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 or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
31 #include <linux/stacktrace.h>
33 #include <trace/events/kmem.h>
37 * 1. slub_lock (Global Semaphore)
39 * 3. slab_lock(page) (Only on some arches and for debugging)
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
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 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
134 /* Enable to log cmpxchg failures */
135 #undef SLUB_DEBUG_CMPXCHG
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 #define MIN_PARTIAL 5
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
148 #define MAX_PARTIAL 10
150 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
158 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
161 * Set of flags that will prevent slab merging
163 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
167 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
171 #define OO_MASK ((1 << OO_SHIFT) - 1)
172 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174 /* Internal SLUB flags */
175 #define __OBJECT_POISON 0x80000000UL /* Poison object */
176 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178 static int kmem_size = sizeof(struct kmem_cache);
181 static struct notifier_block slab_notifier;
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
191 /* A list of all slab caches on the system */
192 static DECLARE_RWSEM(slub_lock);
193 static LIST_HEAD(slab_caches);
196 * Tracking user of a slab.
198 #define TRACK_ADDRS_COUNT 16
200 unsigned long addr; /* Called from address */
201 #ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
209 enum track_item { TRACK_ALLOC, TRACK_FREE };
212 static int sysfs_slab_add(struct kmem_cache *);
213 static int sysfs_slab_alias(struct kmem_cache *, const char *);
214 static void sysfs_slab_remove(struct kmem_cache *);
217 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
220 static inline void sysfs_slab_remove(struct kmem_cache *s)
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 int slab_is_available(void)
241 return slab_state >= UP;
244 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246 return s->node[node];
249 /* Verify that a pointer has an address that is valid within a slab page */
250 static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
267 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 return *(void **)(object + s->offset);
272 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
276 #ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
279 p = get_freepointer(s, object);
284 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
286 *(void **)(object + s->offset) = fp;
289 /* Loop over all objects in a slab */
290 #define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
294 /* Determine object index from a given position */
295 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
297 return (p - addr) / s->size;
300 static inline size_t slab_ksize(const struct kmem_cache *s)
302 #ifdef CONFIG_SLUB_DEBUG
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
319 * Else we can use all the padding etc for the allocation
324 static inline int order_objects(int order, unsigned long size, int reserved)
326 return ((PAGE_SIZE << order) - reserved) / size;
329 static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
339 static inline int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 bit_spin_lock(PG_locked, &page->flags);
357 static __always_inline void slab_unlock(struct page *page)
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
368 VM_BUG_ON(!irqs_disabled());
369 #ifdef CONFIG_CMPXCHG_DOUBLE
370 if (s->flags & __CMPXCHG_DOUBLE) {
371 if (cmpxchg_double(&page->freelist,
372 freelist_old, counters_old,
373 freelist_new, counters_new))
379 if (page->freelist == freelist_old && page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
389 stat(s, CMPXCHG_DOUBLE_FAIL);
391 #ifdef SLUB_DEBUG_CMPXCHG
392 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
398 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
403 #ifdef CONFIG_CMPXCHG_DOUBLE
404 if (s->flags & __CMPXCHG_DOUBLE) {
405 if (cmpxchg_double(&page->freelist,
406 freelist_old, counters_old,
407 freelist_new, counters_new))
414 local_irq_save(flags);
416 if (page->freelist == freelist_old && page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
420 local_irq_restore(flags);
424 local_irq_restore(flags);
428 stat(s, CMPXCHG_DOUBLE_FAIL);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
447 void *addr = page_address(page);
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
456 #ifdef CONFIG_SLUB_DEBUG_ON
457 static int slub_debug = DEBUG_DEFAULT_FLAGS;
459 static int slub_debug;
462 static char *slub_debug_slabs;
463 static int disable_higher_order_debug;
468 static void print_section(char *text, u8 *addr, unsigned int length)
470 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
474 static struct track *get_track(struct kmem_cache *s, void *object,
475 enum track_item alloc)
480 p = object + s->offset + sizeof(void *);
482 p = object + s->inuse;
487 static void set_track(struct kmem_cache *s, void *object,
488 enum track_item alloc, unsigned long addr)
490 struct track *p = get_track(s, object, alloc);
493 #ifdef CONFIG_STACKTRACE
494 struct stack_trace trace;
497 trace.nr_entries = 0;
498 trace.max_entries = TRACK_ADDRS_COUNT;
499 trace.entries = p->addrs;
501 save_stack_trace(&trace);
503 /* See rant in lockdep.c */
504 if (trace.nr_entries != 0 &&
505 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
508 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
512 p->cpu = smp_processor_id();
513 p->pid = current->pid;
516 memset(p, 0, sizeof(struct track));
519 static void init_tracking(struct kmem_cache *s, void *object)
521 if (!(s->flags & SLAB_STORE_USER))
524 set_track(s, object, TRACK_FREE, 0UL);
525 set_track(s, object, TRACK_ALLOC, 0UL);
528 static void print_track(const char *s, struct track *t)
533 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
534 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
535 #ifdef CONFIG_STACKTRACE
538 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
540 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
547 static void print_tracking(struct kmem_cache *s, void *object)
549 if (!(s->flags & SLAB_STORE_USER))
552 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
553 print_track("Freed", get_track(s, object, TRACK_FREE));
556 static void print_page_info(struct page *page)
558 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
559 page, page->objects, page->inuse, page->freelist, page->flags);
563 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
569 vsnprintf(buf, sizeof(buf), fmt, args);
571 printk(KERN_ERR "========================================"
572 "=====================================\n");
573 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
574 printk(KERN_ERR "----------------------------------------"
575 "-------------------------------------\n\n");
578 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
584 vsnprintf(buf, sizeof(buf), fmt, args);
586 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
589 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
591 unsigned int off; /* Offset of last byte */
592 u8 *addr = page_address(page);
594 print_tracking(s, p);
596 print_page_info(page);
598 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
599 p, p - addr, get_freepointer(s, p));
602 print_section("Bytes b4 ", p - 16, 16);
604 print_section("Object ", p, min_t(unsigned long, s->objsize,
606 if (s->flags & SLAB_RED_ZONE)
607 print_section("Redzone ", p + s->objsize,
608 s->inuse - s->objsize);
611 off = s->offset + sizeof(void *);
615 if (s->flags & SLAB_STORE_USER)
616 off += 2 * sizeof(struct track);
619 /* Beginning of the filler is the free pointer */
620 print_section("Padding ", p + off, s->size - off);
625 static void object_err(struct kmem_cache *s, struct page *page,
626 u8 *object, char *reason)
628 slab_bug(s, "%s", reason);
629 print_trailer(s, page, object);
632 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
638 vsnprintf(buf, sizeof(buf), fmt, args);
640 slab_bug(s, "%s", buf);
641 print_page_info(page);
645 static void init_object(struct kmem_cache *s, void *object, u8 val)
649 if (s->flags & __OBJECT_POISON) {
650 memset(p, POISON_FREE, s->objsize - 1);
651 p[s->objsize - 1] = POISON_END;
654 if (s->flags & SLAB_RED_ZONE)
655 memset(p + s->objsize, val, s->inuse - s->objsize);
658 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
659 void *from, void *to)
661 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
662 memset(from, data, to - from);
665 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
666 u8 *object, char *what,
667 u8 *start, unsigned int value, unsigned int bytes)
672 fault = memchr_inv(start, value, bytes);
677 while (end > fault && end[-1] == value)
680 slab_bug(s, "%s overwritten", what);
681 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
682 fault, end - 1, fault[0], value);
683 print_trailer(s, page, object);
685 restore_bytes(s, what, value, fault, end);
693 * Bytes of the object to be managed.
694 * If the freepointer may overlay the object then the free
695 * pointer is the first word of the object.
697 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
700 * object + s->objsize
701 * Padding to reach word boundary. This is also used for Redzoning.
702 * Padding is extended by another word if Redzoning is enabled and
705 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
706 * 0xcc (RED_ACTIVE) for objects in use.
709 * Meta data starts here.
711 * A. Free pointer (if we cannot overwrite object on free)
712 * B. Tracking data for SLAB_STORE_USER
713 * C. Padding to reach required alignment boundary or at mininum
714 * one word if debugging is on to be able to detect writes
715 * before the word boundary.
717 * Padding is done using 0x5a (POISON_INUSE)
720 * Nothing is used beyond s->size.
722 * If slabcaches are merged then the objsize and inuse boundaries are mostly
723 * ignored. And therefore no slab options that rely on these boundaries
724 * may be used with merged slabcaches.
727 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
729 unsigned long off = s->inuse; /* The end of info */
732 /* Freepointer is placed after the object. */
733 off += sizeof(void *);
735 if (s->flags & SLAB_STORE_USER)
736 /* We also have user information there */
737 off += 2 * sizeof(struct track);
742 return check_bytes_and_report(s, page, p, "Object padding",
743 p + off, POISON_INUSE, s->size - off);
746 /* Check the pad bytes at the end of a slab page */
747 static int slab_pad_check(struct kmem_cache *s, struct page *page)
755 if (!(s->flags & SLAB_POISON))
758 start = page_address(page);
759 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
760 end = start + length;
761 remainder = length % s->size;
765 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
768 while (end > fault && end[-1] == POISON_INUSE)
771 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
772 print_section("Padding ", end - remainder, remainder);
774 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
778 static int check_object(struct kmem_cache *s, struct page *page,
779 void *object, u8 val)
782 u8 *endobject = object + s->objsize;
784 if (s->flags & SLAB_RED_ZONE) {
785 if (!check_bytes_and_report(s, page, object, "Redzone",
786 endobject, val, s->inuse - s->objsize))
789 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
790 check_bytes_and_report(s, page, p, "Alignment padding",
791 endobject, POISON_INUSE, s->inuse - s->objsize);
795 if (s->flags & SLAB_POISON) {
796 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
797 (!check_bytes_and_report(s, page, p, "Poison", p,
798 POISON_FREE, s->objsize - 1) ||
799 !check_bytes_and_report(s, page, p, "Poison",
800 p + s->objsize - 1, POISON_END, 1)))
803 * check_pad_bytes cleans up on its own.
805 check_pad_bytes(s, page, p);
808 if (!s->offset && val == SLUB_RED_ACTIVE)
810 * Object and freepointer overlap. Cannot check
811 * freepointer while object is allocated.
815 /* Check free pointer validity */
816 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
817 object_err(s, page, p, "Freepointer corrupt");
819 * No choice but to zap it and thus lose the remainder
820 * of the free objects in this slab. May cause
821 * another error because the object count is now wrong.
823 set_freepointer(s, p, NULL);
829 static int check_slab(struct kmem_cache *s, struct page *page)
833 VM_BUG_ON(!irqs_disabled());
835 if (!PageSlab(page)) {
836 slab_err(s, page, "Not a valid slab page");
840 maxobj = order_objects(compound_order(page), s->size, s->reserved);
841 if (page->objects > maxobj) {
842 slab_err(s, page, "objects %u > max %u",
843 s->name, page->objects, maxobj);
846 if (page->inuse > page->objects) {
847 slab_err(s, page, "inuse %u > max %u",
848 s->name, page->inuse, page->objects);
851 /* Slab_pad_check fixes things up after itself */
852 slab_pad_check(s, page);
857 * Determine if a certain object on a page is on the freelist. Must hold the
858 * slab lock to guarantee that the chains are in a consistent state.
860 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
865 unsigned long max_objects;
868 while (fp && nr <= page->objects) {
871 if (!check_valid_pointer(s, page, fp)) {
873 object_err(s, page, object,
874 "Freechain corrupt");
875 set_freepointer(s, object, NULL);
878 slab_err(s, page, "Freepointer corrupt");
879 page->freelist = NULL;
880 page->inuse = page->objects;
881 slab_fix(s, "Freelist cleared");
887 fp = get_freepointer(s, object);
891 max_objects = order_objects(compound_order(page), s->size, s->reserved);
892 if (max_objects > MAX_OBJS_PER_PAGE)
893 max_objects = MAX_OBJS_PER_PAGE;
895 if (page->objects != max_objects) {
896 slab_err(s, page, "Wrong number of objects. Found %d but "
897 "should be %d", page->objects, max_objects);
898 page->objects = max_objects;
899 slab_fix(s, "Number of objects adjusted.");
901 if (page->inuse != page->objects - nr) {
902 slab_err(s, page, "Wrong object count. Counter is %d but "
903 "counted were %d", page->inuse, page->objects - nr);
904 page->inuse = page->objects - nr;
905 slab_fix(s, "Object count adjusted.");
907 return search == NULL;
910 static void trace(struct kmem_cache *s, struct page *page, void *object,
913 if (s->flags & SLAB_TRACE) {
914 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
916 alloc ? "alloc" : "free",
921 print_section("Object ", (void *)object, s->objsize);
928 * Hooks for other subsystems that check memory allocations. In a typical
929 * production configuration these hooks all should produce no code at all.
931 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
933 flags &= gfp_allowed_mask;
934 lockdep_trace_alloc(flags);
935 might_sleep_if(flags & __GFP_WAIT);
937 return should_failslab(s->objsize, flags, s->flags);
940 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
942 flags &= gfp_allowed_mask;
943 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
944 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
947 static inline void slab_free_hook(struct kmem_cache *s, void *x)
949 kmemleak_free_recursive(x, s->flags);
952 * Trouble is that we may no longer disable interupts in the fast path
953 * So in order to make the debug calls that expect irqs to be
954 * disabled we need to disable interrupts temporarily.
956 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
960 local_irq_save(flags);
961 kmemcheck_slab_free(s, x, s->objsize);
962 debug_check_no_locks_freed(x, s->objsize);
963 local_irq_restore(flags);
966 if (!(s->flags & SLAB_DEBUG_OBJECTS))
967 debug_check_no_obj_freed(x, s->objsize);
971 * Tracking of fully allocated slabs for debugging purposes.
973 * list_lock must be held.
975 static void add_full(struct kmem_cache *s,
976 struct kmem_cache_node *n, struct page *page)
978 if (!(s->flags & SLAB_STORE_USER))
981 list_add(&page->lru, &n->full);
985 * list_lock must be held.
987 static void remove_full(struct kmem_cache *s, struct page *page)
989 if (!(s->flags & SLAB_STORE_USER))
992 list_del(&page->lru);
995 /* Tracking of the number of slabs for debugging purposes */
996 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
998 struct kmem_cache_node *n = get_node(s, node);
1000 return atomic_long_read(&n->nr_slabs);
1003 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1005 return atomic_long_read(&n->nr_slabs);
1008 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1010 struct kmem_cache_node *n = get_node(s, node);
1013 * May be called early in order to allocate a slab for the
1014 * kmem_cache_node structure. Solve the chicken-egg
1015 * dilemma by deferring the increment of the count during
1016 * bootstrap (see early_kmem_cache_node_alloc).
1019 atomic_long_inc(&n->nr_slabs);
1020 atomic_long_add(objects, &n->total_objects);
1023 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1025 struct kmem_cache_node *n = get_node(s, node);
1027 atomic_long_dec(&n->nr_slabs);
1028 atomic_long_sub(objects, &n->total_objects);
1031 /* Object debug checks for alloc/free paths */
1032 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1035 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1038 init_object(s, object, SLUB_RED_INACTIVE);
1039 init_tracking(s, object);
1042 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1043 void *object, unsigned long addr)
1045 if (!check_slab(s, page))
1048 if (!check_valid_pointer(s, page, object)) {
1049 object_err(s, page, object, "Freelist Pointer check fails");
1053 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1056 /* Success perform special debug activities for allocs */
1057 if (s->flags & SLAB_STORE_USER)
1058 set_track(s, object, TRACK_ALLOC, addr);
1059 trace(s, page, object, 1);
1060 init_object(s, object, SLUB_RED_ACTIVE);
1064 if (PageSlab(page)) {
1066 * If this is a slab page then lets do the best we can
1067 * to avoid issues in the future. Marking all objects
1068 * as used avoids touching the remaining objects.
1070 slab_fix(s, "Marking all objects used");
1071 page->inuse = page->objects;
1072 page->freelist = NULL;
1077 static noinline int free_debug_processing(struct kmem_cache *s,
1078 struct page *page, void *object, unsigned long addr)
1080 unsigned long flags;
1083 local_irq_save(flags);
1086 if (!check_slab(s, page))
1089 if (!check_valid_pointer(s, page, object)) {
1090 slab_err(s, page, "Invalid object pointer 0x%p", object);
1094 if (on_freelist(s, page, object)) {
1095 object_err(s, page, object, "Object already free");
1099 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1102 if (unlikely(s != page->slab)) {
1103 if (!PageSlab(page)) {
1104 slab_err(s, page, "Attempt to free object(0x%p) "
1105 "outside of slab", object);
1106 } else if (!page->slab) {
1108 "SLUB <none>: no slab for object 0x%p.\n",
1112 object_err(s, page, object,
1113 "page slab pointer corrupt.");
1117 if (s->flags & SLAB_STORE_USER)
1118 set_track(s, object, TRACK_FREE, addr);
1119 trace(s, page, object, 0);
1120 init_object(s, object, SLUB_RED_INACTIVE);
1124 local_irq_restore(flags);
1128 slab_fix(s, "Object at 0x%p not freed", object);
1132 static int __init setup_slub_debug(char *str)
1134 slub_debug = DEBUG_DEFAULT_FLAGS;
1135 if (*str++ != '=' || !*str)
1137 * No options specified. Switch on full debugging.
1143 * No options but restriction on slabs. This means full
1144 * debugging for slabs matching a pattern.
1148 if (tolower(*str) == 'o') {
1150 * Avoid enabling debugging on caches if its minimum order
1151 * would increase as a result.
1153 disable_higher_order_debug = 1;
1160 * Switch off all debugging measures.
1165 * Determine which debug features should be switched on
1167 for (; *str && *str != ','; str++) {
1168 switch (tolower(*str)) {
1170 slub_debug |= SLAB_DEBUG_FREE;
1173 slub_debug |= SLAB_RED_ZONE;
1176 slub_debug |= SLAB_POISON;
1179 slub_debug |= SLAB_STORE_USER;
1182 slub_debug |= SLAB_TRACE;
1185 slub_debug |= SLAB_FAILSLAB;
1188 printk(KERN_ERR "slub_debug option '%c' "
1189 "unknown. skipped\n", *str);
1195 slub_debug_slabs = str + 1;
1200 __setup("slub_debug", setup_slub_debug);
1202 static unsigned long kmem_cache_flags(unsigned long objsize,
1203 unsigned long flags, const char *name,
1204 void (*ctor)(void *))
1207 * Enable debugging if selected on the kernel commandline.
1209 if (slub_debug && (!slub_debug_slabs ||
1210 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1211 flags |= slub_debug;
1216 static inline void setup_object_debug(struct kmem_cache *s,
1217 struct page *page, void *object) {}
1219 static inline int alloc_debug_processing(struct kmem_cache *s,
1220 struct page *page, void *object, unsigned long addr) { return 0; }
1222 static inline int free_debug_processing(struct kmem_cache *s,
1223 struct page *page, void *object, unsigned long addr) { return 0; }
1225 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1227 static inline int check_object(struct kmem_cache *s, struct page *page,
1228 void *object, u8 val) { return 1; }
1229 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1230 struct page *page) {}
1231 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1232 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1233 unsigned long flags, const char *name,
1234 void (*ctor)(void *))
1238 #define slub_debug 0
1240 #define disable_higher_order_debug 0
1242 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1244 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1246 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1248 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1251 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1254 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1257 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1259 #endif /* CONFIG_SLUB_DEBUG */
1262 * Slab allocation and freeing
1264 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1265 struct kmem_cache_order_objects oo)
1267 int order = oo_order(oo);
1269 flags |= __GFP_NOTRACK;
1271 if (node == NUMA_NO_NODE)
1272 return alloc_pages(flags, order);
1274 return alloc_pages_exact_node(node, flags, order);
1277 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1280 struct kmem_cache_order_objects oo = s->oo;
1283 flags &= gfp_allowed_mask;
1285 if (flags & __GFP_WAIT)
1288 flags |= s->allocflags;
1291 * Let the initial higher-order allocation fail under memory pressure
1292 * so we fall-back to the minimum order allocation.
1294 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1296 page = alloc_slab_page(alloc_gfp, node, oo);
1297 if (unlikely(!page)) {
1300 * Allocation may have failed due to fragmentation.
1301 * Try a lower order alloc if possible
1303 page = alloc_slab_page(flags, node, oo);
1306 stat(s, ORDER_FALLBACK);
1309 if (flags & __GFP_WAIT)
1310 local_irq_disable();
1315 if (kmemcheck_enabled
1316 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1317 int pages = 1 << oo_order(oo);
1319 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1322 * Objects from caches that have a constructor don't get
1323 * cleared when they're allocated, so we need to do it here.
1326 kmemcheck_mark_uninitialized_pages(page, pages);
1328 kmemcheck_mark_unallocated_pages(page, pages);
1331 page->objects = oo_objects(oo);
1332 mod_zone_page_state(page_zone(page),
1333 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1334 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1340 static void setup_object(struct kmem_cache *s, struct page *page,
1343 setup_object_debug(s, page, object);
1344 if (unlikely(s->ctor))
1348 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1355 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1357 page = allocate_slab(s,
1358 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1362 inc_slabs_node(s, page_to_nid(page), page->objects);
1364 page->flags |= 1 << PG_slab;
1366 start = page_address(page);
1368 if (unlikely(s->flags & SLAB_POISON))
1369 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1372 for_each_object(p, s, start, page->objects) {
1373 setup_object(s, page, last);
1374 set_freepointer(s, last, p);
1377 setup_object(s, page, last);
1378 set_freepointer(s, last, NULL);
1380 page->freelist = start;
1381 page->inuse = page->objects;
1387 static void __free_slab(struct kmem_cache *s, struct page *page)
1389 int order = compound_order(page);
1390 int pages = 1 << order;
1392 if (kmem_cache_debug(s)) {
1395 slab_pad_check(s, page);
1396 for_each_object(p, s, page_address(page),
1398 check_object(s, page, p, SLUB_RED_INACTIVE);
1401 kmemcheck_free_shadow(page, compound_order(page));
1403 mod_zone_page_state(page_zone(page),
1404 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1405 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1408 __ClearPageSlab(page);
1409 reset_page_mapcount(page);
1410 if (current->reclaim_state)
1411 current->reclaim_state->reclaimed_slab += pages;
1412 __free_pages(page, order);
1415 #define need_reserve_slab_rcu \
1416 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1418 static void rcu_free_slab(struct rcu_head *h)
1422 if (need_reserve_slab_rcu)
1423 page = virt_to_head_page(h);
1425 page = container_of((struct list_head *)h, struct page, lru);
1427 __free_slab(page->slab, page);
1430 static void free_slab(struct kmem_cache *s, struct page *page)
1432 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1433 struct rcu_head *head;
1435 if (need_reserve_slab_rcu) {
1436 int order = compound_order(page);
1437 int offset = (PAGE_SIZE << order) - s->reserved;
1439 VM_BUG_ON(s->reserved != sizeof(*head));
1440 head = page_address(page) + offset;
1443 * RCU free overloads the RCU head over the LRU
1445 head = (void *)&page->lru;
1448 call_rcu(head, rcu_free_slab);
1450 __free_slab(s, page);
1453 static void discard_slab(struct kmem_cache *s, struct page *page)
1455 dec_slabs_node(s, page_to_nid(page), page->objects);
1460 * Management of partially allocated slabs.
1462 * list_lock must be held.
1464 static inline void add_partial(struct kmem_cache_node *n,
1465 struct page *page, int tail)
1468 if (tail == DEACTIVATE_TO_TAIL)
1469 list_add_tail(&page->lru, &n->partial);
1471 list_add(&page->lru, &n->partial);
1475 * list_lock must be held.
1477 static inline void remove_partial(struct kmem_cache_node *n,
1480 list_del(&page->lru);
1485 * Lock slab, remove from the partial list and put the object into the
1488 * Returns a list of objects or NULL if it fails.
1490 * Must hold list_lock.
1492 static inline void *acquire_slab(struct kmem_cache *s,
1493 struct kmem_cache_node *n, struct page *page,
1497 unsigned long counters;
1501 * Zap the freelist and set the frozen bit.
1502 * The old freelist is the list of objects for the
1503 * per cpu allocation list.
1506 freelist = page->freelist;
1507 counters = page->counters;
1508 new.counters = counters;
1510 new.inuse = page->objects;
1512 VM_BUG_ON(new.frozen);
1515 } while (!__cmpxchg_double_slab(s, page,
1518 "lock and freeze"));
1520 remove_partial(n, page);
1524 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1527 * Try to allocate a partial slab from a specific node.
1529 static void *get_partial_node(struct kmem_cache *s,
1530 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1532 struct page *page, *page2;
1533 void *object = NULL;
1536 * Racy check. If we mistakenly see no partial slabs then we
1537 * just allocate an empty slab. If we mistakenly try to get a
1538 * partial slab and there is none available then get_partials()
1541 if (!n || !n->nr_partial)
1544 spin_lock(&n->list_lock);
1545 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1546 void *t = acquire_slab(s, n, page, object == NULL);
1554 c->node = page_to_nid(page);
1555 stat(s, ALLOC_FROM_PARTIAL);
1557 available = page->objects - page->inuse;
1560 available = put_cpu_partial(s, page, 0);
1562 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1566 spin_unlock(&n->list_lock);
1571 * Get a page from somewhere. Search in increasing NUMA distances.
1573 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1574 struct kmem_cache_cpu *c)
1577 struct zonelist *zonelist;
1580 enum zone_type high_zoneidx = gfp_zone(flags);
1584 * The defrag ratio allows a configuration of the tradeoffs between
1585 * inter node defragmentation and node local allocations. A lower
1586 * defrag_ratio increases the tendency to do local allocations
1587 * instead of attempting to obtain partial slabs from other nodes.
1589 * If the defrag_ratio is set to 0 then kmalloc() always
1590 * returns node local objects. If the ratio is higher then kmalloc()
1591 * may return off node objects because partial slabs are obtained
1592 * from other nodes and filled up.
1594 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1595 * defrag_ratio = 1000) then every (well almost) allocation will
1596 * first attempt to defrag slab caches on other nodes. This means
1597 * scanning over all nodes to look for partial slabs which may be
1598 * expensive if we do it every time we are trying to find a slab
1599 * with available objects.
1601 if (!s->remote_node_defrag_ratio ||
1602 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1606 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1607 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1608 struct kmem_cache_node *n;
1610 n = get_node(s, zone_to_nid(zone));
1612 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1613 n->nr_partial > s->min_partial) {
1614 object = get_partial_node(s, n, c);
1627 * Get a partial page, lock it and return it.
1629 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1630 struct kmem_cache_cpu *c)
1633 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1635 object = get_partial_node(s, get_node(s, searchnode), c);
1636 if (object || node != NUMA_NO_NODE)
1639 return get_any_partial(s, flags, c);
1642 #ifdef CONFIG_PREEMPT
1644 * Calculate the next globally unique transaction for disambiguiation
1645 * during cmpxchg. The transactions start with the cpu number and are then
1646 * incremented by CONFIG_NR_CPUS.
1648 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1651 * No preemption supported therefore also no need to check for
1657 static inline unsigned long next_tid(unsigned long tid)
1659 return tid + TID_STEP;
1662 static inline unsigned int tid_to_cpu(unsigned long tid)
1664 return tid % TID_STEP;
1667 static inline unsigned long tid_to_event(unsigned long tid)
1669 return tid / TID_STEP;
1672 static inline unsigned int init_tid(int cpu)
1677 static inline void note_cmpxchg_failure(const char *n,
1678 const struct kmem_cache *s, unsigned long tid)
1680 #ifdef SLUB_DEBUG_CMPXCHG
1681 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1683 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1685 #ifdef CONFIG_PREEMPT
1686 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1687 printk("due to cpu change %d -> %d\n",
1688 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1691 if (tid_to_event(tid) != tid_to_event(actual_tid))
1692 printk("due to cpu running other code. Event %ld->%ld\n",
1693 tid_to_event(tid), tid_to_event(actual_tid));
1695 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1696 actual_tid, tid, next_tid(tid));
1698 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1701 void init_kmem_cache_cpus(struct kmem_cache *s)
1705 for_each_possible_cpu(cpu)
1706 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1710 * Remove the cpu slab
1712 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1714 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1715 struct page *page = c->page;
1716 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1718 enum slab_modes l = M_NONE, m = M_NONE;
1721 int tail = DEACTIVATE_TO_HEAD;
1725 if (page->freelist) {
1726 stat(s, DEACTIVATE_REMOTE_FREES);
1727 tail = DEACTIVATE_TO_TAIL;
1730 c->tid = next_tid(c->tid);
1732 freelist = c->freelist;
1736 * Stage one: Free all available per cpu objects back
1737 * to the page freelist while it is still frozen. Leave the
1740 * There is no need to take the list->lock because the page
1743 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1745 unsigned long counters;
1748 prior = page->freelist;
1749 counters = page->counters;
1750 set_freepointer(s, freelist, prior);
1751 new.counters = counters;
1753 VM_BUG_ON(!new.frozen);
1755 } while (!__cmpxchg_double_slab(s, page,
1757 freelist, new.counters,
1758 "drain percpu freelist"));
1760 freelist = nextfree;
1764 * Stage two: Ensure that the page is unfrozen while the
1765 * list presence reflects the actual number of objects
1768 * We setup the list membership and then perform a cmpxchg
1769 * with the count. If there is a mismatch then the page
1770 * is not unfrozen but the page is on the wrong list.
1772 * Then we restart the process which may have to remove
1773 * the page from the list that we just put it on again
1774 * because the number of objects in the slab may have
1779 old.freelist = page->freelist;
1780 old.counters = page->counters;
1781 VM_BUG_ON(!old.frozen);
1783 /* Determine target state of the slab */
1784 new.counters = old.counters;
1787 set_freepointer(s, freelist, old.freelist);
1788 new.freelist = freelist;
1790 new.freelist = old.freelist;
1794 if (!new.inuse && n->nr_partial > s->min_partial)
1796 else if (new.freelist) {
1801 * Taking the spinlock removes the possiblity
1802 * that acquire_slab() will see a slab page that
1805 spin_lock(&n->list_lock);
1809 if (kmem_cache_debug(s) && !lock) {
1812 * This also ensures that the scanning of full
1813 * slabs from diagnostic functions will not see
1816 spin_lock(&n->list_lock);
1824 remove_partial(n, page);
1826 else if (l == M_FULL)
1828 remove_full(s, page);
1830 if (m == M_PARTIAL) {
1832 add_partial(n, page, tail);
1835 } else if (m == M_FULL) {
1837 stat(s, DEACTIVATE_FULL);
1838 add_full(s, n, page);
1844 if (!__cmpxchg_double_slab(s, page,
1845 old.freelist, old.counters,
1846 new.freelist, new.counters,
1851 spin_unlock(&n->list_lock);
1854 stat(s, DEACTIVATE_EMPTY);
1855 discard_slab(s, page);
1860 /* Unfreeze all the cpu partial slabs */
1861 static void unfreeze_partials(struct kmem_cache *s)
1863 struct kmem_cache_node *n = NULL;
1864 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1865 struct page *page, *discard_page = NULL;
1867 while ((page = c->partial)) {
1868 enum slab_modes { M_PARTIAL, M_FREE };
1869 enum slab_modes l, m;
1873 c->partial = page->next;
1878 old.freelist = page->freelist;
1879 old.counters = page->counters;
1880 VM_BUG_ON(!old.frozen);
1882 new.counters = old.counters;
1883 new.freelist = old.freelist;
1887 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1890 struct kmem_cache_node *n2 = get_node(s,
1896 spin_unlock(&n->list_lock);
1899 spin_lock(&n->list_lock);
1905 remove_partial(n, page);
1907 add_partial(n, page,
1908 DEACTIVATE_TO_TAIL);
1913 } while (!cmpxchg_double_slab(s, page,
1914 old.freelist, old.counters,
1915 new.freelist, new.counters,
1916 "unfreezing slab"));
1919 page->next = discard_page;
1920 discard_page = page;
1925 spin_unlock(&n->list_lock);
1927 while (discard_page) {
1928 page = discard_page;
1929 discard_page = discard_page->next;
1931 stat(s, DEACTIVATE_EMPTY);
1932 discard_slab(s, page);
1938 * Put a page that was just frozen (in __slab_free) into a partial page
1939 * slot if available. This is done without interrupts disabled and without
1940 * preemption disabled. The cmpxchg is racy and may put the partial page
1941 * onto a random cpus partial slot.
1943 * If we did not find a slot then simply move all the partials to the
1944 * per node partial list.
1946 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1948 struct page *oldpage;
1955 oldpage = this_cpu_read(s->cpu_slab->partial);
1958 pobjects = oldpage->pobjects;
1959 pages = oldpage->pages;
1960 if (drain && pobjects > s->cpu_partial) {
1961 unsigned long flags;
1963 * partial array is full. Move the existing
1964 * set to the per node partial list.
1966 local_irq_save(flags);
1967 unfreeze_partials(s);
1968 local_irq_restore(flags);
1975 pobjects += page->objects - page->inuse;
1977 page->pages = pages;
1978 page->pobjects = pobjects;
1979 page->next = oldpage;
1981 } while (irqsafe_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1982 stat(s, CPU_PARTIAL_FREE);
1986 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1988 stat(s, CPUSLAB_FLUSH);
1989 deactivate_slab(s, c);
1995 * Called from IPI handler with interrupts disabled.
1997 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1999 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2005 unfreeze_partials(s);
2009 static void flush_cpu_slab(void *d)
2011 struct kmem_cache *s = d;
2013 __flush_cpu_slab(s, smp_processor_id());
2016 static void flush_all(struct kmem_cache *s)
2018 on_each_cpu(flush_cpu_slab, s, 1);
2022 * Check if the objects in a per cpu structure fit numa
2023 * locality expectations.
2025 static inline int node_match(struct kmem_cache_cpu *c, int node)
2028 if (node != NUMA_NO_NODE && c->node != node)
2034 static int count_free(struct page *page)
2036 return page->objects - page->inuse;
2039 static unsigned long count_partial(struct kmem_cache_node *n,
2040 int (*get_count)(struct page *))
2042 unsigned long flags;
2043 unsigned long x = 0;
2046 spin_lock_irqsave(&n->list_lock, flags);
2047 list_for_each_entry(page, &n->partial, lru)
2048 x += get_count(page);
2049 spin_unlock_irqrestore(&n->list_lock, flags);
2053 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2055 #ifdef CONFIG_SLUB_DEBUG
2056 return atomic_long_read(&n->total_objects);
2062 static noinline void
2063 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2068 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2070 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2071 "default order: %d, min order: %d\n", s->name, s->objsize,
2072 s->size, oo_order(s->oo), oo_order(s->min));
2074 if (oo_order(s->min) > get_order(s->objsize))
2075 printk(KERN_WARNING " %s debugging increased min order, use "
2076 "slub_debug=O to disable.\n", s->name);
2078 for_each_online_node(node) {
2079 struct kmem_cache_node *n = get_node(s, node);
2080 unsigned long nr_slabs;
2081 unsigned long nr_objs;
2082 unsigned long nr_free;
2087 nr_free = count_partial(n, count_free);
2088 nr_slabs = node_nr_slabs(n);
2089 nr_objs = node_nr_objs(n);
2092 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2093 node, nr_slabs, nr_objs, nr_free);
2097 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2098 int node, struct kmem_cache_cpu **pc)
2101 struct kmem_cache_cpu *c;
2102 struct page *page = new_slab(s, flags, node);
2105 c = __this_cpu_ptr(s->cpu_slab);
2110 * No other reference to the page yet so we can
2111 * muck around with it freely without cmpxchg
2113 object = page->freelist;
2114 page->freelist = NULL;
2116 stat(s, ALLOC_SLAB);
2117 c->node = page_to_nid(page);
2127 * Slow path. The lockless freelist is empty or we need to perform
2130 * Processing is still very fast if new objects have been freed to the
2131 * regular freelist. In that case we simply take over the regular freelist
2132 * as the lockless freelist and zap the regular freelist.
2134 * If that is not working then we fall back to the partial lists. We take the
2135 * first element of the freelist as the object to allocate now and move the
2136 * rest of the freelist to the lockless freelist.
2138 * And if we were unable to get a new slab from the partial slab lists then
2139 * we need to allocate a new slab. This is the slowest path since it involves
2140 * a call to the page allocator and the setup of a new slab.
2142 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2143 unsigned long addr, struct kmem_cache_cpu *c)
2146 unsigned long flags;
2148 unsigned long counters;
2150 local_irq_save(flags);
2151 #ifdef CONFIG_PREEMPT
2153 * We may have been preempted and rescheduled on a different
2154 * cpu before disabling interrupts. Need to reload cpu area
2157 c = this_cpu_ptr(s->cpu_slab);
2163 if (unlikely(!node_match(c, node))) {
2164 stat(s, ALLOC_NODE_MISMATCH);
2165 deactivate_slab(s, c);
2169 stat(s, ALLOC_SLOWPATH);
2172 object = c->page->freelist;
2173 counters = c->page->counters;
2174 new.counters = counters;
2175 VM_BUG_ON(!new.frozen);
2178 * If there is no object left then we use this loop to
2179 * deactivate the slab which is simple since no objects
2180 * are left in the slab and therefore we do not need to
2181 * put the page back onto the partial list.
2183 * If there are objects left then we retrieve them
2184 * and use them to refill the per cpu queue.
2187 new.inuse = c->page->objects;
2188 new.frozen = object != NULL;
2190 } while (!__cmpxchg_double_slab(s, c->page,
2197 stat(s, DEACTIVATE_BYPASS);
2201 stat(s, ALLOC_REFILL);
2204 c->freelist = get_freepointer(s, object);
2205 c->tid = next_tid(c->tid);
2206 local_irq_restore(flags);
2212 c->page = c->partial;
2213 c->partial = c->page->next;
2214 c->node = page_to_nid(c->page);
2215 stat(s, CPU_PARTIAL_ALLOC);
2220 /* Then do expensive stuff like retrieving pages from the partial lists */
2221 object = get_partial(s, gfpflags, node, c);
2223 if (unlikely(!object)) {
2225 object = new_slab_objects(s, gfpflags, node, &c);
2227 if (unlikely(!object)) {
2228 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2229 slab_out_of_memory(s, gfpflags, node);
2231 local_irq_restore(flags);
2236 if (likely(!kmem_cache_debug(s)))
2239 /* Only entered in the debug case */
2240 if (!alloc_debug_processing(s, c->page, object, addr))
2241 goto new_slab; /* Slab failed checks. Next slab needed */
2243 c->freelist = get_freepointer(s, object);
2244 deactivate_slab(s, c);
2245 c->node = NUMA_NO_NODE;
2246 local_irq_restore(flags);
2251 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2252 * have the fastpath folded into their functions. So no function call
2253 * overhead for requests that can be satisfied on the fastpath.
2255 * The fastpath works by first checking if the lockless freelist can be used.
2256 * If not then __slab_alloc is called for slow processing.
2258 * Otherwise we can simply pick the next object from the lockless free list.
2260 static __always_inline void *slab_alloc(struct kmem_cache *s,
2261 gfp_t gfpflags, int node, unsigned long addr)
2264 struct kmem_cache_cpu *c;
2267 if (slab_pre_alloc_hook(s, gfpflags))
2273 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2274 * enabled. We may switch back and forth between cpus while
2275 * reading from one cpu area. That does not matter as long
2276 * as we end up on the original cpu again when doing the cmpxchg.
2278 c = __this_cpu_ptr(s->cpu_slab);
2281 * The transaction ids are globally unique per cpu and per operation on
2282 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2283 * occurs on the right processor and that there was no operation on the
2284 * linked list in between.
2289 object = c->freelist;
2290 if (unlikely(!object || !node_match(c, node)))
2292 object = __slab_alloc(s, gfpflags, node, addr, c);
2296 * The cmpxchg will only match if there was no additional
2297 * operation and if we are on the right processor.
2299 * The cmpxchg does the following atomically (without lock semantics!)
2300 * 1. Relocate first pointer to the current per cpu area.
2301 * 2. Verify that tid and freelist have not been changed
2302 * 3. If they were not changed replace tid and freelist
2304 * Since this is without lock semantics the protection is only against
2305 * code executing on this cpu *not* from access by other cpus.
2307 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2308 s->cpu_slab->freelist, s->cpu_slab->tid,
2310 get_freepointer_safe(s, object), next_tid(tid)))) {
2312 note_cmpxchg_failure("slab_alloc", s, tid);
2315 stat(s, ALLOC_FASTPATH);
2318 if (unlikely(gfpflags & __GFP_ZERO) && object)
2319 memset(object, 0, s->objsize);
2321 slab_post_alloc_hook(s, gfpflags, object);
2326 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2328 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2330 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2334 EXPORT_SYMBOL(kmem_cache_alloc);
2336 #ifdef CONFIG_TRACING
2337 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2339 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2340 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2343 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2345 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2347 void *ret = kmalloc_order(size, flags, order);
2348 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2351 EXPORT_SYMBOL(kmalloc_order_trace);
2355 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2357 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2359 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2360 s->objsize, s->size, gfpflags, node);
2364 EXPORT_SYMBOL(kmem_cache_alloc_node);
2366 #ifdef CONFIG_TRACING
2367 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2369 int node, size_t size)
2371 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2373 trace_kmalloc_node(_RET_IP_, ret,
2374 size, s->size, gfpflags, node);
2377 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2382 * Slow patch handling. This may still be called frequently since objects
2383 * have a longer lifetime than the cpu slabs in most processing loads.
2385 * So we still attempt to reduce cache line usage. Just take the slab
2386 * lock and free the item. If there is no additional partial page
2387 * handling required then we can return immediately.
2389 static void __slab_free(struct kmem_cache *s, struct page *page,
2390 void *x, unsigned long addr)
2393 void **object = (void *)x;
2397 unsigned long counters;
2398 struct kmem_cache_node *n = NULL;
2399 unsigned long uninitialized_var(flags);
2401 stat(s, FREE_SLOWPATH);
2403 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2407 prior = page->freelist;
2408 counters = page->counters;
2409 set_freepointer(s, object, prior);
2410 new.counters = counters;
2411 was_frozen = new.frozen;
2413 if ((!new.inuse || !prior) && !was_frozen && !n) {
2415 if (!kmem_cache_debug(s) && !prior)
2418 * Slab was on no list before and will be partially empty
2419 * We can defer the list move and instead freeze it.
2423 else { /* Needs to be taken off a list */
2425 n = get_node(s, page_to_nid(page));
2427 * Speculatively acquire the list_lock.
2428 * If the cmpxchg does not succeed then we may
2429 * drop the list_lock without any processing.
2431 * Otherwise the list_lock will synchronize with
2432 * other processors updating the list of slabs.
2434 spin_lock_irqsave(&n->list_lock, flags);
2440 } while (!cmpxchg_double_slab(s, page,
2442 object, new.counters,
2448 * If we just froze the page then put it onto the
2449 * per cpu partial list.
2451 if (new.frozen && !was_frozen)
2452 put_cpu_partial(s, page, 1);
2455 * The list lock was not taken therefore no list
2456 * activity can be necessary.
2459 stat(s, FREE_FROZEN);
2464 * was_frozen may have been set after we acquired the list_lock in
2465 * an earlier loop. So we need to check it here again.
2468 stat(s, FREE_FROZEN);
2470 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2474 * Objects left in the slab. If it was not on the partial list before
2477 if (unlikely(!prior)) {
2478 remove_full(s, page);
2479 add_partial(n, page, DEACTIVATE_TO_TAIL);
2480 stat(s, FREE_ADD_PARTIAL);
2483 spin_unlock_irqrestore(&n->list_lock, flags);
2489 * Slab on the partial list.
2491 remove_partial(n, page);
2492 stat(s, FREE_REMOVE_PARTIAL);
2494 /* Slab must be on the full list */
2495 remove_full(s, page);
2497 spin_unlock_irqrestore(&n->list_lock, flags);
2499 discard_slab(s, page);
2503 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2504 * can perform fastpath freeing without additional function calls.
2506 * The fastpath is only possible if we are freeing to the current cpu slab
2507 * of this processor. This typically the case if we have just allocated
2510 * If fastpath is not possible then fall back to __slab_free where we deal
2511 * with all sorts of special processing.
2513 static __always_inline void slab_free(struct kmem_cache *s,
2514 struct page *page, void *x, unsigned long addr)
2516 void **object = (void *)x;
2517 struct kmem_cache_cpu *c;
2520 slab_free_hook(s, x);
2524 * Determine the currently cpus per cpu slab.
2525 * The cpu may change afterward. However that does not matter since
2526 * data is retrieved via this pointer. If we are on the same cpu
2527 * during the cmpxchg then the free will succedd.
2529 c = __this_cpu_ptr(s->cpu_slab);
2534 if (likely(page == c->page)) {
2535 set_freepointer(s, object, c->freelist);
2537 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2538 s->cpu_slab->freelist, s->cpu_slab->tid,
2540 object, next_tid(tid)))) {
2542 note_cmpxchg_failure("slab_free", s, tid);
2545 stat(s, FREE_FASTPATH);
2547 __slab_free(s, page, x, addr);
2551 void kmem_cache_free(struct kmem_cache *s, void *x)
2555 page = virt_to_head_page(x);
2557 slab_free(s, page, x, _RET_IP_);
2559 trace_kmem_cache_free(_RET_IP_, x);
2561 EXPORT_SYMBOL(kmem_cache_free);
2564 * Object placement in a slab is made very easy because we always start at
2565 * offset 0. If we tune the size of the object to the alignment then we can
2566 * get the required alignment by putting one properly sized object after
2569 * Notice that the allocation order determines the sizes of the per cpu
2570 * caches. Each processor has always one slab available for allocations.
2571 * Increasing the allocation order reduces the number of times that slabs
2572 * must be moved on and off the partial lists and is therefore a factor in
2577 * Mininum / Maximum order of slab pages. This influences locking overhead
2578 * and slab fragmentation. A higher order reduces the number of partial slabs
2579 * and increases the number of allocations possible without having to
2580 * take the list_lock.
2582 static int slub_min_order;
2583 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2584 static int slub_min_objects;
2587 * Merge control. If this is set then no merging of slab caches will occur.
2588 * (Could be removed. This was introduced to pacify the merge skeptics.)
2590 static int slub_nomerge;
2593 * Calculate the order of allocation given an slab object size.
2595 * The order of allocation has significant impact on performance and other
2596 * system components. Generally order 0 allocations should be preferred since
2597 * order 0 does not cause fragmentation in the page allocator. Larger objects
2598 * be problematic to put into order 0 slabs because there may be too much
2599 * unused space left. We go to a higher order if more than 1/16th of the slab
2602 * In order to reach satisfactory performance we must ensure that a minimum
2603 * number of objects is in one slab. Otherwise we may generate too much
2604 * activity on the partial lists which requires taking the list_lock. This is
2605 * less a concern for large slabs though which are rarely used.
2607 * slub_max_order specifies the order where we begin to stop considering the
2608 * number of objects in a slab as critical. If we reach slub_max_order then
2609 * we try to keep the page order as low as possible. So we accept more waste
2610 * of space in favor of a small page order.
2612 * Higher order allocations also allow the placement of more objects in a
2613 * slab and thereby reduce object handling overhead. If the user has
2614 * requested a higher mininum order then we start with that one instead of
2615 * the smallest order which will fit the object.
2617 static inline int slab_order(int size, int min_objects,
2618 int max_order, int fract_leftover, int reserved)
2622 int min_order = slub_min_order;
2624 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2625 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2627 for (order = max(min_order,
2628 fls(min_objects * size - 1) - PAGE_SHIFT);
2629 order <= max_order; order++) {
2631 unsigned long slab_size = PAGE_SIZE << order;
2633 if (slab_size < min_objects * size + reserved)
2636 rem = (slab_size - reserved) % size;
2638 if (rem <= slab_size / fract_leftover)
2646 static inline int calculate_order(int size, int reserved)
2654 * Attempt to find best configuration for a slab. This
2655 * works by first attempting to generate a layout with
2656 * the best configuration and backing off gradually.
2658 * First we reduce the acceptable waste in a slab. Then
2659 * we reduce the minimum objects required in a slab.
2661 min_objects = slub_min_objects;
2663 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2664 max_objects = order_objects(slub_max_order, size, reserved);
2665 min_objects = min(min_objects, max_objects);
2667 while (min_objects > 1) {
2669 while (fraction >= 4) {
2670 order = slab_order(size, min_objects,
2671 slub_max_order, fraction, reserved);
2672 if (order <= slub_max_order)
2680 * We were unable to place multiple objects in a slab. Now
2681 * lets see if we can place a single object there.
2683 order = slab_order(size, 1, slub_max_order, 1, reserved);
2684 if (order <= slub_max_order)
2688 * Doh this slab cannot be placed using slub_max_order.
2690 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2691 if (order < MAX_ORDER)
2697 * Figure out what the alignment of the objects will be.
2699 static unsigned long calculate_alignment(unsigned long flags,
2700 unsigned long align, unsigned long size)
2703 * If the user wants hardware cache aligned objects then follow that
2704 * suggestion if the object is sufficiently large.
2706 * The hardware cache alignment cannot override the specified
2707 * alignment though. If that is greater then use it.
2709 if (flags & SLAB_HWCACHE_ALIGN) {
2710 unsigned long ralign = cache_line_size();
2711 while (size <= ralign / 2)
2713 align = max(align, ralign);
2716 if (align < ARCH_SLAB_MINALIGN)
2717 align = ARCH_SLAB_MINALIGN;
2719 return ALIGN(align, sizeof(void *));
2723 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2726 spin_lock_init(&n->list_lock);
2727 INIT_LIST_HEAD(&n->partial);
2728 #ifdef CONFIG_SLUB_DEBUG
2729 atomic_long_set(&n->nr_slabs, 0);
2730 atomic_long_set(&n->total_objects, 0);
2731 INIT_LIST_HEAD(&n->full);
2735 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2737 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2738 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2741 * Must align to double word boundary for the double cmpxchg
2742 * instructions to work; see __pcpu_double_call_return_bool().
2744 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2745 2 * sizeof(void *));
2750 init_kmem_cache_cpus(s);
2755 static struct kmem_cache *kmem_cache_node;
2758 * No kmalloc_node yet so do it by hand. We know that this is the first
2759 * slab on the node for this slabcache. There are no concurrent accesses
2762 * Note that this function only works on the kmalloc_node_cache
2763 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2764 * memory on a fresh node that has no slab structures yet.
2766 static void early_kmem_cache_node_alloc(int node)
2769 struct kmem_cache_node *n;
2771 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2773 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2776 if (page_to_nid(page) != node) {
2777 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2779 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2780 "in order to be able to continue\n");
2785 page->freelist = get_freepointer(kmem_cache_node, n);
2788 kmem_cache_node->node[node] = n;
2789 #ifdef CONFIG_SLUB_DEBUG
2790 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2791 init_tracking(kmem_cache_node, n);
2793 init_kmem_cache_node(n, kmem_cache_node);
2794 inc_slabs_node(kmem_cache_node, node, page->objects);
2796 add_partial(n, page, DEACTIVATE_TO_HEAD);
2799 static void free_kmem_cache_nodes(struct kmem_cache *s)
2803 for_each_node_state(node, N_NORMAL_MEMORY) {
2804 struct kmem_cache_node *n = s->node[node];
2807 kmem_cache_free(kmem_cache_node, n);
2809 s->node[node] = NULL;
2813 static int init_kmem_cache_nodes(struct kmem_cache *s)
2817 for_each_node_state(node, N_NORMAL_MEMORY) {
2818 struct kmem_cache_node *n;
2820 if (slab_state == DOWN) {
2821 early_kmem_cache_node_alloc(node);
2824 n = kmem_cache_alloc_node(kmem_cache_node,
2828 free_kmem_cache_nodes(s);
2833 init_kmem_cache_node(n, s);
2838 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2840 if (min < MIN_PARTIAL)
2842 else if (min > MAX_PARTIAL)
2844 s->min_partial = min;
2848 * calculate_sizes() determines the order and the distribution of data within
2851 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2853 unsigned long flags = s->flags;
2854 unsigned long size = s->objsize;
2855 unsigned long align = s->align;
2859 * Round up object size to the next word boundary. We can only
2860 * place the free pointer at word boundaries and this determines
2861 * the possible location of the free pointer.
2863 size = ALIGN(size, sizeof(void *));
2865 #ifdef CONFIG_SLUB_DEBUG
2867 * Determine if we can poison the object itself. If the user of
2868 * the slab may touch the object after free or before allocation
2869 * then we should never poison the object itself.
2871 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2873 s->flags |= __OBJECT_POISON;
2875 s->flags &= ~__OBJECT_POISON;
2879 * If we are Redzoning then check if there is some space between the
2880 * end of the object and the free pointer. If not then add an
2881 * additional word to have some bytes to store Redzone information.
2883 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2884 size += sizeof(void *);
2888 * With that we have determined the number of bytes in actual use
2889 * by the object. This is the potential offset to the free pointer.
2893 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2896 * Relocate free pointer after the object if it is not
2897 * permitted to overwrite the first word of the object on
2900 * This is the case if we do RCU, have a constructor or
2901 * destructor or are poisoning the objects.
2904 size += sizeof(void *);
2907 #ifdef CONFIG_SLUB_DEBUG
2908 if (flags & SLAB_STORE_USER)
2910 * Need to store information about allocs and frees after
2913 size += 2 * sizeof(struct track);
2915 if (flags & SLAB_RED_ZONE)
2917 * Add some empty padding so that we can catch
2918 * overwrites from earlier objects rather than let
2919 * tracking information or the free pointer be
2920 * corrupted if a user writes before the start
2923 size += sizeof(void *);
2927 * Determine the alignment based on various parameters that the
2928 * user specified and the dynamic determination of cache line size
2931 align = calculate_alignment(flags, align, s->objsize);
2935 * SLUB stores one object immediately after another beginning from
2936 * offset 0. In order to align the objects we have to simply size
2937 * each object to conform to the alignment.
2939 size = ALIGN(size, align);
2941 if (forced_order >= 0)
2942 order = forced_order;
2944 order = calculate_order(size, s->reserved);
2951 s->allocflags |= __GFP_COMP;
2953 if (s->flags & SLAB_CACHE_DMA)
2954 s->allocflags |= SLUB_DMA;
2956 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2957 s->allocflags |= __GFP_RECLAIMABLE;
2960 * Determine the number of objects per slab
2962 s->oo = oo_make(order, size, s->reserved);
2963 s->min = oo_make(get_order(size), size, s->reserved);
2964 if (oo_objects(s->oo) > oo_objects(s->max))
2967 return !!oo_objects(s->oo);
2971 static int kmem_cache_open(struct kmem_cache *s,
2972 const char *name, size_t size,
2973 size_t align, unsigned long flags,
2974 void (*ctor)(void *))
2976 memset(s, 0, kmem_size);
2981 s->flags = kmem_cache_flags(size, flags, name, ctor);
2984 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2985 s->reserved = sizeof(struct rcu_head);
2987 if (!calculate_sizes(s, -1))
2989 if (disable_higher_order_debug) {
2991 * Disable debugging flags that store metadata if the min slab
2994 if (get_order(s->size) > get_order(s->objsize)) {
2995 s->flags &= ~DEBUG_METADATA_FLAGS;
2997 if (!calculate_sizes(s, -1))
3002 #ifdef CONFIG_CMPXCHG_DOUBLE
3003 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3004 /* Enable fast mode */
3005 s->flags |= __CMPXCHG_DOUBLE;
3009 * The larger the object size is, the more pages we want on the partial
3010 * list to avoid pounding the page allocator excessively.
3012 set_min_partial(s, ilog2(s->size) / 2);
3015 * cpu_partial determined the maximum number of objects kept in the
3016 * per cpu partial lists of a processor.
3018 * Per cpu partial lists mainly contain slabs that just have one
3019 * object freed. If they are used for allocation then they can be
3020 * filled up again with minimal effort. The slab will never hit the
3021 * per node partial lists and therefore no locking will be required.
3023 * This setting also determines
3025 * A) The number of objects from per cpu partial slabs dumped to the
3026 * per node list when we reach the limit.
3027 * B) The number of objects in cpu partial slabs to extract from the
3028 * per node list when we run out of per cpu objects. We only fetch 50%
3029 * to keep some capacity around for frees.
3031 if (s->size >= PAGE_SIZE)
3033 else if (s->size >= 1024)
3035 else if (s->size >= 256)
3036 s->cpu_partial = 13;
3038 s->cpu_partial = 30;
3042 s->remote_node_defrag_ratio = 1000;
3044 if (!init_kmem_cache_nodes(s))
3047 if (alloc_kmem_cache_cpus(s))
3050 free_kmem_cache_nodes(s);
3052 if (flags & SLAB_PANIC)
3053 panic("Cannot create slab %s size=%lu realsize=%u "
3054 "order=%u offset=%u flags=%lx\n",
3055 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3061 * Determine the size of a slab object
3063 unsigned int kmem_cache_size(struct kmem_cache *s)
3067 EXPORT_SYMBOL(kmem_cache_size);
3069 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3072 #ifdef CONFIG_SLUB_DEBUG
3073 void *addr = page_address(page);
3075 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3076 sizeof(long), GFP_ATOMIC);
3079 slab_err(s, page, "%s", text);
3082 get_map(s, page, map);
3083 for_each_object(p, s, addr, page->objects) {
3085 if (!test_bit(slab_index(p, s, addr), map)) {
3086 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3088 print_tracking(s, p);
3097 * Attempt to free all partial slabs on a node.
3098 * This is called from kmem_cache_close(). We must be the last thread
3099 * using the cache and therefore we do not need to lock anymore.
3101 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3103 struct page *page, *h;
3105 list_for_each_entry_safe(page, h, &n->partial, lru) {
3107 remove_partial(n, page);
3108 discard_slab(s, page);
3110 list_slab_objects(s, page,
3111 "Objects remaining on kmem_cache_close()");
3117 * Release all resources used by a slab cache.
3119 static inline int kmem_cache_close(struct kmem_cache *s)
3124 free_percpu(s->cpu_slab);
3125 /* Attempt to free all objects */
3126 for_each_node_state(node, N_NORMAL_MEMORY) {
3127 struct kmem_cache_node *n = get_node(s, node);
3130 if (n->nr_partial || slabs_node(s, node))
3133 free_kmem_cache_nodes(s);
3138 * Close a cache and release the kmem_cache structure
3139 * (must be used for caches created using kmem_cache_create)
3141 void kmem_cache_destroy(struct kmem_cache *s)
3143 down_write(&slub_lock);
3147 up_write(&slub_lock);
3148 if (kmem_cache_close(s)) {
3149 printk(KERN_ERR "SLUB %s: %s called for cache that "
3150 "still has objects.\n", s->name, __func__);
3153 if (s->flags & SLAB_DESTROY_BY_RCU)
3155 sysfs_slab_remove(s);
3157 up_write(&slub_lock);
3159 EXPORT_SYMBOL(kmem_cache_destroy);
3161 /********************************************************************
3163 *******************************************************************/
3165 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3166 EXPORT_SYMBOL(kmalloc_caches);
3168 static struct kmem_cache *kmem_cache;
3170 #ifdef CONFIG_ZONE_DMA
3171 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3174 static int __init setup_slub_min_order(char *str)
3176 get_option(&str, &slub_min_order);
3181 __setup("slub_min_order=", setup_slub_min_order);
3183 static int __init setup_slub_max_order(char *str)
3185 get_option(&str, &slub_max_order);
3186 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3191 __setup("slub_max_order=", setup_slub_max_order);
3193 static int __init setup_slub_min_objects(char *str)
3195 get_option(&str, &slub_min_objects);
3200 __setup("slub_min_objects=", setup_slub_min_objects);
3202 static int __init setup_slub_nomerge(char *str)
3208 __setup("slub_nomerge", setup_slub_nomerge);
3210 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3211 int size, unsigned int flags)
3213 struct kmem_cache *s;
3215 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3218 * This function is called with IRQs disabled during early-boot on
3219 * single CPU so there's no need to take slub_lock here.
3221 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3225 list_add(&s->list, &slab_caches);
3229 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3234 * Conversion table for small slabs sizes / 8 to the index in the
3235 * kmalloc array. This is necessary for slabs < 192 since we have non power
3236 * of two cache sizes there. The size of larger slabs can be determined using
3239 static s8 size_index[24] = {
3266 static inline int size_index_elem(size_t bytes)
3268 return (bytes - 1) / 8;
3271 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3277 return ZERO_SIZE_PTR;
3279 index = size_index[size_index_elem(size)];
3281 index = fls(size - 1);
3283 #ifdef CONFIG_ZONE_DMA
3284 if (unlikely((flags & SLUB_DMA)))
3285 return kmalloc_dma_caches[index];
3288 return kmalloc_caches[index];
3291 void *__kmalloc(size_t size, gfp_t flags)
3293 struct kmem_cache *s;
3296 if (unlikely(size > SLUB_MAX_SIZE))
3297 return kmalloc_large(size, flags);
3299 s = get_slab(size, flags);
3301 if (unlikely(ZERO_OR_NULL_PTR(s)))
3304 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3306 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3310 EXPORT_SYMBOL(__kmalloc);
3313 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3318 flags |= __GFP_COMP | __GFP_NOTRACK;
3319 page = alloc_pages_node(node, flags, get_order(size));
3321 ptr = page_address(page);
3323 kmemleak_alloc(ptr, size, 1, flags);
3327 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3329 struct kmem_cache *s;
3332 if (unlikely(size > SLUB_MAX_SIZE)) {
3333 ret = kmalloc_large_node(size, flags, node);
3335 trace_kmalloc_node(_RET_IP_, ret,
3336 size, PAGE_SIZE << get_order(size),
3342 s = get_slab(size, flags);
3344 if (unlikely(ZERO_OR_NULL_PTR(s)))
3347 ret = slab_alloc(s, flags, node, _RET_IP_);
3349 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3353 EXPORT_SYMBOL(__kmalloc_node);
3356 size_t ksize(const void *object)
3360 if (unlikely(object == ZERO_SIZE_PTR))
3363 page = virt_to_head_page(object);
3365 if (unlikely(!PageSlab(page))) {
3366 WARN_ON(!PageCompound(page));
3367 return PAGE_SIZE << compound_order(page);
3370 return slab_ksize(page->slab);
3372 EXPORT_SYMBOL(ksize);
3374 #ifdef CONFIG_SLUB_DEBUG
3375 bool verify_mem_not_deleted(const void *x)
3378 void *object = (void *)x;
3379 unsigned long flags;
3382 if (unlikely(ZERO_OR_NULL_PTR(x)))
3385 local_irq_save(flags);
3387 page = virt_to_head_page(x);
3388 if (unlikely(!PageSlab(page))) {
3389 /* maybe it was from stack? */
3395 if (on_freelist(page->slab, page, object)) {
3396 object_err(page->slab, page, object, "Object is on free-list");
3404 local_irq_restore(flags);
3407 EXPORT_SYMBOL(verify_mem_not_deleted);
3410 void kfree(const void *x)
3413 void *object = (void *)x;
3415 trace_kfree(_RET_IP_, x);
3417 if (unlikely(ZERO_OR_NULL_PTR(x)))
3420 page = virt_to_head_page(x);
3421 if (unlikely(!PageSlab(page))) {
3422 BUG_ON(!PageCompound(page));
3427 slab_free(page->slab, page, object, _RET_IP_);
3429 EXPORT_SYMBOL(kfree);
3432 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3433 * the remaining slabs by the number of items in use. The slabs with the
3434 * most items in use come first. New allocations will then fill those up
3435 * and thus they can be removed from the partial lists.
3437 * The slabs with the least items are placed last. This results in them
3438 * being allocated from last increasing the chance that the last objects
3439 * are freed in them.
3441 int kmem_cache_shrink(struct kmem_cache *s)
3445 struct kmem_cache_node *n;
3448 int objects = oo_objects(s->max);
3449 struct list_head *slabs_by_inuse =
3450 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3451 unsigned long flags;
3453 if (!slabs_by_inuse)
3457 for_each_node_state(node, N_NORMAL_MEMORY) {
3458 n = get_node(s, node);
3463 for (i = 0; i < objects; i++)
3464 INIT_LIST_HEAD(slabs_by_inuse + i);
3466 spin_lock_irqsave(&n->list_lock, flags);
3469 * Build lists indexed by the items in use in each slab.
3471 * Note that concurrent frees may occur while we hold the
3472 * list_lock. page->inuse here is the upper limit.
3474 list_for_each_entry_safe(page, t, &n->partial, lru) {
3475 list_move(&page->lru, slabs_by_inuse + page->inuse);
3481 * Rebuild the partial list with the slabs filled up most
3482 * first and the least used slabs at the end.
3484 for (i = objects - 1; i > 0; i--)
3485 list_splice(slabs_by_inuse + i, n->partial.prev);
3487 spin_unlock_irqrestore(&n->list_lock, flags);
3489 /* Release empty slabs */
3490 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3491 discard_slab(s, page);
3494 kfree(slabs_by_inuse);
3497 EXPORT_SYMBOL(kmem_cache_shrink);
3499 #if defined(CONFIG_MEMORY_HOTPLUG)
3500 static int slab_mem_going_offline_callback(void *arg)
3502 struct kmem_cache *s;
3504 down_read(&slub_lock);
3505 list_for_each_entry(s, &slab_caches, list)
3506 kmem_cache_shrink(s);
3507 up_read(&slub_lock);
3512 static void slab_mem_offline_callback(void *arg)
3514 struct kmem_cache_node *n;
3515 struct kmem_cache *s;
3516 struct memory_notify *marg = arg;
3519 offline_node = marg->status_change_nid;
3522 * If the node still has available memory. we need kmem_cache_node
3525 if (offline_node < 0)
3528 down_read(&slub_lock);
3529 list_for_each_entry(s, &slab_caches, list) {
3530 n = get_node(s, offline_node);
3533 * if n->nr_slabs > 0, slabs still exist on the node
3534 * that is going down. We were unable to free them,
3535 * and offline_pages() function shouldn't call this
3536 * callback. So, we must fail.
3538 BUG_ON(slabs_node(s, offline_node));
3540 s->node[offline_node] = NULL;
3541 kmem_cache_free(kmem_cache_node, n);
3544 up_read(&slub_lock);
3547 static int slab_mem_going_online_callback(void *arg)
3549 struct kmem_cache_node *n;
3550 struct kmem_cache *s;
3551 struct memory_notify *marg = arg;
3552 int nid = marg->status_change_nid;
3556 * If the node's memory is already available, then kmem_cache_node is
3557 * already created. Nothing to do.
3563 * We are bringing a node online. No memory is available yet. We must
3564 * allocate a kmem_cache_node structure in order to bring the node
3567 down_read(&slub_lock);
3568 list_for_each_entry(s, &slab_caches, list) {
3570 * XXX: kmem_cache_alloc_node will fallback to other nodes
3571 * since memory is not yet available from the node that
3574 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3579 init_kmem_cache_node(n, s);
3583 up_read(&slub_lock);
3587 static int slab_memory_callback(struct notifier_block *self,
3588 unsigned long action, void *arg)
3593 case MEM_GOING_ONLINE:
3594 ret = slab_mem_going_online_callback(arg);
3596 case MEM_GOING_OFFLINE:
3597 ret = slab_mem_going_offline_callback(arg);
3600 case MEM_CANCEL_ONLINE:
3601 slab_mem_offline_callback(arg);
3604 case MEM_CANCEL_OFFLINE:
3608 ret = notifier_from_errno(ret);
3614 #endif /* CONFIG_MEMORY_HOTPLUG */
3616 /********************************************************************
3617 * Basic setup of slabs
3618 *******************************************************************/
3621 * Used for early kmem_cache structures that were allocated using
3622 * the page allocator
3625 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3629 list_add(&s->list, &slab_caches);
3632 for_each_node_state(node, N_NORMAL_MEMORY) {
3633 struct kmem_cache_node *n = get_node(s, node);
3637 list_for_each_entry(p, &n->partial, lru)
3640 #ifdef CONFIG_SLUB_DEBUG
3641 list_for_each_entry(p, &n->full, lru)
3648 void __init kmem_cache_init(void)
3652 struct kmem_cache *temp_kmem_cache;
3654 struct kmem_cache *temp_kmem_cache_node;
3655 unsigned long kmalloc_size;
3657 kmem_size = offsetof(struct kmem_cache, node) +
3658 nr_node_ids * sizeof(struct kmem_cache_node *);
3660 /* Allocate two kmem_caches from the page allocator */
3661 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3662 order = get_order(2 * kmalloc_size);
3663 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3666 * Must first have the slab cache available for the allocations of the
3667 * struct kmem_cache_node's. There is special bootstrap code in
3668 * kmem_cache_open for slab_state == DOWN.
3670 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3672 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3673 sizeof(struct kmem_cache_node),
3674 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3676 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3678 /* Able to allocate the per node structures */
3679 slab_state = PARTIAL;
3681 temp_kmem_cache = kmem_cache;
3682 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3683 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3684 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3685 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3688 * Allocate kmem_cache_node properly from the kmem_cache slab.
3689 * kmem_cache_node is separately allocated so no need to
3690 * update any list pointers.
3692 temp_kmem_cache_node = kmem_cache_node;
3694 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3695 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3697 kmem_cache_bootstrap_fixup(kmem_cache_node);
3700 kmem_cache_bootstrap_fixup(kmem_cache);
3702 /* Free temporary boot structure */
3703 free_pages((unsigned long)temp_kmem_cache, order);
3705 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3708 * Patch up the size_index table if we have strange large alignment
3709 * requirements for the kmalloc array. This is only the case for
3710 * MIPS it seems. The standard arches will not generate any code here.
3712 * Largest permitted alignment is 256 bytes due to the way we
3713 * handle the index determination for the smaller caches.
3715 * Make sure that nothing crazy happens if someone starts tinkering
3716 * around with ARCH_KMALLOC_MINALIGN
3718 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3719 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3721 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3722 int elem = size_index_elem(i);
3723 if (elem >= ARRAY_SIZE(size_index))
3725 size_index[elem] = KMALLOC_SHIFT_LOW;
3728 if (KMALLOC_MIN_SIZE == 64) {
3730 * The 96 byte size cache is not used if the alignment
3733 for (i = 64 + 8; i <= 96; i += 8)
3734 size_index[size_index_elem(i)] = 7;
3735 } else if (KMALLOC_MIN_SIZE == 128) {
3737 * The 192 byte sized cache is not used if the alignment
3738 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3741 for (i = 128 + 8; i <= 192; i += 8)
3742 size_index[size_index_elem(i)] = 8;
3745 /* Caches that are not of the two-to-the-power-of size */
3746 if (KMALLOC_MIN_SIZE <= 32) {
3747 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3751 if (KMALLOC_MIN_SIZE <= 64) {
3752 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3756 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3757 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3763 /* Provide the correct kmalloc names now that the caches are up */
3764 if (KMALLOC_MIN_SIZE <= 32) {
3765 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3766 BUG_ON(!kmalloc_caches[1]->name);
3769 if (KMALLOC_MIN_SIZE <= 64) {
3770 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3771 BUG_ON(!kmalloc_caches[2]->name);
3774 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3775 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3778 kmalloc_caches[i]->name = s;
3782 register_cpu_notifier(&slab_notifier);
3785 #ifdef CONFIG_ZONE_DMA
3786 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3787 struct kmem_cache *s = kmalloc_caches[i];
3790 char *name = kasprintf(GFP_NOWAIT,
3791 "dma-kmalloc-%d", s->objsize);
3794 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3795 s->objsize, SLAB_CACHE_DMA);
3800 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3801 " CPUs=%d, Nodes=%d\n",
3802 caches, cache_line_size(),
3803 slub_min_order, slub_max_order, slub_min_objects,
3804 nr_cpu_ids, nr_node_ids);
3807 void __init kmem_cache_init_late(void)
3812 * Find a mergeable slab cache
3814 static int slab_unmergeable(struct kmem_cache *s)
3816 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3823 * We may have set a slab to be unmergeable during bootstrap.
3825 if (s->refcount < 0)
3831 static struct kmem_cache *find_mergeable(size_t size,
3832 size_t align, unsigned long flags, const char *name,
3833 void (*ctor)(void *))
3835 struct kmem_cache *s;
3837 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3843 size = ALIGN(size, sizeof(void *));
3844 align = calculate_alignment(flags, align, size);
3845 size = ALIGN(size, align);
3846 flags = kmem_cache_flags(size, flags, name, NULL);
3848 list_for_each_entry(s, &slab_caches, list) {
3849 if (slab_unmergeable(s))
3855 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3858 * Check if alignment is compatible.
3859 * Courtesy of Adrian Drzewiecki
3861 if ((s->size & ~(align - 1)) != s->size)
3864 if (s->size - size >= sizeof(void *))
3872 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3873 size_t align, unsigned long flags, void (*ctor)(void *))
3875 struct kmem_cache *s;
3881 down_write(&slub_lock);
3882 s = find_mergeable(size, align, flags, name, ctor);
3886 * Adjust the object sizes so that we clear
3887 * the complete object on kzalloc.
3889 s->objsize = max(s->objsize, (int)size);
3890 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3892 if (sysfs_slab_alias(s, name)) {
3896 up_write(&slub_lock);
3900 n = kstrdup(name, GFP_KERNEL);
3904 s = kmalloc(kmem_size, GFP_KERNEL);
3906 if (kmem_cache_open(s, n,
3907 size, align, flags, ctor)) {
3908 list_add(&s->list, &slab_caches);
3909 if (sysfs_slab_add(s)) {
3915 up_write(&slub_lock);
3922 up_write(&slub_lock);
3924 if (flags & SLAB_PANIC)
3925 panic("Cannot create slabcache %s\n", name);
3930 EXPORT_SYMBOL(kmem_cache_create);
3934 * Use the cpu notifier to insure that the cpu slabs are flushed when
3937 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3938 unsigned long action, void *hcpu)
3940 long cpu = (long)hcpu;
3941 struct kmem_cache *s;
3942 unsigned long flags;
3945 case CPU_UP_CANCELED:
3946 case CPU_UP_CANCELED_FROZEN:
3948 case CPU_DEAD_FROZEN:
3949 down_read(&slub_lock);
3950 list_for_each_entry(s, &slab_caches, list) {
3951 local_irq_save(flags);
3952 __flush_cpu_slab(s, cpu);
3953 local_irq_restore(flags);
3955 up_read(&slub_lock);
3963 static struct notifier_block __cpuinitdata slab_notifier = {
3964 .notifier_call = slab_cpuup_callback
3969 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3971 struct kmem_cache *s;
3974 if (unlikely(size > SLUB_MAX_SIZE))
3975 return kmalloc_large(size, gfpflags);
3977 s = get_slab(size, gfpflags);
3979 if (unlikely(ZERO_OR_NULL_PTR(s)))
3982 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3984 /* Honor the call site pointer we received. */
3985 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3991 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3992 int node, unsigned long caller)
3994 struct kmem_cache *s;
3997 if (unlikely(size > SLUB_MAX_SIZE)) {
3998 ret = kmalloc_large_node(size, gfpflags, node);
4000 trace_kmalloc_node(caller, ret,
4001 size, PAGE_SIZE << get_order(size),
4007 s = get_slab(size, gfpflags);
4009 if (unlikely(ZERO_OR_NULL_PTR(s)))
4012 ret = slab_alloc(s, gfpflags, node, caller);
4014 /* Honor the call site pointer we received. */
4015 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4022 static int count_inuse(struct page *page)
4027 static int count_total(struct page *page)
4029 return page->objects;
4033 #ifdef CONFIG_SLUB_DEBUG
4034 static int validate_slab(struct kmem_cache *s, struct page *page,
4038 void *addr = page_address(page);
4040 if (!check_slab(s, page) ||
4041 !on_freelist(s, page, NULL))
4044 /* Now we know that a valid freelist exists */
4045 bitmap_zero(map, page->objects);
4047 get_map(s, page, map);
4048 for_each_object(p, s, addr, page->objects) {
4049 if (test_bit(slab_index(p, s, addr), map))
4050 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4054 for_each_object(p, s, addr, page->objects)
4055 if (!test_bit(slab_index(p, s, addr), map))
4056 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4061 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4065 validate_slab(s, page, map);
4069 static int validate_slab_node(struct kmem_cache *s,
4070 struct kmem_cache_node *n, unsigned long *map)
4072 unsigned long count = 0;
4074 unsigned long flags;
4076 spin_lock_irqsave(&n->list_lock, flags);
4078 list_for_each_entry(page, &n->partial, lru) {
4079 validate_slab_slab(s, page, map);
4082 if (count != n->nr_partial)
4083 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4084 "counter=%ld\n", s->name, count, n->nr_partial);
4086 if (!(s->flags & SLAB_STORE_USER))
4089 list_for_each_entry(page, &n->full, lru) {
4090 validate_slab_slab(s, page, map);
4093 if (count != atomic_long_read(&n->nr_slabs))
4094 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4095 "counter=%ld\n", s->name, count,
4096 atomic_long_read(&n->nr_slabs));
4099 spin_unlock_irqrestore(&n->list_lock, flags);
4103 static long validate_slab_cache(struct kmem_cache *s)
4106 unsigned long count = 0;
4107 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4108 sizeof(unsigned long), GFP_KERNEL);
4114 for_each_node_state(node, N_NORMAL_MEMORY) {
4115 struct kmem_cache_node *n = get_node(s, node);
4117 count += validate_slab_node(s, n, map);
4123 * Generate lists of code addresses where slabcache objects are allocated
4128 unsigned long count;
4135 DECLARE_BITMAP(cpus, NR_CPUS);
4141 unsigned long count;
4142 struct location *loc;
4145 static void free_loc_track(struct loc_track *t)
4148 free_pages((unsigned long)t->loc,
4149 get_order(sizeof(struct location) * t->max));
4152 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4157 order = get_order(sizeof(struct location) * max);
4159 l = (void *)__get_free_pages(flags, order);
4164 memcpy(l, t->loc, sizeof(struct location) * t->count);
4172 static int add_location(struct loc_track *t, struct kmem_cache *s,
4173 const struct track *track)
4175 long start, end, pos;
4177 unsigned long caddr;
4178 unsigned long age = jiffies - track->when;
4184 pos = start + (end - start + 1) / 2;
4187 * There is nothing at "end". If we end up there
4188 * we need to add something to before end.
4193 caddr = t->loc[pos].addr;
4194 if (track->addr == caddr) {
4200 if (age < l->min_time)
4202 if (age > l->max_time)
4205 if (track->pid < l->min_pid)
4206 l->min_pid = track->pid;
4207 if (track->pid > l->max_pid)
4208 l->max_pid = track->pid;
4210 cpumask_set_cpu(track->cpu,
4211 to_cpumask(l->cpus));
4213 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4217 if (track->addr < caddr)
4224 * Not found. Insert new tracking element.
4226 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4232 (t->count - pos) * sizeof(struct location));
4235 l->addr = track->addr;
4239 l->min_pid = track->pid;
4240 l->max_pid = track->pid;
4241 cpumask_clear(to_cpumask(l->cpus));
4242 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4243 nodes_clear(l->nodes);
4244 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4248 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4249 struct page *page, enum track_item alloc,
4252 void *addr = page_address(page);
4255 bitmap_zero(map, page->objects);
4256 get_map(s, page, map);
4258 for_each_object(p, s, addr, page->objects)
4259 if (!test_bit(slab_index(p, s, addr), map))
4260 add_location(t, s, get_track(s, p, alloc));
4263 static int list_locations(struct kmem_cache *s, char *buf,
4264 enum track_item alloc)
4268 struct loc_track t = { 0, 0, NULL };
4270 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4271 sizeof(unsigned long), GFP_KERNEL);
4273 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4276 return sprintf(buf, "Out of memory\n");
4278 /* Push back cpu slabs */
4281 for_each_node_state(node, N_NORMAL_MEMORY) {
4282 struct kmem_cache_node *n = get_node(s, node);
4283 unsigned long flags;
4286 if (!atomic_long_read(&n->nr_slabs))
4289 spin_lock_irqsave(&n->list_lock, flags);
4290 list_for_each_entry(page, &n->partial, lru)
4291 process_slab(&t, s, page, alloc, map);
4292 list_for_each_entry(page, &n->full, lru)
4293 process_slab(&t, s, page, alloc, map);
4294 spin_unlock_irqrestore(&n->list_lock, flags);
4297 for (i = 0; i < t.count; i++) {
4298 struct location *l = &t.loc[i];
4300 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4302 len += sprintf(buf + len, "%7ld ", l->count);
4305 len += sprintf(buf + len, "%pS", (void *)l->addr);
4307 len += sprintf(buf + len, "<not-available>");
4309 if (l->sum_time != l->min_time) {
4310 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4312 (long)div_u64(l->sum_time, l->count),
4315 len += sprintf(buf + len, " age=%ld",
4318 if (l->min_pid != l->max_pid)
4319 len += sprintf(buf + len, " pid=%ld-%ld",
4320 l->min_pid, l->max_pid);
4322 len += sprintf(buf + len, " pid=%ld",
4325 if (num_online_cpus() > 1 &&
4326 !cpumask_empty(to_cpumask(l->cpus)) &&
4327 len < PAGE_SIZE - 60) {
4328 len += sprintf(buf + len, " cpus=");
4329 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4330 to_cpumask(l->cpus));
4333 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4334 len < PAGE_SIZE - 60) {
4335 len += sprintf(buf + len, " nodes=");
4336 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4340 len += sprintf(buf + len, "\n");
4346 len += sprintf(buf, "No data\n");
4351 #ifdef SLUB_RESILIENCY_TEST
4352 static void resiliency_test(void)
4356 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4358 printk(KERN_ERR "SLUB resiliency testing\n");
4359 printk(KERN_ERR "-----------------------\n");
4360 printk(KERN_ERR "A. Corruption after allocation\n");
4362 p = kzalloc(16, GFP_KERNEL);
4364 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4365 " 0x12->0x%p\n\n", p + 16);
4367 validate_slab_cache(kmalloc_caches[4]);
4369 /* Hmmm... The next two are dangerous */
4370 p = kzalloc(32, GFP_KERNEL);
4371 p[32 + sizeof(void *)] = 0x34;
4372 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4373 " 0x34 -> -0x%p\n", p);
4375 "If allocated object is overwritten then not detectable\n\n");
4377 validate_slab_cache(kmalloc_caches[5]);
4378 p = kzalloc(64, GFP_KERNEL);
4379 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4381 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4384 "If allocated object is overwritten then not detectable\n\n");
4385 validate_slab_cache(kmalloc_caches[6]);
4387 printk(KERN_ERR "\nB. Corruption after free\n");
4388 p = kzalloc(128, GFP_KERNEL);
4391 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4392 validate_slab_cache(kmalloc_caches[7]);
4394 p = kzalloc(256, GFP_KERNEL);
4397 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4399 validate_slab_cache(kmalloc_caches[8]);
4401 p = kzalloc(512, GFP_KERNEL);
4404 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4405 validate_slab_cache(kmalloc_caches[9]);
4409 static void resiliency_test(void) {};
4414 enum slab_stat_type {
4415 SL_ALL, /* All slabs */
4416 SL_PARTIAL, /* Only partially allocated slabs */
4417 SL_CPU, /* Only slabs used for cpu caches */
4418 SL_OBJECTS, /* Determine allocated objects not slabs */
4419 SL_TOTAL /* Determine object capacity not slabs */
4422 #define SO_ALL (1 << SL_ALL)
4423 #define SO_PARTIAL (1 << SL_PARTIAL)
4424 #define SO_CPU (1 << SL_CPU)
4425 #define SO_OBJECTS (1 << SL_OBJECTS)
4426 #define SO_TOTAL (1 << SL_TOTAL)
4428 static ssize_t show_slab_objects(struct kmem_cache *s,
4429 char *buf, unsigned long flags)
4431 unsigned long total = 0;
4434 unsigned long *nodes;
4435 unsigned long *per_cpu;
4437 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4440 per_cpu = nodes + nr_node_ids;
4442 if (flags & SO_CPU) {
4445 for_each_possible_cpu(cpu) {
4446 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4449 if (!c || c->node < 0)
4453 if (flags & SO_TOTAL)
4454 x = c->page->objects;
4455 else if (flags & SO_OBJECTS)
4461 nodes[c->node] += x;
4468 nodes[c->node] += x;
4474 lock_memory_hotplug();
4475 #ifdef CONFIG_SLUB_DEBUG
4476 if (flags & SO_ALL) {
4477 for_each_node_state(node, N_NORMAL_MEMORY) {
4478 struct kmem_cache_node *n = get_node(s, node);
4480 if (flags & SO_TOTAL)
4481 x = atomic_long_read(&n->total_objects);
4482 else if (flags & SO_OBJECTS)
4483 x = atomic_long_read(&n->total_objects) -
4484 count_partial(n, count_free);
4487 x = atomic_long_read(&n->nr_slabs);
4494 if (flags & SO_PARTIAL) {
4495 for_each_node_state(node, N_NORMAL_MEMORY) {
4496 struct kmem_cache_node *n = get_node(s, node);
4498 if (flags & SO_TOTAL)
4499 x = count_partial(n, count_total);
4500 else if (flags & SO_OBJECTS)
4501 x = count_partial(n, count_inuse);
4508 x = sprintf(buf, "%lu", total);
4510 for_each_node_state(node, N_NORMAL_MEMORY)
4512 x += sprintf(buf + x, " N%d=%lu",
4515 unlock_memory_hotplug();
4517 return x + sprintf(buf + x, "\n");
4520 #ifdef CONFIG_SLUB_DEBUG
4521 static int any_slab_objects(struct kmem_cache *s)
4525 for_each_online_node(node) {
4526 struct kmem_cache_node *n = get_node(s, node);
4531 if (atomic_long_read(&n->total_objects))
4538 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4539 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4541 struct slab_attribute {
4542 struct attribute attr;
4543 ssize_t (*show)(struct kmem_cache *s, char *buf);
4544 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4547 #define SLAB_ATTR_RO(_name) \
4548 static struct slab_attribute _name##_attr = \
4549 __ATTR(_name, 0400, _name##_show, NULL)
4551 #define SLAB_ATTR(_name) \
4552 static struct slab_attribute _name##_attr = \
4553 __ATTR(_name, 0600, _name##_show, _name##_store)
4555 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4557 return sprintf(buf, "%d\n", s->size);
4559 SLAB_ATTR_RO(slab_size);
4561 static ssize_t align_show(struct kmem_cache *s, char *buf)
4563 return sprintf(buf, "%d\n", s->align);
4565 SLAB_ATTR_RO(align);
4567 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4569 return sprintf(buf, "%d\n", s->objsize);
4571 SLAB_ATTR_RO(object_size);
4573 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4575 return sprintf(buf, "%d\n", oo_objects(s->oo));
4577 SLAB_ATTR_RO(objs_per_slab);
4579 static ssize_t order_store(struct kmem_cache *s,
4580 const char *buf, size_t length)
4582 unsigned long order;
4585 err = strict_strtoul(buf, 10, &order);
4589 if (order > slub_max_order || order < slub_min_order)
4592 calculate_sizes(s, order);
4596 static ssize_t order_show(struct kmem_cache *s, char *buf)
4598 return sprintf(buf, "%d\n", oo_order(s->oo));
4602 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4604 return sprintf(buf, "%lu\n", s->min_partial);
4607 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4613 err = strict_strtoul(buf, 10, &min);
4617 set_min_partial(s, min);
4620 SLAB_ATTR(min_partial);
4622 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4624 return sprintf(buf, "%u\n", s->cpu_partial);
4627 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4630 unsigned long objects;
4633 err = strict_strtoul(buf, 10, &objects);
4637 s->cpu_partial = objects;
4641 SLAB_ATTR(cpu_partial);
4643 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4647 return sprintf(buf, "%pS\n", s->ctor);
4651 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4653 return sprintf(buf, "%d\n", s->refcount - 1);
4655 SLAB_ATTR_RO(aliases);
4657 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4659 return show_slab_objects(s, buf, SO_PARTIAL);
4661 SLAB_ATTR_RO(partial);
4663 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4665 return show_slab_objects(s, buf, SO_CPU);
4667 SLAB_ATTR_RO(cpu_slabs);
4669 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4671 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4673 SLAB_ATTR_RO(objects);
4675 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4677 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4679 SLAB_ATTR_RO(objects_partial);
4681 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4688 for_each_online_cpu(cpu) {
4689 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4692 pages += page->pages;
4693 objects += page->pobjects;
4697 len = sprintf(buf, "%d(%d)", objects, pages);
4700 for_each_online_cpu(cpu) {
4701 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4703 if (page && len < PAGE_SIZE - 20)
4704 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4705 page->pobjects, page->pages);
4708 return len + sprintf(buf + len, "\n");
4710 SLAB_ATTR_RO(slabs_cpu_partial);
4712 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4714 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4717 static ssize_t reclaim_account_store(struct kmem_cache *s,
4718 const char *buf, size_t length)
4720 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4722 s->flags |= SLAB_RECLAIM_ACCOUNT;
4725 SLAB_ATTR(reclaim_account);
4727 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4729 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4731 SLAB_ATTR_RO(hwcache_align);
4733 #ifdef CONFIG_ZONE_DMA
4734 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4736 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4738 SLAB_ATTR_RO(cache_dma);
4741 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4743 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4745 SLAB_ATTR_RO(destroy_by_rcu);
4747 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4749 return sprintf(buf, "%d\n", s->reserved);
4751 SLAB_ATTR_RO(reserved);
4753 #ifdef CONFIG_SLUB_DEBUG
4754 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4756 return show_slab_objects(s, buf, SO_ALL);
4758 SLAB_ATTR_RO(slabs);
4760 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4762 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4764 SLAB_ATTR_RO(total_objects);
4766 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4768 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4771 static ssize_t sanity_checks_store(struct kmem_cache *s,
4772 const char *buf, size_t length)
4774 s->flags &= ~SLAB_DEBUG_FREE;
4775 if (buf[0] == '1') {
4776 s->flags &= ~__CMPXCHG_DOUBLE;
4777 s->flags |= SLAB_DEBUG_FREE;
4781 SLAB_ATTR(sanity_checks);
4783 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4785 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4788 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4791 s->flags &= ~SLAB_TRACE;
4792 if (buf[0] == '1') {
4793 s->flags &= ~__CMPXCHG_DOUBLE;
4794 s->flags |= SLAB_TRACE;
4800 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4802 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4805 static ssize_t red_zone_store(struct kmem_cache *s,
4806 const char *buf, size_t length)
4808 if (any_slab_objects(s))
4811 s->flags &= ~SLAB_RED_ZONE;
4812 if (buf[0] == '1') {
4813 s->flags &= ~__CMPXCHG_DOUBLE;
4814 s->flags |= SLAB_RED_ZONE;
4816 calculate_sizes(s, -1);
4819 SLAB_ATTR(red_zone);
4821 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4823 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4826 static ssize_t poison_store(struct kmem_cache *s,
4827 const char *buf, size_t length)
4829 if (any_slab_objects(s))
4832 s->flags &= ~SLAB_POISON;
4833 if (buf[0] == '1') {
4834 s->flags &= ~__CMPXCHG_DOUBLE;
4835 s->flags |= SLAB_POISON;
4837 calculate_sizes(s, -1);
4842 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4847 static ssize_t store_user_store(struct kmem_cache *s,
4848 const char *buf, size_t length)
4850 if (any_slab_objects(s))
4853 s->flags &= ~SLAB_STORE_USER;
4854 if (buf[0] == '1') {
4855 s->flags &= ~__CMPXCHG_DOUBLE;
4856 s->flags |= SLAB_STORE_USER;
4858 calculate_sizes(s, -1);
4861 SLAB_ATTR(store_user);
4863 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4868 static ssize_t validate_store(struct kmem_cache *s,
4869 const char *buf, size_t length)
4873 if (buf[0] == '1') {
4874 ret = validate_slab_cache(s);
4880 SLAB_ATTR(validate);
4882 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4884 if (!(s->flags & SLAB_STORE_USER))
4886 return list_locations(s, buf, TRACK_ALLOC);
4888 SLAB_ATTR_RO(alloc_calls);
4890 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4892 if (!(s->flags & SLAB_STORE_USER))
4894 return list_locations(s, buf, TRACK_FREE);
4896 SLAB_ATTR_RO(free_calls);
4897 #endif /* CONFIG_SLUB_DEBUG */
4899 #ifdef CONFIG_FAILSLAB
4900 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4902 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4905 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4908 s->flags &= ~SLAB_FAILSLAB;
4910 s->flags |= SLAB_FAILSLAB;
4913 SLAB_ATTR(failslab);
4916 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4921 static ssize_t shrink_store(struct kmem_cache *s,
4922 const char *buf, size_t length)
4924 if (buf[0] == '1') {
4925 int rc = kmem_cache_shrink(s);
4936 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4938 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4941 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4942 const char *buf, size_t length)
4944 unsigned long ratio;
4947 err = strict_strtoul(buf, 10, &ratio);
4952 s->remote_node_defrag_ratio = ratio * 10;
4956 SLAB_ATTR(remote_node_defrag_ratio);
4959 #ifdef CONFIG_SLUB_STATS
4960 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4962 unsigned long sum = 0;
4965 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4970 for_each_online_cpu(cpu) {
4971 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4977 len = sprintf(buf, "%lu", sum);
4980 for_each_online_cpu(cpu) {
4981 if (data[cpu] && len < PAGE_SIZE - 20)
4982 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4986 return len + sprintf(buf + len, "\n");
4989 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4993 for_each_online_cpu(cpu)
4994 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4997 #define STAT_ATTR(si, text) \
4998 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5000 return show_stat(s, buf, si); \
5002 static ssize_t text##_store(struct kmem_cache *s, \
5003 const char *buf, size_t length) \
5005 if (buf[0] != '0') \
5007 clear_stat(s, si); \
5012 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5013 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5014 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5015 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5016 STAT_ATTR(FREE_FROZEN, free_frozen);
5017 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5018 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5019 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5020 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5021 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5022 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5023 STAT_ATTR(FREE_SLAB, free_slab);
5024 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5025 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5026 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5027 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5028 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5029 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5030 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5031 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5032 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5033 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5034 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5035 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5038 static struct attribute *slab_attrs[] = {
5039 &slab_size_attr.attr,
5040 &object_size_attr.attr,
5041 &objs_per_slab_attr.attr,
5043 &min_partial_attr.attr,
5044 &cpu_partial_attr.attr,
5046 &objects_partial_attr.attr,
5048 &cpu_slabs_attr.attr,
5052 &hwcache_align_attr.attr,
5053 &reclaim_account_attr.attr,
5054 &destroy_by_rcu_attr.attr,
5056 &reserved_attr.attr,
5057 &slabs_cpu_partial_attr.attr,
5058 #ifdef CONFIG_SLUB_DEBUG
5059 &total_objects_attr.attr,
5061 &sanity_checks_attr.attr,
5063 &red_zone_attr.attr,
5065 &store_user_attr.attr,
5066 &validate_attr.attr,
5067 &alloc_calls_attr.attr,
5068 &free_calls_attr.attr,
5070 #ifdef CONFIG_ZONE_DMA
5071 &cache_dma_attr.attr,
5074 &remote_node_defrag_ratio_attr.attr,
5076 #ifdef CONFIG_SLUB_STATS
5077 &alloc_fastpath_attr.attr,
5078 &alloc_slowpath_attr.attr,
5079 &free_fastpath_attr.attr,
5080 &free_slowpath_attr.attr,
5081 &free_frozen_attr.attr,
5082 &free_add_partial_attr.attr,
5083 &free_remove_partial_attr.attr,
5084 &alloc_from_partial_attr.attr,
5085 &alloc_slab_attr.attr,
5086 &alloc_refill_attr.attr,
5087 &alloc_node_mismatch_attr.attr,
5088 &free_slab_attr.attr,
5089 &cpuslab_flush_attr.attr,
5090 &deactivate_full_attr.attr,
5091 &deactivate_empty_attr.attr,
5092 &deactivate_to_head_attr.attr,
5093 &deactivate_to_tail_attr.attr,
5094 &deactivate_remote_frees_attr.attr,
5095 &deactivate_bypass_attr.attr,
5096 &order_fallback_attr.attr,
5097 &cmpxchg_double_fail_attr.attr,
5098 &cmpxchg_double_cpu_fail_attr.attr,
5099 &cpu_partial_alloc_attr.attr,
5100 &cpu_partial_free_attr.attr,
5102 #ifdef CONFIG_FAILSLAB
5103 &failslab_attr.attr,
5109 static struct attribute_group slab_attr_group = {
5110 .attrs = slab_attrs,
5113 static ssize_t slab_attr_show(struct kobject *kobj,
5114 struct attribute *attr,
5117 struct slab_attribute *attribute;
5118 struct kmem_cache *s;
5121 attribute = to_slab_attr(attr);
5124 if (!attribute->show)
5127 err = attribute->show(s, buf);
5132 static ssize_t slab_attr_store(struct kobject *kobj,
5133 struct attribute *attr,
5134 const char *buf, size_t len)
5136 struct slab_attribute *attribute;
5137 struct kmem_cache *s;
5140 attribute = to_slab_attr(attr);
5143 if (!attribute->store)
5146 err = attribute->store(s, buf, len);
5151 static void kmem_cache_release(struct kobject *kobj)
5153 struct kmem_cache *s = to_slab(kobj);
5159 static const struct sysfs_ops slab_sysfs_ops = {
5160 .show = slab_attr_show,
5161 .store = slab_attr_store,
5164 static struct kobj_type slab_ktype = {
5165 .sysfs_ops = &slab_sysfs_ops,
5166 .release = kmem_cache_release
5169 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5171 struct kobj_type *ktype = get_ktype(kobj);
5173 if (ktype == &slab_ktype)
5178 static const struct kset_uevent_ops slab_uevent_ops = {
5179 .filter = uevent_filter,
5182 static struct kset *slab_kset;
5184 #define ID_STR_LENGTH 64
5186 /* Create a unique string id for a slab cache:
5188 * Format :[flags-]size
5190 static char *create_unique_id(struct kmem_cache *s)
5192 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5199 * First flags affecting slabcache operations. We will only
5200 * get here for aliasable slabs so we do not need to support
5201 * too many flags. The flags here must cover all flags that
5202 * are matched during merging to guarantee that the id is
5205 if (s->flags & SLAB_CACHE_DMA)
5207 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5209 if (s->flags & SLAB_DEBUG_FREE)
5211 if (!(s->flags & SLAB_NOTRACK))
5215 p += sprintf(p, "%07d", s->size);
5216 BUG_ON(p > name + ID_STR_LENGTH - 1);
5220 static int sysfs_slab_add(struct kmem_cache *s)
5226 if (slab_state < SYSFS)
5227 /* Defer until later */
5230 unmergeable = slab_unmergeable(s);
5233 * Slabcache can never be merged so we can use the name proper.
5234 * This is typically the case for debug situations. In that
5235 * case we can catch duplicate names easily.
5237 sysfs_remove_link(&slab_kset->kobj, s->name);
5241 * Create a unique name for the slab as a target
5244 name = create_unique_id(s);
5247 s->kobj.kset = slab_kset;
5248 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5250 kobject_put(&s->kobj);
5254 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5256 kobject_del(&s->kobj);
5257 kobject_put(&s->kobj);
5260 kobject_uevent(&s->kobj, KOBJ_ADD);
5262 /* Setup first alias */
5263 sysfs_slab_alias(s, s->name);
5269 static void sysfs_slab_remove(struct kmem_cache *s)
5271 if (slab_state < SYSFS)
5273 * Sysfs has not been setup yet so no need to remove the
5278 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5279 kobject_del(&s->kobj);
5280 kobject_put(&s->kobj);
5284 * Need to buffer aliases during bootup until sysfs becomes
5285 * available lest we lose that information.
5287 struct saved_alias {
5288 struct kmem_cache *s;
5290 struct saved_alias *next;
5293 static struct saved_alias *alias_list;
5295 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5297 struct saved_alias *al;
5299 if (slab_state == SYSFS) {
5301 * If we have a leftover link then remove it.
5303 sysfs_remove_link(&slab_kset->kobj, name);
5304 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5307 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5313 al->next = alias_list;
5318 static int __init slab_sysfs_init(void)
5320 struct kmem_cache *s;
5323 down_write(&slub_lock);
5325 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5327 up_write(&slub_lock);
5328 printk(KERN_ERR "Cannot register slab subsystem.\n");
5334 list_for_each_entry(s, &slab_caches, list) {
5335 err = sysfs_slab_add(s);
5337 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5338 " to sysfs\n", s->name);
5341 while (alias_list) {
5342 struct saved_alias *al = alias_list;
5344 alias_list = alias_list->next;
5345 err = sysfs_slab_alias(al->s, al->name);
5347 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5348 " %s to sysfs\n", s->name);
5352 up_write(&slub_lock);
5357 __initcall(slab_sysfs_init);
5358 #endif /* CONFIG_SYSFS */
5361 * The /proc/slabinfo ABI
5363 #ifdef CONFIG_SLABINFO
5364 static void print_slabinfo_header(struct seq_file *m)
5366 seq_puts(m, "slabinfo - version: 2.1\n");
5367 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5368 "<objperslab> <pagesperslab>");
5369 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5370 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5374 static void *s_start(struct seq_file *m, loff_t *pos)
5378 down_read(&slub_lock);
5380 print_slabinfo_header(m);
5382 return seq_list_start(&slab_caches, *pos);
5385 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5387 return seq_list_next(p, &slab_caches, pos);
5390 static void s_stop(struct seq_file *m, void *p)
5392 up_read(&slub_lock);
5395 static int s_show(struct seq_file *m, void *p)
5397 unsigned long nr_partials = 0;
5398 unsigned long nr_slabs = 0;
5399 unsigned long nr_inuse = 0;
5400 unsigned long nr_objs = 0;
5401 unsigned long nr_free = 0;
5402 struct kmem_cache *s;
5405 s = list_entry(p, struct kmem_cache, list);
5407 for_each_online_node(node) {
5408 struct kmem_cache_node *n = get_node(s, node);
5413 nr_partials += n->nr_partial;
5414 nr_slabs += atomic_long_read(&n->nr_slabs);
5415 nr_objs += atomic_long_read(&n->total_objects);
5416 nr_free += count_partial(n, count_free);
5419 nr_inuse = nr_objs - nr_free;
5421 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5422 nr_objs, s->size, oo_objects(s->oo),
5423 (1 << oo_order(s->oo)));
5424 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5425 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5431 static const struct seq_operations slabinfo_op = {
5438 static int slabinfo_open(struct inode *inode, struct file *file)
5440 return seq_open(file, &slabinfo_op);
5443 static const struct file_operations proc_slabinfo_operations = {
5444 .open = slabinfo_open,
5446 .llseek = seq_lseek,
5447 .release = seq_release,
5450 static int __init slab_proc_init(void)
5452 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5455 module_init(slab_proc_init);
5456 #endif /* CONFIG_SLABINFO */