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/kmemleak.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
38 * The slab_lock protects operations on the object of a particular
39 * slab and its metadata in the page struct. If the slab lock
40 * has been taken then no allocations nor frees can be performed
41 * on the objects in the slab nor can the slab be added or removed
42 * from the partial or full lists since this would mean modifying
43 * the page_struct of the slab.
45 * The list_lock protects the partial and full list on each node and
46 * the partial slab counter. If taken then no new slabs may be added or
47 * removed from the lists nor make the number of partial slabs be modified.
48 * (Note that the total number of slabs is an atomic value that may be
49 * modified without taking the list lock).
51 * The list_lock is a centralized lock and thus we avoid taking it as
52 * much as possible. As long as SLUB does not have to handle partial
53 * slabs, operations can continue without any centralized lock. F.e.
54 * allocating a long series of objects that fill up slabs does not require
57 * The lock order is sometimes inverted when we are trying to get a slab
58 * off a list. We take the list_lock and then look for a page on the list
59 * to use. While we do that objects in the slabs may be freed. We can
60 * only operate on the slab if we have also taken the slab_lock. So we use
61 * a slab_trylock() on the slab. If trylock was successful then no frees
62 * can occur anymore and we can use the slab for allocations etc. If the
63 * slab_trylock() does not succeed then frees are in progress in the slab and
64 * we must stay away from it for a while since we may cause a bouncing
65 * cacheline if we try to acquire the lock. So go onto the next slab.
66 * If all pages are busy then we may allocate a new slab instead of reusing
67 * a partial slab. A new slab has noone operating on it and thus there is
68 * no danger of cacheline contention.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #ifdef CONFIG_SLUB_DEBUG
118 * Issues still to be resolved:
120 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
122 * - Variable sizing of the per node arrays
125 /* Enable to test recovery from slab corruption on boot */
126 #undef SLUB_RESILIENCY_TEST
129 * Mininum number of partial slabs. These will be left on the partial
130 * lists even if they are empty. kmem_cache_shrink may reclaim them.
132 #define MIN_PARTIAL 5
135 * Maximum number of desirable partial slabs.
136 * The existence of more partial slabs makes kmem_cache_shrink
137 * sort the partial list by the number of objects in the.
139 #define MAX_PARTIAL 10
141 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
142 SLAB_POISON | SLAB_STORE_USER)
145 * Set of flags that will prevent slab merging
147 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
148 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE)
150 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
151 SLAB_CACHE_DMA | SLAB_NOTRACK)
153 #ifndef ARCH_KMALLOC_MINALIGN
154 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
157 #ifndef ARCH_SLAB_MINALIGN
158 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
162 #define OO_MASK ((1 << OO_SHIFT) - 1)
163 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
165 /* Internal SLUB flags */
166 #define __OBJECT_POISON 0x80000000 /* Poison object */
167 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
169 static int kmem_size = sizeof(struct kmem_cache);
172 static struct notifier_block slab_notifier;
176 DOWN, /* No slab functionality available */
177 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
178 UP, /* Everything works but does not show up in sysfs */
182 /* A list of all slab caches on the system */
183 static DECLARE_RWSEM(slub_lock);
184 static LIST_HEAD(slab_caches);
187 * Tracking user of a slab.
190 unsigned long addr; /* Called from address */
191 int cpu; /* Was running on cpu */
192 int pid; /* Pid context */
193 unsigned long when; /* When did the operation occur */
196 enum track_item { TRACK_ALLOC, TRACK_FREE };
198 #ifdef CONFIG_SLUB_DEBUG
199 static int sysfs_slab_add(struct kmem_cache *);
200 static int sysfs_slab_alias(struct kmem_cache *, const char *);
201 static void sysfs_slab_remove(struct kmem_cache *);
204 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
205 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
207 static inline void sysfs_slab_remove(struct kmem_cache *s)
214 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
216 #ifdef CONFIG_SLUB_STATS
221 /********************************************************************
222 * Core slab cache functions
223 *******************************************************************/
225 int slab_is_available(void)
227 return slab_state >= UP;
230 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
233 return s->node[node];
235 return &s->local_node;
239 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
242 return s->cpu_slab[cpu];
248 /* Verify that a pointer has an address that is valid within a slab page */
249 static inline int check_valid_pointer(struct kmem_cache *s,
250 struct page *page, const void *object)
257 base = page_address(page);
258 if (object < base || object >= base + page->objects * s->size ||
259 (object - base) % s->size) {
267 * Slow version of get and set free pointer.
269 * This version requires touching the cache lines of kmem_cache which
270 * we avoid to do in the fast alloc free paths. There we obtain the offset
271 * from the page struct.
273 static inline void *get_freepointer(struct kmem_cache *s, void *object)
275 return *(void **)(object + s->offset);
278 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
280 *(void **)(object + s->offset) = fp;
283 /* Loop over all objects in a slab */
284 #define for_each_object(__p, __s, __addr, __objects) \
285 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
289 #define for_each_free_object(__p, __s, __free) \
290 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
292 /* Determine object index from a given position */
293 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
295 return (p - addr) / s->size;
298 static inline struct kmem_cache_order_objects oo_make(int order,
301 struct kmem_cache_order_objects x = {
302 (order << OO_SHIFT) + (PAGE_SIZE << order) / size
308 static inline int oo_order(struct kmem_cache_order_objects x)
310 return x.x >> OO_SHIFT;
313 static inline int oo_objects(struct kmem_cache_order_objects x)
315 return x.x & OO_MASK;
318 #ifdef CONFIG_SLUB_DEBUG
322 #ifdef CONFIG_SLUB_DEBUG_ON
323 static int slub_debug = DEBUG_DEFAULT_FLAGS;
325 static int slub_debug;
328 static char *slub_debug_slabs;
333 static void print_section(char *text, u8 *addr, unsigned int length)
341 for (i = 0; i < length; i++) {
343 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
346 printk(KERN_CONT " %02x", addr[i]);
348 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
350 printk(KERN_CONT " %s\n", ascii);
357 printk(KERN_CONT " ");
361 printk(KERN_CONT " %s\n", ascii);
365 static struct track *get_track(struct kmem_cache *s, void *object,
366 enum track_item alloc)
371 p = object + s->offset + sizeof(void *);
373 p = object + s->inuse;
378 static void set_track(struct kmem_cache *s, void *object,
379 enum track_item alloc, unsigned long addr)
381 struct track *p = get_track(s, object, alloc);
385 p->cpu = smp_processor_id();
386 p->pid = current->pid;
389 memset(p, 0, sizeof(struct track));
392 static void init_tracking(struct kmem_cache *s, void *object)
394 if (!(s->flags & SLAB_STORE_USER))
397 set_track(s, object, TRACK_FREE, 0UL);
398 set_track(s, object, TRACK_ALLOC, 0UL);
401 static void print_track(const char *s, struct track *t)
406 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
407 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
410 static void print_tracking(struct kmem_cache *s, void *object)
412 if (!(s->flags & SLAB_STORE_USER))
415 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
416 print_track("Freed", get_track(s, object, TRACK_FREE));
419 static void print_page_info(struct page *page)
421 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
422 page, page->objects, page->inuse, page->freelist, page->flags);
426 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
432 vsnprintf(buf, sizeof(buf), fmt, args);
434 printk(KERN_ERR "========================================"
435 "=====================================\n");
436 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
437 printk(KERN_ERR "----------------------------------------"
438 "-------------------------------------\n\n");
441 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
447 vsnprintf(buf, sizeof(buf), fmt, args);
449 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
452 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
454 unsigned int off; /* Offset of last byte */
455 u8 *addr = page_address(page);
457 print_tracking(s, p);
459 print_page_info(page);
461 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
462 p, p - addr, get_freepointer(s, p));
465 print_section("Bytes b4", p - 16, 16);
467 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
469 if (s->flags & SLAB_RED_ZONE)
470 print_section("Redzone", p + s->objsize,
471 s->inuse - s->objsize);
474 off = s->offset + sizeof(void *);
478 if (s->flags & SLAB_STORE_USER)
479 off += 2 * sizeof(struct track);
482 /* Beginning of the filler is the free pointer */
483 print_section("Padding", p + off, s->size - off);
488 static void object_err(struct kmem_cache *s, struct page *page,
489 u8 *object, char *reason)
491 slab_bug(s, "%s", reason);
492 print_trailer(s, page, object);
495 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
501 vsnprintf(buf, sizeof(buf), fmt, args);
503 slab_bug(s, "%s", buf);
504 print_page_info(page);
508 static void init_object(struct kmem_cache *s, void *object, int active)
512 if (s->flags & __OBJECT_POISON) {
513 memset(p, POISON_FREE, s->objsize - 1);
514 p[s->objsize - 1] = POISON_END;
517 if (s->flags & SLAB_RED_ZONE)
518 memset(p + s->objsize,
519 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
520 s->inuse - s->objsize);
523 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
526 if (*start != (u8)value)
534 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
535 void *from, void *to)
537 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
538 memset(from, data, to - from);
541 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
542 u8 *object, char *what,
543 u8 *start, unsigned int value, unsigned int bytes)
548 fault = check_bytes(start, value, bytes);
553 while (end > fault && end[-1] == value)
556 slab_bug(s, "%s overwritten", what);
557 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
558 fault, end - 1, fault[0], value);
559 print_trailer(s, page, object);
561 restore_bytes(s, what, value, fault, end);
569 * Bytes of the object to be managed.
570 * If the freepointer may overlay the object then the free
571 * pointer is the first word of the object.
573 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
576 * object + s->objsize
577 * Padding to reach word boundary. This is also used for Redzoning.
578 * Padding is extended by another word if Redzoning is enabled and
581 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
582 * 0xcc (RED_ACTIVE) for objects in use.
585 * Meta data starts here.
587 * A. Free pointer (if we cannot overwrite object on free)
588 * B. Tracking data for SLAB_STORE_USER
589 * C. Padding to reach required alignment boundary or at mininum
590 * one word if debugging is on to be able to detect writes
591 * before the word boundary.
593 * Padding is done using 0x5a (POISON_INUSE)
596 * Nothing is used beyond s->size.
598 * If slabcaches are merged then the objsize and inuse boundaries are mostly
599 * ignored. And therefore no slab options that rely on these boundaries
600 * may be used with merged slabcaches.
603 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
605 unsigned long off = s->inuse; /* The end of info */
608 /* Freepointer is placed after the object. */
609 off += sizeof(void *);
611 if (s->flags & SLAB_STORE_USER)
612 /* We also have user information there */
613 off += 2 * sizeof(struct track);
618 return check_bytes_and_report(s, page, p, "Object padding",
619 p + off, POISON_INUSE, s->size - off);
622 /* Check the pad bytes at the end of a slab page */
623 static int slab_pad_check(struct kmem_cache *s, struct page *page)
631 if (!(s->flags & SLAB_POISON))
634 start = page_address(page);
635 length = (PAGE_SIZE << compound_order(page));
636 end = start + length;
637 remainder = length % s->size;
641 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
644 while (end > fault && end[-1] == POISON_INUSE)
647 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
648 print_section("Padding", end - remainder, remainder);
650 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
654 static int check_object(struct kmem_cache *s, struct page *page,
655 void *object, int active)
658 u8 *endobject = object + s->objsize;
660 if (s->flags & SLAB_RED_ZONE) {
662 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
664 if (!check_bytes_and_report(s, page, object, "Redzone",
665 endobject, red, s->inuse - s->objsize))
668 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
669 check_bytes_and_report(s, page, p, "Alignment padding",
670 endobject, POISON_INUSE, s->inuse - s->objsize);
674 if (s->flags & SLAB_POISON) {
675 if (!active && (s->flags & __OBJECT_POISON) &&
676 (!check_bytes_and_report(s, page, p, "Poison", p,
677 POISON_FREE, s->objsize - 1) ||
678 !check_bytes_and_report(s, page, p, "Poison",
679 p + s->objsize - 1, POISON_END, 1)))
682 * check_pad_bytes cleans up on its own.
684 check_pad_bytes(s, page, p);
687 if (!s->offset && active)
689 * Object and freepointer overlap. Cannot check
690 * freepointer while object is allocated.
694 /* Check free pointer validity */
695 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
696 object_err(s, page, p, "Freepointer corrupt");
698 * No choice but to zap it and thus lose the remainder
699 * of the free objects in this slab. May cause
700 * another error because the object count is now wrong.
702 set_freepointer(s, p, NULL);
708 static int check_slab(struct kmem_cache *s, struct page *page)
712 VM_BUG_ON(!irqs_disabled());
714 if (!PageSlab(page)) {
715 slab_err(s, page, "Not a valid slab page");
719 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
720 if (page->objects > maxobj) {
721 slab_err(s, page, "objects %u > max %u",
722 s->name, page->objects, maxobj);
725 if (page->inuse > page->objects) {
726 slab_err(s, page, "inuse %u > max %u",
727 s->name, page->inuse, page->objects);
730 /* Slab_pad_check fixes things up after itself */
731 slab_pad_check(s, page);
736 * Determine if a certain object on a page is on the freelist. Must hold the
737 * slab lock to guarantee that the chains are in a consistent state.
739 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
742 void *fp = page->freelist;
744 unsigned long max_objects;
746 while (fp && nr <= page->objects) {
749 if (!check_valid_pointer(s, page, fp)) {
751 object_err(s, page, object,
752 "Freechain corrupt");
753 set_freepointer(s, object, NULL);
756 slab_err(s, page, "Freepointer corrupt");
757 page->freelist = NULL;
758 page->inuse = page->objects;
759 slab_fix(s, "Freelist cleared");
765 fp = get_freepointer(s, object);
769 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
770 if (max_objects > MAX_OBJS_PER_PAGE)
771 max_objects = MAX_OBJS_PER_PAGE;
773 if (page->objects != max_objects) {
774 slab_err(s, page, "Wrong number of objects. Found %d but "
775 "should be %d", page->objects, max_objects);
776 page->objects = max_objects;
777 slab_fix(s, "Number of objects adjusted.");
779 if (page->inuse != page->objects - nr) {
780 slab_err(s, page, "Wrong object count. Counter is %d but "
781 "counted were %d", page->inuse, page->objects - nr);
782 page->inuse = page->objects - nr;
783 slab_fix(s, "Object count adjusted.");
785 return search == NULL;
788 static void trace(struct kmem_cache *s, struct page *page, void *object,
791 if (s->flags & SLAB_TRACE) {
792 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
794 alloc ? "alloc" : "free",
799 print_section("Object", (void *)object, s->objsize);
806 * Tracking of fully allocated slabs for debugging purposes.
808 static void add_full(struct kmem_cache_node *n, struct page *page)
810 spin_lock(&n->list_lock);
811 list_add(&page->lru, &n->full);
812 spin_unlock(&n->list_lock);
815 static void remove_full(struct kmem_cache *s, struct page *page)
817 struct kmem_cache_node *n;
819 if (!(s->flags & SLAB_STORE_USER))
822 n = get_node(s, page_to_nid(page));
824 spin_lock(&n->list_lock);
825 list_del(&page->lru);
826 spin_unlock(&n->list_lock);
829 /* Tracking of the number of slabs for debugging purposes */
830 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
832 struct kmem_cache_node *n = get_node(s, node);
834 return atomic_long_read(&n->nr_slabs);
837 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
839 struct kmem_cache_node *n = get_node(s, node);
842 * May be called early in order to allocate a slab for the
843 * kmem_cache_node structure. Solve the chicken-egg
844 * dilemma by deferring the increment of the count during
845 * bootstrap (see early_kmem_cache_node_alloc).
847 if (!NUMA_BUILD || n) {
848 atomic_long_inc(&n->nr_slabs);
849 atomic_long_add(objects, &n->total_objects);
852 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
854 struct kmem_cache_node *n = get_node(s, node);
856 atomic_long_dec(&n->nr_slabs);
857 atomic_long_sub(objects, &n->total_objects);
860 /* Object debug checks for alloc/free paths */
861 static void setup_object_debug(struct kmem_cache *s, struct page *page,
864 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
867 init_object(s, object, 0);
868 init_tracking(s, object);
871 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
872 void *object, unsigned long addr)
874 if (!check_slab(s, page))
877 if (!on_freelist(s, page, object)) {
878 object_err(s, page, object, "Object already allocated");
882 if (!check_valid_pointer(s, page, object)) {
883 object_err(s, page, object, "Freelist Pointer check fails");
887 if (!check_object(s, page, object, 0))
890 /* Success perform special debug activities for allocs */
891 if (s->flags & SLAB_STORE_USER)
892 set_track(s, object, TRACK_ALLOC, addr);
893 trace(s, page, object, 1);
894 init_object(s, object, 1);
898 if (PageSlab(page)) {
900 * If this is a slab page then lets do the best we can
901 * to avoid issues in the future. Marking all objects
902 * as used avoids touching the remaining objects.
904 slab_fix(s, "Marking all objects used");
905 page->inuse = page->objects;
906 page->freelist = NULL;
911 static int free_debug_processing(struct kmem_cache *s, struct page *page,
912 void *object, unsigned long addr)
914 if (!check_slab(s, page))
917 if (!check_valid_pointer(s, page, object)) {
918 slab_err(s, page, "Invalid object pointer 0x%p", object);
922 if (on_freelist(s, page, object)) {
923 object_err(s, page, object, "Object already free");
927 if (!check_object(s, page, object, 1))
930 if (unlikely(s != page->slab)) {
931 if (!PageSlab(page)) {
932 slab_err(s, page, "Attempt to free object(0x%p) "
933 "outside of slab", object);
934 } else if (!page->slab) {
936 "SLUB <none>: no slab for object 0x%p.\n",
940 object_err(s, page, object,
941 "page slab pointer corrupt.");
945 /* Special debug activities for freeing objects */
946 if (!PageSlubFrozen(page) && !page->freelist)
947 remove_full(s, page);
948 if (s->flags & SLAB_STORE_USER)
949 set_track(s, object, TRACK_FREE, addr);
950 trace(s, page, object, 0);
951 init_object(s, object, 0);
955 slab_fix(s, "Object at 0x%p not freed", object);
959 static int __init setup_slub_debug(char *str)
961 slub_debug = DEBUG_DEFAULT_FLAGS;
962 if (*str++ != '=' || !*str)
964 * No options specified. Switch on full debugging.
970 * No options but restriction on slabs. This means full
971 * debugging for slabs matching a pattern.
978 * Switch off all debugging measures.
983 * Determine which debug features should be switched on
985 for (; *str && *str != ','; str++) {
986 switch (tolower(*str)) {
988 slub_debug |= SLAB_DEBUG_FREE;
991 slub_debug |= SLAB_RED_ZONE;
994 slub_debug |= SLAB_POISON;
997 slub_debug |= SLAB_STORE_USER;
1000 slub_debug |= SLAB_TRACE;
1003 printk(KERN_ERR "slub_debug option '%c' "
1004 "unknown. skipped\n", *str);
1010 slub_debug_slabs = str + 1;
1015 __setup("slub_debug", setup_slub_debug);
1017 static unsigned long kmem_cache_flags(unsigned long objsize,
1018 unsigned long flags, const char *name,
1019 void (*ctor)(void *))
1022 * Enable debugging if selected on the kernel commandline.
1024 if (slub_debug && (!slub_debug_slabs ||
1025 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1026 flags |= slub_debug;
1031 static inline void setup_object_debug(struct kmem_cache *s,
1032 struct page *page, void *object) {}
1034 static inline int alloc_debug_processing(struct kmem_cache *s,
1035 struct page *page, void *object, unsigned long addr) { return 0; }
1037 static inline int free_debug_processing(struct kmem_cache *s,
1038 struct page *page, void *object, unsigned long addr) { return 0; }
1040 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1042 static inline int check_object(struct kmem_cache *s, struct page *page,
1043 void *object, int active) { return 1; }
1044 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1045 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1046 unsigned long flags, const char *name,
1047 void (*ctor)(void *))
1051 #define slub_debug 0
1053 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1055 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1057 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1062 * Slab allocation and freeing
1064 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1065 struct kmem_cache_order_objects oo)
1067 int order = oo_order(oo);
1069 flags |= __GFP_NOTRACK;
1072 return alloc_pages(flags, order);
1074 return alloc_pages_node(node, flags, order);
1077 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1080 struct kmem_cache_order_objects oo = s->oo;
1082 flags |= s->allocflags;
1084 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1086 if (unlikely(!page)) {
1089 * Allocation may have failed due to fragmentation.
1090 * Try a lower order alloc if possible
1092 page = alloc_slab_page(flags, node, oo);
1096 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1099 if (kmemcheck_enabled
1100 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS)))
1102 int pages = 1 << oo_order(oo);
1104 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1107 * Objects from caches that have a constructor don't get
1108 * cleared when they're allocated, so we need to do it here.
1111 kmemcheck_mark_uninitialized_pages(page, pages);
1113 kmemcheck_mark_unallocated_pages(page, pages);
1116 page->objects = oo_objects(oo);
1117 mod_zone_page_state(page_zone(page),
1118 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1119 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1125 static void setup_object(struct kmem_cache *s, struct page *page,
1128 setup_object_debug(s, page, object);
1129 if (unlikely(s->ctor))
1133 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1140 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1142 page = allocate_slab(s,
1143 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1147 inc_slabs_node(s, page_to_nid(page), page->objects);
1149 page->flags |= 1 << PG_slab;
1150 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1151 SLAB_STORE_USER | SLAB_TRACE))
1152 __SetPageSlubDebug(page);
1154 start = page_address(page);
1156 if (unlikely(s->flags & SLAB_POISON))
1157 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1160 for_each_object(p, s, start, page->objects) {
1161 setup_object(s, page, last);
1162 set_freepointer(s, last, p);
1165 setup_object(s, page, last);
1166 set_freepointer(s, last, NULL);
1168 page->freelist = start;
1174 static void __free_slab(struct kmem_cache *s, struct page *page)
1176 int order = compound_order(page);
1177 int pages = 1 << order;
1179 if (unlikely(SLABDEBUG && PageSlubDebug(page))) {
1182 slab_pad_check(s, page);
1183 for_each_object(p, s, page_address(page),
1185 check_object(s, page, p, 0);
1186 __ClearPageSlubDebug(page);
1189 kmemcheck_free_shadow(page, compound_order(page));
1191 mod_zone_page_state(page_zone(page),
1192 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1193 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1196 __ClearPageSlab(page);
1197 reset_page_mapcount(page);
1198 if (current->reclaim_state)
1199 current->reclaim_state->reclaimed_slab += pages;
1200 __free_pages(page, order);
1203 static void rcu_free_slab(struct rcu_head *h)
1207 page = container_of((struct list_head *)h, struct page, lru);
1208 __free_slab(page->slab, page);
1211 static void free_slab(struct kmem_cache *s, struct page *page)
1213 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1215 * RCU free overloads the RCU head over the LRU
1217 struct rcu_head *head = (void *)&page->lru;
1219 call_rcu(head, rcu_free_slab);
1221 __free_slab(s, page);
1224 static void discard_slab(struct kmem_cache *s, struct page *page)
1226 dec_slabs_node(s, page_to_nid(page), page->objects);
1231 * Per slab locking using the pagelock
1233 static __always_inline void slab_lock(struct page *page)
1235 bit_spin_lock(PG_locked, &page->flags);
1238 static __always_inline void slab_unlock(struct page *page)
1240 __bit_spin_unlock(PG_locked, &page->flags);
1243 static __always_inline int slab_trylock(struct page *page)
1247 rc = bit_spin_trylock(PG_locked, &page->flags);
1252 * Management of partially allocated slabs
1254 static void add_partial(struct kmem_cache_node *n,
1255 struct page *page, int tail)
1257 spin_lock(&n->list_lock);
1260 list_add_tail(&page->lru, &n->partial);
1262 list_add(&page->lru, &n->partial);
1263 spin_unlock(&n->list_lock);
1266 static void remove_partial(struct kmem_cache *s, struct page *page)
1268 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1270 spin_lock(&n->list_lock);
1271 list_del(&page->lru);
1273 spin_unlock(&n->list_lock);
1277 * Lock slab and remove from the partial list.
1279 * Must hold list_lock.
1281 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1284 if (slab_trylock(page)) {
1285 list_del(&page->lru);
1287 __SetPageSlubFrozen(page);
1294 * Try to allocate a partial slab from a specific node.
1296 static struct page *get_partial_node(struct kmem_cache_node *n)
1301 * Racy check. If we mistakenly see no partial slabs then we
1302 * just allocate an empty slab. If we mistakenly try to get a
1303 * partial slab and there is none available then get_partials()
1306 if (!n || !n->nr_partial)
1309 spin_lock(&n->list_lock);
1310 list_for_each_entry(page, &n->partial, lru)
1311 if (lock_and_freeze_slab(n, page))
1315 spin_unlock(&n->list_lock);
1320 * Get a page from somewhere. Search in increasing NUMA distances.
1322 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1325 struct zonelist *zonelist;
1328 enum zone_type high_zoneidx = gfp_zone(flags);
1332 * The defrag ratio allows a configuration of the tradeoffs between
1333 * inter node defragmentation and node local allocations. A lower
1334 * defrag_ratio increases the tendency to do local allocations
1335 * instead of attempting to obtain partial slabs from other nodes.
1337 * If the defrag_ratio is set to 0 then kmalloc() always
1338 * returns node local objects. If the ratio is higher then kmalloc()
1339 * may return off node objects because partial slabs are obtained
1340 * from other nodes and filled up.
1342 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1343 * defrag_ratio = 1000) then every (well almost) allocation will
1344 * first attempt to defrag slab caches on other nodes. This means
1345 * scanning over all nodes to look for partial slabs which may be
1346 * expensive if we do it every time we are trying to find a slab
1347 * with available objects.
1349 if (!s->remote_node_defrag_ratio ||
1350 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1353 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1354 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1355 struct kmem_cache_node *n;
1357 n = get_node(s, zone_to_nid(zone));
1359 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1360 n->nr_partial > s->min_partial) {
1361 page = get_partial_node(n);
1371 * Get a partial page, lock it and return it.
1373 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1376 int searchnode = (node == -1) ? numa_node_id() : node;
1378 page = get_partial_node(get_node(s, searchnode));
1379 if (page || (flags & __GFP_THISNODE))
1382 return get_any_partial(s, flags);
1386 * Move a page back to the lists.
1388 * Must be called with the slab lock held.
1390 * On exit the slab lock will have been dropped.
1392 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1394 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1395 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1397 __ClearPageSlubFrozen(page);
1400 if (page->freelist) {
1401 add_partial(n, page, tail);
1402 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1404 stat(c, DEACTIVATE_FULL);
1405 if (SLABDEBUG && PageSlubDebug(page) &&
1406 (s->flags & SLAB_STORE_USER))
1411 stat(c, DEACTIVATE_EMPTY);
1412 if (n->nr_partial < s->min_partial) {
1414 * Adding an empty slab to the partial slabs in order
1415 * to avoid page allocator overhead. This slab needs
1416 * to come after the other slabs with objects in
1417 * so that the others get filled first. That way the
1418 * size of the partial list stays small.
1420 * kmem_cache_shrink can reclaim any empty slabs from
1423 add_partial(n, page, 1);
1427 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1428 discard_slab(s, page);
1434 * Remove the cpu slab
1436 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1438 struct page *page = c->page;
1442 stat(c, DEACTIVATE_REMOTE_FREES);
1444 * Merge cpu freelist into slab freelist. Typically we get here
1445 * because both freelists are empty. So this is unlikely
1448 while (unlikely(c->freelist)) {
1451 tail = 0; /* Hot objects. Put the slab first */
1453 /* Retrieve object from cpu_freelist */
1454 object = c->freelist;
1455 c->freelist = c->freelist[c->offset];
1457 /* And put onto the regular freelist */
1458 object[c->offset] = page->freelist;
1459 page->freelist = object;
1463 unfreeze_slab(s, page, tail);
1466 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1468 stat(c, CPUSLAB_FLUSH);
1470 deactivate_slab(s, c);
1476 * Called from IPI handler with interrupts disabled.
1478 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1480 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1482 if (likely(c && c->page))
1486 static void flush_cpu_slab(void *d)
1488 struct kmem_cache *s = d;
1490 __flush_cpu_slab(s, smp_processor_id());
1493 static void flush_all(struct kmem_cache *s)
1495 on_each_cpu(flush_cpu_slab, s, 1);
1499 * Check if the objects in a per cpu structure fit numa
1500 * locality expectations.
1502 static inline int node_match(struct kmem_cache_cpu *c, int node)
1505 if (node != -1 && c->node != node)
1512 * Slow path. The lockless freelist is empty or we need to perform
1515 * Interrupts are disabled.
1517 * Processing is still very fast if new objects have been freed to the
1518 * regular freelist. In that case we simply take over the regular freelist
1519 * as the lockless freelist and zap the regular freelist.
1521 * If that is not working then we fall back to the partial lists. We take the
1522 * first element of the freelist as the object to allocate now and move the
1523 * rest of the freelist to the lockless freelist.
1525 * And if we were unable to get a new slab from the partial slab lists then
1526 * we need to allocate a new slab. This is the slowest path since it involves
1527 * a call to the page allocator and the setup of a new slab.
1529 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1530 unsigned long addr, struct kmem_cache_cpu *c)
1535 /* We handle __GFP_ZERO in the caller */
1536 gfpflags &= ~__GFP_ZERO;
1542 if (unlikely(!node_match(c, node)))
1545 stat(c, ALLOC_REFILL);
1548 object = c->page->freelist;
1549 if (unlikely(!object))
1551 if (unlikely(SLABDEBUG && PageSlubDebug(c->page)))
1554 c->freelist = object[c->offset];
1555 c->page->inuse = c->page->objects;
1556 c->page->freelist = NULL;
1557 c->node = page_to_nid(c->page);
1559 slab_unlock(c->page);
1560 stat(c, ALLOC_SLOWPATH);
1564 deactivate_slab(s, c);
1567 new = get_partial(s, gfpflags, node);
1570 stat(c, ALLOC_FROM_PARTIAL);
1574 if (gfpflags & __GFP_WAIT)
1577 new = new_slab(s, gfpflags, node);
1579 if (gfpflags & __GFP_WAIT)
1580 local_irq_disable();
1583 c = get_cpu_slab(s, smp_processor_id());
1584 stat(c, ALLOC_SLAB);
1588 __SetPageSlubFrozen(new);
1594 if (!alloc_debug_processing(s, c->page, object, addr))
1598 c->page->freelist = object[c->offset];
1604 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1605 * have the fastpath folded into their functions. So no function call
1606 * overhead for requests that can be satisfied on the fastpath.
1608 * The fastpath works by first checking if the lockless freelist can be used.
1609 * If not then __slab_alloc is called for slow processing.
1611 * Otherwise we can simply pick the next object from the lockless free list.
1613 static __always_inline void *slab_alloc(struct kmem_cache *s,
1614 gfp_t gfpflags, int node, unsigned long addr)
1617 struct kmem_cache_cpu *c;
1618 unsigned long flags;
1619 unsigned int objsize;
1621 lockdep_trace_alloc(gfpflags);
1622 might_sleep_if(gfpflags & __GFP_WAIT);
1624 if (should_failslab(s->objsize, gfpflags))
1627 local_irq_save(flags);
1628 c = get_cpu_slab(s, smp_processor_id());
1629 objsize = c->objsize;
1630 if (unlikely(!c->freelist || !node_match(c, node)))
1632 object = __slab_alloc(s, gfpflags, node, addr, c);
1635 object = c->freelist;
1636 c->freelist = object[c->offset];
1637 stat(c, ALLOC_FASTPATH);
1639 local_irq_restore(flags);
1641 if (unlikely((gfpflags & __GFP_ZERO) && object))
1642 memset(object, 0, objsize);
1644 kmemcheck_slab_alloc(s, gfpflags, object, c->objsize);
1645 kmemleak_alloc_recursive(object, objsize, 1, s->flags, gfpflags);
1650 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1652 void *ret = slab_alloc(s, gfpflags, -1, _RET_IP_);
1654 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
1658 EXPORT_SYMBOL(kmem_cache_alloc);
1660 #ifdef CONFIG_KMEMTRACE
1661 void *kmem_cache_alloc_notrace(struct kmem_cache *s, gfp_t gfpflags)
1663 return slab_alloc(s, gfpflags, -1, _RET_IP_);
1665 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
1669 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1671 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
1673 trace_kmem_cache_alloc_node(_RET_IP_, ret,
1674 s->objsize, s->size, gfpflags, node);
1678 EXPORT_SYMBOL(kmem_cache_alloc_node);
1681 #ifdef CONFIG_KMEMTRACE
1682 void *kmem_cache_alloc_node_notrace(struct kmem_cache *s,
1686 return slab_alloc(s, gfpflags, node, _RET_IP_);
1688 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
1692 * Slow patch handling. This may still be called frequently since objects
1693 * have a longer lifetime than the cpu slabs in most processing loads.
1695 * So we still attempt to reduce cache line usage. Just take the slab
1696 * lock and free the item. If there is no additional partial page
1697 * handling required then we can return immediately.
1699 static void __slab_free(struct kmem_cache *s, struct page *page,
1700 void *x, unsigned long addr, unsigned int offset)
1703 void **object = (void *)x;
1704 struct kmem_cache_cpu *c;
1706 c = get_cpu_slab(s, raw_smp_processor_id());
1707 stat(c, FREE_SLOWPATH);
1710 if (unlikely(SLABDEBUG && PageSlubDebug(page)))
1714 prior = object[offset] = page->freelist;
1715 page->freelist = object;
1718 if (unlikely(PageSlubFrozen(page))) {
1719 stat(c, FREE_FROZEN);
1723 if (unlikely(!page->inuse))
1727 * Objects left in the slab. If it was not on the partial list before
1730 if (unlikely(!prior)) {
1731 add_partial(get_node(s, page_to_nid(page)), page, 1);
1732 stat(c, FREE_ADD_PARTIAL);
1742 * Slab still on the partial list.
1744 remove_partial(s, page);
1745 stat(c, FREE_REMOVE_PARTIAL);
1749 discard_slab(s, page);
1753 if (!free_debug_processing(s, page, x, addr))
1759 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1760 * can perform fastpath freeing without additional function calls.
1762 * The fastpath is only possible if we are freeing to the current cpu slab
1763 * of this processor. This typically the case if we have just allocated
1766 * If fastpath is not possible then fall back to __slab_free where we deal
1767 * with all sorts of special processing.
1769 static __always_inline void slab_free(struct kmem_cache *s,
1770 struct page *page, void *x, unsigned long addr)
1772 void **object = (void *)x;
1773 struct kmem_cache_cpu *c;
1774 unsigned long flags;
1776 kmemleak_free_recursive(x, s->flags);
1777 local_irq_save(flags);
1778 c = get_cpu_slab(s, smp_processor_id());
1779 kmemcheck_slab_free(s, object, c->objsize);
1780 debug_check_no_locks_freed(object, c->objsize);
1781 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1782 debug_check_no_obj_freed(object, c->objsize);
1783 if (likely(page == c->page && c->node >= 0)) {
1784 object[c->offset] = c->freelist;
1785 c->freelist = object;
1786 stat(c, FREE_FASTPATH);
1788 __slab_free(s, page, x, addr, c->offset);
1790 local_irq_restore(flags);
1793 void kmem_cache_free(struct kmem_cache *s, void *x)
1797 page = virt_to_head_page(x);
1799 slab_free(s, page, x, _RET_IP_);
1801 trace_kmem_cache_free(_RET_IP_, x);
1803 EXPORT_SYMBOL(kmem_cache_free);
1805 /* Figure out on which slab page the object resides */
1806 static struct page *get_object_page(const void *x)
1808 struct page *page = virt_to_head_page(x);
1810 if (!PageSlab(page))
1817 * Object placement in a slab is made very easy because we always start at
1818 * offset 0. If we tune the size of the object to the alignment then we can
1819 * get the required alignment by putting one properly sized object after
1822 * Notice that the allocation order determines the sizes of the per cpu
1823 * caches. Each processor has always one slab available for allocations.
1824 * Increasing the allocation order reduces the number of times that slabs
1825 * must be moved on and off the partial lists and is therefore a factor in
1830 * Mininum / Maximum order of slab pages. This influences locking overhead
1831 * and slab fragmentation. A higher order reduces the number of partial slabs
1832 * and increases the number of allocations possible without having to
1833 * take the list_lock.
1835 static int slub_min_order;
1836 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1837 static int slub_min_objects;
1840 * Merge control. If this is set then no merging of slab caches will occur.
1841 * (Could be removed. This was introduced to pacify the merge skeptics.)
1843 static int slub_nomerge;
1846 * Calculate the order of allocation given an slab object size.
1848 * The order of allocation has significant impact on performance and other
1849 * system components. Generally order 0 allocations should be preferred since
1850 * order 0 does not cause fragmentation in the page allocator. Larger objects
1851 * be problematic to put into order 0 slabs because there may be too much
1852 * unused space left. We go to a higher order if more than 1/16th of the slab
1855 * In order to reach satisfactory performance we must ensure that a minimum
1856 * number of objects is in one slab. Otherwise we may generate too much
1857 * activity on the partial lists which requires taking the list_lock. This is
1858 * less a concern for large slabs though which are rarely used.
1860 * slub_max_order specifies the order where we begin to stop considering the
1861 * number of objects in a slab as critical. If we reach slub_max_order then
1862 * we try to keep the page order as low as possible. So we accept more waste
1863 * of space in favor of a small page order.
1865 * Higher order allocations also allow the placement of more objects in a
1866 * slab and thereby reduce object handling overhead. If the user has
1867 * requested a higher mininum order then we start with that one instead of
1868 * the smallest order which will fit the object.
1870 static inline int slab_order(int size, int min_objects,
1871 int max_order, int fract_leftover)
1875 int min_order = slub_min_order;
1877 if ((PAGE_SIZE << min_order) / size > MAX_OBJS_PER_PAGE)
1878 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
1880 for (order = max(min_order,
1881 fls(min_objects * size - 1) - PAGE_SHIFT);
1882 order <= max_order; order++) {
1884 unsigned long slab_size = PAGE_SIZE << order;
1886 if (slab_size < min_objects * size)
1889 rem = slab_size % size;
1891 if (rem <= slab_size / fract_leftover)
1899 static inline int calculate_order(int size)
1907 * Attempt to find best configuration for a slab. This
1908 * works by first attempting to generate a layout with
1909 * the best configuration and backing off gradually.
1911 * First we reduce the acceptable waste in a slab. Then
1912 * we reduce the minimum objects required in a slab.
1914 min_objects = slub_min_objects;
1916 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1917 max_objects = (PAGE_SIZE << slub_max_order)/size;
1918 min_objects = min(min_objects, max_objects);
1920 while (min_objects > 1) {
1922 while (fraction >= 4) {
1923 order = slab_order(size, min_objects,
1924 slub_max_order, fraction);
1925 if (order <= slub_max_order)
1933 * We were unable to place multiple objects in a slab. Now
1934 * lets see if we can place a single object there.
1936 order = slab_order(size, 1, slub_max_order, 1);
1937 if (order <= slub_max_order)
1941 * Doh this slab cannot be placed using slub_max_order.
1943 order = slab_order(size, 1, MAX_ORDER, 1);
1944 if (order < MAX_ORDER)
1950 * Figure out what the alignment of the objects will be.
1952 static unsigned long calculate_alignment(unsigned long flags,
1953 unsigned long align, unsigned long size)
1956 * If the user wants hardware cache aligned objects then follow that
1957 * suggestion if the object is sufficiently large.
1959 * The hardware cache alignment cannot override the specified
1960 * alignment though. If that is greater then use it.
1962 if (flags & SLAB_HWCACHE_ALIGN) {
1963 unsigned long ralign = cache_line_size();
1964 while (size <= ralign / 2)
1966 align = max(align, ralign);
1969 if (align < ARCH_SLAB_MINALIGN)
1970 align = ARCH_SLAB_MINALIGN;
1972 return ALIGN(align, sizeof(void *));
1975 static void init_kmem_cache_cpu(struct kmem_cache *s,
1976 struct kmem_cache_cpu *c)
1981 c->offset = s->offset / sizeof(void *);
1982 c->objsize = s->objsize;
1983 #ifdef CONFIG_SLUB_STATS
1984 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1989 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
1992 spin_lock_init(&n->list_lock);
1993 INIT_LIST_HEAD(&n->partial);
1994 #ifdef CONFIG_SLUB_DEBUG
1995 atomic_long_set(&n->nr_slabs, 0);
1996 atomic_long_set(&n->total_objects, 0);
1997 INIT_LIST_HEAD(&n->full);
2003 * Per cpu array for per cpu structures.
2005 * The per cpu array places all kmem_cache_cpu structures from one processor
2006 * close together meaning that it becomes possible that multiple per cpu
2007 * structures are contained in one cacheline. This may be particularly
2008 * beneficial for the kmalloc caches.
2010 * A desktop system typically has around 60-80 slabs. With 100 here we are
2011 * likely able to get per cpu structures for all caches from the array defined
2012 * here. We must be able to cover all kmalloc caches during bootstrap.
2014 * If the per cpu array is exhausted then fall back to kmalloc
2015 * of individual cachelines. No sharing is possible then.
2017 #define NR_KMEM_CACHE_CPU 100
2019 static DEFINE_PER_CPU(struct kmem_cache_cpu,
2020 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
2022 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
2023 static DECLARE_BITMAP(kmem_cach_cpu_free_init_once, CONFIG_NR_CPUS);
2025 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
2026 int cpu, gfp_t flags)
2028 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
2031 per_cpu(kmem_cache_cpu_free, cpu) =
2032 (void *)c->freelist;
2034 /* Table overflow: So allocate ourselves */
2036 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2037 flags, cpu_to_node(cpu));
2042 init_kmem_cache_cpu(s, c);
2046 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2048 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2049 c >= per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2053 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2054 per_cpu(kmem_cache_cpu_free, cpu) = c;
2057 static void free_kmem_cache_cpus(struct kmem_cache *s)
2061 for_each_online_cpu(cpu) {
2062 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2065 s->cpu_slab[cpu] = NULL;
2066 free_kmem_cache_cpu(c, cpu);
2071 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2075 for_each_online_cpu(cpu) {
2076 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2081 c = alloc_kmem_cache_cpu(s, cpu, flags);
2083 free_kmem_cache_cpus(s);
2086 s->cpu_slab[cpu] = c;
2092 * Initialize the per cpu array.
2094 static void init_alloc_cpu_cpu(int cpu)
2098 if (cpumask_test_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once)))
2101 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2102 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2104 cpumask_set_cpu(cpu, to_cpumask(kmem_cach_cpu_free_init_once));
2107 static void __init init_alloc_cpu(void)
2111 for_each_online_cpu(cpu)
2112 init_alloc_cpu_cpu(cpu);
2116 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2117 static inline void init_alloc_cpu(void) {}
2119 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2121 init_kmem_cache_cpu(s, &s->cpu_slab);
2128 * No kmalloc_node yet so do it by hand. We know that this is the first
2129 * slab on the node for this slabcache. There are no concurrent accesses
2132 * Note that this function only works on the kmalloc_node_cache
2133 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2134 * memory on a fresh node that has no slab structures yet.
2136 static void early_kmem_cache_node_alloc(gfp_t gfpflags, int node)
2139 struct kmem_cache_node *n;
2140 unsigned long flags;
2142 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2144 page = new_slab(kmalloc_caches, gfpflags, node);
2147 if (page_to_nid(page) != node) {
2148 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2150 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2151 "in order to be able to continue\n");
2156 page->freelist = get_freepointer(kmalloc_caches, n);
2158 kmalloc_caches->node[node] = n;
2159 #ifdef CONFIG_SLUB_DEBUG
2160 init_object(kmalloc_caches, n, 1);
2161 init_tracking(kmalloc_caches, n);
2163 init_kmem_cache_node(n, kmalloc_caches);
2164 inc_slabs_node(kmalloc_caches, node, page->objects);
2167 * lockdep requires consistent irq usage for each lock
2168 * so even though there cannot be a race this early in
2169 * the boot sequence, we still disable irqs.
2171 local_irq_save(flags);
2172 add_partial(n, page, 0);
2173 local_irq_restore(flags);
2176 static void free_kmem_cache_nodes(struct kmem_cache *s)
2180 for_each_node_state(node, N_NORMAL_MEMORY) {
2181 struct kmem_cache_node *n = s->node[node];
2182 if (n && n != &s->local_node)
2183 kmem_cache_free(kmalloc_caches, n);
2184 s->node[node] = NULL;
2188 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2193 if (slab_state >= UP)
2194 local_node = page_to_nid(virt_to_page(s));
2198 for_each_node_state(node, N_NORMAL_MEMORY) {
2199 struct kmem_cache_node *n;
2201 if (local_node == node)
2204 if (slab_state == DOWN) {
2205 early_kmem_cache_node_alloc(gfpflags, node);
2208 n = kmem_cache_alloc_node(kmalloc_caches,
2212 free_kmem_cache_nodes(s);
2218 init_kmem_cache_node(n, s);
2223 static void free_kmem_cache_nodes(struct kmem_cache *s)
2227 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2229 init_kmem_cache_node(&s->local_node, s);
2234 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2236 if (min < MIN_PARTIAL)
2238 else if (min > MAX_PARTIAL)
2240 s->min_partial = min;
2244 * calculate_sizes() determines the order and the distribution of data within
2247 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2249 unsigned long flags = s->flags;
2250 unsigned long size = s->objsize;
2251 unsigned long align = s->align;
2255 * Round up object size to the next word boundary. We can only
2256 * place the free pointer at word boundaries and this determines
2257 * the possible location of the free pointer.
2259 size = ALIGN(size, sizeof(void *));
2261 #ifdef CONFIG_SLUB_DEBUG
2263 * Determine if we can poison the object itself. If the user of
2264 * the slab may touch the object after free or before allocation
2265 * then we should never poison the object itself.
2267 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2269 s->flags |= __OBJECT_POISON;
2271 s->flags &= ~__OBJECT_POISON;
2275 * If we are Redzoning then check if there is some space between the
2276 * end of the object and the free pointer. If not then add an
2277 * additional word to have some bytes to store Redzone information.
2279 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2280 size += sizeof(void *);
2284 * With that we have determined the number of bytes in actual use
2285 * by the object. This is the potential offset to the free pointer.
2289 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2292 * Relocate free pointer after the object if it is not
2293 * permitted to overwrite the first word of the object on
2296 * This is the case if we do RCU, have a constructor or
2297 * destructor or are poisoning the objects.
2300 size += sizeof(void *);
2303 #ifdef CONFIG_SLUB_DEBUG
2304 if (flags & SLAB_STORE_USER)
2306 * Need to store information about allocs and frees after
2309 size += 2 * sizeof(struct track);
2311 if (flags & SLAB_RED_ZONE)
2313 * Add some empty padding so that we can catch
2314 * overwrites from earlier objects rather than let
2315 * tracking information or the free pointer be
2316 * corrupted if a user writes before the start
2319 size += sizeof(void *);
2323 * Determine the alignment based on various parameters that the
2324 * user specified and the dynamic determination of cache line size
2327 align = calculate_alignment(flags, align, s->objsize);
2330 * SLUB stores one object immediately after another beginning from
2331 * offset 0. In order to align the objects we have to simply size
2332 * each object to conform to the alignment.
2334 size = ALIGN(size, align);
2336 if (forced_order >= 0)
2337 order = forced_order;
2339 order = calculate_order(size);
2346 s->allocflags |= __GFP_COMP;
2348 if (s->flags & SLAB_CACHE_DMA)
2349 s->allocflags |= SLUB_DMA;
2351 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2352 s->allocflags |= __GFP_RECLAIMABLE;
2355 * Determine the number of objects per slab
2357 s->oo = oo_make(order, size);
2358 s->min = oo_make(get_order(size), size);
2359 if (oo_objects(s->oo) > oo_objects(s->max))
2362 return !!oo_objects(s->oo);
2366 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2367 const char *name, size_t size,
2368 size_t align, unsigned long flags,
2369 void (*ctor)(void *))
2371 memset(s, 0, kmem_size);
2376 s->flags = kmem_cache_flags(size, flags, name, ctor);
2378 if (!calculate_sizes(s, -1))
2382 * The larger the object size is, the more pages we want on the partial
2383 * list to avoid pounding the page allocator excessively.
2385 set_min_partial(s, ilog2(s->size));
2388 s->remote_node_defrag_ratio = 1000;
2390 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2393 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2395 free_kmem_cache_nodes(s);
2397 if (flags & SLAB_PANIC)
2398 panic("Cannot create slab %s size=%lu realsize=%u "
2399 "order=%u offset=%u flags=%lx\n",
2400 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2406 * Check if a given pointer is valid
2408 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2412 page = get_object_page(object);
2414 if (!page || s != page->slab)
2415 /* No slab or wrong slab */
2418 if (!check_valid_pointer(s, page, object))
2422 * We could also check if the object is on the slabs freelist.
2423 * But this would be too expensive and it seems that the main
2424 * purpose of kmem_ptr_valid() is to check if the object belongs
2425 * to a certain slab.
2429 EXPORT_SYMBOL(kmem_ptr_validate);
2432 * Determine the size of a slab object
2434 unsigned int kmem_cache_size(struct kmem_cache *s)
2438 EXPORT_SYMBOL(kmem_cache_size);
2440 const char *kmem_cache_name(struct kmem_cache *s)
2444 EXPORT_SYMBOL(kmem_cache_name);
2446 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2449 #ifdef CONFIG_SLUB_DEBUG
2450 void *addr = page_address(page);
2452 DECLARE_BITMAP(map, page->objects);
2454 bitmap_zero(map, page->objects);
2455 slab_err(s, page, "%s", text);
2457 for_each_free_object(p, s, page->freelist)
2458 set_bit(slab_index(p, s, addr), map);
2460 for_each_object(p, s, addr, page->objects) {
2462 if (!test_bit(slab_index(p, s, addr), map)) {
2463 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2465 print_tracking(s, p);
2473 * Attempt to free all partial slabs on a node.
2475 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2477 unsigned long flags;
2478 struct page *page, *h;
2480 spin_lock_irqsave(&n->list_lock, flags);
2481 list_for_each_entry_safe(page, h, &n->partial, lru) {
2483 list_del(&page->lru);
2484 discard_slab(s, page);
2487 list_slab_objects(s, page,
2488 "Objects remaining on kmem_cache_close()");
2491 spin_unlock_irqrestore(&n->list_lock, flags);
2495 * Release all resources used by a slab cache.
2497 static inline int kmem_cache_close(struct kmem_cache *s)
2503 /* Attempt to free all objects */
2504 free_kmem_cache_cpus(s);
2505 for_each_node_state(node, N_NORMAL_MEMORY) {
2506 struct kmem_cache_node *n = get_node(s, node);
2509 if (n->nr_partial || slabs_node(s, node))
2512 free_kmem_cache_nodes(s);
2517 * Close a cache and release the kmem_cache structure
2518 * (must be used for caches created using kmem_cache_create)
2520 void kmem_cache_destroy(struct kmem_cache *s)
2522 down_write(&slub_lock);
2526 up_write(&slub_lock);
2527 if (kmem_cache_close(s)) {
2528 printk(KERN_ERR "SLUB %s: %s called for cache that "
2529 "still has objects.\n", s->name, __func__);
2532 sysfs_slab_remove(s);
2534 up_write(&slub_lock);
2536 EXPORT_SYMBOL(kmem_cache_destroy);
2538 /********************************************************************
2540 *******************************************************************/
2542 struct kmem_cache kmalloc_caches[SLUB_PAGE_SHIFT] __cacheline_aligned;
2543 EXPORT_SYMBOL(kmalloc_caches);
2545 static int __init setup_slub_min_order(char *str)
2547 get_option(&str, &slub_min_order);
2552 __setup("slub_min_order=", setup_slub_min_order);
2554 static int __init setup_slub_max_order(char *str)
2556 get_option(&str, &slub_max_order);
2557 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2562 __setup("slub_max_order=", setup_slub_max_order);
2564 static int __init setup_slub_min_objects(char *str)
2566 get_option(&str, &slub_min_objects);
2571 __setup("slub_min_objects=", setup_slub_min_objects);
2573 static int __init setup_slub_nomerge(char *str)
2579 __setup("slub_nomerge", setup_slub_nomerge);
2581 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2582 const char *name, int size, gfp_t gfp_flags)
2584 unsigned int flags = 0;
2586 if (gfp_flags & SLUB_DMA)
2587 flags = SLAB_CACHE_DMA;
2590 * This function is called with IRQs disabled during early-boot on
2591 * single CPU so there's no need to take slub_lock here.
2593 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2597 list_add(&s->list, &slab_caches);
2599 if (sysfs_slab_add(s))
2604 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2607 #ifdef CONFIG_ZONE_DMA
2608 static struct kmem_cache *kmalloc_caches_dma[SLUB_PAGE_SHIFT];
2610 static void sysfs_add_func(struct work_struct *w)
2612 struct kmem_cache *s;
2614 down_write(&slub_lock);
2615 list_for_each_entry(s, &slab_caches, list) {
2616 if (s->flags & __SYSFS_ADD_DEFERRED) {
2617 s->flags &= ~__SYSFS_ADD_DEFERRED;
2621 up_write(&slub_lock);
2624 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2626 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2628 struct kmem_cache *s;
2632 s = kmalloc_caches_dma[index];
2636 /* Dynamically create dma cache */
2637 if (flags & __GFP_WAIT)
2638 down_write(&slub_lock);
2640 if (!down_write_trylock(&slub_lock))
2644 if (kmalloc_caches_dma[index])
2647 realsize = kmalloc_caches[index].objsize;
2648 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2649 (unsigned int)realsize);
2650 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2652 if (!s || !text || !kmem_cache_open(s, flags, text,
2653 realsize, ARCH_KMALLOC_MINALIGN,
2654 SLAB_CACHE_DMA|SLAB_NOTRACK|__SYSFS_ADD_DEFERRED,
2661 list_add(&s->list, &slab_caches);
2662 kmalloc_caches_dma[index] = s;
2664 schedule_work(&sysfs_add_work);
2667 up_write(&slub_lock);
2669 return kmalloc_caches_dma[index];
2674 * Conversion table for small slabs sizes / 8 to the index in the
2675 * kmalloc array. This is necessary for slabs < 192 since we have non power
2676 * of two cache sizes there. The size of larger slabs can be determined using
2679 static s8 size_index[24] = {
2706 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2712 return ZERO_SIZE_PTR;
2714 index = size_index[(size - 1) / 8];
2716 index = fls(size - 1);
2718 #ifdef CONFIG_ZONE_DMA
2719 if (unlikely((flags & SLUB_DMA)))
2720 return dma_kmalloc_cache(index, flags);
2723 return &kmalloc_caches[index];
2726 void *__kmalloc(size_t size, gfp_t flags)
2728 struct kmem_cache *s;
2731 if (unlikely(size > SLUB_MAX_SIZE))
2732 return kmalloc_large(size, flags);
2734 s = get_slab(size, flags);
2736 if (unlikely(ZERO_OR_NULL_PTR(s)))
2739 ret = slab_alloc(s, flags, -1, _RET_IP_);
2741 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2745 EXPORT_SYMBOL(__kmalloc);
2747 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2751 flags |= __GFP_COMP | __GFP_NOTRACK;
2752 page = alloc_pages_node(node, flags, get_order(size));
2754 return page_address(page);
2760 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2762 struct kmem_cache *s;
2765 if (unlikely(size > SLUB_MAX_SIZE)) {
2766 ret = kmalloc_large_node(size, flags, node);
2768 trace_kmalloc_node(_RET_IP_, ret,
2769 size, PAGE_SIZE << get_order(size),
2775 s = get_slab(size, flags);
2777 if (unlikely(ZERO_OR_NULL_PTR(s)))
2780 ret = slab_alloc(s, flags, node, _RET_IP_);
2782 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2786 EXPORT_SYMBOL(__kmalloc_node);
2789 size_t ksize(const void *object)
2792 struct kmem_cache *s;
2794 if (unlikely(object == ZERO_SIZE_PTR))
2797 page = virt_to_head_page(object);
2799 if (unlikely(!PageSlab(page))) {
2800 WARN_ON(!PageCompound(page));
2801 return PAGE_SIZE << compound_order(page);
2805 #ifdef CONFIG_SLUB_DEBUG
2807 * Debugging requires use of the padding between object
2808 * and whatever may come after it.
2810 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2815 * If we have the need to store the freelist pointer
2816 * back there or track user information then we can
2817 * only use the space before that information.
2819 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2822 * Else we can use all the padding etc for the allocation
2826 EXPORT_SYMBOL(ksize);
2828 void kfree(const void *x)
2831 void *object = (void *)x;
2833 trace_kfree(_RET_IP_, x);
2835 if (unlikely(ZERO_OR_NULL_PTR(x)))
2838 page = virt_to_head_page(x);
2839 if (unlikely(!PageSlab(page))) {
2840 BUG_ON(!PageCompound(page));
2844 slab_free(page->slab, page, object, _RET_IP_);
2846 EXPORT_SYMBOL(kfree);
2849 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2850 * the remaining slabs by the number of items in use. The slabs with the
2851 * most items in use come first. New allocations will then fill those up
2852 * and thus they can be removed from the partial lists.
2854 * The slabs with the least items are placed last. This results in them
2855 * being allocated from last increasing the chance that the last objects
2856 * are freed in them.
2858 int kmem_cache_shrink(struct kmem_cache *s)
2862 struct kmem_cache_node *n;
2865 int objects = oo_objects(s->max);
2866 struct list_head *slabs_by_inuse =
2867 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2868 unsigned long flags;
2870 if (!slabs_by_inuse)
2874 for_each_node_state(node, N_NORMAL_MEMORY) {
2875 n = get_node(s, node);
2880 for (i = 0; i < objects; i++)
2881 INIT_LIST_HEAD(slabs_by_inuse + i);
2883 spin_lock_irqsave(&n->list_lock, flags);
2886 * Build lists indexed by the items in use in each slab.
2888 * Note that concurrent frees may occur while we hold the
2889 * list_lock. page->inuse here is the upper limit.
2891 list_for_each_entry_safe(page, t, &n->partial, lru) {
2892 if (!page->inuse && slab_trylock(page)) {
2894 * Must hold slab lock here because slab_free
2895 * may have freed the last object and be
2896 * waiting to release the slab.
2898 list_del(&page->lru);
2901 discard_slab(s, page);
2903 list_move(&page->lru,
2904 slabs_by_inuse + page->inuse);
2909 * Rebuild the partial list with the slabs filled up most
2910 * first and the least used slabs at the end.
2912 for (i = objects - 1; i >= 0; i--)
2913 list_splice(slabs_by_inuse + i, n->partial.prev);
2915 spin_unlock_irqrestore(&n->list_lock, flags);
2918 kfree(slabs_by_inuse);
2921 EXPORT_SYMBOL(kmem_cache_shrink);
2923 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2924 static int slab_mem_going_offline_callback(void *arg)
2926 struct kmem_cache *s;
2928 down_read(&slub_lock);
2929 list_for_each_entry(s, &slab_caches, list)
2930 kmem_cache_shrink(s);
2931 up_read(&slub_lock);
2936 static void slab_mem_offline_callback(void *arg)
2938 struct kmem_cache_node *n;
2939 struct kmem_cache *s;
2940 struct memory_notify *marg = arg;
2943 offline_node = marg->status_change_nid;
2946 * If the node still has available memory. we need kmem_cache_node
2949 if (offline_node < 0)
2952 down_read(&slub_lock);
2953 list_for_each_entry(s, &slab_caches, list) {
2954 n = get_node(s, offline_node);
2957 * if n->nr_slabs > 0, slabs still exist on the node
2958 * that is going down. We were unable to free them,
2959 * and offline_pages() function shoudn't call this
2960 * callback. So, we must fail.
2962 BUG_ON(slabs_node(s, offline_node));
2964 s->node[offline_node] = NULL;
2965 kmem_cache_free(kmalloc_caches, n);
2968 up_read(&slub_lock);
2971 static int slab_mem_going_online_callback(void *arg)
2973 struct kmem_cache_node *n;
2974 struct kmem_cache *s;
2975 struct memory_notify *marg = arg;
2976 int nid = marg->status_change_nid;
2980 * If the node's memory is already available, then kmem_cache_node is
2981 * already created. Nothing to do.
2987 * We are bringing a node online. No memory is available yet. We must
2988 * allocate a kmem_cache_node structure in order to bring the node
2991 down_read(&slub_lock);
2992 list_for_each_entry(s, &slab_caches, list) {
2994 * XXX: kmem_cache_alloc_node will fallback to other nodes
2995 * since memory is not yet available from the node that
2998 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
3003 init_kmem_cache_node(n, s);
3007 up_read(&slub_lock);
3011 static int slab_memory_callback(struct notifier_block *self,
3012 unsigned long action, void *arg)
3017 case MEM_GOING_ONLINE:
3018 ret = slab_mem_going_online_callback(arg);
3020 case MEM_GOING_OFFLINE:
3021 ret = slab_mem_going_offline_callback(arg);
3024 case MEM_CANCEL_ONLINE:
3025 slab_mem_offline_callback(arg);
3028 case MEM_CANCEL_OFFLINE:
3032 ret = notifier_from_errno(ret);
3038 #endif /* CONFIG_MEMORY_HOTPLUG */
3040 /********************************************************************
3041 * Basic setup of slabs
3042 *******************************************************************/
3044 void __init kmem_cache_init(void)
3053 * Must first have the slab cache available for the allocations of the
3054 * struct kmem_cache_node's. There is special bootstrap code in
3055 * kmem_cache_open for slab_state == DOWN.
3057 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
3058 sizeof(struct kmem_cache_node), GFP_NOWAIT);
3059 kmalloc_caches[0].refcount = -1;
3062 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3065 /* Able to allocate the per node structures */
3066 slab_state = PARTIAL;
3068 /* Caches that are not of the two-to-the-power-of size */
3069 if (KMALLOC_MIN_SIZE <= 64) {
3070 create_kmalloc_cache(&kmalloc_caches[1],
3071 "kmalloc-96", 96, GFP_NOWAIT);
3073 create_kmalloc_cache(&kmalloc_caches[2],
3074 "kmalloc-192", 192, GFP_NOWAIT);
3078 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3079 create_kmalloc_cache(&kmalloc_caches[i],
3080 "kmalloc", 1 << i, GFP_NOWAIT);
3086 * Patch up the size_index table if we have strange large alignment
3087 * requirements for the kmalloc array. This is only the case for
3088 * MIPS it seems. The standard arches will not generate any code here.
3090 * Largest permitted alignment is 256 bytes due to the way we
3091 * handle the index determination for the smaller caches.
3093 * Make sure that nothing crazy happens if someone starts tinkering
3094 * around with ARCH_KMALLOC_MINALIGN
3096 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3097 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3099 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3100 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3102 if (KMALLOC_MIN_SIZE == 128) {
3104 * The 192 byte sized cache is not used if the alignment
3105 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3108 for (i = 128 + 8; i <= 192; i += 8)
3109 size_index[(i - 1) / 8] = 8;
3114 /* Provide the correct kmalloc names now that the caches are up */
3115 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++)
3116 kmalloc_caches[i]. name =
3117 kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3120 register_cpu_notifier(&slab_notifier);
3121 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3122 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3124 kmem_size = sizeof(struct kmem_cache);
3128 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3129 " CPUs=%d, Nodes=%d\n",
3130 caches, cache_line_size(),
3131 slub_min_order, slub_max_order, slub_min_objects,
3132 nr_cpu_ids, nr_node_ids);
3136 * Find a mergeable slab cache
3138 static int slab_unmergeable(struct kmem_cache *s)
3140 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3147 * We may have set a slab to be unmergeable during bootstrap.
3149 if (s->refcount < 0)
3155 static struct kmem_cache *find_mergeable(size_t size,
3156 size_t align, unsigned long flags, const char *name,
3157 void (*ctor)(void *))
3159 struct kmem_cache *s;
3161 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3167 size = ALIGN(size, sizeof(void *));
3168 align = calculate_alignment(flags, align, size);
3169 size = ALIGN(size, align);
3170 flags = kmem_cache_flags(size, flags, name, NULL);
3172 list_for_each_entry(s, &slab_caches, list) {
3173 if (slab_unmergeable(s))
3179 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3182 * Check if alignment is compatible.
3183 * Courtesy of Adrian Drzewiecki
3185 if ((s->size & ~(align - 1)) != s->size)
3188 if (s->size - size >= sizeof(void *))
3196 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3197 size_t align, unsigned long flags, void (*ctor)(void *))
3199 struct kmem_cache *s;
3201 down_write(&slub_lock);
3202 s = find_mergeable(size, align, flags, name, ctor);
3208 * Adjust the object sizes so that we clear
3209 * the complete object on kzalloc.
3211 s->objsize = max(s->objsize, (int)size);
3214 * And then we need to update the object size in the
3215 * per cpu structures
3217 for_each_online_cpu(cpu)
3218 get_cpu_slab(s, cpu)->objsize = s->objsize;
3220 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3221 up_write(&slub_lock);
3223 if (sysfs_slab_alias(s, name)) {
3224 down_write(&slub_lock);
3226 up_write(&slub_lock);
3232 s = kmalloc(kmem_size, GFP_KERNEL);
3234 if (kmem_cache_open(s, GFP_KERNEL, name,
3235 size, align, flags, ctor)) {
3236 list_add(&s->list, &slab_caches);
3237 up_write(&slub_lock);
3238 if (sysfs_slab_add(s)) {
3239 down_write(&slub_lock);
3241 up_write(&slub_lock);
3249 up_write(&slub_lock);
3252 if (flags & SLAB_PANIC)
3253 panic("Cannot create slabcache %s\n", name);
3258 EXPORT_SYMBOL(kmem_cache_create);
3262 * Use the cpu notifier to insure that the cpu slabs are flushed when
3265 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3266 unsigned long action, void *hcpu)
3268 long cpu = (long)hcpu;
3269 struct kmem_cache *s;
3270 unsigned long flags;
3273 case CPU_UP_PREPARE:
3274 case CPU_UP_PREPARE_FROZEN:
3275 init_alloc_cpu_cpu(cpu);
3276 down_read(&slub_lock);
3277 list_for_each_entry(s, &slab_caches, list)
3278 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3280 up_read(&slub_lock);
3283 case CPU_UP_CANCELED:
3284 case CPU_UP_CANCELED_FROZEN:
3286 case CPU_DEAD_FROZEN:
3287 down_read(&slub_lock);
3288 list_for_each_entry(s, &slab_caches, list) {
3289 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3291 local_irq_save(flags);
3292 __flush_cpu_slab(s, cpu);
3293 local_irq_restore(flags);
3294 free_kmem_cache_cpu(c, cpu);
3295 s->cpu_slab[cpu] = NULL;
3297 up_read(&slub_lock);
3305 static struct notifier_block __cpuinitdata slab_notifier = {
3306 .notifier_call = slab_cpuup_callback
3311 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3313 struct kmem_cache *s;
3316 if (unlikely(size > SLUB_MAX_SIZE))
3317 return kmalloc_large(size, gfpflags);
3319 s = get_slab(size, gfpflags);
3321 if (unlikely(ZERO_OR_NULL_PTR(s)))
3324 ret = slab_alloc(s, gfpflags, -1, caller);
3326 /* Honor the call site pointer we recieved. */
3327 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3332 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3333 int node, unsigned long caller)
3335 struct kmem_cache *s;
3338 if (unlikely(size > SLUB_MAX_SIZE))
3339 return kmalloc_large_node(size, gfpflags, node);
3341 s = get_slab(size, gfpflags);
3343 if (unlikely(ZERO_OR_NULL_PTR(s)))
3346 ret = slab_alloc(s, gfpflags, node, caller);
3348 /* Honor the call site pointer we recieved. */
3349 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3354 #ifdef CONFIG_SLUB_DEBUG
3355 static unsigned long count_partial(struct kmem_cache_node *n,
3356 int (*get_count)(struct page *))
3358 unsigned long flags;
3359 unsigned long x = 0;
3362 spin_lock_irqsave(&n->list_lock, flags);
3363 list_for_each_entry(page, &n->partial, lru)
3364 x += get_count(page);
3365 spin_unlock_irqrestore(&n->list_lock, flags);
3369 static int count_inuse(struct page *page)
3374 static int count_total(struct page *page)
3376 return page->objects;
3379 static int count_free(struct page *page)
3381 return page->objects - page->inuse;
3384 static int validate_slab(struct kmem_cache *s, struct page *page,
3388 void *addr = page_address(page);
3390 if (!check_slab(s, page) ||
3391 !on_freelist(s, page, NULL))
3394 /* Now we know that a valid freelist exists */
3395 bitmap_zero(map, page->objects);
3397 for_each_free_object(p, s, page->freelist) {
3398 set_bit(slab_index(p, s, addr), map);
3399 if (!check_object(s, page, p, 0))
3403 for_each_object(p, s, addr, page->objects)
3404 if (!test_bit(slab_index(p, s, addr), map))
3405 if (!check_object(s, page, p, 1))
3410 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3413 if (slab_trylock(page)) {
3414 validate_slab(s, page, map);
3417 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3420 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3421 if (!PageSlubDebug(page))
3422 printk(KERN_ERR "SLUB %s: SlubDebug not set "
3423 "on slab 0x%p\n", s->name, page);
3425 if (PageSlubDebug(page))
3426 printk(KERN_ERR "SLUB %s: SlubDebug set on "
3427 "slab 0x%p\n", s->name, page);
3431 static int validate_slab_node(struct kmem_cache *s,
3432 struct kmem_cache_node *n, unsigned long *map)
3434 unsigned long count = 0;
3436 unsigned long flags;
3438 spin_lock_irqsave(&n->list_lock, flags);
3440 list_for_each_entry(page, &n->partial, lru) {
3441 validate_slab_slab(s, page, map);
3444 if (count != n->nr_partial)
3445 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3446 "counter=%ld\n", s->name, count, n->nr_partial);
3448 if (!(s->flags & SLAB_STORE_USER))
3451 list_for_each_entry(page, &n->full, lru) {
3452 validate_slab_slab(s, page, map);
3455 if (count != atomic_long_read(&n->nr_slabs))
3456 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3457 "counter=%ld\n", s->name, count,
3458 atomic_long_read(&n->nr_slabs));
3461 spin_unlock_irqrestore(&n->list_lock, flags);
3465 static long validate_slab_cache(struct kmem_cache *s)
3468 unsigned long count = 0;
3469 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3470 sizeof(unsigned long), GFP_KERNEL);
3476 for_each_node_state(node, N_NORMAL_MEMORY) {
3477 struct kmem_cache_node *n = get_node(s, node);
3479 count += validate_slab_node(s, n, map);
3485 #ifdef SLUB_RESILIENCY_TEST
3486 static void resiliency_test(void)
3490 printk(KERN_ERR "SLUB resiliency testing\n");
3491 printk(KERN_ERR "-----------------------\n");
3492 printk(KERN_ERR "A. Corruption after allocation\n");
3494 p = kzalloc(16, GFP_KERNEL);
3496 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3497 " 0x12->0x%p\n\n", p + 16);
3499 validate_slab_cache(kmalloc_caches + 4);
3501 /* Hmmm... The next two are dangerous */
3502 p = kzalloc(32, GFP_KERNEL);
3503 p[32 + sizeof(void *)] = 0x34;
3504 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3505 " 0x34 -> -0x%p\n", p);
3507 "If allocated object is overwritten then not detectable\n\n");
3509 validate_slab_cache(kmalloc_caches + 5);
3510 p = kzalloc(64, GFP_KERNEL);
3511 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3513 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3516 "If allocated object is overwritten then not detectable\n\n");
3517 validate_slab_cache(kmalloc_caches + 6);
3519 printk(KERN_ERR "\nB. Corruption after free\n");
3520 p = kzalloc(128, GFP_KERNEL);
3523 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3524 validate_slab_cache(kmalloc_caches + 7);
3526 p = kzalloc(256, GFP_KERNEL);
3529 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3531 validate_slab_cache(kmalloc_caches + 8);
3533 p = kzalloc(512, GFP_KERNEL);
3536 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3537 validate_slab_cache(kmalloc_caches + 9);
3540 static void resiliency_test(void) {};
3544 * Generate lists of code addresses where slabcache objects are allocated
3549 unsigned long count;
3556 DECLARE_BITMAP(cpus, NR_CPUS);
3562 unsigned long count;
3563 struct location *loc;
3566 static void free_loc_track(struct loc_track *t)
3569 free_pages((unsigned long)t->loc,
3570 get_order(sizeof(struct location) * t->max));
3573 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3578 order = get_order(sizeof(struct location) * max);
3580 l = (void *)__get_free_pages(flags, order);
3585 memcpy(l, t->loc, sizeof(struct location) * t->count);
3593 static int add_location(struct loc_track *t, struct kmem_cache *s,
3594 const struct track *track)
3596 long start, end, pos;
3598 unsigned long caddr;
3599 unsigned long age = jiffies - track->when;
3605 pos = start + (end - start + 1) / 2;
3608 * There is nothing at "end". If we end up there
3609 * we need to add something to before end.
3614 caddr = t->loc[pos].addr;
3615 if (track->addr == caddr) {
3621 if (age < l->min_time)
3623 if (age > l->max_time)
3626 if (track->pid < l->min_pid)
3627 l->min_pid = track->pid;
3628 if (track->pid > l->max_pid)
3629 l->max_pid = track->pid;
3631 cpumask_set_cpu(track->cpu,
3632 to_cpumask(l->cpus));
3634 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3638 if (track->addr < caddr)
3645 * Not found. Insert new tracking element.
3647 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3653 (t->count - pos) * sizeof(struct location));
3656 l->addr = track->addr;
3660 l->min_pid = track->pid;
3661 l->max_pid = track->pid;
3662 cpumask_clear(to_cpumask(l->cpus));
3663 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3664 nodes_clear(l->nodes);
3665 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3669 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3670 struct page *page, enum track_item alloc)
3672 void *addr = page_address(page);
3673 DECLARE_BITMAP(map, page->objects);
3676 bitmap_zero(map, page->objects);
3677 for_each_free_object(p, s, page->freelist)
3678 set_bit(slab_index(p, s, addr), map);
3680 for_each_object(p, s, addr, page->objects)
3681 if (!test_bit(slab_index(p, s, addr), map))
3682 add_location(t, s, get_track(s, p, alloc));
3685 static int list_locations(struct kmem_cache *s, char *buf,
3686 enum track_item alloc)
3690 struct loc_track t = { 0, 0, NULL };
3693 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3695 return sprintf(buf, "Out of memory\n");
3697 /* Push back cpu slabs */
3700 for_each_node_state(node, N_NORMAL_MEMORY) {
3701 struct kmem_cache_node *n = get_node(s, node);
3702 unsigned long flags;
3705 if (!atomic_long_read(&n->nr_slabs))
3708 spin_lock_irqsave(&n->list_lock, flags);
3709 list_for_each_entry(page, &n->partial, lru)
3710 process_slab(&t, s, page, alloc);
3711 list_for_each_entry(page, &n->full, lru)
3712 process_slab(&t, s, page, alloc);
3713 spin_unlock_irqrestore(&n->list_lock, flags);
3716 for (i = 0; i < t.count; i++) {
3717 struct location *l = &t.loc[i];
3719 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3721 len += sprintf(buf + len, "%7ld ", l->count);
3724 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3726 len += sprintf(buf + len, "<not-available>");
3728 if (l->sum_time != l->min_time) {
3729 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3731 (long)div_u64(l->sum_time, l->count),
3734 len += sprintf(buf + len, " age=%ld",
3737 if (l->min_pid != l->max_pid)
3738 len += sprintf(buf + len, " pid=%ld-%ld",
3739 l->min_pid, l->max_pid);
3741 len += sprintf(buf + len, " pid=%ld",
3744 if (num_online_cpus() > 1 &&
3745 !cpumask_empty(to_cpumask(l->cpus)) &&
3746 len < PAGE_SIZE - 60) {
3747 len += sprintf(buf + len, " cpus=");
3748 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3749 to_cpumask(l->cpus));
3752 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3753 len < PAGE_SIZE - 60) {
3754 len += sprintf(buf + len, " nodes=");
3755 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3759 len += sprintf(buf + len, "\n");
3764 len += sprintf(buf, "No data\n");
3768 enum slab_stat_type {
3769 SL_ALL, /* All slabs */
3770 SL_PARTIAL, /* Only partially allocated slabs */
3771 SL_CPU, /* Only slabs used for cpu caches */
3772 SL_OBJECTS, /* Determine allocated objects not slabs */
3773 SL_TOTAL /* Determine object capacity not slabs */
3776 #define SO_ALL (1 << SL_ALL)
3777 #define SO_PARTIAL (1 << SL_PARTIAL)
3778 #define SO_CPU (1 << SL_CPU)
3779 #define SO_OBJECTS (1 << SL_OBJECTS)
3780 #define SO_TOTAL (1 << SL_TOTAL)
3782 static ssize_t show_slab_objects(struct kmem_cache *s,
3783 char *buf, unsigned long flags)
3785 unsigned long total = 0;
3788 unsigned long *nodes;
3789 unsigned long *per_cpu;
3791 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3794 per_cpu = nodes + nr_node_ids;
3796 if (flags & SO_CPU) {
3799 for_each_possible_cpu(cpu) {
3800 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3802 if (!c || c->node < 0)
3806 if (flags & SO_TOTAL)
3807 x = c->page->objects;
3808 else if (flags & SO_OBJECTS)
3814 nodes[c->node] += x;
3820 if (flags & SO_ALL) {
3821 for_each_node_state(node, N_NORMAL_MEMORY) {
3822 struct kmem_cache_node *n = get_node(s, node);
3824 if (flags & SO_TOTAL)
3825 x = atomic_long_read(&n->total_objects);
3826 else if (flags & SO_OBJECTS)
3827 x = atomic_long_read(&n->total_objects) -
3828 count_partial(n, count_free);
3831 x = atomic_long_read(&n->nr_slabs);
3836 } else if (flags & SO_PARTIAL) {
3837 for_each_node_state(node, N_NORMAL_MEMORY) {
3838 struct kmem_cache_node *n = get_node(s, node);
3840 if (flags & SO_TOTAL)
3841 x = count_partial(n, count_total);
3842 else if (flags & SO_OBJECTS)
3843 x = count_partial(n, count_inuse);
3850 x = sprintf(buf, "%lu", total);
3852 for_each_node_state(node, N_NORMAL_MEMORY)
3854 x += sprintf(buf + x, " N%d=%lu",
3858 return x + sprintf(buf + x, "\n");
3861 static int any_slab_objects(struct kmem_cache *s)
3865 for_each_online_node(node) {
3866 struct kmem_cache_node *n = get_node(s, node);
3871 if (atomic_long_read(&n->total_objects))
3877 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3878 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3880 struct slab_attribute {
3881 struct attribute attr;
3882 ssize_t (*show)(struct kmem_cache *s, char *buf);
3883 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3886 #define SLAB_ATTR_RO(_name) \
3887 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3889 #define SLAB_ATTR(_name) \
3890 static struct slab_attribute _name##_attr = \
3891 __ATTR(_name, 0644, _name##_show, _name##_store)
3893 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3895 return sprintf(buf, "%d\n", s->size);
3897 SLAB_ATTR_RO(slab_size);
3899 static ssize_t align_show(struct kmem_cache *s, char *buf)
3901 return sprintf(buf, "%d\n", s->align);
3903 SLAB_ATTR_RO(align);
3905 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3907 return sprintf(buf, "%d\n", s->objsize);
3909 SLAB_ATTR_RO(object_size);
3911 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3913 return sprintf(buf, "%d\n", oo_objects(s->oo));
3915 SLAB_ATTR_RO(objs_per_slab);
3917 static ssize_t order_store(struct kmem_cache *s,
3918 const char *buf, size_t length)
3920 unsigned long order;
3923 err = strict_strtoul(buf, 10, &order);
3927 if (order > slub_max_order || order < slub_min_order)
3930 calculate_sizes(s, order);
3934 static ssize_t order_show(struct kmem_cache *s, char *buf)
3936 return sprintf(buf, "%d\n", oo_order(s->oo));
3940 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
3942 return sprintf(buf, "%lu\n", s->min_partial);
3945 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
3951 err = strict_strtoul(buf, 10, &min);
3955 set_min_partial(s, min);
3958 SLAB_ATTR(min_partial);
3960 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3963 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3965 return n + sprintf(buf + n, "\n");
3971 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3973 return sprintf(buf, "%d\n", s->refcount - 1);
3975 SLAB_ATTR_RO(aliases);
3977 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3979 return show_slab_objects(s, buf, SO_ALL);
3981 SLAB_ATTR_RO(slabs);
3983 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3985 return show_slab_objects(s, buf, SO_PARTIAL);
3987 SLAB_ATTR_RO(partial);
3989 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3991 return show_slab_objects(s, buf, SO_CPU);
3993 SLAB_ATTR_RO(cpu_slabs);
3995 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3997 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3999 SLAB_ATTR_RO(objects);
4001 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4003 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4005 SLAB_ATTR_RO(objects_partial);
4007 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4009 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4011 SLAB_ATTR_RO(total_objects);
4013 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4015 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4018 static ssize_t sanity_checks_store(struct kmem_cache *s,
4019 const char *buf, size_t length)
4021 s->flags &= ~SLAB_DEBUG_FREE;
4023 s->flags |= SLAB_DEBUG_FREE;
4026 SLAB_ATTR(sanity_checks);
4028 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4030 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4033 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4036 s->flags &= ~SLAB_TRACE;
4038 s->flags |= SLAB_TRACE;
4043 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4045 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4048 static ssize_t reclaim_account_store(struct kmem_cache *s,
4049 const char *buf, size_t length)
4051 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4053 s->flags |= SLAB_RECLAIM_ACCOUNT;
4056 SLAB_ATTR(reclaim_account);
4058 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4060 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4062 SLAB_ATTR_RO(hwcache_align);
4064 #ifdef CONFIG_ZONE_DMA
4065 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4067 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4069 SLAB_ATTR_RO(cache_dma);
4072 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4074 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4076 SLAB_ATTR_RO(destroy_by_rcu);
4078 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4080 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4083 static ssize_t red_zone_store(struct kmem_cache *s,
4084 const char *buf, size_t length)
4086 if (any_slab_objects(s))
4089 s->flags &= ~SLAB_RED_ZONE;
4091 s->flags |= SLAB_RED_ZONE;
4092 calculate_sizes(s, -1);
4095 SLAB_ATTR(red_zone);
4097 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4099 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4102 static ssize_t poison_store(struct kmem_cache *s,
4103 const char *buf, size_t length)
4105 if (any_slab_objects(s))
4108 s->flags &= ~SLAB_POISON;
4110 s->flags |= SLAB_POISON;
4111 calculate_sizes(s, -1);
4116 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4118 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4121 static ssize_t store_user_store(struct kmem_cache *s,
4122 const char *buf, size_t length)
4124 if (any_slab_objects(s))
4127 s->flags &= ~SLAB_STORE_USER;
4129 s->flags |= SLAB_STORE_USER;
4130 calculate_sizes(s, -1);
4133 SLAB_ATTR(store_user);
4135 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4140 static ssize_t validate_store(struct kmem_cache *s,
4141 const char *buf, size_t length)
4145 if (buf[0] == '1') {
4146 ret = validate_slab_cache(s);
4152 SLAB_ATTR(validate);
4154 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4159 static ssize_t shrink_store(struct kmem_cache *s,
4160 const char *buf, size_t length)
4162 if (buf[0] == '1') {
4163 int rc = kmem_cache_shrink(s);
4173 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4175 if (!(s->flags & SLAB_STORE_USER))
4177 return list_locations(s, buf, TRACK_ALLOC);
4179 SLAB_ATTR_RO(alloc_calls);
4181 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4183 if (!(s->flags & SLAB_STORE_USER))
4185 return list_locations(s, buf, TRACK_FREE);
4187 SLAB_ATTR_RO(free_calls);
4190 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4192 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4195 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4196 const char *buf, size_t length)
4198 unsigned long ratio;
4201 err = strict_strtoul(buf, 10, &ratio);
4206 s->remote_node_defrag_ratio = ratio * 10;
4210 SLAB_ATTR(remote_node_defrag_ratio);
4213 #ifdef CONFIG_SLUB_STATS
4214 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4216 unsigned long sum = 0;
4219 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4224 for_each_online_cpu(cpu) {
4225 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4231 len = sprintf(buf, "%lu", sum);
4234 for_each_online_cpu(cpu) {
4235 if (data[cpu] && len < PAGE_SIZE - 20)
4236 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4240 return len + sprintf(buf + len, "\n");
4243 #define STAT_ATTR(si, text) \
4244 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4246 return show_stat(s, buf, si); \
4248 SLAB_ATTR_RO(text); \
4250 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4251 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4252 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4253 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4254 STAT_ATTR(FREE_FROZEN, free_frozen);
4255 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4256 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4257 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4258 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4259 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4260 STAT_ATTR(FREE_SLAB, free_slab);
4261 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4262 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4263 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4264 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4265 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4266 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4267 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4270 static struct attribute *slab_attrs[] = {
4271 &slab_size_attr.attr,
4272 &object_size_attr.attr,
4273 &objs_per_slab_attr.attr,
4275 &min_partial_attr.attr,
4277 &objects_partial_attr.attr,
4278 &total_objects_attr.attr,
4281 &cpu_slabs_attr.attr,
4285 &sanity_checks_attr.attr,
4287 &hwcache_align_attr.attr,
4288 &reclaim_account_attr.attr,
4289 &destroy_by_rcu_attr.attr,
4290 &red_zone_attr.attr,
4292 &store_user_attr.attr,
4293 &validate_attr.attr,
4295 &alloc_calls_attr.attr,
4296 &free_calls_attr.attr,
4297 #ifdef CONFIG_ZONE_DMA
4298 &cache_dma_attr.attr,
4301 &remote_node_defrag_ratio_attr.attr,
4303 #ifdef CONFIG_SLUB_STATS
4304 &alloc_fastpath_attr.attr,
4305 &alloc_slowpath_attr.attr,
4306 &free_fastpath_attr.attr,
4307 &free_slowpath_attr.attr,
4308 &free_frozen_attr.attr,
4309 &free_add_partial_attr.attr,
4310 &free_remove_partial_attr.attr,
4311 &alloc_from_partial_attr.attr,
4312 &alloc_slab_attr.attr,
4313 &alloc_refill_attr.attr,
4314 &free_slab_attr.attr,
4315 &cpuslab_flush_attr.attr,
4316 &deactivate_full_attr.attr,
4317 &deactivate_empty_attr.attr,
4318 &deactivate_to_head_attr.attr,
4319 &deactivate_to_tail_attr.attr,
4320 &deactivate_remote_frees_attr.attr,
4321 &order_fallback_attr.attr,
4326 static struct attribute_group slab_attr_group = {
4327 .attrs = slab_attrs,
4330 static ssize_t slab_attr_show(struct kobject *kobj,
4331 struct attribute *attr,
4334 struct slab_attribute *attribute;
4335 struct kmem_cache *s;
4338 attribute = to_slab_attr(attr);
4341 if (!attribute->show)
4344 err = attribute->show(s, buf);
4349 static ssize_t slab_attr_store(struct kobject *kobj,
4350 struct attribute *attr,
4351 const char *buf, size_t len)
4353 struct slab_attribute *attribute;
4354 struct kmem_cache *s;
4357 attribute = to_slab_attr(attr);
4360 if (!attribute->store)
4363 err = attribute->store(s, buf, len);
4368 static void kmem_cache_release(struct kobject *kobj)
4370 struct kmem_cache *s = to_slab(kobj);
4375 static struct sysfs_ops slab_sysfs_ops = {
4376 .show = slab_attr_show,
4377 .store = slab_attr_store,
4380 static struct kobj_type slab_ktype = {
4381 .sysfs_ops = &slab_sysfs_ops,
4382 .release = kmem_cache_release
4385 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4387 struct kobj_type *ktype = get_ktype(kobj);
4389 if (ktype == &slab_ktype)
4394 static struct kset_uevent_ops slab_uevent_ops = {
4395 .filter = uevent_filter,
4398 static struct kset *slab_kset;
4400 #define ID_STR_LENGTH 64
4402 /* Create a unique string id for a slab cache:
4404 * Format :[flags-]size
4406 static char *create_unique_id(struct kmem_cache *s)
4408 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4415 * First flags affecting slabcache operations. We will only
4416 * get here for aliasable slabs so we do not need to support
4417 * too many flags. The flags here must cover all flags that
4418 * are matched during merging to guarantee that the id is
4421 if (s->flags & SLAB_CACHE_DMA)
4423 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4425 if (s->flags & SLAB_DEBUG_FREE)
4427 if (!(s->flags & SLAB_NOTRACK))
4431 p += sprintf(p, "%07d", s->size);
4432 BUG_ON(p > name + ID_STR_LENGTH - 1);
4436 static int sysfs_slab_add(struct kmem_cache *s)
4442 if (slab_state < SYSFS)
4443 /* Defer until later */
4446 unmergeable = slab_unmergeable(s);
4449 * Slabcache can never be merged so we can use the name proper.
4450 * This is typically the case for debug situations. In that
4451 * case we can catch duplicate names easily.
4453 sysfs_remove_link(&slab_kset->kobj, s->name);
4457 * Create a unique name for the slab as a target
4460 name = create_unique_id(s);
4463 s->kobj.kset = slab_kset;
4464 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4466 kobject_put(&s->kobj);
4470 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4473 kobject_uevent(&s->kobj, KOBJ_ADD);
4475 /* Setup first alias */
4476 sysfs_slab_alias(s, s->name);
4482 static void sysfs_slab_remove(struct kmem_cache *s)
4484 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4485 kobject_del(&s->kobj);
4486 kobject_put(&s->kobj);
4490 * Need to buffer aliases during bootup until sysfs becomes
4491 * available lest we lose that information.
4493 struct saved_alias {
4494 struct kmem_cache *s;
4496 struct saved_alias *next;
4499 static struct saved_alias *alias_list;
4501 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4503 struct saved_alias *al;
4505 if (slab_state == SYSFS) {
4507 * If we have a leftover link then remove it.
4509 sysfs_remove_link(&slab_kset->kobj, name);
4510 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4513 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4519 al->next = alias_list;
4524 static int __init slab_sysfs_init(void)
4526 struct kmem_cache *s;
4529 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4531 printk(KERN_ERR "Cannot register slab subsystem.\n");
4537 list_for_each_entry(s, &slab_caches, list) {
4538 err = sysfs_slab_add(s);
4540 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4541 " to sysfs\n", s->name);
4544 while (alias_list) {
4545 struct saved_alias *al = alias_list;
4547 alias_list = alias_list->next;
4548 err = sysfs_slab_alias(al->s, al->name);
4550 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4551 " %s to sysfs\n", s->name);
4559 __initcall(slab_sysfs_init);
4563 * The /proc/slabinfo ABI
4565 #ifdef CONFIG_SLABINFO
4566 static void print_slabinfo_header(struct seq_file *m)
4568 seq_puts(m, "slabinfo - version: 2.1\n");
4569 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4570 "<objperslab> <pagesperslab>");
4571 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4572 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4576 static void *s_start(struct seq_file *m, loff_t *pos)
4580 down_read(&slub_lock);
4582 print_slabinfo_header(m);
4584 return seq_list_start(&slab_caches, *pos);
4587 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4589 return seq_list_next(p, &slab_caches, pos);
4592 static void s_stop(struct seq_file *m, void *p)
4594 up_read(&slub_lock);
4597 static int s_show(struct seq_file *m, void *p)
4599 unsigned long nr_partials = 0;
4600 unsigned long nr_slabs = 0;
4601 unsigned long nr_inuse = 0;
4602 unsigned long nr_objs = 0;
4603 unsigned long nr_free = 0;
4604 struct kmem_cache *s;
4607 s = list_entry(p, struct kmem_cache, list);
4609 for_each_online_node(node) {
4610 struct kmem_cache_node *n = get_node(s, node);
4615 nr_partials += n->nr_partial;
4616 nr_slabs += atomic_long_read(&n->nr_slabs);
4617 nr_objs += atomic_long_read(&n->total_objects);
4618 nr_free += count_partial(n, count_free);
4621 nr_inuse = nr_objs - nr_free;
4623 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4624 nr_objs, s->size, oo_objects(s->oo),
4625 (1 << oo_order(s->oo)));
4626 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4627 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4633 static const struct seq_operations slabinfo_op = {
4640 static int slabinfo_open(struct inode *inode, struct file *file)
4642 return seq_open(file, &slabinfo_op);
4645 static const struct file_operations proc_slabinfo_operations = {
4646 .open = slabinfo_open,
4648 .llseek = seq_lseek,
4649 .release = seq_release,
4652 static int __init slab_proc_init(void)
4654 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4657 module_init(slab_proc_init);
4658 #endif /* CONFIG_SLABINFO */