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 <clameter@sgi.com>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
69 * Slabs with free elements are kept on a partial list and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
79 * Overloading of page flags that are otherwise used for LRU management.
81 * PageActive The slab is frozen and exempt from list processing.
82 * This means that the slab is dedicated to a purpose
83 * such as satisfying allocations for a specific
84 * processor. Objects may be freed in the slab while
85 * it is frozen but slab_free will then skip the usual
86 * list operations. It is up to the processor holding
87 * the slab to integrate the slab into the slab lists
88 * when the slab is no longer needed.
90 * One use of this flag is to mark slabs that are
91 * used for allocations. Then such a slab becomes a cpu
92 * slab. The cpu slab may be equipped with an additional
93 * lockless_freelist that allows lockless access to
94 * free objects in addition to the regular freelist
95 * that requires the slab lock.
97 * PageError Slab requires special handling due to debug
98 * options set. This moves slab handling out of
99 * the fast path and disables lockless freelists.
102 #define FROZEN (1 << PG_active)
104 #ifdef CONFIG_SLUB_DEBUG
105 #define SLABDEBUG (1 << PG_error)
110 static inline int SlabFrozen(struct page *page)
112 return page->flags & FROZEN;
115 static inline void SetSlabFrozen(struct page *page)
117 page->flags |= FROZEN;
120 static inline void ClearSlabFrozen(struct page *page)
122 page->flags &= ~FROZEN;
125 static inline int SlabDebug(struct page *page)
127 return page->flags & SLABDEBUG;
130 static inline void SetSlabDebug(struct page *page)
132 page->flags |= SLABDEBUG;
135 static inline void ClearSlabDebug(struct page *page)
137 page->flags &= ~SLABDEBUG;
141 * Issues still to be resolved:
143 * - The per cpu array is updated for each new slab and and is a remote
144 * cacheline for most nodes. This could become a bouncing cacheline given
145 * enough frequent updates. There are 16 pointers in a cacheline, so at
146 * max 16 cpus could compete for the cacheline which may be okay.
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
159 * Small page size. Make sure that we do not fragment memory
161 #define DEFAULT_MAX_ORDER 1
162 #define DEFAULT_MIN_OBJECTS 4
167 * Large page machines are customarily able to handle larger
170 #define DEFAULT_MAX_ORDER 2
171 #define DEFAULT_MIN_OBJECTS 8
176 * Mininum number of partial slabs. These will be left on the partial
177 * lists even if they are empty. kmem_cache_shrink may reclaim them.
179 #define MIN_PARTIAL 2
182 * Maximum number of desirable partial slabs.
183 * The existence of more partial slabs makes kmem_cache_shrink
184 * sort the partial list by the number of objects in the.
186 #define MAX_PARTIAL 10
188 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
189 SLAB_POISON | SLAB_STORE_USER)
192 * Set of flags that will prevent slab merging
194 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
195 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
197 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
200 #ifndef ARCH_KMALLOC_MINALIGN
201 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
204 #ifndef ARCH_SLAB_MINALIGN
205 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
209 * The page->inuse field is 16 bit thus we have this limitation
211 #define MAX_OBJECTS_PER_SLAB 65535
213 /* Internal SLUB flags */
214 #define __OBJECT_POISON 0x80000000 /* Poison object */
215 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
217 /* Not all arches define cache_line_size */
218 #ifndef cache_line_size
219 #define cache_line_size() L1_CACHE_BYTES
222 static int kmem_size = sizeof(struct kmem_cache);
225 static struct notifier_block slab_notifier;
229 DOWN, /* No slab functionality available */
230 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
231 UP, /* Everything works but does not show up in sysfs */
235 /* A list of all slab caches on the system */
236 static DECLARE_RWSEM(slub_lock);
237 static LIST_HEAD(slab_caches);
240 * Tracking user of a slab.
243 void *addr; /* Called from address */
244 int cpu; /* Was running on cpu */
245 int pid; /* Pid context */
246 unsigned long when; /* When did the operation occur */
249 enum track_item { TRACK_ALLOC, TRACK_FREE };
251 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
252 static int sysfs_slab_add(struct kmem_cache *);
253 static int sysfs_slab_alias(struct kmem_cache *, const char *);
254 static void sysfs_slab_remove(struct kmem_cache *);
256 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
257 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
259 static inline void sysfs_slab_remove(struct kmem_cache *s) {}
262 /********************************************************************
263 * Core slab cache functions
264 *******************************************************************/
266 int slab_is_available(void)
268 return slab_state >= UP;
271 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
274 return s->node[node];
276 return &s->local_node;
280 static inline int check_valid_pointer(struct kmem_cache *s,
281 struct page *page, const void *object)
288 base = page_address(page);
289 if (object < base || object >= base + s->objects * s->size ||
290 (object - base) % s->size) {
298 * Slow version of get and set free pointer.
300 * This version requires touching the cache lines of kmem_cache which
301 * we avoid to do in the fast alloc free paths. There we obtain the offset
302 * from the page struct.
304 static inline void *get_freepointer(struct kmem_cache *s, void *object)
306 return *(void **)(object + s->offset);
309 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
311 *(void **)(object + s->offset) = fp;
314 /* Loop over all objects in a slab */
315 #define for_each_object(__p, __s, __addr) \
316 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
320 #define for_each_free_object(__p, __s, __free) \
321 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
323 /* Determine object index from a given position */
324 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
326 return (p - addr) / s->size;
329 #ifdef CONFIG_SLUB_DEBUG
333 #ifdef CONFIG_SLUB_DEBUG_ON
334 static int slub_debug = DEBUG_DEFAULT_FLAGS;
336 static int slub_debug;
339 static char *slub_debug_slabs;
344 static void print_section(char *text, u8 *addr, unsigned int length)
352 for (i = 0; i < length; i++) {
354 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
357 printk(" %02x", addr[i]);
359 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
361 printk(" %s\n",ascii);
372 printk(" %s\n", ascii);
376 static struct track *get_track(struct kmem_cache *s, void *object,
377 enum track_item alloc)
382 p = object + s->offset + sizeof(void *);
384 p = object + s->inuse;
389 static void set_track(struct kmem_cache *s, void *object,
390 enum track_item alloc, void *addr)
395 p = object + s->offset + sizeof(void *);
397 p = object + s->inuse;
402 p->cpu = smp_processor_id();
403 p->pid = current ? current->pid : -1;
406 memset(p, 0, sizeof(struct track));
409 static void init_tracking(struct kmem_cache *s, void *object)
411 if (!(s->flags & SLAB_STORE_USER))
414 set_track(s, object, TRACK_FREE, NULL);
415 set_track(s, object, TRACK_ALLOC, NULL);
418 static void print_track(const char *s, struct track *t)
423 printk(KERN_ERR "INFO: %s in ", s);
424 __print_symbol("%s", (unsigned long)t->addr);
425 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
428 static void print_tracking(struct kmem_cache *s, void *object)
430 if (!(s->flags & SLAB_STORE_USER))
433 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
434 print_track("Freed", get_track(s, object, TRACK_FREE));
437 static void print_page_info(struct page *page)
439 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
440 page, page->inuse, page->freelist, page->flags);
444 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
450 vsnprintf(buf, sizeof(buf), fmt, args);
452 printk(KERN_ERR "========================================"
453 "=====================================\n");
454 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
455 printk(KERN_ERR "----------------------------------------"
456 "-------------------------------------\n\n");
459 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
465 vsnprintf(buf, sizeof(buf), fmt, args);
467 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
470 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
472 unsigned int off; /* Offset of last byte */
473 u8 *addr = page_address(page);
475 print_tracking(s, p);
477 print_page_info(page);
479 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
480 p, p - addr, get_freepointer(s, p));
483 print_section("Bytes b4", p - 16, 16);
485 print_section("Object", p, min(s->objsize, 128));
487 if (s->flags & SLAB_RED_ZONE)
488 print_section("Redzone", p + s->objsize,
489 s->inuse - s->objsize);
492 off = s->offset + sizeof(void *);
496 if (s->flags & SLAB_STORE_USER)
497 off += 2 * sizeof(struct track);
500 /* Beginning of the filler is the free pointer */
501 print_section("Padding", p + off, s->size - off);
506 static void object_err(struct kmem_cache *s, struct page *page,
507 u8 *object, char *reason)
510 print_trailer(s, page, object);
513 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
519 vsnprintf(buf, sizeof(buf), fmt, args);
522 print_page_info(page);
526 static void init_object(struct kmem_cache *s, void *object, int active)
530 if (s->flags & __OBJECT_POISON) {
531 memset(p, POISON_FREE, s->objsize - 1);
532 p[s->objsize -1] = POISON_END;
535 if (s->flags & SLAB_RED_ZONE)
536 memset(p + s->objsize,
537 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
538 s->inuse - s->objsize);
541 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
544 if (*start != (u8)value)
552 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
553 void *from, void *to)
555 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
556 memset(from, data, to - from);
559 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
560 u8 *object, char *what,
561 u8* start, unsigned int value, unsigned int bytes)
566 fault = check_bytes(start, value, bytes);
571 while (end > fault && end[-1] == value)
574 slab_bug(s, "%s overwritten", what);
575 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
576 fault, end - 1, fault[0], value);
577 print_trailer(s, page, object);
579 restore_bytes(s, what, value, fault, end);
587 * Bytes of the object to be managed.
588 * If the freepointer may overlay the object then the free
589 * pointer is the first word of the object.
591 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
594 * object + s->objsize
595 * Padding to reach word boundary. This is also used for Redzoning.
596 * Padding is extended by another word if Redzoning is enabled and
599 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
600 * 0xcc (RED_ACTIVE) for objects in use.
603 * Meta data starts here.
605 * A. Free pointer (if we cannot overwrite object on free)
606 * B. Tracking data for SLAB_STORE_USER
607 * C. Padding to reach required alignment boundary or at mininum
608 * one word if debuggin is on to be able to detect writes
609 * before the word boundary.
611 * Padding is done using 0x5a (POISON_INUSE)
614 * Nothing is used beyond s->size.
616 * If slabcaches are merged then the objsize and inuse boundaries are mostly
617 * ignored. And therefore no slab options that rely on these boundaries
618 * may be used with merged slabcaches.
621 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
623 unsigned long off = s->inuse; /* The end of info */
626 /* Freepointer is placed after the object. */
627 off += sizeof(void *);
629 if (s->flags & SLAB_STORE_USER)
630 /* We also have user information there */
631 off += 2 * sizeof(struct track);
636 return check_bytes_and_report(s, page, p, "Object padding",
637 p + off, POISON_INUSE, s->size - off);
640 static int slab_pad_check(struct kmem_cache *s, struct page *page)
648 if (!(s->flags & SLAB_POISON))
651 start = page_address(page);
652 end = start + (PAGE_SIZE << s->order);
653 length = s->objects * s->size;
654 remainder = end - (start + length);
658 fault = check_bytes(start + length, POISON_INUSE, remainder);
661 while (end > fault && end[-1] == POISON_INUSE)
664 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
665 print_section("Padding", start, length);
667 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
671 static int check_object(struct kmem_cache *s, struct page *page,
672 void *object, int active)
675 u8 *endobject = object + s->objsize;
677 if (s->flags & SLAB_RED_ZONE) {
679 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
681 if (!check_bytes_and_report(s, page, object, "Redzone",
682 endobject, red, s->inuse - s->objsize))
685 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
686 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
687 POISON_INUSE, s->inuse - s->objsize);
690 if (s->flags & SLAB_POISON) {
691 if (!active && (s->flags & __OBJECT_POISON) &&
692 (!check_bytes_and_report(s, page, p, "Poison", p,
693 POISON_FREE, s->objsize - 1) ||
694 !check_bytes_and_report(s, page, p, "Poison",
695 p + s->objsize -1, POISON_END, 1)))
698 * check_pad_bytes cleans up on its own.
700 check_pad_bytes(s, page, p);
703 if (!s->offset && active)
705 * Object and freepointer overlap. Cannot check
706 * freepointer while object is allocated.
710 /* Check free pointer validity */
711 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
712 object_err(s, page, p, "Freepointer corrupt");
714 * No choice but to zap it and thus loose the remainder
715 * of the free objects in this slab. May cause
716 * another error because the object count is now wrong.
718 set_freepointer(s, p, NULL);
724 static int check_slab(struct kmem_cache *s, struct page *page)
726 VM_BUG_ON(!irqs_disabled());
728 if (!PageSlab(page)) {
729 slab_err(s, page, "Not a valid slab page");
732 if (page->offset * sizeof(void *) != s->offset) {
733 slab_err(s, page, "Corrupted offset %lu",
734 (unsigned long)(page->offset * sizeof(void *)));
737 if (page->inuse > s->objects) {
738 slab_err(s, page, "inuse %u > max %u",
739 s->name, page->inuse, s->objects);
742 /* Slab_pad_check fixes things up after itself */
743 slab_pad_check(s, page);
748 * Determine if a certain object on a page is on the freelist. Must hold the
749 * slab lock to guarantee that the chains are in a consistent state.
751 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
754 void *fp = page->freelist;
757 while (fp && nr <= s->objects) {
760 if (!check_valid_pointer(s, page, fp)) {
762 object_err(s, page, object,
763 "Freechain corrupt");
764 set_freepointer(s, object, NULL);
767 slab_err(s, page, "Freepointer corrupt");
768 page->freelist = NULL;
769 page->inuse = s->objects;
770 slab_fix(s, "Freelist cleared");
776 fp = get_freepointer(s, object);
780 if (page->inuse != s->objects - nr) {
781 slab_err(s, page, "Wrong object count. Counter is %d but "
782 "counted were %d", page->inuse, s->objects - nr);
783 page->inuse = s->objects - nr;
784 slab_fix(s, "Object count adjusted.");
786 return search == NULL;
789 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
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 static void setup_object_debug(struct kmem_cache *s, struct page *page,
832 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
835 init_object(s, object, 0);
836 init_tracking(s, object);
839 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
840 void *object, void *addr)
842 if (!check_slab(s, page))
845 if (object && !on_freelist(s, page, object)) {
846 object_err(s, page, object, "Object already allocated");
850 if (!check_valid_pointer(s, page, object)) {
851 object_err(s, page, object, "Freelist Pointer check fails");
855 if (object && !check_object(s, page, object, 0))
858 /* Success perform special debug activities for allocs */
859 if (s->flags & SLAB_STORE_USER)
860 set_track(s, object, TRACK_ALLOC, addr);
861 trace(s, page, object, 1);
862 init_object(s, object, 1);
866 if (PageSlab(page)) {
868 * If this is a slab page then lets do the best we can
869 * to avoid issues in the future. Marking all objects
870 * as used avoids touching the remaining objects.
872 slab_fix(s, "Marking all objects used");
873 page->inuse = s->objects;
874 page->freelist = NULL;
875 /* Fix up fields that may be corrupted */
876 page->offset = s->offset / sizeof(void *);
881 static int free_debug_processing(struct kmem_cache *s, struct page *page,
882 void *object, void *addr)
884 if (!check_slab(s, page))
887 if (!check_valid_pointer(s, page, object)) {
888 slab_err(s, page, "Invalid object pointer 0x%p", object);
892 if (on_freelist(s, page, object)) {
893 object_err(s, page, object, "Object already free");
897 if (!check_object(s, page, object, 1))
900 if (unlikely(s != page->slab)) {
902 slab_err(s, page, "Attempt to free object(0x%p) "
903 "outside of slab", object);
907 "SLUB <none>: no slab for object 0x%p.\n",
912 object_err(s, page, object,
913 "page slab pointer corrupt.");
917 /* Special debug activities for freeing objects */
918 if (!SlabFrozen(page) && !page->freelist)
919 remove_full(s, page);
920 if (s->flags & SLAB_STORE_USER)
921 set_track(s, object, TRACK_FREE, addr);
922 trace(s, page, object, 0);
923 init_object(s, object, 0);
927 slab_fix(s, "Object at 0x%p not freed", object);
931 static int __init setup_slub_debug(char *str)
933 slub_debug = DEBUG_DEFAULT_FLAGS;
934 if (*str++ != '=' || !*str)
936 * No options specified. Switch on full debugging.
942 * No options but restriction on slabs. This means full
943 * debugging for slabs matching a pattern.
950 * Switch off all debugging measures.
955 * Determine which debug features should be switched on
957 for ( ;*str && *str != ','; str++) {
958 switch (tolower(*str)) {
960 slub_debug |= SLAB_DEBUG_FREE;
963 slub_debug |= SLAB_RED_ZONE;
966 slub_debug |= SLAB_POISON;
969 slub_debug |= SLAB_STORE_USER;
972 slub_debug |= SLAB_TRACE;
975 printk(KERN_ERR "slub_debug option '%c' "
976 "unknown. skipped\n",*str);
982 slub_debug_slabs = str + 1;
987 __setup("slub_debug", setup_slub_debug);
989 static unsigned long kmem_cache_flags(unsigned long objsize,
990 unsigned long flags, const char *name,
991 void (*ctor)(void *, struct kmem_cache *, unsigned long))
994 * The page->offset field is only 16 bit wide. This is an offset
995 * in units of words from the beginning of an object. If the slab
996 * size is bigger then we cannot move the free pointer behind the
999 * On 32 bit platforms the limit is 256k. On 64bit platforms
1000 * the limit is 512k.
1002 * Debugging or ctor may create a need to move the free
1003 * pointer. Fail if this happens.
1005 if (objsize >= 65535 * sizeof(void *)) {
1006 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1007 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1011 * Enable debugging if selected on the kernel commandline.
1013 if (slub_debug && (!slub_debug_slabs ||
1014 strncmp(slub_debug_slabs, name,
1015 strlen(slub_debug_slabs)) == 0))
1016 flags |= slub_debug;
1022 static inline void setup_object_debug(struct kmem_cache *s,
1023 struct page *page, void *object) {}
1025 static inline int alloc_debug_processing(struct kmem_cache *s,
1026 struct page *page, void *object, void *addr) { return 0; }
1028 static inline int free_debug_processing(struct kmem_cache *s,
1029 struct page *page, void *object, void *addr) { return 0; }
1031 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1033 static inline int check_object(struct kmem_cache *s, struct page *page,
1034 void *object, int active) { return 1; }
1035 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1036 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1037 unsigned long flags, const char *name,
1038 void (*ctor)(void *, struct kmem_cache *, unsigned long))
1042 #define slub_debug 0
1045 * Slab allocation and freeing
1047 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1050 int pages = 1 << s->order;
1053 flags |= __GFP_COMP;
1055 if (s->flags & SLAB_CACHE_DMA)
1058 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1059 flags |= __GFP_RECLAIMABLE;
1062 page = alloc_pages(flags, s->order);
1064 page = alloc_pages_node(node, flags, s->order);
1069 mod_zone_page_state(page_zone(page),
1070 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1071 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1077 static void setup_object(struct kmem_cache *s, struct page *page,
1080 setup_object_debug(s, page, object);
1081 if (unlikely(s->ctor))
1082 s->ctor(object, s, 0);
1085 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1088 struct kmem_cache_node *n;
1094 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1096 if (flags & __GFP_WAIT)
1099 page = allocate_slab(s,
1100 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1104 n = get_node(s, page_to_nid(page));
1106 atomic_long_inc(&n->nr_slabs);
1107 page->offset = s->offset / sizeof(void *);
1109 page->flags |= 1 << PG_slab;
1110 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1111 SLAB_STORE_USER | SLAB_TRACE))
1114 start = page_address(page);
1115 end = start + s->objects * s->size;
1117 if (unlikely(s->flags & SLAB_POISON))
1118 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1121 for_each_object(p, s, start) {
1122 setup_object(s, page, last);
1123 set_freepointer(s, last, p);
1126 setup_object(s, page, last);
1127 set_freepointer(s, last, NULL);
1129 page->freelist = start;
1130 page->lockless_freelist = NULL;
1133 if (flags & __GFP_WAIT)
1134 local_irq_disable();
1138 static void __free_slab(struct kmem_cache *s, struct page *page)
1140 int pages = 1 << s->order;
1142 if (unlikely(SlabDebug(page))) {
1145 slab_pad_check(s, page);
1146 for_each_object(p, s, page_address(page))
1147 check_object(s, page, p, 0);
1148 ClearSlabDebug(page);
1151 mod_zone_page_state(page_zone(page),
1152 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1153 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1156 page->mapping = NULL;
1157 __free_pages(page, s->order);
1160 static void rcu_free_slab(struct rcu_head *h)
1164 page = container_of((struct list_head *)h, struct page, lru);
1165 __free_slab(page->slab, page);
1168 static void free_slab(struct kmem_cache *s, struct page *page)
1170 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1172 * RCU free overloads the RCU head over the LRU
1174 struct rcu_head *head = (void *)&page->lru;
1176 call_rcu(head, rcu_free_slab);
1178 __free_slab(s, page);
1181 static void discard_slab(struct kmem_cache *s, struct page *page)
1183 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1185 atomic_long_dec(&n->nr_slabs);
1186 reset_page_mapcount(page);
1187 __ClearPageSlab(page);
1192 * Per slab locking using the pagelock
1194 static __always_inline void slab_lock(struct page *page)
1196 bit_spin_lock(PG_locked, &page->flags);
1199 static __always_inline void slab_unlock(struct page *page)
1201 bit_spin_unlock(PG_locked, &page->flags);
1204 static __always_inline int slab_trylock(struct page *page)
1208 rc = bit_spin_trylock(PG_locked, &page->flags);
1213 * Management of partially allocated slabs
1215 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1217 spin_lock(&n->list_lock);
1219 list_add_tail(&page->lru, &n->partial);
1220 spin_unlock(&n->list_lock);
1223 static void add_partial(struct kmem_cache_node *n, struct page *page)
1225 spin_lock(&n->list_lock);
1227 list_add(&page->lru, &n->partial);
1228 spin_unlock(&n->list_lock);
1231 static void remove_partial(struct kmem_cache *s,
1234 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1236 spin_lock(&n->list_lock);
1237 list_del(&page->lru);
1239 spin_unlock(&n->list_lock);
1243 * Lock slab and remove from the partial list.
1245 * Must hold list_lock.
1247 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1249 if (slab_trylock(page)) {
1250 list_del(&page->lru);
1252 SetSlabFrozen(page);
1259 * Try to allocate a partial slab from a specific node.
1261 static struct page *get_partial_node(struct kmem_cache_node *n)
1266 * Racy check. If we mistakenly see no partial slabs then we
1267 * just allocate an empty slab. If we mistakenly try to get a
1268 * partial slab and there is none available then get_partials()
1271 if (!n || !n->nr_partial)
1274 spin_lock(&n->list_lock);
1275 list_for_each_entry(page, &n->partial, lru)
1276 if (lock_and_freeze_slab(n, page))
1280 spin_unlock(&n->list_lock);
1285 * Get a page from somewhere. Search in increasing NUMA distances.
1287 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1290 struct zonelist *zonelist;
1295 * The defrag ratio allows a configuration of the tradeoffs between
1296 * inter node defragmentation and node local allocations. A lower
1297 * defrag_ratio increases the tendency to do local allocations
1298 * instead of attempting to obtain partial slabs from other nodes.
1300 * If the defrag_ratio is set to 0 then kmalloc() always
1301 * returns node local objects. If the ratio is higher then kmalloc()
1302 * may return off node objects because partial slabs are obtained
1303 * from other nodes and filled up.
1305 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1306 * defrag_ratio = 1000) then every (well almost) allocation will
1307 * first attempt to defrag slab caches on other nodes. This means
1308 * scanning over all nodes to look for partial slabs which may be
1309 * expensive if we do it every time we are trying to find a slab
1310 * with available objects.
1312 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1315 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1316 ->node_zonelists[gfp_zone(flags)];
1317 for (z = zonelist->zones; *z; z++) {
1318 struct kmem_cache_node *n;
1320 n = get_node(s, zone_to_nid(*z));
1322 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1323 n->nr_partial > MIN_PARTIAL) {
1324 page = get_partial_node(n);
1334 * Get a partial page, lock it and return it.
1336 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1339 int searchnode = (node == -1) ? numa_node_id() : node;
1341 page = get_partial_node(get_node(s, searchnode));
1342 if (page || (flags & __GFP_THISNODE))
1345 return get_any_partial(s, flags);
1349 * Move a page back to the lists.
1351 * Must be called with the slab lock held.
1353 * On exit the slab lock will have been dropped.
1355 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1357 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1359 ClearSlabFrozen(page);
1363 add_partial(n, page);
1364 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1369 if (n->nr_partial < MIN_PARTIAL) {
1371 * Adding an empty slab to the partial slabs in order
1372 * to avoid page allocator overhead. This slab needs
1373 * to come after the other slabs with objects in
1374 * order to fill them up. That way the size of the
1375 * partial list stays small. kmem_cache_shrink can
1376 * reclaim empty slabs from the partial list.
1378 add_partial_tail(n, page);
1382 discard_slab(s, page);
1388 * Remove the cpu slab
1390 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1393 * Merge cpu freelist into freelist. Typically we get here
1394 * because both freelists are empty. So this is unlikely
1397 while (unlikely(page->lockless_freelist)) {
1400 /* Retrieve object from cpu_freelist */
1401 object = page->lockless_freelist;
1402 page->lockless_freelist = page->lockless_freelist[page->offset];
1404 /* And put onto the regular freelist */
1405 object[page->offset] = page->freelist;
1406 page->freelist = object;
1409 s->cpu_slab[cpu] = NULL;
1410 unfreeze_slab(s, page);
1413 static inline void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1416 deactivate_slab(s, page, cpu);
1421 * Called from IPI handler with interrupts disabled.
1423 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1425 struct page *page = s->cpu_slab[cpu];
1428 flush_slab(s, page, cpu);
1431 static void flush_cpu_slab(void *d)
1433 struct kmem_cache *s = d;
1434 int cpu = smp_processor_id();
1436 __flush_cpu_slab(s, cpu);
1439 static void flush_all(struct kmem_cache *s)
1442 on_each_cpu(flush_cpu_slab, s, 1, 1);
1444 unsigned long flags;
1446 local_irq_save(flags);
1448 local_irq_restore(flags);
1453 * Slow path. The lockless freelist is empty or we need to perform
1456 * Interrupts are disabled.
1458 * Processing is still very fast if new objects have been freed to the
1459 * regular freelist. In that case we simply take over the regular freelist
1460 * as the lockless freelist and zap the regular freelist.
1462 * If that is not working then we fall back to the partial lists. We take the
1463 * first element of the freelist as the object to allocate now and move the
1464 * rest of the freelist to the lockless freelist.
1466 * And if we were unable to get a new slab from the partial slab lists then
1467 * we need to allocate a new slab. This is slowest path since we may sleep.
1469 static void *__slab_alloc(struct kmem_cache *s,
1470 gfp_t gfpflags, int node, void *addr, struct page *page)
1473 int cpu = smp_processor_id();
1479 if (unlikely(node != -1 && page_to_nid(page) != node))
1482 object = page->freelist;
1483 if (unlikely(!object))
1485 if (unlikely(SlabDebug(page)))
1488 object = page->freelist;
1489 page->lockless_freelist = object[page->offset];
1490 page->inuse = s->objects;
1491 page->freelist = NULL;
1496 deactivate_slab(s, page, cpu);
1499 page = get_partial(s, gfpflags, node);
1501 s->cpu_slab[cpu] = page;
1505 page = new_slab(s, gfpflags, node);
1507 cpu = smp_processor_id();
1508 if (s->cpu_slab[cpu]) {
1510 * Someone else populated the cpu_slab while we
1511 * enabled interrupts, or we have gotten scheduled
1512 * on another cpu. The page may not be on the
1513 * requested node even if __GFP_THISNODE was
1514 * specified. So we need to recheck.
1517 page_to_nid(s->cpu_slab[cpu]) == node) {
1519 * Current cpuslab is acceptable and we
1520 * want the current one since its cache hot
1522 discard_slab(s, page);
1523 page = s->cpu_slab[cpu];
1527 /* New slab does not fit our expectations */
1528 flush_slab(s, s->cpu_slab[cpu], cpu);
1531 SetSlabFrozen(page);
1532 s->cpu_slab[cpu] = page;
1537 object = page->freelist;
1538 if (!alloc_debug_processing(s, page, object, addr))
1542 page->freelist = object[page->offset];
1548 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1549 * have the fastpath folded into their functions. So no function call
1550 * overhead for requests that can be satisfied on the fastpath.
1552 * The fastpath works by first checking if the lockless freelist can be used.
1553 * If not then __slab_alloc is called for slow processing.
1555 * Otherwise we can simply pick the next object from the lockless free list.
1557 static void __always_inline *slab_alloc(struct kmem_cache *s,
1558 gfp_t gfpflags, int node, void *addr)
1562 unsigned long flags;
1564 local_irq_save(flags);
1565 page = s->cpu_slab[smp_processor_id()];
1566 if (unlikely(!page || !page->lockless_freelist ||
1567 (node != -1 && page_to_nid(page) != node)))
1569 object = __slab_alloc(s, gfpflags, node, addr, page);
1572 object = page->lockless_freelist;
1573 page->lockless_freelist = object[page->offset];
1575 local_irq_restore(flags);
1577 if (unlikely((gfpflags & __GFP_ZERO) && object))
1578 memset(object, 0, s->objsize);
1583 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1585 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1587 EXPORT_SYMBOL(kmem_cache_alloc);
1590 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1592 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1594 EXPORT_SYMBOL(kmem_cache_alloc_node);
1598 * Slow patch handling. This may still be called frequently since objects
1599 * have a longer lifetime than the cpu slabs in most processing loads.
1601 * So we still attempt to reduce cache line usage. Just take the slab
1602 * lock and free the item. If there is no additional partial page
1603 * handling required then we can return immediately.
1605 static void __slab_free(struct kmem_cache *s, struct page *page,
1606 void *x, void *addr)
1609 void **object = (void *)x;
1613 if (unlikely(SlabDebug(page)))
1616 prior = object[page->offset] = page->freelist;
1617 page->freelist = object;
1620 if (unlikely(SlabFrozen(page)))
1623 if (unlikely(!page->inuse))
1627 * Objects left in the slab. If it
1628 * was not on the partial list before
1631 if (unlikely(!prior))
1632 add_partial(get_node(s, page_to_nid(page)), page);
1641 * Slab still on the partial list.
1643 remove_partial(s, page);
1646 discard_slab(s, page);
1650 if (!free_debug_processing(s, page, x, addr))
1656 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1657 * can perform fastpath freeing without additional function calls.
1659 * The fastpath is only possible if we are freeing to the current cpu slab
1660 * of this processor. This typically the case if we have just allocated
1663 * If fastpath is not possible then fall back to __slab_free where we deal
1664 * with all sorts of special processing.
1666 static void __always_inline slab_free(struct kmem_cache *s,
1667 struct page *page, void *x, void *addr)
1669 void **object = (void *)x;
1670 unsigned long flags;
1672 local_irq_save(flags);
1673 debug_check_no_locks_freed(object, s->objsize);
1674 if (likely(page == s->cpu_slab[smp_processor_id()] &&
1675 !SlabDebug(page))) {
1676 object[page->offset] = page->lockless_freelist;
1677 page->lockless_freelist = object;
1679 __slab_free(s, page, x, addr);
1681 local_irq_restore(flags);
1684 void kmem_cache_free(struct kmem_cache *s, void *x)
1688 page = virt_to_head_page(x);
1690 slab_free(s, page, x, __builtin_return_address(0));
1692 EXPORT_SYMBOL(kmem_cache_free);
1694 /* Figure out on which slab object the object resides */
1695 static struct page *get_object_page(const void *x)
1697 struct page *page = virt_to_head_page(x);
1699 if (!PageSlab(page))
1706 * Object placement in a slab is made very easy because we always start at
1707 * offset 0. If we tune the size of the object to the alignment then we can
1708 * get the required alignment by putting one properly sized object after
1711 * Notice that the allocation order determines the sizes of the per cpu
1712 * caches. Each processor has always one slab available for allocations.
1713 * Increasing the allocation order reduces the number of times that slabs
1714 * must be moved on and off the partial lists and is therefore a factor in
1719 * Mininum / Maximum order of slab pages. This influences locking overhead
1720 * and slab fragmentation. A higher order reduces the number of partial slabs
1721 * and increases the number of allocations possible without having to
1722 * take the list_lock.
1724 static int slub_min_order;
1725 static int slub_max_order = DEFAULT_MAX_ORDER;
1726 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1729 * Merge control. If this is set then no merging of slab caches will occur.
1730 * (Could be removed. This was introduced to pacify the merge skeptics.)
1732 static int slub_nomerge;
1735 * Calculate the order of allocation given an slab object size.
1737 * The order of allocation has significant impact on performance and other
1738 * system components. Generally order 0 allocations should be preferred since
1739 * order 0 does not cause fragmentation in the page allocator. Larger objects
1740 * be problematic to put into order 0 slabs because there may be too much
1741 * unused space left. We go to a higher order if more than 1/8th of the slab
1744 * In order to reach satisfactory performance we must ensure that a minimum
1745 * number of objects is in one slab. Otherwise we may generate too much
1746 * activity on the partial lists which requires taking the list_lock. This is
1747 * less a concern for large slabs though which are rarely used.
1749 * slub_max_order specifies the order where we begin to stop considering the
1750 * number of objects in a slab as critical. If we reach slub_max_order then
1751 * we try to keep the page order as low as possible. So we accept more waste
1752 * of space in favor of a small page order.
1754 * Higher order allocations also allow the placement of more objects in a
1755 * slab and thereby reduce object handling overhead. If the user has
1756 * requested a higher mininum order then we start with that one instead of
1757 * the smallest order which will fit the object.
1759 static inline int slab_order(int size, int min_objects,
1760 int max_order, int fract_leftover)
1764 int min_order = slub_min_order;
1767 * If we would create too many object per slab then reduce
1768 * the slab order even if it goes below slub_min_order.
1770 while (min_order > 0 &&
1771 (PAGE_SIZE << min_order) >= MAX_OBJECTS_PER_SLAB * size)
1774 for (order = max(min_order,
1775 fls(min_objects * size - 1) - PAGE_SHIFT);
1776 order <= max_order; order++) {
1778 unsigned long slab_size = PAGE_SIZE << order;
1780 if (slab_size < min_objects * size)
1783 rem = slab_size % size;
1785 if (rem <= slab_size / fract_leftover)
1788 /* If the next size is too high then exit now */
1789 if (slab_size * 2 >= MAX_OBJECTS_PER_SLAB * size)
1796 static inline int calculate_order(int size)
1803 * Attempt to find best configuration for a slab. This
1804 * works by first attempting to generate a layout with
1805 * the best configuration and backing off gradually.
1807 * First we reduce the acceptable waste in a slab. Then
1808 * we reduce the minimum objects required in a slab.
1810 min_objects = slub_min_objects;
1811 while (min_objects > 1) {
1813 while (fraction >= 4) {
1814 order = slab_order(size, min_objects,
1815 slub_max_order, fraction);
1816 if (order <= slub_max_order)
1824 * We were unable to place multiple objects in a slab. Now
1825 * lets see if we can place a single object there.
1827 order = slab_order(size, 1, slub_max_order, 1);
1828 if (order <= slub_max_order)
1832 * Doh this slab cannot be placed using slub_max_order.
1834 order = slab_order(size, 1, MAX_ORDER, 1);
1835 if (order <= MAX_ORDER)
1841 * Figure out what the alignment of the objects will be.
1843 static unsigned long calculate_alignment(unsigned long flags,
1844 unsigned long align, unsigned long size)
1847 * If the user wants hardware cache aligned objects then
1848 * follow that suggestion if the object is sufficiently
1851 * The hardware cache alignment cannot override the
1852 * specified alignment though. If that is greater
1855 if ((flags & SLAB_HWCACHE_ALIGN) &&
1856 size > cache_line_size() / 2)
1857 return max_t(unsigned long, align, cache_line_size());
1859 if (align < ARCH_SLAB_MINALIGN)
1860 return ARCH_SLAB_MINALIGN;
1862 return ALIGN(align, sizeof(void *));
1865 static void init_kmem_cache_node(struct kmem_cache_node *n)
1868 atomic_long_set(&n->nr_slabs, 0);
1869 spin_lock_init(&n->list_lock);
1870 INIT_LIST_HEAD(&n->partial);
1871 #ifdef CONFIG_SLUB_DEBUG
1872 INIT_LIST_HEAD(&n->full);
1878 * No kmalloc_node yet so do it by hand. We know that this is the first
1879 * slab on the node for this slabcache. There are no concurrent accesses
1882 * Note that this function only works on the kmalloc_node_cache
1883 * when allocating for the kmalloc_node_cache.
1885 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
1889 struct kmem_cache_node *n;
1891 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1893 page = new_slab(kmalloc_caches, gfpflags, node);
1896 if (page_to_nid(page) != node) {
1897 printk(KERN_ERR "SLUB: Unable to allocate memory from "
1899 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
1900 "in order to be able to continue\n");
1905 page->freelist = get_freepointer(kmalloc_caches, n);
1907 kmalloc_caches->node[node] = n;
1908 #ifdef CONFIG_SLUB_DEBUG
1909 init_object(kmalloc_caches, n, 1);
1910 init_tracking(kmalloc_caches, n);
1912 init_kmem_cache_node(n);
1913 atomic_long_inc(&n->nr_slabs);
1914 add_partial(n, page);
1917 * new_slab() disables interupts. If we do not reenable interrupts here
1918 * then bootup would continue with interrupts disabled.
1924 static void free_kmem_cache_nodes(struct kmem_cache *s)
1928 for_each_node_state(node, N_NORMAL_MEMORY) {
1929 struct kmem_cache_node *n = s->node[node];
1930 if (n && n != &s->local_node)
1931 kmem_cache_free(kmalloc_caches, n);
1932 s->node[node] = NULL;
1936 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1941 if (slab_state >= UP)
1942 local_node = page_to_nid(virt_to_page(s));
1946 for_each_node_state(node, N_NORMAL_MEMORY) {
1947 struct kmem_cache_node *n;
1949 if (local_node == node)
1952 if (slab_state == DOWN) {
1953 n = early_kmem_cache_node_alloc(gfpflags,
1957 n = kmem_cache_alloc_node(kmalloc_caches,
1961 free_kmem_cache_nodes(s);
1967 init_kmem_cache_node(n);
1972 static void free_kmem_cache_nodes(struct kmem_cache *s)
1976 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1978 init_kmem_cache_node(&s->local_node);
1984 * calculate_sizes() determines the order and the distribution of data within
1987 static int calculate_sizes(struct kmem_cache *s)
1989 unsigned long flags = s->flags;
1990 unsigned long size = s->objsize;
1991 unsigned long align = s->align;
1994 * Determine if we can poison the object itself. If the user of
1995 * the slab may touch the object after free or before allocation
1996 * then we should never poison the object itself.
1998 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2000 s->flags |= __OBJECT_POISON;
2002 s->flags &= ~__OBJECT_POISON;
2005 * Round up object size to the next word boundary. We can only
2006 * place the free pointer at word boundaries and this determines
2007 * the possible location of the free pointer.
2009 size = ALIGN(size, sizeof(void *));
2011 #ifdef CONFIG_SLUB_DEBUG
2013 * If we are Redzoning then check if there is some space between the
2014 * end of the object and the free pointer. If not then add an
2015 * additional word to have some bytes to store Redzone information.
2017 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2018 size += sizeof(void *);
2022 * With that we have determined the number of bytes in actual use
2023 * by the object. This is the potential offset to the free pointer.
2027 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2030 * Relocate free pointer after the object if it is not
2031 * permitted to overwrite the first word of the object on
2034 * This is the case if we do RCU, have a constructor or
2035 * destructor or are poisoning the objects.
2038 size += sizeof(void *);
2041 #ifdef CONFIG_SLUB_DEBUG
2042 if (flags & SLAB_STORE_USER)
2044 * Need to store information about allocs and frees after
2047 size += 2 * sizeof(struct track);
2049 if (flags & SLAB_RED_ZONE)
2051 * Add some empty padding so that we can catch
2052 * overwrites from earlier objects rather than let
2053 * tracking information or the free pointer be
2054 * corrupted if an user writes before the start
2057 size += sizeof(void *);
2061 * Determine the alignment based on various parameters that the
2062 * user specified and the dynamic determination of cache line size
2065 align = calculate_alignment(flags, align, s->objsize);
2068 * SLUB stores one object immediately after another beginning from
2069 * offset 0. In order to align the objects we have to simply size
2070 * each object to conform to the alignment.
2072 size = ALIGN(size, align);
2075 s->order = calculate_order(size);
2080 * Determine the number of objects per slab
2082 s->objects = (PAGE_SIZE << s->order) / size;
2085 * Verify that the number of objects is within permitted limits.
2086 * The page->inuse field is only 16 bit wide! So we cannot have
2087 * more than 64k objects per slab.
2089 if (!s->objects || s->objects > MAX_OBJECTS_PER_SLAB)
2095 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2096 const char *name, size_t size,
2097 size_t align, unsigned long flags,
2098 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2100 memset(s, 0, kmem_size);
2105 s->flags = kmem_cache_flags(size, flags, name, ctor);
2107 if (!calculate_sizes(s))
2112 s->defrag_ratio = 100;
2115 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2118 if (flags & SLAB_PANIC)
2119 panic("Cannot create slab %s size=%lu realsize=%u "
2120 "order=%u offset=%u flags=%lx\n",
2121 s->name, (unsigned long)size, s->size, s->order,
2127 * Check if a given pointer is valid
2129 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2133 page = get_object_page(object);
2135 if (!page || s != page->slab)
2136 /* No slab or wrong slab */
2139 if (!check_valid_pointer(s, page, object))
2143 * We could also check if the object is on the slabs freelist.
2144 * But this would be too expensive and it seems that the main
2145 * purpose of kmem_ptr_valid is to check if the object belongs
2146 * to a certain slab.
2150 EXPORT_SYMBOL(kmem_ptr_validate);
2153 * Determine the size of a slab object
2155 unsigned int kmem_cache_size(struct kmem_cache *s)
2159 EXPORT_SYMBOL(kmem_cache_size);
2161 const char *kmem_cache_name(struct kmem_cache *s)
2165 EXPORT_SYMBOL(kmem_cache_name);
2168 * Attempt to free all slabs on a node. Return the number of slabs we
2169 * were unable to free.
2171 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2172 struct list_head *list)
2174 int slabs_inuse = 0;
2175 unsigned long flags;
2176 struct page *page, *h;
2178 spin_lock_irqsave(&n->list_lock, flags);
2179 list_for_each_entry_safe(page, h, list, lru)
2181 list_del(&page->lru);
2182 discard_slab(s, page);
2185 spin_unlock_irqrestore(&n->list_lock, flags);
2190 * Release all resources used by a slab cache.
2192 static inline int kmem_cache_close(struct kmem_cache *s)
2198 /* Attempt to free all objects */
2199 for_each_node_state(node, N_NORMAL_MEMORY) {
2200 struct kmem_cache_node *n = get_node(s, node);
2202 n->nr_partial -= free_list(s, n, &n->partial);
2203 if (atomic_long_read(&n->nr_slabs))
2206 free_kmem_cache_nodes(s);
2211 * Close a cache and release the kmem_cache structure
2212 * (must be used for caches created using kmem_cache_create)
2214 void kmem_cache_destroy(struct kmem_cache *s)
2216 down_write(&slub_lock);
2220 up_write(&slub_lock);
2221 if (kmem_cache_close(s))
2223 sysfs_slab_remove(s);
2226 up_write(&slub_lock);
2228 EXPORT_SYMBOL(kmem_cache_destroy);
2230 /********************************************************************
2232 *******************************************************************/
2234 struct kmem_cache kmalloc_caches[PAGE_SHIFT] __cacheline_aligned;
2235 EXPORT_SYMBOL(kmalloc_caches);
2237 #ifdef CONFIG_ZONE_DMA
2238 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT];
2241 static int __init setup_slub_min_order(char *str)
2243 get_option (&str, &slub_min_order);
2248 __setup("slub_min_order=", setup_slub_min_order);
2250 static int __init setup_slub_max_order(char *str)
2252 get_option (&str, &slub_max_order);
2257 __setup("slub_max_order=", setup_slub_max_order);
2259 static int __init setup_slub_min_objects(char *str)
2261 get_option (&str, &slub_min_objects);
2266 __setup("slub_min_objects=", setup_slub_min_objects);
2268 static int __init setup_slub_nomerge(char *str)
2274 __setup("slub_nomerge", setup_slub_nomerge);
2276 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2277 const char *name, int size, gfp_t gfp_flags)
2279 unsigned int flags = 0;
2281 if (gfp_flags & SLUB_DMA)
2282 flags = SLAB_CACHE_DMA;
2284 down_write(&slub_lock);
2285 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2289 list_add(&s->list, &slab_caches);
2290 up_write(&slub_lock);
2291 if (sysfs_slab_add(s))
2296 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2299 #ifdef CONFIG_ZONE_DMA
2301 static void sysfs_add_func(struct work_struct *w)
2303 struct kmem_cache *s;
2305 down_write(&slub_lock);
2306 list_for_each_entry(s, &slab_caches, list) {
2307 if (s->flags & __SYSFS_ADD_DEFERRED) {
2308 s->flags &= ~__SYSFS_ADD_DEFERRED;
2312 up_write(&slub_lock);
2315 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2317 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2319 struct kmem_cache *s;
2323 s = kmalloc_caches_dma[index];
2327 /* Dynamically create dma cache */
2328 if (flags & __GFP_WAIT)
2329 down_write(&slub_lock);
2331 if (!down_write_trylock(&slub_lock))
2335 if (kmalloc_caches_dma[index])
2338 realsize = kmalloc_caches[index].objsize;
2339 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d", (unsigned int)realsize),
2340 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2342 if (!s || !text || !kmem_cache_open(s, flags, text,
2343 realsize, ARCH_KMALLOC_MINALIGN,
2344 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2350 list_add(&s->list, &slab_caches);
2351 kmalloc_caches_dma[index] = s;
2353 schedule_work(&sysfs_add_work);
2356 up_write(&slub_lock);
2358 return kmalloc_caches_dma[index];
2363 * Conversion table for small slabs sizes / 8 to the index in the
2364 * kmalloc array. This is necessary for slabs < 192 since we have non power
2365 * of two cache sizes there. The size of larger slabs can be determined using
2368 static s8 size_index[24] = {
2395 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2401 return ZERO_SIZE_PTR;
2403 index = size_index[(size - 1) / 8];
2405 index = fls(size - 1);
2407 #ifdef CONFIG_ZONE_DMA
2408 if (unlikely((flags & SLUB_DMA)))
2409 return dma_kmalloc_cache(index, flags);
2412 return &kmalloc_caches[index];
2415 void *__kmalloc(size_t size, gfp_t flags)
2417 struct kmem_cache *s;
2419 if (unlikely(size > PAGE_SIZE / 2))
2420 return (void *)__get_free_pages(flags | __GFP_COMP,
2423 s = get_slab(size, flags);
2425 if (unlikely(ZERO_OR_NULL_PTR(s)))
2428 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2430 EXPORT_SYMBOL(__kmalloc);
2433 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2435 struct kmem_cache *s;
2437 if (unlikely(size > PAGE_SIZE / 2))
2438 return (void *)__get_free_pages(flags | __GFP_COMP,
2441 s = get_slab(size, flags);
2443 if (unlikely(ZERO_OR_NULL_PTR(s)))
2446 return slab_alloc(s, flags, node, __builtin_return_address(0));
2448 EXPORT_SYMBOL(__kmalloc_node);
2451 size_t ksize(const void *object)
2454 struct kmem_cache *s;
2457 if (unlikely(object == ZERO_SIZE_PTR))
2460 page = get_object_page(object);
2466 * Debugging requires use of the padding between object
2467 * and whatever may come after it.
2469 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2473 * If we have the need to store the freelist pointer
2474 * back there or track user information then we can
2475 * only use the space before that information.
2477 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2481 * Else we can use all the padding etc for the allocation
2485 EXPORT_SYMBOL(ksize);
2487 void kfree(const void *x)
2491 if (unlikely(ZERO_OR_NULL_PTR(x)))
2494 page = virt_to_head_page(x);
2495 if (unlikely(!PageSlab(page))) {
2499 slab_free(page->slab, page, (void *)x, __builtin_return_address(0));
2501 EXPORT_SYMBOL(kfree);
2504 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2505 * the remaining slabs by the number of items in use. The slabs with the
2506 * most items in use come first. New allocations will then fill those up
2507 * and thus they can be removed from the partial lists.
2509 * The slabs with the least items are placed last. This results in them
2510 * being allocated from last increasing the chance that the last objects
2511 * are freed in them.
2513 int kmem_cache_shrink(struct kmem_cache *s)
2517 struct kmem_cache_node *n;
2520 struct list_head *slabs_by_inuse =
2521 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2522 unsigned long flags;
2524 if (!slabs_by_inuse)
2528 for_each_node_state(node, N_NORMAL_MEMORY) {
2529 n = get_node(s, node);
2534 for (i = 0; i < s->objects; i++)
2535 INIT_LIST_HEAD(slabs_by_inuse + i);
2537 spin_lock_irqsave(&n->list_lock, flags);
2540 * Build lists indexed by the items in use in each slab.
2542 * Note that concurrent frees may occur while we hold the
2543 * list_lock. page->inuse here is the upper limit.
2545 list_for_each_entry_safe(page, t, &n->partial, lru) {
2546 if (!page->inuse && slab_trylock(page)) {
2548 * Must hold slab lock here because slab_free
2549 * may have freed the last object and be
2550 * waiting to release the slab.
2552 list_del(&page->lru);
2555 discard_slab(s, page);
2557 list_move(&page->lru,
2558 slabs_by_inuse + page->inuse);
2563 * Rebuild the partial list with the slabs filled up most
2564 * first and the least used slabs at the end.
2566 for (i = s->objects - 1; i >= 0; i--)
2567 list_splice(slabs_by_inuse + i, n->partial.prev);
2569 spin_unlock_irqrestore(&n->list_lock, flags);
2572 kfree(slabs_by_inuse);
2575 EXPORT_SYMBOL(kmem_cache_shrink);
2577 /********************************************************************
2578 * Basic setup of slabs
2579 *******************************************************************/
2581 void __init kmem_cache_init(void)
2588 * Must first have the slab cache available for the allocations of the
2589 * struct kmem_cache_node's. There is special bootstrap code in
2590 * kmem_cache_open for slab_state == DOWN.
2592 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2593 sizeof(struct kmem_cache_node), GFP_KERNEL);
2594 kmalloc_caches[0].refcount = -1;
2598 /* Able to allocate the per node structures */
2599 slab_state = PARTIAL;
2601 /* Caches that are not of the two-to-the-power-of size */
2602 if (KMALLOC_MIN_SIZE <= 64) {
2603 create_kmalloc_cache(&kmalloc_caches[1],
2604 "kmalloc-96", 96, GFP_KERNEL);
2607 if (KMALLOC_MIN_SIZE <= 128) {
2608 create_kmalloc_cache(&kmalloc_caches[2],
2609 "kmalloc-192", 192, GFP_KERNEL);
2613 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++) {
2614 create_kmalloc_cache(&kmalloc_caches[i],
2615 "kmalloc", 1 << i, GFP_KERNEL);
2621 * Patch up the size_index table if we have strange large alignment
2622 * requirements for the kmalloc array. This is only the case for
2623 * mips it seems. The standard arches will not generate any code here.
2625 * Largest permitted alignment is 256 bytes due to the way we
2626 * handle the index determination for the smaller caches.
2628 * Make sure that nothing crazy happens if someone starts tinkering
2629 * around with ARCH_KMALLOC_MINALIGN
2631 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2632 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2634 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2635 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2639 /* Provide the correct kmalloc names now that the caches are up */
2640 for (i = KMALLOC_SHIFT_LOW; i < PAGE_SHIFT; i++)
2641 kmalloc_caches[i]. name =
2642 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2645 register_cpu_notifier(&slab_notifier);
2648 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2649 nr_cpu_ids * sizeof(struct page *);
2651 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2652 " CPUs=%d, Nodes=%d\n",
2653 caches, cache_line_size(),
2654 slub_min_order, slub_max_order, slub_min_objects,
2655 nr_cpu_ids, nr_node_ids);
2659 * Find a mergeable slab cache
2661 static int slab_unmergeable(struct kmem_cache *s)
2663 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2670 * We may have set a slab to be unmergeable during bootstrap.
2672 if (s->refcount < 0)
2678 static struct kmem_cache *find_mergeable(size_t size,
2679 size_t align, unsigned long flags, const char *name,
2680 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2682 struct kmem_cache *s;
2684 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2690 size = ALIGN(size, sizeof(void *));
2691 align = calculate_alignment(flags, align, size);
2692 size = ALIGN(size, align);
2693 flags = kmem_cache_flags(size, flags, name, NULL);
2695 list_for_each_entry(s, &slab_caches, list) {
2696 if (slab_unmergeable(s))
2702 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
2705 * Check if alignment is compatible.
2706 * Courtesy of Adrian Drzewiecki
2708 if ((s->size & ~(align -1)) != s->size)
2711 if (s->size - size >= sizeof(void *))
2719 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2720 size_t align, unsigned long flags,
2721 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2723 struct kmem_cache *s;
2725 down_write(&slub_lock);
2726 s = find_mergeable(size, align, flags, name, ctor);
2730 * Adjust the object sizes so that we clear
2731 * the complete object on kzalloc.
2733 s->objsize = max(s->objsize, (int)size);
2734 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2735 up_write(&slub_lock);
2736 if (sysfs_slab_alias(s, name))
2740 s = kmalloc(kmem_size, GFP_KERNEL);
2742 if (kmem_cache_open(s, GFP_KERNEL, name,
2743 size, align, flags, ctor)) {
2744 list_add(&s->list, &slab_caches);
2745 up_write(&slub_lock);
2746 if (sysfs_slab_add(s))
2752 up_write(&slub_lock);
2755 if (flags & SLAB_PANIC)
2756 panic("Cannot create slabcache %s\n", name);
2761 EXPORT_SYMBOL(kmem_cache_create);
2765 * Use the cpu notifier to insure that the cpu slabs are flushed when
2768 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2769 unsigned long action, void *hcpu)
2771 long cpu = (long)hcpu;
2772 struct kmem_cache *s;
2773 unsigned long flags;
2776 case CPU_UP_CANCELED:
2777 case CPU_UP_CANCELED_FROZEN:
2779 case CPU_DEAD_FROZEN:
2780 down_read(&slub_lock);
2781 list_for_each_entry(s, &slab_caches, list) {
2782 local_irq_save(flags);
2783 __flush_cpu_slab(s, cpu);
2784 local_irq_restore(flags);
2786 up_read(&slub_lock);
2794 static struct notifier_block __cpuinitdata slab_notifier =
2795 { &slab_cpuup_callback, NULL, 0 };
2799 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2801 struct kmem_cache *s;
2803 if (unlikely(size > PAGE_SIZE / 2))
2804 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
2806 s = get_slab(size, gfpflags);
2808 if (unlikely(ZERO_OR_NULL_PTR(s)))
2811 return slab_alloc(s, gfpflags, -1, caller);
2814 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2815 int node, void *caller)
2817 struct kmem_cache *s;
2819 if (unlikely(size > PAGE_SIZE / 2))
2820 return (void *)__get_free_pages(gfpflags | __GFP_COMP,
2822 s = get_slab(size, gfpflags);
2824 if (unlikely(ZERO_OR_NULL_PTR(s)))
2827 return slab_alloc(s, gfpflags, node, caller);
2830 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2831 static int validate_slab(struct kmem_cache *s, struct page *page,
2835 void *addr = page_address(page);
2837 if (!check_slab(s, page) ||
2838 !on_freelist(s, page, NULL))
2841 /* Now we know that a valid freelist exists */
2842 bitmap_zero(map, s->objects);
2844 for_each_free_object(p, s, page->freelist) {
2845 set_bit(slab_index(p, s, addr), map);
2846 if (!check_object(s, page, p, 0))
2850 for_each_object(p, s, addr)
2851 if (!test_bit(slab_index(p, s, addr), map))
2852 if (!check_object(s, page, p, 1))
2857 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
2860 if (slab_trylock(page)) {
2861 validate_slab(s, page, map);
2864 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2867 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2868 if (!SlabDebug(page))
2869 printk(KERN_ERR "SLUB %s: SlabDebug not set "
2870 "on slab 0x%p\n", s->name, page);
2872 if (SlabDebug(page))
2873 printk(KERN_ERR "SLUB %s: SlabDebug set on "
2874 "slab 0x%p\n", s->name, page);
2878 static int validate_slab_node(struct kmem_cache *s,
2879 struct kmem_cache_node *n, unsigned long *map)
2881 unsigned long count = 0;
2883 unsigned long flags;
2885 spin_lock_irqsave(&n->list_lock, flags);
2887 list_for_each_entry(page, &n->partial, lru) {
2888 validate_slab_slab(s, page, map);
2891 if (count != n->nr_partial)
2892 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2893 "counter=%ld\n", s->name, count, n->nr_partial);
2895 if (!(s->flags & SLAB_STORE_USER))
2898 list_for_each_entry(page, &n->full, lru) {
2899 validate_slab_slab(s, page, map);
2902 if (count != atomic_long_read(&n->nr_slabs))
2903 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2904 "counter=%ld\n", s->name, count,
2905 atomic_long_read(&n->nr_slabs));
2908 spin_unlock_irqrestore(&n->list_lock, flags);
2912 static long validate_slab_cache(struct kmem_cache *s)
2915 unsigned long count = 0;
2916 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
2917 sizeof(unsigned long), GFP_KERNEL);
2923 for_each_node_state(node, N_NORMAL_MEMORY) {
2924 struct kmem_cache_node *n = get_node(s, node);
2926 count += validate_slab_node(s, n, map);
2932 #ifdef SLUB_RESILIENCY_TEST
2933 static void resiliency_test(void)
2937 printk(KERN_ERR "SLUB resiliency testing\n");
2938 printk(KERN_ERR "-----------------------\n");
2939 printk(KERN_ERR "A. Corruption after allocation\n");
2941 p = kzalloc(16, GFP_KERNEL);
2943 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2944 " 0x12->0x%p\n\n", p + 16);
2946 validate_slab_cache(kmalloc_caches + 4);
2948 /* Hmmm... The next two are dangerous */
2949 p = kzalloc(32, GFP_KERNEL);
2950 p[32 + sizeof(void *)] = 0x34;
2951 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2952 " 0x34 -> -0x%p\n", p);
2953 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2955 validate_slab_cache(kmalloc_caches + 5);
2956 p = kzalloc(64, GFP_KERNEL);
2957 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2959 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2961 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2962 validate_slab_cache(kmalloc_caches + 6);
2964 printk(KERN_ERR "\nB. Corruption after free\n");
2965 p = kzalloc(128, GFP_KERNEL);
2968 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2969 validate_slab_cache(kmalloc_caches + 7);
2971 p = kzalloc(256, GFP_KERNEL);
2974 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2975 validate_slab_cache(kmalloc_caches + 8);
2977 p = kzalloc(512, GFP_KERNEL);
2980 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2981 validate_slab_cache(kmalloc_caches + 9);
2984 static void resiliency_test(void) {};
2988 * Generate lists of code addresses where slabcache objects are allocated
2993 unsigned long count;
3006 unsigned long count;
3007 struct location *loc;
3010 static void free_loc_track(struct loc_track *t)
3013 free_pages((unsigned long)t->loc,
3014 get_order(sizeof(struct location) * t->max));
3017 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3022 order = get_order(sizeof(struct location) * max);
3024 l = (void *)__get_free_pages(flags, order);
3029 memcpy(l, t->loc, sizeof(struct location) * t->count);
3037 static int add_location(struct loc_track *t, struct kmem_cache *s,
3038 const struct track *track)
3040 long start, end, pos;
3043 unsigned long age = jiffies - track->when;
3049 pos = start + (end - start + 1) / 2;
3052 * There is nothing at "end". If we end up there
3053 * we need to add something to before end.
3058 caddr = t->loc[pos].addr;
3059 if (track->addr == caddr) {
3065 if (age < l->min_time)
3067 if (age > l->max_time)
3070 if (track->pid < l->min_pid)
3071 l->min_pid = track->pid;
3072 if (track->pid > l->max_pid)
3073 l->max_pid = track->pid;
3075 cpu_set(track->cpu, l->cpus);
3077 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3081 if (track->addr < caddr)
3088 * Not found. Insert new tracking element.
3090 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3096 (t->count - pos) * sizeof(struct location));
3099 l->addr = track->addr;
3103 l->min_pid = track->pid;
3104 l->max_pid = track->pid;
3105 cpus_clear(l->cpus);
3106 cpu_set(track->cpu, l->cpus);
3107 nodes_clear(l->nodes);
3108 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3112 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3113 struct page *page, enum track_item alloc)
3115 void *addr = page_address(page);
3116 DECLARE_BITMAP(map, s->objects);
3119 bitmap_zero(map, s->objects);
3120 for_each_free_object(p, s, page->freelist)
3121 set_bit(slab_index(p, s, addr), map);
3123 for_each_object(p, s, addr)
3124 if (!test_bit(slab_index(p, s, addr), map))
3125 add_location(t, s, get_track(s, p, alloc));
3128 static int list_locations(struct kmem_cache *s, char *buf,
3129 enum track_item alloc)
3133 struct loc_track t = { 0, 0, NULL };
3136 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3138 return sprintf(buf, "Out of memory\n");
3140 /* Push back cpu slabs */
3143 for_each_node_state(node, N_NORMAL_MEMORY) {
3144 struct kmem_cache_node *n = get_node(s, node);
3145 unsigned long flags;
3148 if (!atomic_long_read(&n->nr_slabs))
3151 spin_lock_irqsave(&n->list_lock, flags);
3152 list_for_each_entry(page, &n->partial, lru)
3153 process_slab(&t, s, page, alloc);
3154 list_for_each_entry(page, &n->full, lru)
3155 process_slab(&t, s, page, alloc);
3156 spin_unlock_irqrestore(&n->list_lock, flags);
3159 for (i = 0; i < t.count; i++) {
3160 struct location *l = &t.loc[i];
3162 if (n > PAGE_SIZE - 100)
3164 n += sprintf(buf + n, "%7ld ", l->count);
3167 n += sprint_symbol(buf + n, (unsigned long)l->addr);
3169 n += sprintf(buf + n, "<not-available>");
3171 if (l->sum_time != l->min_time) {
3172 unsigned long remainder;
3174 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3176 div_long_long_rem(l->sum_time, l->count, &remainder),
3179 n += sprintf(buf + n, " age=%ld",
3182 if (l->min_pid != l->max_pid)
3183 n += sprintf(buf + n, " pid=%ld-%ld",
3184 l->min_pid, l->max_pid);
3186 n += sprintf(buf + n, " pid=%ld",
3189 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3190 n < PAGE_SIZE - 60) {
3191 n += sprintf(buf + n, " cpus=");
3192 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3196 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3197 n < PAGE_SIZE - 60) {
3198 n += sprintf(buf + n, " nodes=");
3199 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3203 n += sprintf(buf + n, "\n");
3208 n += sprintf(buf, "No data\n");
3212 static unsigned long count_partial(struct kmem_cache_node *n)
3214 unsigned long flags;
3215 unsigned long x = 0;
3218 spin_lock_irqsave(&n->list_lock, flags);
3219 list_for_each_entry(page, &n->partial, lru)
3221 spin_unlock_irqrestore(&n->list_lock, flags);
3225 enum slab_stat_type {
3232 #define SO_FULL (1 << SL_FULL)
3233 #define SO_PARTIAL (1 << SL_PARTIAL)
3234 #define SO_CPU (1 << SL_CPU)
3235 #define SO_OBJECTS (1 << SL_OBJECTS)
3237 static unsigned long slab_objects(struct kmem_cache *s,
3238 char *buf, unsigned long flags)
3240 unsigned long total = 0;
3244 unsigned long *nodes;
3245 unsigned long *per_cpu;
3247 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3248 per_cpu = nodes + nr_node_ids;
3250 for_each_possible_cpu(cpu) {
3251 struct page *page = s->cpu_slab[cpu];
3255 node = page_to_nid(page);
3256 if (flags & SO_CPU) {
3259 if (flags & SO_OBJECTS)
3270 for_each_node_state(node, N_NORMAL_MEMORY) {
3271 struct kmem_cache_node *n = get_node(s, node);
3273 if (flags & SO_PARTIAL) {
3274 if (flags & SO_OBJECTS)
3275 x = count_partial(n);
3282 if (flags & SO_FULL) {
3283 int full_slabs = atomic_long_read(&n->nr_slabs)
3287 if (flags & SO_OBJECTS)
3288 x = full_slabs * s->objects;
3296 x = sprintf(buf, "%lu", total);
3298 for_each_node_state(node, N_NORMAL_MEMORY)
3300 x += sprintf(buf + x, " N%d=%lu",
3304 return x + sprintf(buf + x, "\n");
3307 static int any_slab_objects(struct kmem_cache *s)
3312 for_each_possible_cpu(cpu)
3313 if (s->cpu_slab[cpu])
3316 for_each_node(node) {
3317 struct kmem_cache_node *n = get_node(s, node);
3319 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
3325 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3326 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3328 struct slab_attribute {
3329 struct attribute attr;
3330 ssize_t (*show)(struct kmem_cache *s, char *buf);
3331 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3334 #define SLAB_ATTR_RO(_name) \
3335 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3337 #define SLAB_ATTR(_name) \
3338 static struct slab_attribute _name##_attr = \
3339 __ATTR(_name, 0644, _name##_show, _name##_store)
3341 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3343 return sprintf(buf, "%d\n", s->size);
3345 SLAB_ATTR_RO(slab_size);
3347 static ssize_t align_show(struct kmem_cache *s, char *buf)
3349 return sprintf(buf, "%d\n", s->align);
3351 SLAB_ATTR_RO(align);
3353 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3355 return sprintf(buf, "%d\n", s->objsize);
3357 SLAB_ATTR_RO(object_size);
3359 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3361 return sprintf(buf, "%d\n", s->objects);
3363 SLAB_ATTR_RO(objs_per_slab);
3365 static ssize_t order_show(struct kmem_cache *s, char *buf)
3367 return sprintf(buf, "%d\n", s->order);
3369 SLAB_ATTR_RO(order);
3371 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3374 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3376 return n + sprintf(buf + n, "\n");
3382 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3384 return sprintf(buf, "%d\n", s->refcount - 1);
3386 SLAB_ATTR_RO(aliases);
3388 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3390 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3392 SLAB_ATTR_RO(slabs);
3394 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3396 return slab_objects(s, buf, SO_PARTIAL);
3398 SLAB_ATTR_RO(partial);
3400 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3402 return slab_objects(s, buf, SO_CPU);
3404 SLAB_ATTR_RO(cpu_slabs);
3406 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3408 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3410 SLAB_ATTR_RO(objects);
3412 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3414 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3417 static ssize_t sanity_checks_store(struct kmem_cache *s,
3418 const char *buf, size_t length)
3420 s->flags &= ~SLAB_DEBUG_FREE;
3422 s->flags |= SLAB_DEBUG_FREE;
3425 SLAB_ATTR(sanity_checks);
3427 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3429 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3432 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3435 s->flags &= ~SLAB_TRACE;
3437 s->flags |= SLAB_TRACE;
3442 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3444 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3447 static ssize_t reclaim_account_store(struct kmem_cache *s,
3448 const char *buf, size_t length)
3450 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3452 s->flags |= SLAB_RECLAIM_ACCOUNT;
3455 SLAB_ATTR(reclaim_account);
3457 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3459 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3461 SLAB_ATTR_RO(hwcache_align);
3463 #ifdef CONFIG_ZONE_DMA
3464 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3466 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3468 SLAB_ATTR_RO(cache_dma);
3471 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3473 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3475 SLAB_ATTR_RO(destroy_by_rcu);
3477 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3479 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3482 static ssize_t red_zone_store(struct kmem_cache *s,
3483 const char *buf, size_t length)
3485 if (any_slab_objects(s))
3488 s->flags &= ~SLAB_RED_ZONE;
3490 s->flags |= SLAB_RED_ZONE;
3494 SLAB_ATTR(red_zone);
3496 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3498 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3501 static ssize_t poison_store(struct kmem_cache *s,
3502 const char *buf, size_t length)
3504 if (any_slab_objects(s))
3507 s->flags &= ~SLAB_POISON;
3509 s->flags |= SLAB_POISON;
3515 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3517 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3520 static ssize_t store_user_store(struct kmem_cache *s,
3521 const char *buf, size_t length)
3523 if (any_slab_objects(s))
3526 s->flags &= ~SLAB_STORE_USER;
3528 s->flags |= SLAB_STORE_USER;
3532 SLAB_ATTR(store_user);
3534 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3539 static ssize_t validate_store(struct kmem_cache *s,
3540 const char *buf, size_t length)
3544 if (buf[0] == '1') {
3545 ret = validate_slab_cache(s);
3551 SLAB_ATTR(validate);
3553 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3558 static ssize_t shrink_store(struct kmem_cache *s,
3559 const char *buf, size_t length)
3561 if (buf[0] == '1') {
3562 int rc = kmem_cache_shrink(s);
3572 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3574 if (!(s->flags & SLAB_STORE_USER))
3576 return list_locations(s, buf, TRACK_ALLOC);
3578 SLAB_ATTR_RO(alloc_calls);
3580 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3582 if (!(s->flags & SLAB_STORE_USER))
3584 return list_locations(s, buf, TRACK_FREE);
3586 SLAB_ATTR_RO(free_calls);
3589 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3591 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3594 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3595 const char *buf, size_t length)
3597 int n = simple_strtoul(buf, NULL, 10);
3600 s->defrag_ratio = n * 10;
3603 SLAB_ATTR(defrag_ratio);
3606 static struct attribute * slab_attrs[] = {
3607 &slab_size_attr.attr,
3608 &object_size_attr.attr,
3609 &objs_per_slab_attr.attr,
3614 &cpu_slabs_attr.attr,
3618 &sanity_checks_attr.attr,
3620 &hwcache_align_attr.attr,
3621 &reclaim_account_attr.attr,
3622 &destroy_by_rcu_attr.attr,
3623 &red_zone_attr.attr,
3625 &store_user_attr.attr,
3626 &validate_attr.attr,
3628 &alloc_calls_attr.attr,
3629 &free_calls_attr.attr,
3630 #ifdef CONFIG_ZONE_DMA
3631 &cache_dma_attr.attr,
3634 &defrag_ratio_attr.attr,
3639 static struct attribute_group slab_attr_group = {
3640 .attrs = slab_attrs,
3643 static ssize_t slab_attr_show(struct kobject *kobj,
3644 struct attribute *attr,
3647 struct slab_attribute *attribute;
3648 struct kmem_cache *s;
3651 attribute = to_slab_attr(attr);
3654 if (!attribute->show)
3657 err = attribute->show(s, buf);
3662 static ssize_t slab_attr_store(struct kobject *kobj,
3663 struct attribute *attr,
3664 const char *buf, size_t len)
3666 struct slab_attribute *attribute;
3667 struct kmem_cache *s;
3670 attribute = to_slab_attr(attr);
3673 if (!attribute->store)
3676 err = attribute->store(s, buf, len);
3681 static struct sysfs_ops slab_sysfs_ops = {
3682 .show = slab_attr_show,
3683 .store = slab_attr_store,
3686 static struct kobj_type slab_ktype = {
3687 .sysfs_ops = &slab_sysfs_ops,
3690 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3692 struct kobj_type *ktype = get_ktype(kobj);
3694 if (ktype == &slab_ktype)
3699 static struct kset_uevent_ops slab_uevent_ops = {
3700 .filter = uevent_filter,
3703 static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3705 #define ID_STR_LENGTH 64
3707 /* Create a unique string id for a slab cache:
3709 * :[flags-]size:[memory address of kmemcache]
3711 static char *create_unique_id(struct kmem_cache *s)
3713 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3720 * First flags affecting slabcache operations. We will only
3721 * get here for aliasable slabs so we do not need to support
3722 * too many flags. The flags here must cover all flags that
3723 * are matched during merging to guarantee that the id is
3726 if (s->flags & SLAB_CACHE_DMA)
3728 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3730 if (s->flags & SLAB_DEBUG_FREE)
3734 p += sprintf(p, "%07d", s->size);
3735 BUG_ON(p > name + ID_STR_LENGTH - 1);
3739 static int sysfs_slab_add(struct kmem_cache *s)
3745 if (slab_state < SYSFS)
3746 /* Defer until later */
3749 unmergeable = slab_unmergeable(s);
3752 * Slabcache can never be merged so we can use the name proper.
3753 * This is typically the case for debug situations. In that
3754 * case we can catch duplicate names easily.
3756 sysfs_remove_link(&slab_subsys.kobj, s->name);
3760 * Create a unique name for the slab as a target
3763 name = create_unique_id(s);
3766 kobj_set_kset_s(s, slab_subsys);
3767 kobject_set_name(&s->kobj, name);
3768 kobject_init(&s->kobj);
3769 err = kobject_add(&s->kobj);
3773 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3776 kobject_uevent(&s->kobj, KOBJ_ADD);
3778 /* Setup first alias */
3779 sysfs_slab_alias(s, s->name);
3785 static void sysfs_slab_remove(struct kmem_cache *s)
3787 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3788 kobject_del(&s->kobj);
3792 * Need to buffer aliases during bootup until sysfs becomes
3793 * available lest we loose that information.
3795 struct saved_alias {
3796 struct kmem_cache *s;
3798 struct saved_alias *next;
3801 static struct saved_alias *alias_list;
3803 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3805 struct saved_alias *al;
3807 if (slab_state == SYSFS) {
3809 * If we have a leftover link then remove it.
3811 sysfs_remove_link(&slab_subsys.kobj, name);
3812 return sysfs_create_link(&slab_subsys.kobj,
3816 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3822 al->next = alias_list;
3827 static int __init slab_sysfs_init(void)
3829 struct kmem_cache *s;
3832 err = subsystem_register(&slab_subsys);
3834 printk(KERN_ERR "Cannot register slab subsystem.\n");
3840 list_for_each_entry(s, &slab_caches, list) {
3841 err = sysfs_slab_add(s);
3843 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
3844 " to sysfs\n", s->name);
3847 while (alias_list) {
3848 struct saved_alias *al = alias_list;
3850 alias_list = alias_list->next;
3851 err = sysfs_slab_alias(al->s, al->name);
3853 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
3854 " %s to sysfs\n", s->name);
3862 __initcall(slab_sysfs_init);