_ZONE)
469 print_section("Redzone", p + s->objsize,
470 s->inuse - s->objsize);
473 off = s->offset + sizeof(void *);
477 if (s->flags & SLAB_STORE_USER)
478 off += 2 * sizeof(struct track);
481 /* Beginning of the filler is the free pointer */
482 print_section("Padding", p + off, s->size - off);
487 static void object_err(struct kmem_cache *s, struct page *page,
488 u8 *object, char *reason)
490 slab_bug(s, "%s", reason);
491 print_trailer(s, page, object);
494 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
500 vsnprintf(buf, sizeof(buf), fmt, args);
502 slab_bug(s, "%s", buf);
503 print_page_info(page);
507 static void init_object(struct kmem_cache *s, void *object, int active)
511 if (s->flags & __OBJECT_POISON) {
512 memset(p, POISON_FREE, s->objsize - 1);
513 p[s->objsize - 1] = POISON_END;
516 if (s->flags & SLAB_RED_ZONE)
517 memset(p + s->objsize,
518 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
519 s->inuse - s->objsize);
522 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
525 if (*start != (u8)value)
533 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
534 void *from, void *to)
536 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
537 memset(from, data, to - from);
540 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
541 u8 *object, char *what,
542 u8 *start, unsigned int value, unsigned int bytes)
547 fault = check_bytes(start, value, bytes);
552 while (end > fault && end[-1] == value)
555 slab_bug(s, "%s overwritten", what);
556 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
557 fault, end - 1, fault[0], value);
558 print_trailer(s, page, object);
560 restore_bytes(s, what, value, fault, end);
568 * Bytes of the object to be managed.
569 * If the freepointer may overlay the object then the free
570 * pointer is the first word of the object.
572 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
575 * object + s->objsize
576 * Padding to reach word boundary. This is also used for Redzoning.
577 * Padding is extended by another word if Redzoning is enabled and
580 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
581 * 0xcc (RED_ACTIVE) for objects in use.
584 * Meta data starts here.
586 * A. Free pointer (if we cannot overwrite object on free)
587 * B. Tracking data for SLAB_STORE_USER
588 * C. Padding to reach required alignment boundary or at mininum
589 * one word if debugging is on to be able to detect writes
590 * before the word boundary.
592 * Padding is done using 0x5a (POISON_INUSE)
595 * Nothing is used beyond s->size.
597 * If slabcaches are merged then the objsize and inuse boundaries are mostly
598 * ignored. And therefore no slab options that rely on these boundaries
599 * may be used with merged slabcaches.
602 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
604 unsigned long off = s->inuse; /* The end of info */
607 /* Freepointer is placed after the object. */
608 off += sizeof(void *);
610 if (s->flags & SLAB_STORE_USER)
611 /* We also have user information there */
612 off += 2 * sizeof(struct track);
617 return check_bytes_and_report(s, page, p, "Object padding",
618 p + off, POISON_INUSE, s->size - off);
621 /* Check the pad bytes at the end of a slab page */
622 static int slab_pad_check(struct kmem_cache *s, struct page *page)
630 if (!(s->flags & SLAB_POISON))
633 start = page_address(page);
634 length = (PAGE_SIZE << compound_order(page));
635 end = start + length;
636 remainder = length % s->size;
640 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
643 while (end > fault && end[-1] == POISON_INUSE)
646 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
647 print_section("Padding", end - remainder, remainder);
649 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
653 static int check_object(struct kmem_cache *s, struct page *page,
654 void *object, int active)
657 u8 *endobject = object + s->objsize;
659 if (s->flags & SLAB_RED_ZONE) {
661 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
663 if (!check_bytes_and_report(s, page, object, "Redzone",
664 endobject, red, s->inuse - s->objsize))
667 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
668 check_bytes_and_report(s, page, p, "Alignment padding",
669 endobject, POISON_INUSE, s->inuse - s->objsize);
673 if (s->flags & SLAB_POISON) {
674 if (!active && (s->flags & __OBJECT_POISON) &&
675 (!check_bytes_and_report(s, page, p, "Poison", p,
676 POISON_FREE, s->objsize - 1) ||
677 !check_bytes_and_report(s, page, p, "Poison",
678 p + s->objsize - 1, POISON_END, 1)))
681 * check_pad_bytes cleans up on its own.
683 check_pad_bytes(s, page, p);
686 if (!s->offset && active)
688 * Object and freepointer overlap. Cannot check
689 * freepointer while object is allocated.
693 /* Check free pointer validity */
694 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
695 object_err(s, page, p, "Freepointer corrupt");
697 * No choice but to zap it and thus lose the remainder
698 * of the free objects in this slab. May cause
699 * another error because the object count is now wrong.
701 set_freepointer(s, p, NULL);
707 static int check_slab(struct kmem_cache *s, struct page *page)
711 VM_BUG_ON(!irqs_disabled());
713 if (!PageSlab(page)) {
714 slab_err(s, page, "Not a valid slab page");
718 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
719 if (page->objects > maxobj) {
720 slab_err(s, page, "objects %u > max %u",
721 s->name, page->objects, maxobj);
724 if (page->inuse > page->objects) {
725 slab_err(s, page, "inuse %u > max %u",
726 s->name, page->inuse, page->objects);
729 /* Slab_pad_check fixes things up after itself */
730 slab_pad_check(s, page);
735 * Determine if a certain object on a page is on the freelist. Must hold the
736 * slab lock to guarantee that the chains are in a consistent state.
738 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
741 void *fp = page->freelist;
743 unsigned long max_objects;
745 while (fp && nr <= page->objects) {
748 if (!check_valid_pointer(s, page, fp)) {
750 object_err(s, page, object,
751 "Freechain corrupt");
752 set_freepointer(s, object, NULL);
755 slab_err(s, page, "Freepointer corrupt");
756 page->freelist = NULL;
757 page->inuse = page->objects;
758 slab_fix(s, "Freelist cleared");
764 fp = get_freepointer(s, object);
768 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
769 if (max_objects > MAX_OBJS_PER_PAGE)
770 max_objects = MAX_OBJS_PER_PAGE;
772 if (page->objects != max_objects) {
773 slab_err(s, page, "Wrong number of objects. Found %d but "
774 "should be %d", page->objects, max_objects);
775 page->objects = max_objects;
776 slab_fix(s, "Number of objects adjusted.");
778 if (page->inuse != page->objects - nr) {
779 slab_err(s, page, "Wrong object count. Counter is %d but "
780 "counted were %d", page->inuse, page->objects - nr);
781 page->inuse = page->objects - nr;
782 slab_fix(s, "Object count adjusted.");
784 return search == NULL;
787 static void trace(struct kmem_cache *s, struct page *page, void *object,
790 if (s->flags & SLAB_TRACE) {
791 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
793 alloc ? "alloc" : "free",
798 print_section("Object", (void *)object, s->objsize);
805 * Tracking of fully allocated slabs for debugging purposes.
807 static void add_full(struct kmem_cache_node *n, struct page *page)
809 spin_lock(&n->list_lock);
810 list_add(&page->lru, &n->full);
811 spin_unlock(&n->list_lock);
814 static void remove_full(struct kmem_cache *s, struct page *page)
816 struct kmem_cache_node *n;
818 if (!(s->flags & SLAB_STORE_USER))
821 n = get_node(s, page_to_nid(page));
823 spin_lock(&n->list_lock);
824 list_del(&page->lru);
825 spin_unlock(&n->list_lock);
828 /* Tracking of the number of slabs for debugging purposes */
829 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
831 struct kmem_cache_node *n = get_node(s, node);
833 return atomic_long_read(&n->nr_slabs);
836 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
838 return atomic_long_read(&n->nr_slabs);
841 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
843 struct kmem_cache_node *n = get_node(s, node);
846 * May be called early in order to allocate a slab for the
847 * kmem_cache_node structure. Solve the chicken-egg
848 * dilemma by deferring the increment of the count during
849 * bootstrap (see early_kmem_cache_node_alloc).
851 if (!NUMA_BUILD || n) {
852 atomic_long_inc(&n->nr_slabs);
853 atomic_long_add(objects, &n->total_objects);
856 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
858 struct kmem_cache_node *n = get_node(s, node);
860 atomic_long_dec(&n->nr_slabs);
861 atomic_long_sub(objects, &n->total_objects);
864 /* Object debug checks for alloc/free paths */
865 static void setup_object_debug(struct kmem_cache *s, struct page *page,
868 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
871 init_object(s, object, 0);
872 init_tracking(s, object);
875 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
876 void *object, unsigned long addr)
878 if (!check_slab(s, page))
881 if (!on_freelist(s, page, object)) {
882 object_err(s, page, object, "Object already allocated");
886 if (!check_valid_pointer(s, page, object)) {
887 object_err(s, page, object, "Freelist Pointer check fails");
891 if (!check_object(s, page, object, 0))
894 /* Success perform special debug activities for allocs */
895 if (s->flags & SLAB_STORE_USER)
896 set_track(s, object, TRACK_ALLOC, addr);
897 trace(s, page, object, 1);
898 init_object(s, object, 1);
902 if (PageSlab(page)) {
904 * If this is a slab page then lets do the best we can
905 * to avoid issues in the future. Marking all objects
906 * as used avoids touching the remaining objects.
908 slab_fix(s, "Marking all objects used");
909 page->inuse = page->objects;
910 page->freelist = NULL;
915 static int free_debug_processing(struct kmem_cache *s, struct page *page,
916 void *object, unsigned long addr)
918 if (!check_slab(s, page))
921 if (!check_valid_pointer(s, page, object)) {
922 slab_err(s, page, "Invalid object pointer 0x%p", object);
926 if (on_freelist(s, page, object)) {
927 object_err(s, page, object, "Object already free");
931 if (!check_object(s, page, object, 1))
934 if (unlikely(s != page->slab)) {
935 if (!PageSlab(page)) {
936 slab_err(s, page, "Attempt to free object(0x%p) "
937 "outside of slab", object);
938 } else if (!page->slab) {
940 "SLUB <none>: no slab for object 0x%p.\n",
944 object_err(s, page, object,
945 "page slab pointer corrupt.");
949 /* Special debug activities for freeing objects */
950 if (!PageSlubFrozen(page) && !page->freelist)
951 remove_full(s, page);
952 if (s->flags & SLAB_STORE_USER)
953 set_track(s, object, TRACK_FREE, addr);
954 trace(s, page, object, 0);
955 init_object(s, object, 0);
959 slab_fix(s, "Object at 0x%p not freed", object);
963 static int __init setup_slub_debug(char *str)
965 slub_debug = DEBUG_DEFAULT_FLAGS;
966 if (*str++ != '=' || !*str)
968 * No options specified. Switch on full debugging.
974 * No options but restriction on slabs. This means full
975 * debugging for slabs matching a pattern.
982 * Switch off all debugging measures.
987 * Determine which debug features should be switched on
989 for (; *str && *str != ','; str++) {
990 switch (tolower(*str)) {
992 slub_debug |= SLAB_DEBUG_FREE;
995 slub_debug |= SLAB_RED_ZONE;
998 slub_debug |= SLAB_POISON;
1001 slub_debug |= SLAB_STORE_USER;
1004 slub_debug |= SLAB_TRACE;
1007 printk(KERN_ERR "slub_debug option '%c' "
1008 "unknown. skipped\n", *str);
1014 slub_debug_slabs = str + 1;
1019 __setup("slub_debug", setup_slub_debug);
1021 static unsigned long kmem_cache_flags(unsigned long objsize,
1022 unsigned long flags, const char *name,
1023 void (*ctor)(void *))
1026 * Enable debugging if selected on the kernel commandline.
1028 if (slub_debug && (!slub_debug_slabs ||
1029 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1030 flags |= slub_debug;
1035 static inline void setup_object_debug(struct kmem_cache *s,
1036 struct page *page, void *object) {}
1038 static inline int alloc_debug_processing(struct kmem_cache *s,
1039 struct page *page, void *object, unsigned long addr) { return 0; }
1041 static inline int free_debug_processing(struct kmem_cache *s,
1042 struct page *page, void *object, unsigned long addr) { return 0; }
1044 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1046 static inline int check_object(struct kmem_cache *s, struct page *page,
1047 void *object, int active) { return 1; }
1048 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1049 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1050 unsigned long flags, const char *name,
1051 void (*ctor)(void *))
1055 #define slub_debug 0
1057 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1059 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1061 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1063 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1068 * Slab allocation and freeing
1070 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1071 struct kmem_cache_order_objects oo)
1073 int order = oo_order(oo);
1075 flags |= __GFP_NOTRACK;
1078 return alloc_pages(flags, order);
1080 return alloc_pages_node(node, flags, order);
1083 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1086 struct kmem_cache_order_objects oo = s->oo;
1089 flags |= s->allocflags;
1092 * Let the initial higher-order allocation fail under memory pressure
1093 * so we fall-back to the minimum order allocation.
1095 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1097 page = alloc_slab_page(alloc_gfp, node, oo);
1098 if (unlikely(!page)) {
1101 * Allocation may have failed due to fragmentation.
1102 * Try a lower order alloc if possible
1104 page = alloc_slab_page(flags, node, oo);
1108 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1111 if (kmemcheck_enabled
1112 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1113 int pages = 1 << oo_order(oo);
1115 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1118 * Objects from caches that have a constructor don't get
1119 * cleared when they're allocated, so we need to do it here.
1122 kmemcheck_mark_uninitialized_pages(page, pages);
1124 kmemcheck_mark_unallocated_pages(page, pages);
1127 page->objects = oo_objects(oo);
1128 mod_zone_page_state(page_zone(page),
1129 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1130 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1136 static void setup_object(struct kmem_cache *s, struct page *page,
1139 setup_object_debug(s, page, object);
1140 if (unlikely(s->ctor))
1144 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1151 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1153 page = allocate_slab(s,
1154 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1158 inc_slabs_node(s, page_to_nid(page), page->objects);
1160 page->flags |= 1 << PG_slab;
1161 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1162 SLAB_STORE_USER | SLAB_TRACE))
1163 __SetPageSlubDebug(page);
1165 start = page_address(page);
1167 if (unlikely(s->flags & SLAB_POISON))
1168 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1171 for_each_object(p, s, start, page->objects) {
1172 setup_object(s, page, last);
1173 set_freepointer(s, last, p);
1176 setup_object(s, page, last);
1177 set_freepointer(s, last, NULL);
1179 page->freelist = start;
1185 static void __free_slab(struct kmem_cache *s, struct page *page)
1187 int order = compound_order(page);
1188 int pages = 1 << order;
1190 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1193 slab_pad_check(s, page);
1194 for_each_object(p, s, page_address(page),
1196 check_object(s, page, p, 0);
1197 __ClearPageSlubDebug(page);
1200 kmemcheck_free_shadow(page, compound_order(page));
1202 mod_zone_page_state(page_zone(page),
1203 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1204 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1207 __ClearPageSlab(page);
1208 reset_page_mapcount(page);
1209 if (current->reclaim_state)
1210 current->reclaim_state->reclaimed_slab += pages;
1211 __free_pages(page, order);
1214 static void rcu_free_slab(struct rcu_head *h)
1218 page = container_of((struct list_head *)h, struct page, lru);
1219 __free_slab(page->slab, page);
1222 static void free_slab(struct kmem_cache *s, struct page *page)
1224 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1226 * RCU free overloads the RCU head over the LRU
1228 struct rcu_head *head = (void *)&page->lru;
1230 call_rcu(head, rcu_free_slab);
1232 __free_slab(s, page);
1235 static void discard_slab(struct kmem_cache *s, struct page *page)
1237 dec_slabs_node(s, page_to_nid(page), page->objects);
1242 * Per slab locking using the pagelock
1244 static __always_inline void slab_lock(struct page *page)
1246 bit_spin_lock(PG_locked, &page->flags);
1249 static __always_inline void slab_unlock(struct page *page)
1251 __bit_spin_unlock(PG_locked, &page->flags);
1254 static __always_inline int slab_trylock(struct page *page)
1258 rc = bit_spin_trylock(PG_locked, &page->flags);
1263 * Management of partially allocated slabs
1265 static void add_partial(struct kmem_cache_node *n,
1266 struct page *page, int tail)
1268 spin_lock(&n->list_lock);
1271 list_add_tail(&page->lru, &n->partial);
1273 list_add(&page->lru, &n->partial);
1274 spin_unlock(&n->list_lock);
1277 static void remove_partial(struct kmem_cache *s, struct page *page)
1279 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1281 spin_lock(&n->list_lock);
1282 list_del(&page->lru);
1284 spin_unlock(&n->list_lock);
1288 * Lock slab and remove from the partial list.
1290 * Must hold list_lock.
1292 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1295 if (slab_trylock(page)) {
1296 list_del(&page->lru);
1298 __SetPageSlubFrozen(page);
1305 * Try to allocate a partial slab from a specific node.
1307 static struct page *get_partial_node(struct kmem_cache_node *n)
1312 * Racy check. If we mistakenly see no partial slabs then we
1313 * just allocate an empty slab. If we mistakenly try to get a
1314 * partial slab and there is none available then get_partials()
1317 if (!n || !n->nr_partial)
1320 spin_lock(&n->list_lock);
1321 list_for_each_entry(page, &n->partial, lru)
1322 if (lock_and_freeze_slab(n, page))
1326 spin_unlock(&n->list_lock);
1331 * Get a page from somewhere. Search in increasing NUMA distances.
1333 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1336 struct zonelist *zonelist;
1339 enum zone_type high_zoneidx = gfp_zone(flags);
1343 * The defrag ratio allows a configuration of the tradeoffs between
1344 * inter node defragmentation and node local allocations. A lower
1345 * defrag_ratio increases the tendency to do local allocations
1346 * instead of attempting to obtain partial slabs from other nodes.
1348 * If the defrag_ratio is set to 0 then kmalloc() always
1349 * returns node local objects. If the ratio is higher then kmalloc()
1350 * may return off node objects because partial slabs are obtained
1351 * from other nodes and filled up.
1353 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1354 * defrag_ratio = 1000) then every (well almost) allocation will
1355 * first attempt to defrag slab caches on other nodes. This means
1356 * scanning over all nodes to look for partial slabs which may be
1357 * expensive if we do it every time we are trying to find a slab
1358 * with available objects.
1360 if (!s->remote_node_defrag_ratio ||
1361 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1364 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1365 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1366 struct kmem_cache_node *n;
1368 n = get_node(s, zone_to_nid(zone));
1370 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1371 n->nr_partial > s->min_partial) {
1372 page = get_partial_node(n);
1382 * Get a partial page, lock it and return it.
1384 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1387 int searchnode = (node == -1) ? numa_node_id() : node;
1389 page = get_partial_node(get_node(s, searchnode));
1390 if (page || (flags & __GFP_THISNODE))
1393 return get_any_partial(s, flags);
1397 * Move a page back to the lists.
1399 * Must be called with the slab lock held.
1401 * On exit the slab lock will have been dropped.
1403 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1405 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1406 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1408 __ClearPageSlubFrozen(page);
1411 if (page->freelist) {
1412 add_partial(n, page, tail);
1413 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1415 stat(c, DEACTIVATE_FULL);
1416 if (SLABDEBUG && PageSlubDebug(page) &&
1417 (s->flags & SLAB_STORE_USER))
1422 stat(c, DEACTIVATE_EMPTY);
1423 if (n->nr_partial < s->min_partial) {
1425 * Adding an empty slab to the partial slabs in order
1426 * to avoid page allocator overhead. This slab needs
1427 * to come after the other slabs with objects in
1428 * so that the others get filled first. That way the
1429 * size of the partial list stays small.
1431 * kmem_cache_shrink can reclaim any empty slabs from
1434 add_partial(n, page, 1);
1438 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1439 discard_slab(s, page);
1445 * Remove the cpu slab
1447 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1449 struct page *page = c->page;
1453 stat(c, DEACTIVATE_REMOTE_FREES);
1455 * Merge cpu freelist into slab freelist. Typically we get here
1456 * because both freelists are empty. So this is unlikely
1459 while (unlikely(c->freelist)) {
1462 tail = 0; /* Hot objects. Put the slab first */
1464 /* Retrieve object from cpu_freelist */
1465 object = c->freelist;
1466 c->freelist = c->freelist[c->offset];
1468 /* And put onto the regular freelist */
1469 object[c->offset] = page->freelist;
1470 page->freelist = object;
1474 unfreeze_slab(s, page, tail);
1477 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1479 stat(c, CPUSLAB_FLUSH);
1481 deactivate_slab(s, c);
1487 * Called from IPI handler with interrupts disabled.
1489 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1491 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1493 if (likely(c && c->page))
1497 static void flush_cpu_slab(void *d)
1499 struct kmem_cache *s = d;
1501 __flush_cpu_slab(s, smp_processor_id());
1504 static void flush_all(struct kmem_cache *s)
1506 on_each_cpu(flush_cpu_slab, s, 1);
1510 * Check if the objects in a per cpu structure fit numa
1511 * locality expectations.
1513 static inline int node_match(struct kmem_cache_cpu *c, int node)
1516 if (node != -1 && c->node != node)
1522 static int count_free(struct page *page)
1524 return page->objects - page->inuse;
1527 static unsigned long count_partial(struct kmem_cache_node *n,
1528 int (*get_count)(struct page *))
1530 unsigned long flags;
1531 unsigned long x = 0;
1534 spin_lock_irqsave(&n->list_lock, flags);
1535 list_for_each_entry(page, &n->partial, lru)
1536 x += get_count(page);
1537 spin_unlock_irqrestore(&n->list_lock, flags);
1541 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1543 #ifdef CONFIG_SLUB_DEBUG
1544 return atomic_long_read(&n->total_objects);
1550 static noinline void
1551 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1556 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1558 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1559 "default order: %d, min order: %d\n", s->name, s->objsize,
1560 s->size, oo_order(s->oo), oo_order(s->min));
1562 for_each_online_node(node) {
1563 struct kmem_cache_node *n = get_node(s, node);
1564 unsigned long nr_slabs;
1565 unsigned long nr_objs;
1566 unsigned long nr_free;
1571 nr_free = count_partial(n, count_free);
1572 nr_slabs = node_nr_slabs(n);
1573 nr_objs = node_nr_objs(n);
1576 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1577 node, nr_slabs, nr_objs, nr_free);
1582 * Slow path. The lockless freelist is empty or we need to perform
1585 * Interrupts are disabled.
1587 * Processing is still very fast if new objects have been freed to the
1588 * regular freelist. In that case we simply take over the regular freelist
1589 * as the lockless freelist and zap the regular freelist.
1591 * If that is not working then we fall back to the partial lists. We take the
1592 * first element of the freelist as the object to allocate now and move the
1593 * rest of the freelist to the lockless freelist.
1595 * And if we were unable to get a new slab from the partial slab lists then
1596 * we need to allocate a new slab. This is the slowest path since it involves
1597 * a call to the page allocator and the setup of a new slab.
1599 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1600 unsigned long addr, struct kmem_cache_cpu *c)
1605 /* We handle __GFP_ZERO in the caller */
1606 gfpflags &= ~__GFP_ZERO;
1612 if (unlikely(!node_match(c, node)))
1615 stat(c, ALLOC_REFILL);
1618 object = c->page->freelist;
1619 if (unlikely(!object))
1621 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1624 c->freelist = object[c->offset];
1625 c->page->inuse = c->page->objects;
1626 c->page->freelist = NULL;
1627 c->node = page_to_nid(c->page);
1629 slab_unlock(c->page);
1630 stat(c, ALLOC_SLOWPATH);
1634 deactivate_slab(s, c);
1637 new = get_partial(s, gfpflags, node);
1640 stat(c, ALLOC_FROM_PARTIAL);
1644 if (gfpflags & __GFP_WAIT)
1647 new = new_slab(s, gfpflags, node);
1649 if (gfpflags & __GFP_WAIT)
1650 local_irq_disable();
1653 c = get_cpu_slab(s, smp_processor_id());
1654 stat(c, ALLOC_SLAB);
1658 __SetPageSlubFrozen(new);
1662 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1663 slab_out_of_memory(s, gfpflags, node);
1666 if (!alloc_debug_processing(s, c->page, object, addr))
1670 c->page->freelist = object[c->offset];
1676 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1677 * have the fastpath folded into their functions. So no function call
1678 * overhead for requests that can be satisfied on the fastpath.
1680 * The fastpath works by first checking if the lockless freelist can be used.
1681 * If not then __slab_alloc is called for slow processing.
1683 * Otherwise we can simply pick the next object from the lockless free list.
1685 static __always_inline void *slab_alloc(struct kmem_cache *s,
1686 gfp_t gfpflags, int node, unsigned long addr)
1689 struct kmem_cache_cpu *c;
1690 unsigned long flags;
1691 unsigned int objsize;
1693 gfpflags &= gfp_allowed_mask;
1695 lockdep_trace_alloc(gfpflags);
1696 might_sleep_if(gfpflags & __GFP_WAIT);
1698 if (should_failslab(s->objsize, gfpflags))
1701 local_irq_save(flags);
1702 c = get_cpu_slab(s, smp_processor_id());
1703 objsize = c->objsize;
1704 if (unlikely(!c->freelist || !node_match(c, node)))
1706 object = __slab_alloc(s, gfpflags, node, addr, c);
1709 object = c->freelist;
1710 c->freelist = object[c->offset];
1711 stat(c, ALLOC_FASTPATH);
1713 local_irq_restore(flags);
1715 if (unlikely((gfpflags & __GFP_ZERO) && object))
1716 memset(object, 0, objsize);
1718 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize);
1719 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
1724 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1726 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1728 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1732 EXPORT_SYMBOL(kmem_cache_alloc);
1734 #ifdef CONFIG_KMEMTRACE
1735 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1737 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1739 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1743 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1745 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1747 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1748 s->objsize, s->size, gfpflags, node);
1752 EXPORT_SYMBOL(kmem_cache_alloc_node);
1755 #ifdef CONFIG_KMEMTRACE
1756 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1760 return slab_alloc(s, gfpflags, node, _RET_IP_);
1762 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1766 * Slow patch handling. This may still be called frequently since objects
1767 * have a longer lifetime than the cpu slabs in most processing loads.
1769 * So we still attempt to reduce cache line usage. Just take the slab
1770 * lock and free the item. If there is no additional partial page
1771 * handling required then we can return immediately.
1773 static void __slab_free(struct kmem_cache *s, struct page *page,
1774 void *x, unsigned long addr, unsigned int offset)
1777 void **object = (void *)x;
1778 struct kmem_cache_cpu *c;
1780 c = get_cpu_slab(s, raw_smp_processor_id());
1781 stat(c, FREE_SLOWPATH);
1784 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1788 prior = object[offset] = page->freelist;
1789 page->freelist = object;
1792 if (unlikely(PageSlubFrozen(page))) {
1793 stat(c, FREE_FROZEN);
1797 if (unlikely(!page->inuse))
1801 * Objects left in the slab. If it was not on the partial list before
1804 if (unlikely(!prior)) {
1805 add_partial(get_node(s, page_to_nid(page)), page, 1);
1806 stat(c, FREE_ADD_PARTIAL);
1816 * Slab still on the partial list.
1818 remove_partial(s, page);
1819 stat(c, FREE_REMOVE_PARTIAL);
1823 discard_slab(s, page);
1827 if (!free_debug_processing(s, page, x, addr))
1833 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1834 * can perform fastpath freeing without additional function calls.
1836 * The fastpath is only possible if we are freeing to the current cpu slab
1837 * of this processor. This typically the case if we have just allocated
1840 * If fastpath is not possible then fall back to __slab_free where we deal
1841 * with all sorts of special processing.
1843 static __always_inline void slab_free(struct kmem_cache *s,
1844 struct page *page, void *x, unsigned long addr)
1846 void **object = (void *)x;
1847 struct kmem_cache_cpu *c;
1848 unsigned long flags;
1850 kmemleak_free_recursive(x, s->flags);
1851 local_irq_save(flags);
1852 c = get_cpu_slab(s, smp_processor_id());
1853 kmemcheck_slab_free(s, object, c->objsize);
1854 debug_check_no_locks_freed(object, c->objsize);
1855 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1856 debug_check_no_obj_freed(object, c->objsize);
1857 if (likely(page == c->page && c->node >= 0)) {
1858 object[c->offset] = c->freelist;
1859 c->freelist = object;
1860 stat(c, FREE_FASTPATH);
1862 __slab_free(s, page, x, addr, c->offset);
1864 local_irq_restore(flags);
1867 void kmem_cache_free(struct kmem_cache *s, void *x)
1871 page = virt_to_head_page(x);
1873 slab_free(s, page, x, _RET_IP_);
1875 trace_kmem_cache_free(_RET_IP_, x);
1877 EXPORT_SYMBOL(kmem_cache_free);
1879 /* Figure out on which slab page the object resides */
1880 static struct page *get_object_page(const void *x)
1882 struct page *page = virt_to_head_page(x);
1884 if (!PageSlab(page))
1891 * Object placement in a slab is made very easy because we always start at
1892 * offset 0. If we tune the size of the object to the alignment then we can
1893 * get the required alignment by putting one properly sized object after
1896 * Notice that the allocation order determines the sizes of the per cpu
1897 * caches. Each processor has always one slab available for allocations.
1898 * Increasing the allocation order reduces the number of times that slabs
1899 * must be moved on and off the partial lists and is therefore a factor in
1904 * Mininum / Maximum order of slab pages. This influences locking overhead
1905 * and slab fragmentation. A higher order reduces the number of partial slabs
1906 * and increases the number of allocations possible without having to
1907 * take the list_lock.
1909 static int slub_min_order;
1910 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1911 static int slub_min_objects;
1914 * Merge control. If this is set then no merging of slab caches will occur.
1915 * (Could be removed. This was introduced to pacify the merge skeptics.)
1917 static int slub_nomerge;
1920 * Calculate the order of allocation given an slab object size.
1922 * The order of allocation has significant impact on performance and other
1923 * system components. Generally order 0 allocations should be preferred since
1924 * order 0 does not cause fragmentation in the page allocator. Larger objects
1925 * be problematic to put into order 0 slabs because there may be too much
1926 * unused space left. We go to a higher order if more than 1/16th of the slab
1929 * In order to reach satisfactory performance we must ensure that a minimum
1930 * number of objects is in one slab. Otherwise we may generate too much
1931 * activity on the partial lists which requires taking the list_lock. This is
1932 * less a concern for large slabs though which are rarely used.
1934 * slub_max_order specifies the order where we begin to stop considering the
1935 * number of objects in a slab as critical. If we reach slub_max_order then
1936 * we try to keep the page order as low as possible. So we accept more waste
1937 * of space in favor of a small page order.
1939 * Higher order allocations also allow the placement of more objects in a
1940 * slab and thereby reduce object handling overhead. If the user has
1941 * requested a higher mininum order then we start with that one instead of
1942 * the smallest order which will fit the object.
1944 static inline int slab_order(int size, int min_objects,
1945 int max_order, int fract_leftover)
1949 int min_order = slub_min_order;
1951 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1952 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1954 for (order = max(min_order,
1955 fls(min_objects * size - 1) - PAGE_SHIFT);
1956 order <= max_order; order++) {
1958 unsigned long slab_size = PAGE_SIZE << order;
1960 if (slab_size < min_objects * size)
1963 rem = slab_size % size;
1965 if (rem <= slab_size / fract_leftover)
1973 static inline int calculate_order(int size)
1981 * Attempt to find best configuration for a slab. This
1982 * works by first attempting to generate a layout with
1983 * the best configuration and backing off gradually.
1985 * First we reduce the acceptable waste in a slab. Then
1986 * we reduce the minimum objects required in a slab.
1988 min_objects = slub_min_objects;
1990 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1991 max_objects = (PAGE_SIZE << slub_max_order)/size;
1992 min_objects = min(min_objects, max_objects);
1994 while (min_objects > 1) {
1996 while (fraction >= 4) {
1997 order = slab_order(size, min_objects,
1998 slub_max_order, fraction);
1999 if (order <= slub_max_order)
2007 * We were unable to place multiple objects in a slab. Now
2008 * lets see if we can place a single object there.
2010 order = slab_order(size, 1, slub_max_order, 1);
2011 if (order <= slub_max_order)
2015 * Doh this slab cannot be placed using slub_max_order.
2017 order = slab_order(size, 1, MAX_ORDER, 1);
2018 if (order < MAX_ORDER)
2024 * Figure out what the alignment of the objects will be.
2026 static unsigned long calculate_alignment(unsigned long flags,
2027 unsigned long align, unsigned long size)
2030 * If the user wants hardware cache aligned objects then follow that
2031 * suggestion if the object is sufficiently large.
2033 * The hardware cache alignment cannot override the specified
2034 * alignment though. If that is greater then use it.
2036 if (flags & SLAB_HWCACHE_ALIGN) {
2037 unsigned long ralign = cache_line_size();
2038 while (size <= ralign / 2)
2040 align = max(align, ralign);
2043 if (align < ARCH_SLAB_MINALIGN)
2044 align = ARCH_SLAB_MINALIGN;
2046 return ALIGN(align, sizeof(void *));
2049 static void init_kmem_cache_cpu(struct kmem_cache *s,
2050 struct kmem_cache_cpu *c)
2055 c->offset = s->offset / sizeof(void *);
2056 c->objsize = s->objsize;
2057 #ifdef CONFIG_SLUB_STATS
2058 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
2063 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2066 spin_lock_init(&n->list_lock);
2067 INIT_LIST_HEAD(&n->partial);
2068 #ifdef CONFIG_SLUB_DEBUG
2069 atomic_long_set(&n->nr_slabs, 0);
2070 atomic_long_set(&n->total_objects, 0);
2071 INIT_LIST_HEAD(&n->full);
2077 * Per cpu array for per cpu structures.
2079 * The per cpu array places all kmem_cache_cpu structures from one processor
2080 * close together meaning that it becomes possible that multiple per cpu
2081 * structures are contained in one cacheline. This may be particularly
2082 * beneficial for the kmalloc caches.
2084 * A desktop system typically has around 60-80 slabs. With 100 here we are
2085 * likely able to get per cpu structures for all caches from the array defined
2086 * here. We must be able to cover all kmalloc caches during bootstrap.
2088 * If the per cpu array is exhausted then fall back to kmalloc
2089 * of individual cachelines. No sharing is possible then.
2091 #define NR_KMEM_CACHE_CPU 100
2093 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2094 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2096 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2097 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
2099 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2100 int cpu, gfp_t flags)
2102 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2105 per_cpu(kmem_cache_cpu_free, cpu) =
2106 (void *)c->freelist;
2108 /* Table overflow: So allocate ourselves */
2110 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2111 flags, cpu_to_node(cpu));
2116 init_kmem_cache_cpu(s, c);
2120 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2122 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2123 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2127 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2128 per_cpu(kmem_cache_cpu_free, cpu) = c;
2131 static void free_kmem_cache_cpus(struct kmem_cache *s)
2135 for_each_online_cpu(cpu) {
2136 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2139 s->cpu_slab[cpu] = NULL;
2140 free_kmem_cache_cpu(c, cpu);
2145 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2149 for_each_online_cpu(cpu) {
2150 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2155 c = alloc_kmem_cache_cpu(s, cpu, flags);
2157 free_kmem_cache_cpus(s);
2160 s->cpu_slab[cpu] = c;
2166 * Initialize the per cpu array.
2168 static void init_alloc_cpu_cpu(int cpu)
2172 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2175 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2176 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2178 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2181 static void __init init_alloc_cpu(void)
2185 for_each_online_cpu(cpu)
2186 init_alloc_cpu_cpu(cpu);
2190 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2191 static inline void init_alloc_cpu(void) {}
2193 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2195 init_kmem_cache_cpu(s, &s->cpu_slab);
2202 * No kmalloc_node yet so do it by hand. We know that this is the first
2203 * slab on the node for this slabcache. There are no concurrent accesses
2206 * Note that this function only works on the kmalloc_node_cache
2207 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2208 * memory on a fresh node that has no slab structures yet.
2210 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2213 struct kmem_cache_node *n;
2214 unsigned long flags;
2216 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2218 page = new_slab(kmalloc_caches, gfpflags, node);
2221 if (page_to_nid(page) != node) {
2222 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2224 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2225 "in order to be able to continue\n");
2230 page->freelist = get_freepointer(kmalloc_caches, n);
2232 kmalloc_caches->node[node] = n;
2233 #ifdef CONFIG_SLUB_DEBUG
2234 init_object(kmalloc_caches, n, 1);
2235 init_tracking(kmalloc_caches, n);
2237 init_kmem_cache_node(n, kmalloc_caches);
2238 inc_slabs_node(kmalloc_caches, node, page->objects);
2241 * lockdep requires consistent irq usage for each lock
2242 * so even though there cannot be a race this early in
2243 * the boot sequence, we still disable irqs.
2245 local_irq_save(flags);
2246 add_partial(n, page, 0);
2247 local_irq_restore(flags);
2250 static void free_kmem_cache_nodes(struct kmem_cache *s)
2254 for_each_node_state(node, N_NORMAL_MEMORY) {
2255 struct kmem_cache_node *n = s->node[node];
2256 if (n && n != &s->local_node)
2257 kmem_cache_free(kmalloc_caches, n);
2258 s->node[node] = NULL;
2262 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2267 if (slab_state >= UP)
2268 local_node = page_to_nid(virt_to_page(s));
2272 for_each_node_state(node, N_NORMAL_MEMORY) {
2273 struct kmem_cache_node *n;
2275 if (local_node == node)
2278 if (slab_state == DOWN) {
2279 early_kmem_cache_node_alloc(gfpflags, node);
2282 n = kmem_cache_alloc_node(kmalloc_caches,
2286 free_kmem_cache_nodes(s);
2292 init_kmem_cache_node(n, s);
2297 static void free_kmem_cache_nodes(struct kmem_cache *s)
2301 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2303 init_kmem_cache_node(&s->local_node, s);
2308 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2310 if (min < MIN_PARTIAL)
2312 else if (min > MAX_PARTIAL)
2314 s->min_partial = min;
2318 * calculate_sizes() determines the order and the distribution of data within
2321 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2323 unsigned long flags = s->flags;
2324 unsigned long size = s->objsize;
2325 unsigned long align = s->align;
2329 * Round up object size to the next word boundary. We can only
2330 * place the free pointer at word boundaries and this determines
2331 * the possible location of the free pointer.
2333 size = ALIGN(size, sizeof(void *));
2335 #ifdef CONFIG_SLUB_DEBUG
2337 * Determine if we can poison the object itself. If the user of
2338 * the slab may touch the object after free or before allocation
2339 * then we should never poison the object itself.
2341 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2343 s->flags |= __OBJECT_POISON;
2345 s->flags &= ~__OBJECT_POISON;
2349 * If we are Redzoning then check if there is some space between the
2350 * end of the object and the free pointer. If not then add an
2351 * additional word to have some bytes to store Redzone information.
2353 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2354 size += sizeof(void *);
2358 * With that we have determined the number of bytes in actual use
2359 * by the object. This is the potential offset to the free pointer.
2363 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2366 * Relocate free pointer after the object if it is not
2367 * permitted to overwrite the first word of the object on
2370 * This is the case if we do RCU, have a constructor or
2371 * destructor or are poisoning the objects.
2374 size += sizeof(void *);
2377 #ifdef CONFIG_SLUB_DEBUG
2378 if (flags & SLAB_STORE_USER)
2380 * Need to store information about allocs and frees after
2383 size += 2 * sizeof(struct track);
2385 if (flags & SLAB_RED_ZONE)
2387 * Add some empty padding so that we can catch
2388 * overwrites from earlier objects rather than let
2389 * tracking information or the free pointer be
2390 * corrupted if a user writes before the start
2393 size += sizeof(void *);
2397 * Determine the alignment based on various parameters that the
2398 * user specified and the dynamic determination of cache line size
2401 align = calculate_alignment(flags, align, s->objsize);
2404 * SLUB stores one object immediately after another beginning from
2405 * offset 0. In order to align the objects we have to simply size
2406 * each object to conform to the alignment.
2408 size = ALIGN(size, align);
2410 if (forced_order >= 0)
2411 order = forced_order;
2413 order = calculate_order(size);
2420 s->allocflags |= __GFP_COMP;
2422 if (s->flags & SLAB_CACHE_DMA)
2423 s->allocflags |= SLUB_DMA;
2425 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2426 s->allocflags |= __GFP_RECLAIMABLE;
2429 * Determine the number of objects per slab
2431 s->oo = oo_make(order, size);
2432 s->min = oo_make(get_order(size), size);
2433 if (oo_objects(s->oo) > oo_objects(s->max))
2436 return !!oo_objects(s->oo);
2440 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2441 const char *name, size_t size,
2442 size_t align, unsigned long flags,
2443 void (*ctor)(void *))
2445 memset(s, 0, kmem_size);
2450 s->flags = kmem_cache_flags(size, flags, name, ctor);
2452 if (!calculate_sizes(s, -1))
2456 * The larger the object size is, the more pages we want on the partial
2457 * list to avoid pounding the page allocator excessively.
2459 set_min_partial(s, ilog2(s->size));
2462 s->remote_node_defrag_ratio = 1000;
2464 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2467 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2469 free_kmem_cache_nodes(s);
2471 if (flags & SLAB_PANIC)
2472 panic("Cannot create slab %s size=%lu realsize=%u "
2473 "order=%u offset=%u flags=%lx\n",
2474 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2480 * Check if a given pointer is valid
2482 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2486 page = get_object_page(object);
2488 if (!page || s != page->slab)
2489 /* No slab or wrong slab */
2492 if (!check_valid_pointer(s, page, object))
2496 * We could also check if the object is on the slabs freelist.
2497 * But this would be too expensive and it seems that the main
2498 * purpose of kmem_ptr_valid() is to check if the object belongs
2499 * to a certain slab.
2503 EXPORT_SYMBOL(kmem_ptr_validate);
2506 * Determine the size of a slab object
2508 unsigned int kmem_cache_size(struct kmem_cache *s)
2512 EXPORT_SYMBOL(kmem_cache_size);
2514 const char *kmem_cache_name(struct kmem_cache *s)
2518 EXPORT_SYMBOL(kmem_cache_name);
2520 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2523 #ifdef CONFIG_SLUB_DEBUG
2524 void *addr = page_address(page);
2526 DECLARE_BITMAP(map, page->objects);
2528 bitmap_zero(map, page->objects);
2529 slab_err(s, page, "%s", text);
2531 for_each_free_object(p, s, page->freelist)
2532 set_bit(slab_index(p, s, addr), map);
2534 for_each_object(p, s, addr, page->objects) {
2536 if (!test_bit(slab_index(p, s, addr), map)) {
2537 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2539 print_tracking(s, p);
2547 * Attempt to free all partial slabs on a node.
2549 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2551 unsigned long flags;
2552 struct page *page, *h;
2554 spin_lock_irqsave(&n->list_lock, flags);
2555 list_for_each_entry_safe(page, h, &n->partial, lru) {
2557 list_del(&page->lru);
2558 discard_slab(s, page);
2561 list_slab_objects(s, page,
2562 "Objects remaining on kmem_cache_close()");
2565 spin_unlock_irqrestore(&n->list_lock, flags);
2569 * Release all resources used by a slab cache.
2571 static inline int kmem_cache_close(struct kmem_cache *s)
2577 /* Attempt to free all objects */
2578 free_kmem_cache_cpus(s);
2579 for_each_node_state(node, N_NORMAL_MEMORY) {
2580 struct kmem_cache_node *n = get_node(s, node);
2583 if (n->nr_partial || slabs_node(s, node))
2586 free_kmem_cache_nodes(s);
2591 * Close a cache and release the kmem_cache structure
2592 * (must be used for caches created using kmem_cache_create)
2594 void kmem_cache_destroy(struct kmem_cache *s)
2596 if (s->flags & SLAB_DESTROY_BY_RCU)
2598 down_write(&slub_lock);
2602 up_write(&slub_lock);
2603 if (kmem_cache_close(s)) {
2604 printk(KERN_ERR "SLUB %s: %s called for cache that "
2605 "still has objects.\n", s->name, __func__);
2608 sysfs_slab_remove(s);
2610 up_write(&slub_lock);
2612 EXPORT_SYMBOL(kmem_cache_destroy);
2614 /********************************************************************
2616 *******************************************************************/
2618 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2619 EXPORT_SYMBOL(kmalloc_caches);
2621 static int __init setup_slub_min_order(char *str)
2623 get_option(&str, &slub_min_order);
2628 __setup("slub_min_order=", setup_slub_min_order);
2630 static int __init setup_slub_max_order(char *str)
2632 get_option(&str, &slub_max_order);
2633 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2638 __setup("slub_max_order=", setup_slub_max_order);
2640 static int __init setup_slub_min_objects(char *str)
2642 get_option(&str, &slub_min_objects);
2647 __setup("slub_min_objects=", setup_slub_min_objects);
2649 static int __init setup_slub_nomerge(char *str)
2655 __setup("slub_nomerge", setup_slub_nomerge);
2657 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2658 const char *name, int size, gfp_t gfp_flags)
2660 unsigned int flags = 0;
2662 if (gfp_flags & SLUB_DMA)
2663 flags = SLAB_CACHE_DMA;
2666 * This function is called with IRQs disabled during early-boot on
2667 * single CPU so there's no need to take slub_lock here.
2669 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2673 list_add(&s->list, &slab_caches);
2675 if (sysfs_slab_add(s))
2680 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2683 #ifdef CONFIG_ZONE_DMA
2684 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2686 static void sysfs_add_func(struct work_struct *w)
2688 struct kmem_cache *s;
2690 down_write(&slub_lock);
2691 list_for_each_entry(s, &slab_caches, list) {
2692 if (s->flags & __SYSFS_ADD_DEFERRED) {
2693 s->flags &= ~__SYSFS_ADD_DEFERRED;
2697 up_write(&slub_lock);
2700 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2702 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2704 struct kmem_cache *s;
2707 unsigned long slabflags;
2709 s = kmalloc_caches_dma[index];
2713 /* Dynamically create dma cache */
2714 if (flags & __GFP_WAIT)
2715 down_write(&slub_lock);
2717 if (!down_write_trylock(&slub_lock))
2721 if (kmalloc_caches_dma[index])
2724 realsize = kmalloc_caches[index].objsize;
2725 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2726 (unsigned int)realsize);
2727 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2730 * Must defer sysfs creation to a workqueue because we don't know
2731 * what context we are called from. Before sysfs comes up, we don't
2732 * need to do anything because our sysfs initcall will start by
2733 * adding all existing slabs to sysfs.
2735 slabflags = SLAB_CACHE_DMA|SLAB_NOTRACK;
2736 if (slab_state >= SYSFS)
2737 slabflags |= __SYSFS_ADD_DEFERRED;
2739 if (!s || !text || !kmem_cache_open(s, flags, text,
2740 realsize, ARCH_KMALLOC_MINALIGN, slabflags, NULL)) {
2746 list_add(&s->list, &slab_caches);
2747 kmalloc_caches_dma[index] = s;
2749 if (slab_state >= SYSFS)
2750 schedule_work(&sysfs_add_work);
2753 up_write(&slub_lock);
2755 return kmalloc_caches_dma[index];
2760 * Conversion table for small slabs sizes / 8 to the index in the
2761 * kmalloc array. This is necessary for slabs < 192 since we have non power
2762 * of two cache sizes there. The size of larger slabs can be determined using
2765 static s8 size_index[24] = {
2792 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2798 return ZERO_SIZE_PTR;
2800 index = size_index[(size - 1) / 8];
2802 index = fls(size - 1);
2804 #ifdef CONFIG_ZONE_DMA
2805 if (unlikely((flags & SLUB_DMA)))
2806 return dma_kmalloc_cache(index, flags);
2809 return &kmalloc_caches[index];
2812 void *__kmalloc(size_t size, gfp_t flags)
2814 struct kmem_cache *s;
2817 if (unlikely(size > SLUB_MAX_SIZE))
2818 return kmalloc_large(size, flags);
2820 s = get_slab(size, flags);
2822 if (unlikely(ZERO_OR_NULL_PTR(s)))
2825 ret = slab_alloc(s, flags, -1, _RET_IP_);
2827 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2831 EXPORT_SYMBOL(__kmalloc);
2833 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2838 flags |= __GFP_COMP | __GFP_NOTRACK;
2839 page = alloc_pages_node(node, flags, get_order(size));
2841 ptr = page_address(page);
2843 kmemleak_alloc(ptr, size, 1, flags);
2848 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2850 struct kmem_cache *s;
2853 if (unlikely(size > SLUB_MAX_SIZE)) {
2854 ret = kmalloc_large_node(size, flags, node);
2856 trace_kmalloc_node(_RET_IP_, ret,
2857 size, PAGE_SIZE << get_order(size),
2863 s = get_slab(size, flags);
2865 if (unlikely(ZERO_OR_NULL_PTR(s)))
2868 ret = slab_alloc(s, flags, node, _RET_IP_);
2870 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2874 EXPORT_SYMBOL(__kmalloc_node);
2877 size_t ksize(const void *object)
2880 struct kmem_cache *s;
2882 if (unlikely(object == ZERO_SIZE_PTR))
2885 page = virt_to_head_page(object);
2887 if (unlikely(!PageSlab(page))) {
2888 WARN_ON(!PageCompound(page));
2889 return PAGE_SIZE << compound_order(page);
2893 #ifdef CONFIG_SLUB_DEBUG
2895 * Debugging requires use of the padding between object
2896 * and whatever may come after it.
2898 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2903 * If we have the need to store the freelist pointer
2904 * back there or track user information then we can
2905 * only use the space before that information.
2907 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2910 * Else we can use all the padding etc for the allocation
2914 EXPORT_SYMBOL(ksize);
2916 void kfree(const void *x)
2919 void *object = (void *)x;
2921 trace_kfree(_RET_IP_, x);
2923 if (unlikely(ZERO_OR_NULL_PTR(x)))
2926 page = virt_to_head_page(x);
2927 if (unlikely(!PageSlab(page))) {
2928 BUG_ON(!PageCompound(page));
2933 slab_free(page->slab, page, object, _RET_IP_);
2935 EXPORT_SYMBOL(kfree);
2938 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2939 * the remaining slabs by the number of items in use. The slabs with the
2940 * most items in use come first. New allocations will then fill those up
2941 * and thus they can be removed from the partial lists.
2943 * The slabs with the least items are placed last. This results in them
2944 * being allocated from last increasing the chance that the last objects
2945 * are freed in them.
2947 int kmem_cache_shrink(struct kmem_cache *s)
2951 struct kmem_cache_node *n;
2954 int objects = oo_objects(s->max);
2955 struct list_head *slabs_by_inuse =
2956 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2957 unsigned long flags;
2959 if (!slabs_by_inuse)
2963 for_each_node_state(node, N_NORMAL_MEMORY) {
2964 n = get_node(s, node);
2969 for (i = 0; i < objects; i++)
2970 INIT_LIST_HEAD(slabs_by_inuse + i);
2972 spin_lock_irqsave(&n->list_lock, flags);
2975 * Build lists indexed by the items in use in each slab.
2977 * Note that concurrent frees may occur while we hold the
2978 * list_lock. page->inuse here is the upper limit.
2980 list_for_each_entry_safe(page, t, &n->partial, lru) {
2981 if (!page->inuse && slab_trylock(page)) {
2983 * Must hold slab lock here because slab_free
2984 * may have freed the last object and be
2985 * waiting to release the slab.
2987 list_del(&page->lru);
2990 discard_slab(s, page);
2992 list_move(&page->lru,
2993 slabs_by_inuse + page->inuse);
2998 * Rebuild the partial list with the slabs filled up most
2999 * first and the least used slabs at the end.
3001 for (i = objects - 1; i >= 0; i--)
3002 list_splice(slabs_by_inuse + i, n->partial.prev);
3004 spin_unlock_irqrestore(&n->list_lock, flags);
3007 kfree(slabs_by_inuse);
3010 EXPORT_SYMBOL(kmem_cache_shrink);
3012 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
3013 static int slab_mem_going_offline_callback(void *arg)
3015 struct kmem_cache *s;
3017 down_read(&slub_lock);
3018 list_for_each_entry(s, &slab_caches, list)
3019 kmem_cache_shrink(s);
3020 up_read(&slub_lock);
3025 static void slab_mem_offline_callback(void *arg)
3027 struct kmem_cache_node *n;
3028 struct kmem_cache *s;
3029 struct memory_notify *marg = arg;
3032 offline_node = marg->status_change_nid;
3035 * If the node still has available memory. we need kmem_cache_node
3038 if (offline_node < 0)
3041 down_read(&slub_lock);
3042 list_for_each_entry(s, &slab_caches, list) {
3043 n = get_node(s, offline_node);
3046 * if n->nr_slabs > 0, slabs still exist on the node
3047 * that is going down. We were unable to free them,
3048 * and offline_pages() function shoudn't call this
3049 * callback. So, we must fail.
3051 BUG_ON(slabs_node(s, offline_node));
3053 s->node[offline_node] = NULL;
3054 kmem_cache_free(kmalloc_caches, n);
3057 up_read(&slub_lock);
3060 static int slab_mem_going_online_callback(void *arg)
3062 struct kmem_cache_node *n;
3063 struct kmem_cache *s;
3064 struct memory_notify *marg = arg;
3065 int nid = marg->status_change_nid;
3069 * If the node's memory is already available, then kmem_cache_node is
3070 * already created. Nothing to do.
3076 * We are bringing a node online. No memory is available yet. We must
3077 * allocate a kmem_cache_node structure in order to bring the node
3080 down_read(&slub_lock);
3081 list_for_each_entry(s, &slab_caches, list) {
3083 * XXX: kmem_cache_alloc_node will fallback to other nodes
3084 * since memory is not yet available from the node that
3087 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3092 init_kmem_cache_node(n, s);
3096 up_read(&slub_lock);
3100 static int slab_memory_callback(struct notifier_block *self,
3101 unsigned long action, void *arg)
3106 case MEM_GOING_ONLINE:
3107 ret = slab_mem_going_online_callback(arg);
3109 case MEM_GOING_OFFLINE:
3110 ret = slab_mem_going_offline_callback(arg);
3113 case MEM_CANCEL_ONLINE:
3114 slab_mem_offline_callback(arg);
3117 case MEM_CANCEL_OFFLINE:
3121 ret = notifier_from_errno(ret);
3127 #endif /* CONFIG_MEMORY_HOTPLUG */
3129 /********************************************************************
3130 * Basic setup of slabs
3131 *******************************************************************/
3133 void __init kmem_cache_init(void)
3142 * Must first have the slab cache available for the allocations of the
3143 * struct kmem_cache_node's. There is special bootstrap code in
3144 * kmem_cache_open for slab_state == DOWN.
3146 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3147 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3148 kmalloc_caches[0].refcount = -1;
3151 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3154 /* Able to allocate the per node structures */
3155 slab_state = PARTIAL;
3157 /* Caches that are not of the two-to-the-power-of size */
3158 if (KMALLOC_MIN_SIZE <= 64) {
3159 create_kmalloc_cache(&kmalloc_caches[1],
3160 "kmalloc-96", 96, GFP_NOWAIT);
3162 create_kmalloc_cache(&kmalloc_caches[2],
3163 "kmalloc-192", 192, GFP_NOWAIT);
3167 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3168 create_kmalloc_cache(&kmalloc_caches[i],
3169 "kmalloc", 1 << i, GFP_NOWAIT);
3175 * Patch up the size_index table if we have strange large alignment
3176 * requirements for the kmalloc array. This is only the case for
3177 * MIPS it seems. The standard arches will not generate any code here.
3179 * Largest permitted alignment is 256 bytes due to the way we
3180 * handle the index determination for the smaller caches.
3182 * Make sure that nothing crazy happens if someone starts tinkering
3183 * around with ARCH_KMALLOC_MINALIGN
3185 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3186 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3188 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3189 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3191 if (KMALLOC_MIN_SIZE == 128) {
3193 * The 192 byte sized cache is not used if the alignment
3194 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3197 for (i = 128 + 8; i <= 192; i += 8)
3198 size_index[(i - 1) / 8] = 8;
3203 /* Provide the correct kmalloc names now that the caches are up */
3204 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3205 kmalloc_caches[i]. name =
3206 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3209 register_cpu_notifier(&slab_notifier);
3210 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3211 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3213 kmem_size = sizeof(struct kmem_cache);
3217 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3218 " CPUs=%d, Nodes=%d\n",
3219 caches, cache_line_size(),
3220 slub_min_order, slub_max_order, slub_min_objects,
3221 nr_cpu_ids, nr_node_ids);
3224 void __init kmem_cache_init_late(void)
3229 * Find a mergeable slab cache
3231 static int slab_unmergeable(struct kmem_cache *s)
3233 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3240 * We may have set a slab to be unmergeable during bootstrap.
3242 if (s->refcount < 0)
3248 static struct kmem_cache *find_mergeable(size_t size,
3249 size_t align, unsigned long flags, const char *name,
3250 void (*ctor)(void *))
3252 struct kmem_cache *s;
3254 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3260 size = ALIGN(size, sizeof(void *));
3261 align = calculate_alignment(flags, align, size);
3262 size = ALIGN(size, align);
3263 flags = kmem_cache_flags(size, flags, name, NULL);
3265 list_for_each_entry(s, &slab_caches, list) {
3266 if (slab_unmergeable(s))
3272 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3275 * Check if alignment is compatible.
3276 * Courtesy of Adrian Drzewiecki
3278 if ((s->size & ~(align - 1)) != s->size)
3281 if (s->size - size >= sizeof(void *))
3289 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3290 size_t align, unsigned long flags, void (*ctor)(void *))
3292 struct kmem_cache *s;
3294 down_write(&slub_lock);
3295 s = find_mergeable(size, align, flags, name, ctor);
3301 * Adjust the object sizes so that we clear
3302 * the complete object on kzalloc.
3304 s->objsize = max(s->objsize, (int)size);
3307 * And then we need to update the object size in the
3308 * per cpu structures
3310 for_each_online_cpu(cpu)
3311 get_cpu_slab(s, cpu)->objsize = s->objsize;
3313 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3314 up_write(&slub_lock);
3316 if (sysfs_slab_alias(s, name)) {
3317 down_write(&slub_lock);
3319 up_write(&slub_lock);
3325 s = kmalloc(kmem_size, GFP_KERNEL);
3327 if (kmem_cache_open(s, GFP_KERNEL, name,
3328 size, align, flags, ctor)) {
3329 list_add(&s->list, &slab_caches);
3330 up_write(&slub_lock);
3331 if (sysfs_slab_add(s)) {
3332 down_write(&slub_lock);
3334 up_write(&slub_lock);
3342 up_write(&slub_lock);
3345 if (flags & SLAB_PANIC)
3346 panic("Cannot create slabcache %s\n", name);
3351 EXPORT_SYMBOL(kmem_cache_create);
3355 * Use the cpu notifier to insure that the cpu slabs are flushed when
3358 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3359 unsigned long action, void *hcpu)
3361 long cpu = (long)hcpu;
3362 struct kmem_cache *s;
3363 unsigned long flags;
3366 case CPU_UP_PREPARE:
3367 case CPU_UP_PREPARE_FROZEN:
3368 init_alloc_cpu_cpu(cpu);
3369 down_read(&slub_lock);
3370 list_for_each_entry(s, &slab_caches, list)
3371 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3373 up_read(&slub_lock);
3376 case CPU_UP_CANCELED:
3377 case CPU_UP_CANCELED_FROZEN:
3379 case CPU_DEAD_FROZEN:
3380 down_read(&slub_lock);
3381 list_for_each_entry(s, &slab_caches, list) {
3382 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3384 local_irq_save(flags);
3385 __flush_cpu_slab(s, cpu);
3386 local_irq_restore(flags);
3387 free_kmem_cache_cpu(c, cpu);
3388 s->cpu_slab[cpu] = NULL;
3390 up_read(&slub_lock);
3398 static struct notifier_block __cpuinitdata slab_notifier = {
3399 .notifier_call = slab_cpuup_callback
3404 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3406 struct kmem_cache *s;
3409 if (unlikely(size > SLUB_MAX_SIZE))
3410 return kmalloc_large(size, gfpflags);
3412 s = get_slab(size, gfpflags);
3414 if (unlikely(ZERO_OR_NULL_PTR(s)))
3417 ret = slab_alloc(s, gfpflags, -1, caller);
3419 /* Honor the call site pointer we recieved. */
3420 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3425 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3426 int node, unsigned long caller)
3428 struct kmem_cache *s;
3431 if (unlikely(size > SLUB_MAX_SIZE))
3432 return kmalloc_large_node(size, gfpflags, node);
3434 s = get_slab(size, gfpflags);
3436 if (unlikely(ZERO_OR_NULL_PTR(s)))
3439 ret = slab_alloc(s, gfpflags, node, caller);
3441 /* Honor the call site pointer we recieved. */
3442 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3447 #ifdef CONFIG_SLUB_DEBUG
3448 static int count_inuse(struct page *page)
3453 static int count_total(struct page *page)
3455 return page->objects;
3458 static int validate_slab(struct kmem_cache *s, struct page *page,
3462 void *addr = page_address(page);
3464 if (!check_slab(s, page) ||
3465 !on_freelist(s, page, NULL))
3468 /* Now we know that a valid freelist exists */
3469 bitmap_zero(map, page->objects);
3471 for_each_free_object(p, s, page->freelist) {
3472 set_bit(slab_index(p, s, addr), map);
3473 if (!check_object(s, page, p, 0))
3477 for_each_object(p, s, addr, page->objects)
3478 if (!test_bit(slab_index(p, s, addr), map))
3479 if (!check_object(s, page, p, 1))
3484 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3487 if (slab_trylock(page)) {
3488 validate_slab(s, page, map);
3491 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3494 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3495 if (!PageSlubDebug(page))
3496 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3497 "on slab 0x%p\n", s->name, page);
3499 if (PageSlubDebug(page))
3500 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3501 "slab 0x%p\n", s->name, page);
3505 static int validate_slab_node(struct kmem_cache *s,
3506 struct kmem_cache_node *n, unsigned long *map)
3508 unsigned long count = 0;
3510 unsigned long flags;
3512 spin_lock_irqsave(&n->list_lock, flags);
3514 list_for_each_entry(page, &n->partial, lru) {
3515 validate_slab_slab(s, page, map);
3518 if (count != n->nr_partial)
3519 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3520 "counter=%ld\n", s->name, count, n->nr_partial);
3522 if (!(s->flags & SLAB_STORE_USER))
3525 list_for_each_entry(page, &n->full, lru) {
3526 validate_slab_slab(s, page, map);
3529 if (count != atomic_long_read(&n->nr_slabs))
3530 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3531 "counter=%ld\n", s->name, count,
3532 atomic_long_read(&n->nr_slabs));
3535 spin_unlock_irqrestore(&n->list_lock, flags);
3539 static long validate_slab_cache(struct kmem_cache *s)
3542 unsigned long count = 0;
3543 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3544 sizeof(unsigned long), GFP_KERNEL);
3550 for_each_node_state(node, N_NORMAL_MEMORY) {
3551 struct kmem_cache_node *n = get_node(s, node);
3553 count += validate_slab_node(s, n, map);
3559 #ifdef SLUB_RESILIENCY_TEST
3560 static void resiliency_test(void)
3564 printk(KERN_ERR "SLUB resiliency testing\n");
3565 printk(KERN_ERR "-----------------------\n");
3566 printk(KERN_ERR "A. Corruption after allocation\n");
3568 p = kzalloc(16, GFP_KERNEL);
3570 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3571 " 0x12->0x%p\n\n", p + 16);
3573 validate_slab_cache(kmalloc_caches + 4);
3575 /* Hmmm... The next two are dangerous */
3576 p = kzalloc(32, GFP_KERNEL);
3577 p[32 + sizeof(void *)] = 0x34;
3578 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3579 " 0x34 -> -0x%p\n", p);
3581 "If allocated object is overwritten then not detectable\n\n");
3583 validate_slab_cache(kmalloc_caches + 5);
3584 p = kzalloc(64, GFP_KERNEL);
3585 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3587 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3590 "If allocated object is overwritten then not detectable\n\n");
3591 validate_slab_cache(kmalloc_caches + 6);
3593 printk(KERN_ERR "\nB. Corruption after free\n");
3594 p = kzalloc(128, GFP_KERNEL);
3597 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3598 validate_slab_cache(kmalloc_caches + 7);
3600 p = kzalloc(256, GFP_KERNEL);
3603 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3605 validate_slab_cache(kmalloc_caches + 8);
3607 p = kzalloc(512, GFP_KERNEL);
3610 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3611 validate_slab_cache(kmalloc_caches + 9);
3614 static void resiliency_test(void) {};
3618 * Generate lists of code addresses where slabcache objects are allocated
3623 unsigned long count;
3630 DECLARE_BITMAP(cpus, NR_CPUS);
3636 unsigned long count;
3637 struct location *loc;
3640 static void free_loc_track(struct loc_track *t)
3643 free_pages((unsigned long)t->loc,
3644 get_order(sizeof(struct location) * t->max));
3647 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3652 order = get_order(sizeof(struct location) * max);
3654 l = (void *)__get_free_pages(flags, order);
3659 memcpy(l, t->loc, sizeof(struct location) * t->count);
3667 static int add_location(struct loc_track *t, struct kmem_cache *s,
3668 const struct track *track)
3670 long start, end, pos;
3672 unsigned long caddr;
3673 unsigned long age = jiffies - track->when;
3679 pos = start + (end - start + 1) / 2;
3682 * There is nothing at "end". If we end up there
3683 * we need to add something to before end.
3688 caddr = t->loc[pos].addr;
3689 if (track->addr == caddr) {
3695 if (age < l->min_time)
3697 if (age > l->max_time)
3700 if (track->pid < l->min_pid)
3701 l->min_pid = track->pid;
3702 if (track->pid > l->max_pid)
3703 l->max_pid = track->pid;
3705 cpumask_set_cpu(track->cpu,
3706 to_cpumask(l->cpus));
3708 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3712 if (track->addr < caddr)
3719 * Not found. Insert new tracking element.
3721 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3727 (t->count - pos) * sizeof(struct location));
3730 l->addr = track->addr;
3734 l->min_pid = track->pid;
3735 l->max_pid = track->pid;
3736 cpumask_clear(to_cpumask(l->cpus));
3737 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3738 nodes_clear(l->nodes);
3739 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3743 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3744 struct page *page, enum track_item alloc)
3746 void *addr = page_address(page);
3747 DECLARE_BITMAP(map, page->objects);
3750 bitmap_zero(map, page->objects);
3751 for_each_free_object(p, s, page->freelist)
3752 set_bit(slab_index(p, s, addr), map);
3754 for_each_object(p, s, addr, page->objects)
3755 if (!test_bit(slab_index(p, s, addr), map))
3756 add_location(t, s, get_track(s, p, alloc));
3759 static int list_locations(struct kmem_cache *s, char *buf,
3760 enum track_item alloc)
3764 struct loc_track t = { 0, 0, NULL };
3767 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3769 return sprintf(buf, "Out of memory\n");
3771 /* Push back cpu slabs */
3774 for_each_node_state(node, N_NORMAL_MEMORY) {
3775 struct kmem_cache_node *n = get_node(s, node);
3776 unsigned long flags;
3779 if (!atomic_long_read(&n->nr_slabs))
3782 spin_lock_irqsave(&n->list_lock, flags);
3783 list_for_each_entry(page, &n->partial, lru)
3784 process_slab(&t, s, page, alloc);
3785 list_for_each_entry(page, &n->full, lru)
3786 process_slab(&t, s, page, alloc);
3787 spin_unlock_irqrestore(&n->list_lock, flags);
3790 for (i = 0; i < t.count; i++) {
3791 struct location *l = &t.loc[i];
3793 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3795 len += sprintf(buf + len, "%7ld ", l->count);
3798 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3800 len += sprintf(buf + len, "<not-available>");
3802 if (l->sum_time != l->min_time) {
3803 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3805 (long)div_u64(l->sum_time, l->count),
3808 len += sprintf(buf + len, " age=%ld",
3811 if (l->min_pid != l->max_pid)
3812 len += sprintf(buf + len, " pid=%ld-%ld",
3813 l->min_pid, l->max_pid);
3815 len += sprintf(buf + len, " pid=%ld",
3818 if (num_online_cpus() > 1 &&
3819 !cpumask_empty(to_cpumask(l->cpus)) &&
3820 len < PAGE_SIZE - 60) {
3821 len += sprintf(buf + len, " cpus=");
3822 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3823 to_cpumask(l->cpus));
3826 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3827 len < PAGE_SIZE - 60) {
3828 len += sprintf(buf + len, " nodes=");
3829 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3833 len += sprintf(buf + len, "\n");
3838 len += sprintf(buf, "No data\n");
3842 enum slab_stat_type {
3843 SL_ALL, /* All slabs */
3844 SL_PARTIAL, /* Only partially allocated slabs */
3845 SL_CPU, /* Only slabs used for cpu caches */
3846 SL_OBJECTS, /* Determine allocated objects not slabs */
3847 SL_TOTAL /* Determine object capacity not slabs */
3850 #define SO_ALL (1 << SL_ALL)
3851 #define SO_PARTIAL (1 << SL_PARTIAL)
3852 #define SO_CPU (1 << SL_CPU)
3853 #define SO_OBJECTS (1 << SL_OBJECTS)
3854 #define SO_TOTAL (1 << SL_TOTAL)
3856 static ssize_t show_slab_objects(struct kmem_cache *s,
3857 char *buf, unsigned long flags)
3859 unsigned long total = 0;
3862 unsigned long *nodes;
3863 unsigned long *per_cpu;
3865 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3868 per_cpu = nodes + nr_node_ids;
3870 if (flags & SO_CPU) {
3873 for_each_possible_cpu(cpu) {
3874 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3876 if (!c || c->node < 0)
3880 if (flags & SO_TOTAL)
3881 x = c->page->objects;
3882 else if (flags & SO_OBJECTS)
3888 nodes[c->node] += x;
3894 if (flags & SO_ALL) {
3895 for_each_node_state(node, N_NORMAL_MEMORY) {
3896 struct kmem_cache_node *n = get_node(s, node);
3898 if (flags & SO_TOTAL)
3899 x = atomic_long_read(&n->total_objects);
3900 else if (flags & SO_OBJECTS)
3901 x = atomic_long_read(&n->total_objects) -
3902 count_partial(n, count_free);
3905 x = atomic_long_read(&n->nr_slabs);
3910 } else if (flags & SO_PARTIAL) {
3911 for_each_node_state(node, N_NORMAL_MEMORY) {
3912 struct kmem_cache_node *n = get_node(s, node);
3914 if (flags & SO_TOTAL)
3915 x = count_partial(n, count_total);
3916 else if (flags & SO_OBJECTS)
3917 x = count_partial(n, count_inuse);
3924 x = sprintf(buf, "%lu", total);
3926 for_each_node_state(node, N_NORMAL_MEMORY)
3928 x += sprintf(buf + x, " N%d=%lu",
3932 return x + sprintf(buf + x, "\n");
3935 static int any_slab_objects(struct kmem_cache *s)
3939 for_each_online_node(node) {
3940 struct kmem_cache_node *n = get_node(s, node);
3945 if (atomic_long_read(&n->total_objects))
3951 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3952 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3954 struct slab_attribute {
3955 struct attribute attr;
3956 ssize_t (*show)(struct kmem_cache *s, char *buf);
3957 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3960 #define SLAB_ATTR_RO(_name) \
3961 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3963 #define SLAB_ATTR(_name) \
3964 static struct slab_attribute _name##_attr = \
3965 __ATTR(_name, 0644, _name##_show, _name##_store)
3967 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3969 return sprintf(buf, "%d\n", s->size);
3971 SLAB_ATTR_RO(slab_size);
3973 static ssize_t align_show(struct kmem_cache *s, char *buf)
3975 return sprintf(buf, "%d\n", s->align);
3977 SLAB_ATTR_RO(align);
3979 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3981 return sprintf(buf, "%d\n", s->objsize);
3983 SLAB_ATTR_RO(object_size);
3985 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3987 return sprintf(buf, "%d\n", oo_objects(s->oo));
3989 SLAB_ATTR_RO(objs_per_slab);
3991 static ssize_t order_store(struct kmem_cache *s,
3992 const char *buf, size_t length)
3994 unsigned long order;
3997 err = strict_strtoul(buf, 10, &order);
4001 if (order > slub_max_order || order < slub_min_order)
4004 calculate_sizes(s, order);
4008 static ssize_t order_show(struct kmem_cache *s, char *buf)
4010 return sprintf(buf, "%d\n", oo_order(s->oo));
4014 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4016 return sprintf(buf, "%lu\n", s->min_partial);
4019 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4025 err = strict_strtoul(buf, 10, &min);
4029 set_min_partial(s, min);
4032 SLAB_ATTR(min_partial);
4034 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4037 int n = sprint_symbol(buf, (unsigned long)s->ctor);
4039 return n + sprintf(buf + n, "\n");
4045 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4047 return sprintf(buf, "%d\n", s->refcount - 1);
4049 SLAB_ATTR_RO(aliases);
4051 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4053 return show_slab_objects(s, buf, SO_ALL);
4055 SLAB_ATTR_RO(slabs);
4057 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4059 return show_slab_objects(s, buf, SO_PARTIAL);
4061 SLAB_ATTR_RO(partial);
4063 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4065 return show_slab_objects(s, buf, SO_CPU);
4067 SLAB_ATTR_RO(cpu_slabs);
4069 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4071 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4073 SLAB_ATTR_RO(objects);
4075 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4077 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4079 SLAB_ATTR_RO(objects_partial);
4081 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4083 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4085 SLAB_ATTR_RO(total_objects);
4087 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4089 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4092 static ssize_t sanity_checks_store(struct kmem_cache *s,
4093 const char *buf, size_t length)
4095 s->flags &= ~SLAB_DEBUG_FREE;
4097 s->flags |= SLAB_DEBUG_FREE;
4100 SLAB_ATTR(sanity_checks);
4102 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4104 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4107 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4110 s->flags &= ~SLAB_TRACE;
4112 s->flags |= SLAB_TRACE;
4117 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4119 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4122 static ssize_t reclaim_account_store(struct kmem_cache *s,
4123 const char *buf, size_t length)
4125 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4127 s->flags |= SLAB_RECLAIM_ACCOUNT;
4130 SLAB_ATTR(reclaim_account);
4132 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4134 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4136 SLAB_ATTR_RO(hwcache_align);
4138 #ifdef CONFIG_ZONE_DMA
4139 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4141 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4143 SLAB_ATTR_RO(cache_dma);
4146 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4148 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4150 SLAB_ATTR_RO(destroy_by_rcu);
4152 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4154 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4157 static ssize_t red_zone_store(struct kmem_cache *s,
4158 const char *buf, size_t length)
4160 if (any_slab_objects(s))
4163 s->flags &= ~SLAB_RED_ZONE;
4165 s->flags |= SLAB_RED_ZONE;
4166 calculate_sizes(s, -1);
4169 SLAB_ATTR(red_zone);
4171 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4173 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4176 static ssize_t poison_store(struct kmem_cache *s,
4177 const char *buf, size_t length)
4179 if (any_slab_objects(s))
4182 s->flags &= ~SLAB_POISON;
4184 s->flags |= SLAB_POISON;
4185 calculate_sizes(s, -1);
4190 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4192 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4195 static ssize_t store_user_store(struct kmem_cache *s,
4196 const char *buf, size_t length)
4198 if (any_slab_objects(s))
4201 s->flags &= ~SLAB_STORE_USER;
4203 s->flags |= SLAB_STORE_USER;
4204 calculate_sizes(s, -1);
4207 SLAB_ATTR(store_user);
4209 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4214 static ssize_t validate_store(struct kmem_cache *s,
4215 const char *buf, size_t length)
4219 if (buf[0] == '1') {
4220 ret = validate_slab_cache(s);
4226 SLAB_ATTR(validate);
4228 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4233 static ssize_t shrink_store(struct kmem_cache *s,
4234 const char *buf, size_t length)
4236 if (buf[0] == '1') {
4237 int rc = kmem_cache_shrink(s);
4247 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4249 if (!(s->flags & SLAB_STORE_USER))
4251 return list_locations(s, buf, TRACK_ALLOC);
4253 SLAB_ATTR_RO(alloc_calls);
4255 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4257 if (!(s->flags & SLAB_STORE_USER))
4259 return list_locations(s, buf, TRACK_FREE);
4261 SLAB_ATTR_RO(free_calls);
4264 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4266 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4269 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4270 const char *buf, size_t length)
4272 unsigned long ratio;
4275 err = strict_strtoul(buf, 10, &ratio);
4280 s->remote_node_defrag_ratio = ratio * 10;
4284 SLAB_ATTR(remote_node_defrag_ratio);
4287 #ifdef CONFIG_SLUB_STATS
4288 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4290 unsigned long sum = 0;
4293 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4298 for_each_online_cpu(cpu) {
4299 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4305 len = sprintf(buf, "%lu", sum);
4308 for_each_online_cpu(cpu) {
4309 if (data[cpu] && len < PAGE_SIZE - 20)
4310 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4314 return len + sprintf(buf + len, "\n");
4317 #define STAT_ATTR(si, text) \
4318 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4320 return show_stat(s, buf, si); \
4322 SLAB_ATTR_RO(text); \
4324 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4325 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4326 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4327 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4328 STAT_ATTR(FREE_FROZEN, free_frozen);
4329 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4330 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4331 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4332 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4333 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4334 STAT_ATTR(FREE_SLAB, free_slab);
4335 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4336 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4337 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4338 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4339 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4340 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4341 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4344 static struct attribute *slab_attrs[] = {
4345 &slab_size_attr.attr,
4346 &object_size_attr.attr,
4347 &objs_per_slab_attr.attr,
4349 &min_partial_attr.attr,
4351 &objects_partial_attr.attr,
4352 &total_objects_attr.attr,
4355 &cpu_slabs_attr.attr,
4359 &sanity_checks_attr.attr,
4361 &hwcache_align_attr.attr,
4362 &reclaim_account_attr.attr,
4363 &destroy_by_rcu_attr.attr,
4364 &red_zone_attr.attr,
4366 &store_user_attr.attr,
4367 &validate_attr.attr,
4369 &alloc_calls_attr.attr,
4370 &free_calls_attr.attr,
4371 #ifdef CONFIG_ZONE_DMA
4372 &cache_dma_attr.attr,
4375 &remote_node_defrag_ratio_attr.attr,
4377 #ifdef CONFIG_SLUB_STATS
4378 &alloc_fastpath_attr.attr,
4379 &alloc_slowpath_attr.attr,
4380 &free_fastpath_attr.attr,
4381 &free_slowpath_attr.attr,
4382 &free_frozen_attr.attr,
4383 &free_add_partial_attr.attr,
4384 &free_remove_partial_attr.attr,
4385 &alloc_from_partial_attr.attr,
4386 &alloc_slab_attr.attr,
4387 &alloc_refill_attr.attr,
4388 &free_slab_attr.attr,
4389 &cpuslab_flush_attr.attr,
4390 &deactivate_full_attr.attr,
4391 &deactivate_empty_attr.attr,
4392 &deactivate_to_head_attr.attr,
4393 &deactivate_to_tail_attr.attr,
4394 &deactivate_remote_frees_attr.attr,
4395 &order_fallback_attr.attr,
4400 static struct attribute_group slab_attr_group = {
4401 .attrs = slab_attrs,
4404 static ssize_t slab_attr_show(struct kobject *kobj,
4405 struct attribute *attr,
4408 struct slab_attribute *attribute;
4409 struct kmem_cache *s;
4412 attribute = to_slab_attr(attr);
4415 if (!attribute->show)
4418 err = attribute->show(s, buf);
4423 static ssize_t slab_attr_store(struct kobject *kobj,
4424 struct attribute *attr,
4425 const char *buf, size_t len)
4427 struct slab_attribute *attribute;
4428 struct kmem_cache *s;
4431 attribute = to_slab_attr(attr);
4434 if (!attribute->store)
4437 err = attribute->store(s, buf, len);
4442 static void kmem_cache_release(struct kobject *kobj)
4444 struct kmem_cache *s = to_slab(kobj);
4449 static struct sysfs_ops slab_sysfs_ops = {
4450 .show = slab_attr_show,
4451 .store = slab_attr_store,
4454 static struct kobj_type slab_ktype = {
4455 .sysfs_ops = &slab_sysfs_ops,
4456 .release = kmem_cache_release
4459 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4461 struct kobj_type *ktype = get_ktype(kobj);
4463 if (ktype == &slab_ktype)
4468 static struct kset_uevent_ops slab_uevent_ops = {
4469 .filter = uevent_filter,
4472 static struct kset *slab_kset;
4474 #define ID_STR_LENGTH 64
4476 /* Create a unique string id for a slab cache:
4478 * Format :[flags-]size
4480 static char *create_unique_id(struct kmem_cache *s)
4482 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4489 * First flags affecting slabcache operations. We will only
4490 * get here for aliasable slabs so we do not need to support
4491 * too many flags. The flags here must cover all flags that
4492 * are matched during merging to guarantee that the id is
4495 if (s->flags & SLAB_CACHE_DMA)
4497 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4499 if (s->flags & SLAB_DEBUG_FREE)
4501 if (!(s->flags & SLAB_NOTRACK))
4505 p += sprintf(p, "%07d", s->size);
4506 BUG_ON(p > name + ID_STR_LENGTH - 1);
4510 static int sysfs_slab_add(struct kmem_cache *s)
4516 if (slab_state < SYSFS)
4517 /* Defer until later */
4520 unmergeable = slab_unmergeable(s);
4523 * Slabcache can never be merged so we can use the name proper.
4524 * This is typically the case for debug situations. In that
4525 * case we can catch duplicate names easily.
4527 sysfs_remove_link(&slab_kset->kobj, s->name);
4531 * Create a unique name for the slab as a target
4534 name = create_unique_id(s);
4537 s->kobj.kset = slab_kset;
4538 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4540 kobject_put(&s->kobj);
4544 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4547 kobject_uevent(&s->kobj, KOBJ_ADD);
4549 /* Setup first alias */
4550 sysfs_slab_alias(s, s->name);
4556 static void sysfs_slab_remove(struct kmem_cache *s)
4558 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4559 kobject_del(&s->kobj);
4560 kobject_put(&s->kobj);
4564 * Need to buffer aliases during bootup until sysfs becomes
4565 * available lest we lose that information.
4567 struct saved_alias {
4568 struct kmem_cache *s;
4570 struct saved_alias *next;
4573 static struct saved_alias *alias_list;
4575 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4577 struct saved_alias *al;
4579 if (slab_state == SYSFS) {
4581 * If we have a leftover link then remove it.
4583 sysfs_remove_link(&slab_kset->kobj, name);
4584 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4587 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4593 al->next = alias_list;
4598 static int __init slab_sysfs_init(void)
4600 struct kmem_cache *s;
4603 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4605 printk(KERN_ERR "Cannot register slab subsystem.\n");
4611 list_for_each_entry(s, &slab_caches, list) {
4612 err = sysfs_slab_add(s);
4614 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4615 " to sysfs\n", s->name);
4618 while (alias_list) {
4619 struct saved_alias *al = alias_list;
4621 alias_list = alias_list->next;
4622 err = sysfs_slab_alias(al->s, al->name);
4624 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4625 " %s to sysfs\n", s->name);
4633 __initcall(slab_sysfs_init);
4637 * The /proc/slabinfo ABI
4639 #ifdef CONFIG_SLABINFO
4640 static void print_slabinfo_header(struct seq_file *m)
4642 seq_puts(m, "slabinfo - version: 2.1\n");
4643 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4644 "<objperslab> <pagesperslab>");
4645 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4646 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4650 static void *s_start(struct seq_file *m, loff_t *pos)
4654 down_read(&slub_lock);
4656 print_slabinfo_header(m);
4658 return seq_list_start(&slab_caches, *pos);
4661 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4663 return seq_list_next(p, &slab_caches, pos);
4666 static void s_stop(struct seq_file *m, void *p)
4668 up_read(&slub_lock);
4671 static int s_show(struct seq_file *m, void *p)
4673 unsigned long nr_partials = 0;
4674 unsigned long nr_slabs = 0;
4675 unsigned long nr_inuse = 0;
4676 unsigned long nr_objs = 0;
4677 unsigned long nr_free = 0;
4678 struct kmem_cache *s;
4681 s = list_entry(p, struct kmem_cache, list);
4683 for_each_online_node(node) {
4684 struct kmem_cache_node *n = get_node(s, node);
4689 nr_partials += n->nr_partial;
4690 nr_slabs += atomic_long_read(&n->nr_slabs);
4691 nr_objs += atomic_long_read(&n->total_objects);
4692 nr_free += count_partial(n, count_free);
4695 nr_inuse = nr_objs - nr_free;
4697 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4698 nr_objs, s->size, oo_objects(s->oo),
4699 (1 << oo_order(s->oo)));
4700 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4701 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4707 static const struct seq_operations slabinfo_op = {
4714 static int slabinfo_open(struct inode *inode, struct file *file)
4716 return seq_open(file, &slabinfo_op);
4719 static const struct file_operations proc_slabinfo_operations = {
4720 .open = slabinfo_open,
4722 .llseek = seq_lseek,
4723 .release = seq_release,
4726 static int __init slab_proc_init(void)
4728 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4731 module_init(slab_proc_init);
4732 #endif /* CONFIG_SLABINFO */