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;
1511 new.freelist = NULL;
1513 new.freelist = freelist;
1516 VM_BUG_ON(new.frozen);
1519 } while (!__cmpxchg_double_slab(s, page,
1521 new.freelist, new.counters,
1522 "lock and freeze"));
1524 remove_partial(n, page);
1528 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1531 * Try to allocate a partial slab from a specific node.
1533 static void *get_partial_node(struct kmem_cache *s,
1534 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1536 struct page *page, *page2;
1537 void *object = NULL;
1540 * Racy check. If we mistakenly see no partial slabs then we
1541 * just allocate an empty slab. If we mistakenly try to get a
1542 * partial slab and there is none available then get_partials()
1545 if (!n || !n->nr_partial)
1548 spin_lock(&n->list_lock);
1549 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1550 void *t = acquire_slab(s, n, page, object == NULL);
1558 c->node = page_to_nid(page);
1559 stat(s, ALLOC_FROM_PARTIAL);
1561 available = page->objects - page->inuse;
1563 available = put_cpu_partial(s, page, 0);
1565 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1569 spin_unlock(&n->list_lock);
1574 * Get a page from somewhere. Search in increasing NUMA distances.
1576 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1577 struct kmem_cache_cpu *c)
1580 struct zonelist *zonelist;
1583 enum zone_type high_zoneidx = gfp_zone(flags);
1585 unsigned int cpuset_mems_cookie;
1588 * The defrag ratio allows a configuration of the tradeoffs between
1589 * inter node defragmentation and node local allocations. A lower
1590 * defrag_ratio increases the tendency to do local allocations
1591 * instead of attempting to obtain partial slabs from other nodes.
1593 * If the defrag_ratio is set to 0 then kmalloc() always
1594 * returns node local objects. If the ratio is higher then kmalloc()
1595 * may return off node objects because partial slabs are obtained
1596 * from other nodes and filled up.
1598 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1599 * defrag_ratio = 1000) then every (well almost) allocation will
1600 * first attempt to defrag slab caches on other nodes. This means
1601 * scanning over all nodes to look for partial slabs which may be
1602 * expensive if we do it every time we are trying to find a slab
1603 * with available objects.
1605 if (!s->remote_node_defrag_ratio ||
1606 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1610 cpuset_mems_cookie = get_mems_allowed();
1611 zonelist = node_zonelist(slab_node(), flags);
1612 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1613 struct kmem_cache_node *n;
1615 n = get_node(s, zone_to_nid(zone));
1617 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1618 n->nr_partial > s->min_partial) {
1619 object = get_partial_node(s, n, c);
1622 * Return the object even if
1623 * put_mems_allowed indicated that
1624 * the cpuset mems_allowed was
1625 * updated in parallel. It's a
1626 * harmless race between the alloc
1627 * and the cpuset update.
1629 put_mems_allowed(cpuset_mems_cookie);
1634 } while (!put_mems_allowed(cpuset_mems_cookie));
1640 * Get a partial page, lock it and return it.
1642 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1643 struct kmem_cache_cpu *c)
1646 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1648 object = get_partial_node(s, get_node(s, searchnode), c);
1649 if (object || node != NUMA_NO_NODE)
1652 return get_any_partial(s, flags, c);
1655 #ifdef CONFIG_PREEMPT
1657 * Calculate the next globally unique transaction for disambiguiation
1658 * during cmpxchg. The transactions start with the cpu number and are then
1659 * incremented by CONFIG_NR_CPUS.
1661 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1664 * No preemption supported therefore also no need to check for
1670 static inline unsigned long next_tid(unsigned long tid)
1672 return tid + TID_STEP;
1675 static inline unsigned int tid_to_cpu(unsigned long tid)
1677 return tid % TID_STEP;
1680 static inline unsigned long tid_to_event(unsigned long tid)
1682 return tid / TID_STEP;
1685 static inline unsigned int init_tid(int cpu)
1690 static inline void note_cmpxchg_failure(const char *n,
1691 const struct kmem_cache *s, unsigned long tid)
1693 #ifdef SLUB_DEBUG_CMPXCHG
1694 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1696 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1698 #ifdef CONFIG_PREEMPT
1699 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1700 printk("due to cpu change %d -> %d\n",
1701 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1704 if (tid_to_event(tid) != tid_to_event(actual_tid))
1705 printk("due to cpu running other code. Event %ld->%ld\n",
1706 tid_to_event(tid), tid_to_event(actual_tid));
1708 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1709 actual_tid, tid, next_tid(tid));
1711 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1714 void init_kmem_cache_cpus(struct kmem_cache *s)
1718 for_each_possible_cpu(cpu)
1719 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1723 * Remove the cpu slab
1725 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1727 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1728 struct page *page = c->page;
1729 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1731 enum slab_modes l = M_NONE, m = M_NONE;
1734 int tail = DEACTIVATE_TO_HEAD;
1738 if (page->freelist) {
1739 stat(s, DEACTIVATE_REMOTE_FREES);
1740 tail = DEACTIVATE_TO_TAIL;
1743 c->tid = next_tid(c->tid);
1745 freelist = c->freelist;
1749 * Stage one: Free all available per cpu objects back
1750 * to the page freelist while it is still frozen. Leave the
1753 * There is no need to take the list->lock because the page
1756 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1758 unsigned long counters;
1761 prior = page->freelist;
1762 counters = page->counters;
1763 set_freepointer(s, freelist, prior);
1764 new.counters = counters;
1766 VM_BUG_ON(!new.frozen);
1768 } while (!__cmpxchg_double_slab(s, page,
1770 freelist, new.counters,
1771 "drain percpu freelist"));
1773 freelist = nextfree;
1777 * Stage two: Ensure that the page is unfrozen while the
1778 * list presence reflects the actual number of objects
1781 * We setup the list membership and then perform a cmpxchg
1782 * with the count. If there is a mismatch then the page
1783 * is not unfrozen but the page is on the wrong list.
1785 * Then we restart the process which may have to remove
1786 * the page from the list that we just put it on again
1787 * because the number of objects in the slab may have
1792 old.freelist = page->freelist;
1793 old.counters = page->counters;
1794 VM_BUG_ON(!old.frozen);
1796 /* Determine target state of the slab */
1797 new.counters = old.counters;
1800 set_freepointer(s, freelist, old.freelist);
1801 new.freelist = freelist;
1803 new.freelist = old.freelist;
1807 if (!new.inuse && n->nr_partial > s->min_partial)
1809 else if (new.freelist) {
1814 * Taking the spinlock removes the possiblity
1815 * that acquire_slab() will see a slab page that
1818 spin_lock(&n->list_lock);
1822 if (kmem_cache_debug(s) && !lock) {
1825 * This also ensures that the scanning of full
1826 * slabs from diagnostic functions will not see
1829 spin_lock(&n->list_lock);
1837 remove_partial(n, page);
1839 else if (l == M_FULL)
1841 remove_full(s, page);
1843 if (m == M_PARTIAL) {
1845 add_partial(n, page, tail);
1848 } else if (m == M_FULL) {
1850 stat(s, DEACTIVATE_FULL);
1851 add_full(s, n, page);
1857 if (!__cmpxchg_double_slab(s, page,
1858 old.freelist, old.counters,
1859 new.freelist, new.counters,
1864 spin_unlock(&n->list_lock);
1867 stat(s, DEACTIVATE_EMPTY);
1868 discard_slab(s, page);
1873 /* Unfreeze all the cpu partial slabs */
1874 static void unfreeze_partials(struct kmem_cache *s)
1876 struct kmem_cache_node *n = NULL, *n2 = NULL;
1877 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1878 struct page *page, *discard_page = NULL;
1880 while ((page = c->partial)) {
1884 c->partial = page->next;
1886 n2 = get_node(s, page_to_nid(page));
1889 spin_unlock(&n->list_lock);
1892 spin_lock(&n->list_lock);
1897 old.freelist = page->freelist;
1898 old.counters = page->counters;
1899 VM_BUG_ON(!old.frozen);
1901 new.counters = old.counters;
1902 new.freelist = old.freelist;
1906 } while (!cmpxchg_double_slab(s, page,
1907 old.freelist, old.counters,
1908 new.freelist, new.counters,
1909 "unfreezing slab"));
1911 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1912 page->next = discard_page;
1913 discard_page = page;
1915 add_partial(n, page, DEACTIVATE_TO_TAIL);
1916 stat(s, FREE_ADD_PARTIAL);
1921 spin_unlock(&n->list_lock);
1923 while (discard_page) {
1924 page = discard_page;
1925 discard_page = discard_page->next;
1927 stat(s, DEACTIVATE_EMPTY);
1928 discard_slab(s, page);
1934 * Put a page that was just frozen (in __slab_free) into a partial page
1935 * slot if available. This is done without interrupts disabled and without
1936 * preemption disabled. The cmpxchg is racy and may put the partial page
1937 * onto a random cpus partial slot.
1939 * If we did not find a slot then simply move all the partials to the
1940 * per node partial list.
1942 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1944 struct page *oldpage;
1951 oldpage = this_cpu_read(s->cpu_slab->partial);
1954 pobjects = oldpage->pobjects;
1955 pages = oldpage->pages;
1956 if (drain && pobjects > s->cpu_partial) {
1957 unsigned long flags;
1959 * partial array is full. Move the existing
1960 * set to the per node partial list.
1962 local_irq_save(flags);
1963 unfreeze_partials(s);
1964 local_irq_restore(flags);
1971 pobjects += page->objects - page->inuse;
1973 page->pages = pages;
1974 page->pobjects = pobjects;
1975 page->next = oldpage;
1977 } while (irqsafe_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1978 stat(s, CPU_PARTIAL_FREE);
1982 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1984 stat(s, CPUSLAB_FLUSH);
1985 deactivate_slab(s, c);
1991 * Called from IPI handler with interrupts disabled.
1993 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1995 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2001 unfreeze_partials(s);
2005 static void flush_cpu_slab(void *d)
2007 struct kmem_cache *s = d;
2009 __flush_cpu_slab(s, smp_processor_id());
2012 static void flush_all(struct kmem_cache *s)
2014 on_each_cpu(flush_cpu_slab, s, 1);
2018 * Check if the objects in a per cpu structure fit numa
2019 * locality expectations.
2021 static inline int node_match(struct kmem_cache_cpu *c, int node)
2024 if (node != NUMA_NO_NODE && c->node != node)
2030 static int count_free(struct page *page)
2032 return page->objects - page->inuse;
2035 static unsigned long count_partial(struct kmem_cache_node *n,
2036 int (*get_count)(struct page *))
2038 unsigned long flags;
2039 unsigned long x = 0;
2042 spin_lock_irqsave(&n->list_lock, flags);
2043 list_for_each_entry(page, &n->partial, lru)
2044 x += get_count(page);
2045 spin_unlock_irqrestore(&n->list_lock, flags);
2049 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2051 #ifdef CONFIG_SLUB_DEBUG
2052 return atomic_long_read(&n->total_objects);
2058 static noinline void
2059 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2064 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2066 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2067 "default order: %d, min order: %d\n", s->name, s->objsize,
2068 s->size, oo_order(s->oo), oo_order(s->min));
2070 if (oo_order(s->min) > get_order(s->objsize))
2071 printk(KERN_WARNING " %s debugging increased min order, use "
2072 "slub_debug=O to disable.\n", s->name);
2074 for_each_online_node(node) {
2075 struct kmem_cache_node *n = get_node(s, node);
2076 unsigned long nr_slabs;
2077 unsigned long nr_objs;
2078 unsigned long nr_free;
2083 nr_free = count_partial(n, count_free);
2084 nr_slabs = node_nr_slabs(n);
2085 nr_objs = node_nr_objs(n);
2088 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2089 node, nr_slabs, nr_objs, nr_free);
2093 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2094 int node, struct kmem_cache_cpu **pc)
2097 struct kmem_cache_cpu *c;
2098 struct page *page = new_slab(s, flags, node);
2101 c = __this_cpu_ptr(s->cpu_slab);
2106 * No other reference to the page yet so we can
2107 * muck around with it freely without cmpxchg
2109 object = page->freelist;
2110 page->freelist = NULL;
2112 stat(s, ALLOC_SLAB);
2113 c->node = page_to_nid(page);
2123 * Slow path. The lockless freelist is empty or we need to perform
2126 * Processing is still very fast if new objects have been freed to the
2127 * regular freelist. In that case we simply take over the regular freelist
2128 * as the lockless freelist and zap the regular freelist.
2130 * If that is not working then we fall back to the partial lists. We take the
2131 * first element of the freelist as the object to allocate now and move the
2132 * rest of the freelist to the lockless freelist.
2134 * And if we were unable to get a new slab from the partial slab lists then
2135 * we need to allocate a new slab. This is the slowest path since it involves
2136 * a call to the page allocator and the setup of a new slab.
2138 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2139 unsigned long addr, struct kmem_cache_cpu *c)
2142 unsigned long flags;
2144 unsigned long counters;
2146 local_irq_save(flags);
2147 #ifdef CONFIG_PREEMPT
2149 * We may have been preempted and rescheduled on a different
2150 * cpu before disabling interrupts. Need to reload cpu area
2153 c = this_cpu_ptr(s->cpu_slab);
2159 if (unlikely(!node_match(c, node))) {
2160 stat(s, ALLOC_NODE_MISMATCH);
2161 deactivate_slab(s, c);
2165 /* must check again c->freelist in case of cpu migration or IRQ */
2166 object = c->freelist;
2170 stat(s, ALLOC_SLOWPATH);
2173 object = c->page->freelist;
2174 counters = c->page->counters;
2175 new.counters = counters;
2176 VM_BUG_ON(!new.frozen);
2179 * If there is no object left then we use this loop to
2180 * deactivate the slab which is simple since no objects
2181 * are left in the slab and therefore we do not need to
2182 * put the page back onto the partial list.
2184 * If there are objects left then we retrieve them
2185 * and use them to refill the per cpu queue.
2188 new.inuse = c->page->objects;
2189 new.frozen = object != NULL;
2191 } while (!__cmpxchg_double_slab(s, c->page,
2198 stat(s, DEACTIVATE_BYPASS);
2202 stat(s, ALLOC_REFILL);
2205 c->freelist = get_freepointer(s, object);
2206 c->tid = next_tid(c->tid);
2207 local_irq_restore(flags);
2213 c->page = c->partial;
2214 c->partial = c->page->next;
2215 c->node = page_to_nid(c->page);
2216 stat(s, CPU_PARTIAL_ALLOC);
2221 /* Then do expensive stuff like retrieving pages from the partial lists */
2222 object = get_partial(s, gfpflags, node, c);
2224 if (unlikely(!object)) {
2226 object = new_slab_objects(s, gfpflags, node, &c);
2228 if (unlikely(!object)) {
2229 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2230 slab_out_of_memory(s, gfpflags, node);
2232 local_irq_restore(flags);
2237 if (likely(!kmem_cache_debug(s)))
2240 /* Only entered in the debug case */
2241 if (!alloc_debug_processing(s, c->page, object, addr))
2242 goto new_slab; /* Slab failed checks. Next slab needed */
2244 c->freelist = get_freepointer(s, object);
2245 deactivate_slab(s, c);
2246 c->node = NUMA_NO_NODE;
2247 local_irq_restore(flags);
2252 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2253 * have the fastpath folded into their functions. So no function call
2254 * overhead for requests that can be satisfied on the fastpath.
2256 * The fastpath works by first checking if the lockless freelist can be used.
2257 * If not then __slab_alloc is called for slow processing.
2259 * Otherwise we can simply pick the next object from the lockless free list.
2261 static __always_inline void *slab_alloc(struct kmem_cache *s,
2262 gfp_t gfpflags, int node, unsigned long addr)
2265 struct kmem_cache_cpu *c;
2268 if (slab_pre_alloc_hook(s, gfpflags))
2274 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2275 * enabled. We may switch back and forth between cpus while
2276 * reading from one cpu area. That does not matter as long
2277 * as we end up on the original cpu again when doing the cmpxchg.
2279 c = __this_cpu_ptr(s->cpu_slab);
2282 * The transaction ids are globally unique per cpu and per operation on
2283 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2284 * occurs on the right processor and that there was no operation on the
2285 * linked list in between.
2290 object = c->freelist;
2291 if (unlikely(!object || !node_match(c, node)))
2293 object = __slab_alloc(s, gfpflags, node, addr, c);
2297 * The cmpxchg will only match if there was no additional
2298 * operation and if we are on the right processor.
2300 * The cmpxchg does the following atomically (without lock semantics!)
2301 * 1. Relocate first pointer to the current per cpu area.
2302 * 2. Verify that tid and freelist have not been changed
2303 * 3. If they were not changed replace tid and freelist
2305 * Since this is without lock semantics the protection is only against
2306 * code executing on this cpu *not* from access by other cpus.
2308 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2309 s->cpu_slab->freelist, s->cpu_slab->tid,
2311 get_freepointer_safe(s, object), next_tid(tid)))) {
2313 note_cmpxchg_failure("slab_alloc", s, tid);
2316 stat(s, ALLOC_FASTPATH);
2319 if (unlikely(gfpflags & __GFP_ZERO) && object)
2320 memset(object, 0, s->objsize);
2322 slab_post_alloc_hook(s, gfpflags, object);
2327 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2329 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2331 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2335 EXPORT_SYMBOL(kmem_cache_alloc);
2337 #ifdef CONFIG_TRACING
2338 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2340 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2341 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2344 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2346 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2348 void *ret = kmalloc_order(size, flags, order);
2349 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2352 EXPORT_SYMBOL(kmalloc_order_trace);
2356 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2358 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2360 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2361 s->objsize, s->size, gfpflags, node);
2365 EXPORT_SYMBOL(kmem_cache_alloc_node);
2367 #ifdef CONFIG_TRACING
2368 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2370 int node, size_t size)
2372 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2374 trace_kmalloc_node(_RET_IP_, ret,
2375 size, s->size, gfpflags, node);
2378 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2383 * Slow patch handling. This may still be called frequently since objects
2384 * have a longer lifetime than the cpu slabs in most processing loads.
2386 * So we still attempt to reduce cache line usage. Just take the slab
2387 * lock and free the item. If there is no additional partial page
2388 * handling required then we can return immediately.
2390 static void __slab_free(struct kmem_cache *s, struct page *page,
2391 void *x, unsigned long addr)
2394 void **object = (void *)x;
2398 unsigned long counters;
2399 struct kmem_cache_node *n = NULL;
2400 unsigned long uninitialized_var(flags);
2402 stat(s, FREE_SLOWPATH);
2404 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2408 prior = page->freelist;
2409 counters = page->counters;
2410 set_freepointer(s, object, prior);
2411 new.counters = counters;
2412 was_frozen = new.frozen;
2414 if ((!new.inuse || !prior) && !was_frozen && !n) {
2416 if (!kmem_cache_debug(s) && !prior)
2419 * Slab was on no list before and will be partially empty
2420 * We can defer the list move and instead freeze it.
2424 else { /* Needs to be taken off a list */
2426 n = get_node(s, page_to_nid(page));
2428 * Speculatively acquire the list_lock.
2429 * If the cmpxchg does not succeed then we may
2430 * drop the list_lock without any processing.
2432 * Otherwise the list_lock will synchronize with
2433 * other processors updating the list of slabs.
2435 spin_lock_irqsave(&n->list_lock, flags);
2441 } while (!cmpxchg_double_slab(s, page,
2443 object, new.counters,
2449 * If we just froze the page then put it onto the
2450 * per cpu partial list.
2452 if (new.frozen && !was_frozen)
2453 put_cpu_partial(s, page, 1);
2456 * The list lock was not taken therefore no list
2457 * activity can be necessary.
2460 stat(s, FREE_FROZEN);
2465 * was_frozen may have been set after we acquired the list_lock in
2466 * an earlier loop. So we need to check it here again.
2469 stat(s, FREE_FROZEN);
2471 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2475 * Objects left in the slab. If it was not on the partial list before
2478 if (unlikely(!prior)) {
2479 remove_full(s, page);
2480 add_partial(n, page, DEACTIVATE_TO_TAIL);
2481 stat(s, FREE_ADD_PARTIAL);
2484 spin_unlock_irqrestore(&n->list_lock, flags);
2490 * Slab on the partial list.
2492 remove_partial(n, page);
2493 stat(s, FREE_REMOVE_PARTIAL);
2495 /* Slab must be on the full list */
2496 remove_full(s, page);
2498 spin_unlock_irqrestore(&n->list_lock, flags);
2500 discard_slab(s, page);
2504 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2505 * can perform fastpath freeing without additional function calls.
2507 * The fastpath is only possible if we are freeing to the current cpu slab
2508 * of this processor. This typically the case if we have just allocated
2511 * If fastpath is not possible then fall back to __slab_free where we deal
2512 * with all sorts of special processing.
2514 static __always_inline void slab_free(struct kmem_cache *s,
2515 struct page *page, void *x, unsigned long addr)
2517 void **object = (void *)x;
2518 struct kmem_cache_cpu *c;
2521 slab_free_hook(s, x);
2525 * Determine the currently cpus per cpu slab.
2526 * The cpu may change afterward. However that does not matter since
2527 * data is retrieved via this pointer. If we are on the same cpu
2528 * during the cmpxchg then the free will succedd.
2530 c = __this_cpu_ptr(s->cpu_slab);
2535 if (likely(page == c->page)) {
2536 set_freepointer(s, object, c->freelist);
2538 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2539 s->cpu_slab->freelist, s->cpu_slab->tid,
2541 object, next_tid(tid)))) {
2543 note_cmpxchg_failure("slab_free", s, tid);
2546 stat(s, FREE_FASTPATH);
2548 __slab_free(s, page, x, addr);
2552 void kmem_cache_free(struct kmem_cache *s, void *x)
2556 page = virt_to_head_page(x);
2558 slab_free(s, page, x, _RET_IP_);
2560 trace_kmem_cache_free(_RET_IP_, x);
2562 EXPORT_SYMBOL(kmem_cache_free);
2565 * Object placement in a slab is made very easy because we always start at
2566 * offset 0. If we tune the size of the object to the alignment then we can
2567 * get the required alignment by putting one properly sized object after
2570 * Notice that the allocation order determines the sizes of the per cpu
2571 * caches. Each processor has always one slab available for allocations.
2572 * Increasing the allocation order reduces the number of times that slabs
2573 * must be moved on and off the partial lists and is therefore a factor in
2578 * Mininum / Maximum order of slab pages. This influences locking overhead
2579 * and slab fragmentation. A higher order reduces the number of partial slabs
2580 * and increases the number of allocations possible without having to
2581 * take the list_lock.
2583 static int slub_min_order;
2584 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2585 static int slub_min_objects;
2588 * Merge control. If this is set then no merging of slab caches will occur.
2589 * (Could be removed. This was introduced to pacify the merge skeptics.)
2591 static int slub_nomerge;
2594 * Calculate the order of allocation given an slab object size.
2596 * The order of allocation has significant impact on performance and other
2597 * system components. Generally order 0 allocations should be preferred since
2598 * order 0 does not cause fragmentation in the page allocator. Larger objects
2599 * be problematic to put into order 0 slabs because there may be too much
2600 * unused space left. We go to a higher order if more than 1/16th of the slab
2603 * In order to reach satisfactory performance we must ensure that a minimum
2604 * number of objects is in one slab. Otherwise we may generate too much
2605 * activity on the partial lists which requires taking the list_lock. This is
2606 * less a concern for large slabs though which are rarely used.
2608 * slub_max_order specifies the order where we begin to stop considering the
2609 * number of objects in a slab as critical. If we reach slub_max_order then
2610 * we try to keep the page order as low as possible. So we accept more waste
2611 * of space in favor of a small page order.
2613 * Higher order allocations also allow the placement of more objects in a
2614 * slab and thereby reduce object handling overhead. If the user has
2615 * requested a higher mininum order then we start with that one instead of
2616 * the smallest order which will fit the object.
2618 static inline int slab_order(int size, int min_objects,
2619 int max_order, int fract_leftover, int reserved)
2623 int min_order = slub_min_order;
2625 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2626 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2628 for (order = max(min_order,
2629 fls(min_objects * size - 1) - PAGE_SHIFT);
2630 order <= max_order; order++) {
2632 unsigned long slab_size = PAGE_SIZE << order;
2634 if (slab_size < min_objects * size + reserved)
2637 rem = (slab_size - reserved) % size;
2639 if (rem <= slab_size / fract_leftover)
2647 static inline int calculate_order(int size, int reserved)
2655 * Attempt to find best configuration for a slab. This
2656 * works by first attempting to generate a layout with
2657 * the best configuration and backing off gradually.
2659 * First we reduce the acceptable waste in a slab. Then
2660 * we reduce the minimum objects required in a slab.
2662 min_objects = slub_min_objects;
2664 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2665 max_objects = order_objects(slub_max_order, size, reserved);
2666 min_objects = min(min_objects, max_objects);
2668 while (min_objects > 1) {
2670 while (fraction >= 4) {
2671 order = slab_order(size, min_objects,
2672 slub_max_order, fraction, reserved);
2673 if (order <= slub_max_order)
2681 * We were unable to place multiple objects in a slab. Now
2682 * lets see if we can place a single object there.
2684 order = slab_order(size, 1, slub_max_order, 1, reserved);
2685 if (order <= slub_max_order)
2689 * Doh this slab cannot be placed using slub_max_order.
2691 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2692 if (order < MAX_ORDER)
2698 * Figure out what the alignment of the objects will be.
2700 static unsigned long calculate_alignment(unsigned long flags,
2701 unsigned long align, unsigned long size)
2704 * If the user wants hardware cache aligned objects then follow that
2705 * suggestion if the object is sufficiently large.
2707 * The hardware cache alignment cannot override the specified
2708 * alignment though. If that is greater then use it.
2710 if (flags & SLAB_HWCACHE_ALIGN) {
2711 unsigned long ralign = cache_line_size();
2712 while (size <= ralign / 2)
2714 align = max(align, ralign);
2717 if (align < ARCH_SLAB_MINALIGN)
2718 align = ARCH_SLAB_MINALIGN;
2720 return ALIGN(align, sizeof(void *));
2724 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2727 spin_lock_init(&n->list_lock);
2728 INIT_LIST_HEAD(&n->partial);
2729 #ifdef CONFIG_SLUB_DEBUG
2730 atomic_long_set(&n->nr_slabs, 0);
2731 atomic_long_set(&n->total_objects, 0);
2732 INIT_LIST_HEAD(&n->full);
2736 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2738 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2739 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2742 * Must align to double word boundary for the double cmpxchg
2743 * instructions to work; see __pcpu_double_call_return_bool().
2745 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2746 2 * sizeof(void *));
2751 init_kmem_cache_cpus(s);
2756 static struct kmem_cache *kmem_cache_node;
2759 * No kmalloc_node yet so do it by hand. We know that this is the first
2760 * slab on the node for this slabcache. There are no concurrent accesses
2763 * Note that this function only works on the kmalloc_node_cache
2764 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2765 * memory on a fresh node that has no slab structures yet.
2767 static void early_kmem_cache_node_alloc(int node)
2770 struct kmem_cache_node *n;
2772 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2774 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2777 if (page_to_nid(page) != node) {
2778 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2780 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2781 "in order to be able to continue\n");
2786 page->freelist = get_freepointer(kmem_cache_node, n);
2789 kmem_cache_node->node[node] = n;
2790 #ifdef CONFIG_SLUB_DEBUG
2791 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2792 init_tracking(kmem_cache_node, n);
2794 init_kmem_cache_node(n, kmem_cache_node);
2795 inc_slabs_node(kmem_cache_node, node, page->objects);
2797 add_partial(n, page, DEACTIVATE_TO_HEAD);
2800 static void free_kmem_cache_nodes(struct kmem_cache *s)
2804 for_each_node_state(node, N_NORMAL_MEMORY) {
2805 struct kmem_cache_node *n = s->node[node];
2808 kmem_cache_free(kmem_cache_node, n);
2810 s->node[node] = NULL;
2814 static int init_kmem_cache_nodes(struct kmem_cache *s)
2818 for_each_node_state(node, N_NORMAL_MEMORY) {
2819 struct kmem_cache_node *n;
2821 if (slab_state == DOWN) {
2822 early_kmem_cache_node_alloc(node);
2825 n = kmem_cache_alloc_node(kmem_cache_node,
2829 free_kmem_cache_nodes(s);
2834 init_kmem_cache_node(n, s);
2839 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2841 if (min < MIN_PARTIAL)
2843 else if (min > MAX_PARTIAL)
2845 s->min_partial = min;
2849 * calculate_sizes() determines the order and the distribution of data within
2852 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2854 unsigned long flags = s->flags;
2855 unsigned long size = s->objsize;
2856 unsigned long align = s->align;
2860 * Round up object size to the next word boundary. We can only
2861 * place the free pointer at word boundaries and this determines
2862 * the possible location of the free pointer.
2864 size = ALIGN(size, sizeof(void *));
2866 #ifdef CONFIG_SLUB_DEBUG
2868 * Determine if we can poison the object itself. If the user of
2869 * the slab may touch the object after free or before allocation
2870 * then we should never poison the object itself.
2872 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2874 s->flags |= __OBJECT_POISON;
2876 s->flags &= ~__OBJECT_POISON;
2880 * If we are Redzoning then check if there is some space between the
2881 * end of the object and the free pointer. If not then add an
2882 * additional word to have some bytes to store Redzone information.
2884 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2885 size += sizeof(void *);
2889 * With that we have determined the number of bytes in actual use
2890 * by the object. This is the potential offset to the free pointer.
2894 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2897 * Relocate free pointer after the object if it is not
2898 * permitted to overwrite the first word of the object on
2901 * This is the case if we do RCU, have a constructor or
2902 * destructor or are poisoning the objects.
2905 size += sizeof(void *);
2908 #ifdef CONFIG_SLUB_DEBUG
2909 if (flags & SLAB_STORE_USER)
2911 * Need to store information about allocs and frees after
2914 size += 2 * sizeof(struct track);
2916 if (flags & SLAB_RED_ZONE)
2918 * Add some empty padding so that we can catch
2919 * overwrites from earlier objects rather than let
2920 * tracking information or the free pointer be
2921 * corrupted if a user writes before the start
2924 size += sizeof(void *);
2928 * Determine the alignment based on various parameters that the
2929 * user specified and the dynamic determination of cache line size
2932 align = calculate_alignment(flags, align, s->objsize);
2936 * SLUB stores one object immediately after another beginning from
2937 * offset 0. In order to align the objects we have to simply size
2938 * each object to conform to the alignment.
2940 size = ALIGN(size, align);
2942 if (forced_order >= 0)
2943 order = forced_order;
2945 order = calculate_order(size, s->reserved);
2952 s->allocflags |= __GFP_COMP;
2954 if (s->flags & SLAB_CACHE_DMA)
2955 s->allocflags |= SLUB_DMA;
2957 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2958 s->allocflags |= __GFP_RECLAIMABLE;
2961 * Determine the number of objects per slab
2963 s->oo = oo_make(order, size, s->reserved);
2964 s->min = oo_make(get_order(size), size, s->reserved);
2965 if (oo_objects(s->oo) > oo_objects(s->max))
2968 return !!oo_objects(s->oo);
2972 static int kmem_cache_open(struct kmem_cache *s,
2973 const char *name, size_t size,
2974 size_t align, unsigned long flags,
2975 void (*ctor)(void *))
2977 memset(s, 0, kmem_size);
2982 s->flags = kmem_cache_flags(size, flags, name, ctor);
2985 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2986 s->reserved = sizeof(struct rcu_head);
2988 if (!calculate_sizes(s, -1))
2990 if (disable_higher_order_debug) {
2992 * Disable debugging flags that store metadata if the min slab
2995 if (get_order(s->size) > get_order(s->objsize)) {
2996 s->flags &= ~DEBUG_METADATA_FLAGS;
2998 if (!calculate_sizes(s, -1))
3003 #ifdef CONFIG_CMPXCHG_DOUBLE
3004 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3005 /* Enable fast mode */
3006 s->flags |= __CMPXCHG_DOUBLE;
3010 * The larger the object size is, the more pages we want on the partial
3011 * list to avoid pounding the page allocator excessively.
3013 set_min_partial(s, ilog2(s->size) / 2);
3016 * cpu_partial determined the maximum number of objects kept in the
3017 * per cpu partial lists of a processor.
3019 * Per cpu partial lists mainly contain slabs that just have one
3020 * object freed. If they are used for allocation then they can be
3021 * filled up again with minimal effort. The slab will never hit the
3022 * per node partial lists and therefore no locking will be required.
3024 * This setting also determines
3026 * A) The number of objects from per cpu partial slabs dumped to the
3027 * per node list when we reach the limit.
3028 * B) The number of objects in cpu partial slabs to extract from the
3029 * per node list when we run out of per cpu objects. We only fetch 50%
3030 * to keep some capacity around for frees.
3032 if (s->size >= PAGE_SIZE)
3034 else if (s->size >= 1024)
3036 else if (s->size >= 256)
3037 s->cpu_partial = 13;
3039 s->cpu_partial = 30;
3043 s->remote_node_defrag_ratio = 1000;
3045 if (!init_kmem_cache_nodes(s))
3048 if (alloc_kmem_cache_cpus(s))
3051 free_kmem_cache_nodes(s);
3053 if (flags & SLAB_PANIC)
3054 panic("Cannot create slab %s size=%lu realsize=%u "
3055 "order=%u offset=%u flags=%lx\n",
3056 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3062 * Determine the size of a slab object
3064 unsigned int kmem_cache_size(struct kmem_cache *s)
3068 EXPORT_SYMBOL(kmem_cache_size);
3070 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3073 #ifdef CONFIG_SLUB_DEBUG
3074 void *addr = page_address(page);
3076 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3077 sizeof(long), GFP_ATOMIC);
3080 slab_err(s, page, "%s", text);
3083 get_map(s, page, map);
3084 for_each_object(p, s, addr, page->objects) {
3086 if (!test_bit(slab_index(p, s, addr), map)) {
3087 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3089 print_tracking(s, p);
3098 * Attempt to free all partial slabs on a node.
3099 * This is called from kmem_cache_close(). We must be the last thread
3100 * using the cache and therefore we do not need to lock anymore.
3102 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3104 struct page *page, *h;
3106 list_for_each_entry_safe(page, h, &n->partial, lru) {
3108 remove_partial(n, page);
3109 discard_slab(s, page);
3111 list_slab_objects(s, page,
3112 "Objects remaining on kmem_cache_close()");
3118 * Release all resources used by a slab cache.
3120 static inline int kmem_cache_close(struct kmem_cache *s)
3125 free_percpu(s->cpu_slab);
3126 /* Attempt to free all objects */
3127 for_each_node_state(node, N_NORMAL_MEMORY) {
3128 struct kmem_cache_node *n = get_node(s, node);
3131 if (n->nr_partial || slabs_node(s, node))
3134 free_kmem_cache_nodes(s);
3139 * Close a cache and release the kmem_cache structure
3140 * (must be used for caches created using kmem_cache_create)
3142 void kmem_cache_destroy(struct kmem_cache *s)
3144 down_write(&slub_lock);
3148 up_write(&slub_lock);
3149 if (kmem_cache_close(s)) {
3150 printk(KERN_ERR "SLUB %s: %s called for cache that "
3151 "still has objects.\n", s->name, __func__);
3154 if (s->flags & SLAB_DESTROY_BY_RCU)
3156 sysfs_slab_remove(s);
3158 up_write(&slub_lock);
3160 EXPORT_SYMBOL(kmem_cache_destroy);
3162 /********************************************************************
3164 *******************************************************************/
3166 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3167 EXPORT_SYMBOL(kmalloc_caches);
3169 static struct kmem_cache *kmem_cache;
3171 #ifdef CONFIG_ZONE_DMA
3172 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3175 static int __init setup_slub_min_order(char *str)
3177 get_option(&str, &slub_min_order);
3182 __setup("slub_min_order=", setup_slub_min_order);
3184 static int __init setup_slub_max_order(char *str)
3186 get_option(&str, &slub_max_order);
3187 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3192 __setup("slub_max_order=", setup_slub_max_order);
3194 static int __init setup_slub_min_objects(char *str)
3196 get_option(&str, &slub_min_objects);
3201 __setup("slub_min_objects=", setup_slub_min_objects);
3203 static int __init setup_slub_nomerge(char *str)
3209 __setup("slub_nomerge", setup_slub_nomerge);
3211 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3212 int size, unsigned int flags)
3214 struct kmem_cache *s;
3216 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3219 * This function is called with IRQs disabled during early-boot on
3220 * single CPU so there's no need to take slub_lock here.
3222 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3226 list_add(&s->list, &slab_caches);
3230 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3235 * Conversion table for small slabs sizes / 8 to the index in the
3236 * kmalloc array. This is necessary for slabs < 192 since we have non power
3237 * of two cache sizes there. The size of larger slabs can be determined using
3240 static s8 size_index[24] = {
3267 static inline int size_index_elem(size_t bytes)
3269 return (bytes - 1) / 8;
3272 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3278 return ZERO_SIZE_PTR;
3280 index = size_index[size_index_elem(size)];
3282 index = fls(size - 1);
3284 #ifdef CONFIG_ZONE_DMA
3285 if (unlikely((flags & SLUB_DMA)))
3286 return kmalloc_dma_caches[index];
3289 return kmalloc_caches[index];
3292 void *__kmalloc(size_t size, gfp_t flags)
3294 struct kmem_cache *s;
3297 if (unlikely(size > SLUB_MAX_SIZE))
3298 return kmalloc_large(size, flags);
3300 s = get_slab(size, flags);
3302 if (unlikely(ZERO_OR_NULL_PTR(s)))
3305 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3307 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3311 EXPORT_SYMBOL(__kmalloc);
3314 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3319 flags |= __GFP_COMP | __GFP_NOTRACK;
3320 page = alloc_pages_node(node, flags, get_order(size));
3322 ptr = page_address(page);
3324 kmemleak_alloc(ptr, size, 1, flags);
3328 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3330 struct kmem_cache *s;
3333 if (unlikely(size > SLUB_MAX_SIZE)) {
3334 ret = kmalloc_large_node(size, flags, node);
3336 trace_kmalloc_node(_RET_IP_, ret,
3337 size, PAGE_SIZE << get_order(size),
3343 s = get_slab(size, flags);
3345 if (unlikely(ZERO_OR_NULL_PTR(s)))
3348 ret = slab_alloc(s, flags, node, _RET_IP_);
3350 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3354 EXPORT_SYMBOL(__kmalloc_node);
3357 size_t ksize(const void *object)
3361 if (unlikely(object == ZERO_SIZE_PTR))
3364 page = virt_to_head_page(object);
3366 if (unlikely(!PageSlab(page))) {
3367 WARN_ON(!PageCompound(page));
3368 return PAGE_SIZE << compound_order(page);
3371 return slab_ksize(page->slab);
3373 EXPORT_SYMBOL(ksize);
3375 #ifdef CONFIG_SLUB_DEBUG
3376 bool verify_mem_not_deleted(const void *x)
3379 void *object = (void *)x;
3380 unsigned long flags;
3383 if (unlikely(ZERO_OR_NULL_PTR(x)))
3386 local_irq_save(flags);
3388 page = virt_to_head_page(x);
3389 if (unlikely(!PageSlab(page))) {
3390 /* maybe it was from stack? */
3396 if (on_freelist(page->slab, page, object)) {
3397 object_err(page->slab, page, object, "Object is on free-list");
3405 local_irq_restore(flags);
3408 EXPORT_SYMBOL(verify_mem_not_deleted);
3411 void kfree(const void *x)
3414 void *object = (void *)x;
3416 trace_kfree(_RET_IP_, x);
3418 if (unlikely(ZERO_OR_NULL_PTR(x)))
3421 page = virt_to_head_page(x);
3422 if (unlikely(!PageSlab(page))) {
3423 BUG_ON(!PageCompound(page));
3428 slab_free(page->slab, page, object, _RET_IP_);
3430 EXPORT_SYMBOL(kfree);
3433 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3434 * the remaining slabs by the number of items in use. The slabs with the
3435 * most items in use come first. New allocations will then fill those up
3436 * and thus they can be removed from the partial lists.
3438 * The slabs with the least items are placed last. This results in them
3439 * being allocated from last increasing the chance that the last objects
3440 * are freed in them.
3442 int kmem_cache_shrink(struct kmem_cache *s)
3446 struct kmem_cache_node *n;
3449 int objects = oo_objects(s->max);
3450 struct list_head *slabs_by_inuse =
3451 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3452 unsigned long flags;
3454 if (!slabs_by_inuse)
3458 for_each_node_state(node, N_NORMAL_MEMORY) {
3459 n = get_node(s, node);
3464 for (i = 0; i < objects; i++)
3465 INIT_LIST_HEAD(slabs_by_inuse + i);
3467 spin_lock_irqsave(&n->list_lock, flags);
3470 * Build lists indexed by the items in use in each slab.
3472 * Note that concurrent frees may occur while we hold the
3473 * list_lock. page->inuse here is the upper limit.
3475 list_for_each_entry_safe(page, t, &n->partial, lru) {
3476 list_move(&page->lru, slabs_by_inuse + page->inuse);
3482 * Rebuild the partial list with the slabs filled up most
3483 * first and the least used slabs at the end.
3485 for (i = objects - 1; i > 0; i--)
3486 list_splice(slabs_by_inuse + i, n->partial.prev);
3488 spin_unlock_irqrestore(&n->list_lock, flags);
3490 /* Release empty slabs */
3491 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3492 discard_slab(s, page);
3495 kfree(slabs_by_inuse);
3498 EXPORT_SYMBOL(kmem_cache_shrink);
3500 #if defined(CONFIG_MEMORY_HOTPLUG)
3501 static int slab_mem_going_offline_callback(void *arg)
3503 struct kmem_cache *s;
3505 down_read(&slub_lock);
3506 list_for_each_entry(s, &slab_caches, list)
3507 kmem_cache_shrink(s);
3508 up_read(&slub_lock);
3513 static void slab_mem_offline_callback(void *arg)
3515 struct kmem_cache_node *n;
3516 struct kmem_cache *s;
3517 struct memory_notify *marg = arg;
3520 offline_node = marg->status_change_nid;
3523 * If the node still has available memory. we need kmem_cache_node
3526 if (offline_node < 0)
3529 down_read(&slub_lock);
3530 list_for_each_entry(s, &slab_caches, list) {
3531 n = get_node(s, offline_node);
3534 * if n->nr_slabs > 0, slabs still exist on the node
3535 * that is going down. We were unable to free them,
3536 * and offline_pages() function shouldn't call this
3537 * callback. So, we must fail.
3539 BUG_ON(slabs_node(s, offline_node));
3541 s->node[offline_node] = NULL;
3542 kmem_cache_free(kmem_cache_node, n);
3545 up_read(&slub_lock);
3548 static int slab_mem_going_online_callback(void *arg)
3550 struct kmem_cache_node *n;
3551 struct kmem_cache *s;
3552 struct memory_notify *marg = arg;
3553 int nid = marg->status_change_nid;
3557 * If the node's memory is already available, then kmem_cache_node is
3558 * already created. Nothing to do.
3564 * We are bringing a node online. No memory is available yet. We must
3565 * allocate a kmem_cache_node structure in order to bring the node
3568 down_read(&slub_lock);
3569 list_for_each_entry(s, &slab_caches, list) {
3571 * XXX: kmem_cache_alloc_node will fallback to other nodes
3572 * since memory is not yet available from the node that
3575 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3580 init_kmem_cache_node(n, s);
3584 up_read(&slub_lock);
3588 static int slab_memory_callback(struct notifier_block *self,
3589 unsigned long action, void *arg)
3594 case MEM_GOING_ONLINE:
3595 ret = slab_mem_going_online_callback(arg);
3597 case MEM_GOING_OFFLINE:
3598 ret = slab_mem_going_offline_callback(arg);
3601 case MEM_CANCEL_ONLINE:
3602 slab_mem_offline_callback(arg);
3605 case MEM_CANCEL_OFFLINE:
3609 ret = notifier_from_errno(ret);
3615 #endif /* CONFIG_MEMORY_HOTPLUG */
3617 /********************************************************************
3618 * Basic setup of slabs
3619 *******************************************************************/
3622 * Used for early kmem_cache structures that were allocated using
3623 * the page allocator
3626 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3630 list_add(&s->list, &slab_caches);
3633 for_each_node_state(node, N_NORMAL_MEMORY) {
3634 struct kmem_cache_node *n = get_node(s, node);
3638 list_for_each_entry(p, &n->partial, lru)
3641 #ifdef CONFIG_SLUB_DEBUG
3642 list_for_each_entry(p, &n->full, lru)
3649 void __init kmem_cache_init(void)
3653 struct kmem_cache *temp_kmem_cache;
3655 struct kmem_cache *temp_kmem_cache_node;
3656 unsigned long kmalloc_size;
3658 kmem_size = offsetof(struct kmem_cache, node) +
3659 nr_node_ids * sizeof(struct kmem_cache_node *);
3661 /* Allocate two kmem_caches from the page allocator */
3662 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3663 order = get_order(2 * kmalloc_size);
3664 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3667 * Must first have the slab cache available for the allocations of the
3668 * struct kmem_cache_node's. There is special bootstrap code in
3669 * kmem_cache_open for slab_state == DOWN.
3671 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3673 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3674 sizeof(struct kmem_cache_node),
3675 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3677 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3679 /* Able to allocate the per node structures */
3680 slab_state = PARTIAL;
3682 temp_kmem_cache = kmem_cache;
3683 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3684 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3685 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3686 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3689 * Allocate kmem_cache_node properly from the kmem_cache slab.
3690 * kmem_cache_node is separately allocated so no need to
3691 * update any list pointers.
3693 temp_kmem_cache_node = kmem_cache_node;
3695 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3696 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3698 kmem_cache_bootstrap_fixup(kmem_cache_node);
3701 kmem_cache_bootstrap_fixup(kmem_cache);
3703 /* Free temporary boot structure */
3704 free_pages((unsigned long)temp_kmem_cache, order);
3706 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3709 * Patch up the size_index table if we have strange large alignment
3710 * requirements for the kmalloc array. This is only the case for
3711 * MIPS it seems. The standard arches will not generate any code here.
3713 * Largest permitted alignment is 256 bytes due to the way we
3714 * handle the index determination for the smaller caches.
3716 * Make sure that nothing crazy happens if someone starts tinkering
3717 * around with ARCH_KMALLOC_MINALIGN
3719 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3720 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3722 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3723 int elem = size_index_elem(i);
3724 if (elem >= ARRAY_SIZE(size_index))
3726 size_index[elem] = KMALLOC_SHIFT_LOW;
3729 if (KMALLOC_MIN_SIZE == 64) {
3731 * The 96 byte size cache is not used if the alignment
3734 for (i = 64 + 8; i <= 96; i += 8)
3735 size_index[size_index_elem(i)] = 7;
3736 } else if (KMALLOC_MIN_SIZE == 128) {
3738 * The 192 byte sized cache is not used if the alignment
3739 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3742 for (i = 128 + 8; i <= 192; i += 8)
3743 size_index[size_index_elem(i)] = 8;
3746 /* Caches that are not of the two-to-the-power-of size */
3747 if (KMALLOC_MIN_SIZE <= 32) {
3748 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3752 if (KMALLOC_MIN_SIZE <= 64) {
3753 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3757 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3758 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3764 /* Provide the correct kmalloc names now that the caches are up */
3765 if (KMALLOC_MIN_SIZE <= 32) {
3766 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3767 BUG_ON(!kmalloc_caches[1]->name);
3770 if (KMALLOC_MIN_SIZE <= 64) {
3771 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3772 BUG_ON(!kmalloc_caches[2]->name);
3775 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3776 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3779 kmalloc_caches[i]->name = s;
3783 register_cpu_notifier(&slab_notifier);
3786 #ifdef CONFIG_ZONE_DMA
3787 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3788 struct kmem_cache *s = kmalloc_caches[i];
3791 char *name = kasprintf(GFP_NOWAIT,
3792 "dma-kmalloc-%d", s->objsize);
3795 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3796 s->objsize, SLAB_CACHE_DMA);
3801 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3802 " CPUs=%d, Nodes=%d\n",
3803 caches, cache_line_size(),
3804 slub_min_order, slub_max_order, slub_min_objects,
3805 nr_cpu_ids, nr_node_ids);
3808 void __init kmem_cache_init_late(void)
3813 * Find a mergeable slab cache
3815 static int slab_unmergeable(struct kmem_cache *s)
3817 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3824 * We may have set a slab to be unmergeable during bootstrap.
3826 if (s->refcount < 0)
3832 static struct kmem_cache *find_mergeable(size_t size,
3833 size_t align, unsigned long flags, const char *name,
3834 void (*ctor)(void *))
3836 struct kmem_cache *s;
3838 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3844 size = ALIGN(size, sizeof(void *));
3845 align = calculate_alignment(flags, align, size);
3846 size = ALIGN(size, align);
3847 flags = kmem_cache_flags(size, flags, name, NULL);
3849 list_for_each_entry(s, &slab_caches, list) {
3850 if (slab_unmergeable(s))
3856 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3859 * Check if alignment is compatible.
3860 * Courtesy of Adrian Drzewiecki
3862 if ((s->size & ~(align - 1)) != s->size)
3865 if (s->size - size >= sizeof(void *))
3873 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3874 size_t align, unsigned long flags, void (*ctor)(void *))
3876 struct kmem_cache *s;
3882 down_write(&slub_lock);
3883 s = find_mergeable(size, align, flags, name, ctor);
3887 * Adjust the object sizes so that we clear
3888 * the complete object on kzalloc.
3890 s->objsize = max(s->objsize, (int)size);
3891 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3893 if (sysfs_slab_alias(s, name)) {
3897 up_write(&slub_lock);
3901 n = kstrdup(name, GFP_KERNEL);
3905 s = kmalloc(kmem_size, GFP_KERNEL);
3907 if (kmem_cache_open(s, n,
3908 size, align, flags, ctor)) {
3909 list_add(&s->list, &slab_caches);
3910 up_write(&slub_lock);
3911 if (sysfs_slab_add(s)) {
3912 down_write(&slub_lock);
3924 up_write(&slub_lock);
3926 if (flags & SLAB_PANIC)
3927 panic("Cannot create slabcache %s\n", name);
3932 EXPORT_SYMBOL(kmem_cache_create);
3936 * Use the cpu notifier to insure that the cpu slabs are flushed when
3939 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3940 unsigned long action, void *hcpu)
3942 long cpu = (long)hcpu;
3943 struct kmem_cache *s;
3944 unsigned long flags;
3947 case CPU_UP_CANCELED:
3948 case CPU_UP_CANCELED_FROZEN:
3950 case CPU_DEAD_FROZEN:
3951 down_read(&slub_lock);
3952 list_for_each_entry(s, &slab_caches, list) {
3953 local_irq_save(flags);
3954 __flush_cpu_slab(s, cpu);
3955 local_irq_restore(flags);
3957 up_read(&slub_lock);
3965 static struct notifier_block __cpuinitdata slab_notifier = {
3966 .notifier_call = slab_cpuup_callback
3971 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3973 struct kmem_cache *s;
3976 if (unlikely(size > SLUB_MAX_SIZE))
3977 return kmalloc_large(size, gfpflags);
3979 s = get_slab(size, gfpflags);
3981 if (unlikely(ZERO_OR_NULL_PTR(s)))
3984 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3986 /* Honor the call site pointer we received. */
3987 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3993 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3994 int node, unsigned long caller)
3996 struct kmem_cache *s;
3999 if (unlikely(size > SLUB_MAX_SIZE)) {
4000 ret = kmalloc_large_node(size, gfpflags, node);
4002 trace_kmalloc_node(caller, ret,
4003 size, PAGE_SIZE << get_order(size),
4009 s = get_slab(size, gfpflags);
4011 if (unlikely(ZERO_OR_NULL_PTR(s)))
4014 ret = slab_alloc(s, gfpflags, node, caller);
4016 /* Honor the call site pointer we received. */
4017 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4024 static int count_inuse(struct page *page)
4029 static int count_total(struct page *page)
4031 return page->objects;
4035 #ifdef CONFIG_SLUB_DEBUG
4036 static int validate_slab(struct kmem_cache *s, struct page *page,
4040 void *addr = page_address(page);
4042 if (!check_slab(s, page) ||
4043 !on_freelist(s, page, NULL))
4046 /* Now we know that a valid freelist exists */
4047 bitmap_zero(map, page->objects);
4049 get_map(s, page, map);
4050 for_each_object(p, s, addr, page->objects) {
4051 if (test_bit(slab_index(p, s, addr), map))
4052 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4056 for_each_object(p, s, addr, page->objects)
4057 if (!test_bit(slab_index(p, s, addr), map))
4058 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4063 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4067 validate_slab(s, page, map);
4071 static int validate_slab_node(struct kmem_cache *s,
4072 struct kmem_cache_node *n, unsigned long *map)
4074 unsigned long count = 0;
4076 unsigned long flags;
4078 spin_lock_irqsave(&n->list_lock, flags);
4080 list_for_each_entry(page, &n->partial, lru) {
4081 validate_slab_slab(s, page, map);
4084 if (count != n->nr_partial)
4085 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4086 "counter=%ld\n", s->name, count, n->nr_partial);
4088 if (!(s->flags & SLAB_STORE_USER))
4091 list_for_each_entry(page, &n->full, lru) {
4092 validate_slab_slab(s, page, map);
4095 if (count != atomic_long_read(&n->nr_slabs))
4096 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4097 "counter=%ld\n", s->name, count,
4098 atomic_long_read(&n->nr_slabs));
4101 spin_unlock_irqrestore(&n->list_lock, flags);
4105 static long validate_slab_cache(struct kmem_cache *s)
4108 unsigned long count = 0;
4109 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4110 sizeof(unsigned long), GFP_KERNEL);
4116 for_each_node_state(node, N_NORMAL_MEMORY) {
4117 struct kmem_cache_node *n = get_node(s, node);
4119 count += validate_slab_node(s, n, map);
4125 * Generate lists of code addresses where slabcache objects are allocated
4130 unsigned long count;
4137 DECLARE_BITMAP(cpus, NR_CPUS);
4143 unsigned long count;
4144 struct location *loc;
4147 static void free_loc_track(struct loc_track *t)
4150 free_pages((unsigned long)t->loc,
4151 get_order(sizeof(struct location) * t->max));
4154 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4159 order = get_order(sizeof(struct location) * max);
4161 l = (void *)__get_free_pages(flags, order);
4166 memcpy(l, t->loc, sizeof(struct location) * t->count);
4174 static int add_location(struct loc_track *t, struct kmem_cache *s,
4175 const struct track *track)
4177 long start, end, pos;
4179 unsigned long caddr;
4180 unsigned long age = jiffies - track->when;
4186 pos = start + (end - start + 1) / 2;
4189 * There is nothing at "end". If we end up there
4190 * we need to add something to before end.
4195 caddr = t->loc[pos].addr;
4196 if (track->addr == caddr) {
4202 if (age < l->min_time)
4204 if (age > l->max_time)
4207 if (track->pid < l->min_pid)
4208 l->min_pid = track->pid;
4209 if (track->pid > l->max_pid)
4210 l->max_pid = track->pid;
4212 cpumask_set_cpu(track->cpu,
4213 to_cpumask(l->cpus));
4215 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4219 if (track->addr < caddr)
4226 * Not found. Insert new tracking element.
4228 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4234 (t->count - pos) * sizeof(struct location));
4237 l->addr = track->addr;
4241 l->min_pid = track->pid;
4242 l->max_pid = track->pid;
4243 cpumask_clear(to_cpumask(l->cpus));
4244 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4245 nodes_clear(l->nodes);
4246 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4250 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4251 struct page *page, enum track_item alloc,
4254 void *addr = page_address(page);
4257 bitmap_zero(map, page->objects);
4258 get_map(s, page, map);
4260 for_each_object(p, s, addr, page->objects)
4261 if (!test_bit(slab_index(p, s, addr), map))
4262 add_location(t, s, get_track(s, p, alloc));
4265 static int list_locations(struct kmem_cache *s, char *buf,
4266 enum track_item alloc)
4270 struct loc_track t = { 0, 0, NULL };
4272 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4273 sizeof(unsigned long), GFP_KERNEL);
4275 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4278 return sprintf(buf, "Out of memory\n");
4280 /* Push back cpu slabs */
4283 for_each_node_state(node, N_NORMAL_MEMORY) {
4284 struct kmem_cache_node *n = get_node(s, node);
4285 unsigned long flags;
4288 if (!atomic_long_read(&n->nr_slabs))
4291 spin_lock_irqsave(&n->list_lock, flags);
4292 list_for_each_entry(page, &n->partial, lru)
4293 process_slab(&t, s, page, alloc, map);
4294 list_for_each_entry(page, &n->full, lru)
4295 process_slab(&t, s, page, alloc, map);
4296 spin_unlock_irqrestore(&n->list_lock, flags);
4299 for (i = 0; i < t.count; i++) {
4300 struct location *l = &t.loc[i];
4302 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4304 len += sprintf(buf + len, "%7ld ", l->count);
4307 len += sprintf(buf + len, "%pS", (void *)l->addr);
4309 len += sprintf(buf + len, "<not-available>");
4311 if (l->sum_time != l->min_time) {
4312 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4314 (long)div_u64(l->sum_time, l->count),
4317 len += sprintf(buf + len, " age=%ld",
4320 if (l->min_pid != l->max_pid)
4321 len += sprintf(buf + len, " pid=%ld-%ld",
4322 l->min_pid, l->max_pid);
4324 len += sprintf(buf + len, " pid=%ld",
4327 if (num_online_cpus() > 1 &&
4328 !cpumask_empty(to_cpumask(l->cpus)) &&
4329 len < PAGE_SIZE - 60) {
4330 len += sprintf(buf + len, " cpus=");
4331 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4332 to_cpumask(l->cpus));
4335 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4336 len < PAGE_SIZE - 60) {
4337 len += sprintf(buf + len, " nodes=");
4338 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4342 len += sprintf(buf + len, "\n");
4348 len += sprintf(buf, "No data\n");
4353 #ifdef SLUB_RESILIENCY_TEST
4354 static void resiliency_test(void)
4358 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4360 printk(KERN_ERR "SLUB resiliency testing\n");
4361 printk(KERN_ERR "-----------------------\n");
4362 printk(KERN_ERR "A. Corruption after allocation\n");
4364 p = kzalloc(16, GFP_KERNEL);
4366 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4367 " 0x12->0x%p\n\n", p + 16);
4369 validate_slab_cache(kmalloc_caches[4]);
4371 /* Hmmm... The next two are dangerous */
4372 p = kzalloc(32, GFP_KERNEL);
4373 p[32 + sizeof(void *)] = 0x34;
4374 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4375 " 0x34 -> -0x%p\n", p);
4377 "If allocated object is overwritten then not detectable\n\n");
4379 validate_slab_cache(kmalloc_caches[5]);
4380 p = kzalloc(64, GFP_KERNEL);
4381 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4383 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4386 "If allocated object is overwritten then not detectable\n\n");
4387 validate_slab_cache(kmalloc_caches[6]);
4389 printk(KERN_ERR "\nB. Corruption after free\n");
4390 p = kzalloc(128, GFP_KERNEL);
4393 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4394 validate_slab_cache(kmalloc_caches[7]);
4396 p = kzalloc(256, GFP_KERNEL);
4399 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4401 validate_slab_cache(kmalloc_caches[8]);
4403 p = kzalloc(512, GFP_KERNEL);
4406 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4407 validate_slab_cache(kmalloc_caches[9]);
4411 static void resiliency_test(void) {};
4416 enum slab_stat_type {
4417 SL_ALL, /* All slabs */
4418 SL_PARTIAL, /* Only partially allocated slabs */
4419 SL_CPU, /* Only slabs used for cpu caches */
4420 SL_OBJECTS, /* Determine allocated objects not slabs */
4421 SL_TOTAL /* Determine object capacity not slabs */
4424 #define SO_ALL (1 << SL_ALL)
4425 #define SO_PARTIAL (1 << SL_PARTIAL)
4426 #define SO_CPU (1 << SL_CPU)
4427 #define SO_OBJECTS (1 << SL_OBJECTS)
4428 #define SO_TOTAL (1 << SL_TOTAL)
4430 static ssize_t show_slab_objects(struct kmem_cache *s,
4431 char *buf, unsigned long flags)
4433 unsigned long total = 0;
4436 unsigned long *nodes;
4437 unsigned long *per_cpu;
4439 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4442 per_cpu = nodes + nr_node_ids;
4444 if (flags & SO_CPU) {
4447 for_each_possible_cpu(cpu) {
4448 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4449 int node = ACCESS_ONCE(c->node);
4454 page = ACCESS_ONCE(c->page);
4456 if (flags & SO_TOTAL)
4458 else if (flags & SO_OBJECTS)
4469 node = page_to_nid(page);
4470 if (flags & SO_TOTAL)
4472 else if (flags & SO_OBJECTS)
4483 lock_memory_hotplug();
4484 #ifdef CONFIG_SLUB_DEBUG
4485 if (flags & SO_ALL) {
4486 for_each_node_state(node, N_NORMAL_MEMORY) {
4487 struct kmem_cache_node *n = get_node(s, node);
4489 if (flags & SO_TOTAL)
4490 x = atomic_long_read(&n->total_objects);
4491 else if (flags & SO_OBJECTS)
4492 x = atomic_long_read(&n->total_objects) -
4493 count_partial(n, count_free);
4496 x = atomic_long_read(&n->nr_slabs);
4503 if (flags & SO_PARTIAL) {
4504 for_each_node_state(node, N_NORMAL_MEMORY) {
4505 struct kmem_cache_node *n = get_node(s, node);
4507 if (flags & SO_TOTAL)
4508 x = count_partial(n, count_total);
4509 else if (flags & SO_OBJECTS)
4510 x = count_partial(n, count_inuse);
4517 x = sprintf(buf, "%lu", total);
4519 for_each_node_state(node, N_NORMAL_MEMORY)
4521 x += sprintf(buf + x, " N%d=%lu",
4524 unlock_memory_hotplug();
4526 return x + sprintf(buf + x, "\n");
4529 #ifdef CONFIG_SLUB_DEBUG
4530 static int any_slab_objects(struct kmem_cache *s)
4534 for_each_online_node(node) {
4535 struct kmem_cache_node *n = get_node(s, node);
4540 if (atomic_long_read(&n->total_objects))
4547 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4548 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4550 struct slab_attribute {
4551 struct attribute attr;
4552 ssize_t (*show)(struct kmem_cache *s, char *buf);
4553 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4556 #define SLAB_ATTR_RO(_name) \
4557 static struct slab_attribute _name##_attr = \
4558 __ATTR(_name, 0400, _name##_show, NULL)
4560 #define SLAB_ATTR(_name) \
4561 static struct slab_attribute _name##_attr = \
4562 __ATTR(_name, 0600, _name##_show, _name##_store)
4564 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4566 return sprintf(buf, "%d\n", s->size);
4568 SLAB_ATTR_RO(slab_size);
4570 static ssize_t align_show(struct kmem_cache *s, char *buf)
4572 return sprintf(buf, "%d\n", s->align);
4574 SLAB_ATTR_RO(align);
4576 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4578 return sprintf(buf, "%d\n", s->objsize);
4580 SLAB_ATTR_RO(object_size);
4582 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4584 return sprintf(buf, "%d\n", oo_objects(s->oo));
4586 SLAB_ATTR_RO(objs_per_slab);
4588 static ssize_t order_store(struct kmem_cache *s,
4589 const char *buf, size_t length)
4591 unsigned long order;
4594 err = strict_strtoul(buf, 10, &order);
4598 if (order > slub_max_order || order < slub_min_order)
4601 calculate_sizes(s, order);
4605 static ssize_t order_show(struct kmem_cache *s, char *buf)
4607 return sprintf(buf, "%d\n", oo_order(s->oo));
4611 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4613 return sprintf(buf, "%lu\n", s->min_partial);
4616 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4622 err = strict_strtoul(buf, 10, &min);
4626 set_min_partial(s, min);
4629 SLAB_ATTR(min_partial);
4631 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4633 return sprintf(buf, "%u\n", s->cpu_partial);
4636 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4639 unsigned long objects;
4642 err = strict_strtoul(buf, 10, &objects);
4646 s->cpu_partial = objects;
4650 SLAB_ATTR(cpu_partial);
4652 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4656 return sprintf(buf, "%pS\n", s->ctor);
4660 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4662 return sprintf(buf, "%d\n", s->refcount - 1);
4664 SLAB_ATTR_RO(aliases);
4666 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4668 return show_slab_objects(s, buf, SO_PARTIAL);
4670 SLAB_ATTR_RO(partial);
4672 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4674 return show_slab_objects(s, buf, SO_CPU);
4676 SLAB_ATTR_RO(cpu_slabs);
4678 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4680 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4682 SLAB_ATTR_RO(objects);
4684 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4686 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4688 SLAB_ATTR_RO(objects_partial);
4690 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4697 for_each_online_cpu(cpu) {
4698 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4701 pages += page->pages;
4702 objects += page->pobjects;
4706 len = sprintf(buf, "%d(%d)", objects, pages);
4709 for_each_online_cpu(cpu) {
4710 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4712 if (page && len < PAGE_SIZE - 20)
4713 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4714 page->pobjects, page->pages);
4717 return len + sprintf(buf + len, "\n");
4719 SLAB_ATTR_RO(slabs_cpu_partial);
4721 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4723 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4726 static ssize_t reclaim_account_store(struct kmem_cache *s,
4727 const char *buf, size_t length)
4729 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4731 s->flags |= SLAB_RECLAIM_ACCOUNT;
4734 SLAB_ATTR(reclaim_account);
4736 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4738 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4740 SLAB_ATTR_RO(hwcache_align);
4742 #ifdef CONFIG_ZONE_DMA
4743 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4745 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4747 SLAB_ATTR_RO(cache_dma);
4750 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4752 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4754 SLAB_ATTR_RO(destroy_by_rcu);
4756 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4758 return sprintf(buf, "%d\n", s->reserved);
4760 SLAB_ATTR_RO(reserved);
4762 #ifdef CONFIG_SLUB_DEBUG
4763 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4765 return show_slab_objects(s, buf, SO_ALL);
4767 SLAB_ATTR_RO(slabs);
4769 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4771 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4773 SLAB_ATTR_RO(total_objects);
4775 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4777 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4780 static ssize_t sanity_checks_store(struct kmem_cache *s,
4781 const char *buf, size_t length)
4783 s->flags &= ~SLAB_DEBUG_FREE;
4784 if (buf[0] == '1') {
4785 s->flags &= ~__CMPXCHG_DOUBLE;
4786 s->flags |= SLAB_DEBUG_FREE;
4790 SLAB_ATTR(sanity_checks);
4792 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4794 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4797 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4800 s->flags &= ~SLAB_TRACE;
4801 if (buf[0] == '1') {
4802 s->flags &= ~__CMPXCHG_DOUBLE;
4803 s->flags |= SLAB_TRACE;
4809 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4811 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4814 static ssize_t red_zone_store(struct kmem_cache *s,
4815 const char *buf, size_t length)
4817 if (any_slab_objects(s))
4820 s->flags &= ~SLAB_RED_ZONE;
4821 if (buf[0] == '1') {
4822 s->flags &= ~__CMPXCHG_DOUBLE;
4823 s->flags |= SLAB_RED_ZONE;
4825 calculate_sizes(s, -1);
4828 SLAB_ATTR(red_zone);
4830 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4832 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4835 static ssize_t poison_store(struct kmem_cache *s,
4836 const char *buf, size_t length)
4838 if (any_slab_objects(s))
4841 s->flags &= ~SLAB_POISON;
4842 if (buf[0] == '1') {
4843 s->flags &= ~__CMPXCHG_DOUBLE;
4844 s->flags |= SLAB_POISON;
4846 calculate_sizes(s, -1);
4851 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4853 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4856 static ssize_t store_user_store(struct kmem_cache *s,
4857 const char *buf, size_t length)
4859 if (any_slab_objects(s))
4862 s->flags &= ~SLAB_STORE_USER;
4863 if (buf[0] == '1') {
4864 s->flags &= ~__CMPXCHG_DOUBLE;
4865 s->flags |= SLAB_STORE_USER;
4867 calculate_sizes(s, -1);
4870 SLAB_ATTR(store_user);
4872 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4877 static ssize_t validate_store(struct kmem_cache *s,
4878 const char *buf, size_t length)
4882 if (buf[0] == '1') {
4883 ret = validate_slab_cache(s);
4889 SLAB_ATTR(validate);
4891 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4893 if (!(s->flags & SLAB_STORE_USER))
4895 return list_locations(s, buf, TRACK_ALLOC);
4897 SLAB_ATTR_RO(alloc_calls);
4899 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4901 if (!(s->flags & SLAB_STORE_USER))
4903 return list_locations(s, buf, TRACK_FREE);
4905 SLAB_ATTR_RO(free_calls);
4906 #endif /* CONFIG_SLUB_DEBUG */
4908 #ifdef CONFIG_FAILSLAB
4909 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4911 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4914 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4917 s->flags &= ~SLAB_FAILSLAB;
4919 s->flags |= SLAB_FAILSLAB;
4922 SLAB_ATTR(failslab);
4925 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4930 static ssize_t shrink_store(struct kmem_cache *s,
4931 const char *buf, size_t length)
4933 if (buf[0] == '1') {
4934 int rc = kmem_cache_shrink(s);
4945 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4947 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4950 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4951 const char *buf, size_t length)
4953 unsigned long ratio;
4956 err = strict_strtoul(buf, 10, &ratio);
4961 s->remote_node_defrag_ratio = ratio * 10;
4965 SLAB_ATTR(remote_node_defrag_ratio);
4968 #ifdef CONFIG_SLUB_STATS
4969 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4971 unsigned long sum = 0;
4974 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4979 for_each_online_cpu(cpu) {
4980 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4986 len = sprintf(buf, "%lu", sum);
4989 for_each_online_cpu(cpu) {
4990 if (data[cpu] && len < PAGE_SIZE - 20)
4991 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4995 return len + sprintf(buf + len, "\n");
4998 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5002 for_each_online_cpu(cpu)
5003 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5006 #define STAT_ATTR(si, text) \
5007 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5009 return show_stat(s, buf, si); \
5011 static ssize_t text##_store(struct kmem_cache *s, \
5012 const char *buf, size_t length) \
5014 if (buf[0] != '0') \
5016 clear_stat(s, si); \
5021 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5022 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5023 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5024 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5025 STAT_ATTR(FREE_FROZEN, free_frozen);
5026 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5027 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5028 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5029 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5030 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5031 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5032 STAT_ATTR(FREE_SLAB, free_slab);
5033 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5034 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5035 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5036 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5037 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5038 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5039 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5040 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5041 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5042 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5043 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5044 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5047 static struct attribute *slab_attrs[] = {
5048 &slab_size_attr.attr,
5049 &object_size_attr.attr,
5050 &objs_per_slab_attr.attr,
5052 &min_partial_attr.attr,
5053 &cpu_partial_attr.attr,
5055 &objects_partial_attr.attr,
5057 &cpu_slabs_attr.attr,
5061 &hwcache_align_attr.attr,
5062 &reclaim_account_attr.attr,
5063 &destroy_by_rcu_attr.attr,
5065 &reserved_attr.attr,
5066 &slabs_cpu_partial_attr.attr,
5067 #ifdef CONFIG_SLUB_DEBUG
5068 &total_objects_attr.attr,
5070 &sanity_checks_attr.attr,
5072 &red_zone_attr.attr,
5074 &store_user_attr.attr,
5075 &validate_attr.attr,
5076 &alloc_calls_attr.attr,
5077 &free_calls_attr.attr,
5079 #ifdef CONFIG_ZONE_DMA
5080 &cache_dma_attr.attr,
5083 &remote_node_defrag_ratio_attr.attr,
5085 #ifdef CONFIG_SLUB_STATS
5086 &alloc_fastpath_attr.attr,
5087 &alloc_slowpath_attr.attr,
5088 &free_fastpath_attr.attr,
5089 &free_slowpath_attr.attr,
5090 &free_frozen_attr.attr,
5091 &free_add_partial_attr.attr,
5092 &free_remove_partial_attr.attr,
5093 &alloc_from_partial_attr.attr,
5094 &alloc_slab_attr.attr,
5095 &alloc_refill_attr.attr,
5096 &alloc_node_mismatch_attr.attr,
5097 &free_slab_attr.attr,
5098 &cpuslab_flush_attr.attr,
5099 &deactivate_full_attr.attr,
5100 &deactivate_empty_attr.attr,
5101 &deactivate_to_head_attr.attr,
5102 &deactivate_to_tail_attr.attr,
5103 &deactivate_remote_frees_attr.attr,
5104 &deactivate_bypass_attr.attr,
5105 &order_fallback_attr.attr,
5106 &cmpxchg_double_fail_attr.attr,
5107 &cmpxchg_double_cpu_fail_attr.attr,
5108 &cpu_partial_alloc_attr.attr,
5109 &cpu_partial_free_attr.attr,
5111 #ifdef CONFIG_FAILSLAB
5112 &failslab_attr.attr,
5118 static struct attribute_group slab_attr_group = {
5119 .attrs = slab_attrs,
5122 static ssize_t slab_attr_show(struct kobject *kobj,
5123 struct attribute *attr,
5126 struct slab_attribute *attribute;
5127 struct kmem_cache *s;
5130 attribute = to_slab_attr(attr);
5133 if (!attribute->show)
5136 err = attribute->show(s, buf);
5141 static ssize_t slab_attr_store(struct kobject *kobj,
5142 struct attribute *attr,
5143 const char *buf, size_t len)
5145 struct slab_attribute *attribute;
5146 struct kmem_cache *s;
5149 attribute = to_slab_attr(attr);
5152 if (!attribute->store)
5155 err = attribute->store(s, buf, len);
5160 static void kmem_cache_release(struct kobject *kobj)
5162 struct kmem_cache *s = to_slab(kobj);
5168 static const struct sysfs_ops slab_sysfs_ops = {
5169 .show = slab_attr_show,
5170 .store = slab_attr_store,
5173 static struct kobj_type slab_ktype = {
5174 .sysfs_ops = &slab_sysfs_ops,
5175 .release = kmem_cache_release
5178 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5180 struct kobj_type *ktype = get_ktype(kobj);
5182 if (ktype == &slab_ktype)
5187 static const struct kset_uevent_ops slab_uevent_ops = {
5188 .filter = uevent_filter,
5191 static struct kset *slab_kset;
5193 #define ID_STR_LENGTH 64
5195 /* Create a unique string id for a slab cache:
5197 * Format :[flags-]size
5199 static char *create_unique_id(struct kmem_cache *s)
5201 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5208 * First flags affecting slabcache operations. We will only
5209 * get here for aliasable slabs so we do not need to support
5210 * too many flags. The flags here must cover all flags that
5211 * are matched during merging to guarantee that the id is
5214 if (s->flags & SLAB_CACHE_DMA)
5216 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5218 if (s->flags & SLAB_DEBUG_FREE)
5220 if (!(s->flags & SLAB_NOTRACK))
5224 p += sprintf(p, "%07d", s->size);
5225 BUG_ON(p > name + ID_STR_LENGTH - 1);
5229 static int sysfs_slab_add(struct kmem_cache *s)
5235 if (slab_state < SYSFS)
5236 /* Defer until later */
5239 unmergeable = slab_unmergeable(s);
5242 * Slabcache can never be merged so we can use the name proper.
5243 * This is typically the case for debug situations. In that
5244 * case we can catch duplicate names easily.
5246 sysfs_remove_link(&slab_kset->kobj, s->name);
5250 * Create a unique name for the slab as a target
5253 name = create_unique_id(s);
5256 s->kobj.kset = slab_kset;
5257 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5259 kobject_put(&s->kobj);
5263 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5265 kobject_del(&s->kobj);
5266 kobject_put(&s->kobj);
5269 kobject_uevent(&s->kobj, KOBJ_ADD);
5271 /* Setup first alias */
5272 sysfs_slab_alias(s, s->name);
5278 static void sysfs_slab_remove(struct kmem_cache *s)
5280 if (slab_state < SYSFS)
5282 * Sysfs has not been setup yet so no need to remove the
5287 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5288 kobject_del(&s->kobj);
5289 kobject_put(&s->kobj);
5293 * Need to buffer aliases during bootup until sysfs becomes
5294 * available lest we lose that information.
5296 struct saved_alias {
5297 struct kmem_cache *s;
5299 struct saved_alias *next;
5302 static struct saved_alias *alias_list;
5304 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5306 struct saved_alias *al;
5308 if (slab_state == SYSFS) {
5310 * If we have a leftover link then remove it.
5312 sysfs_remove_link(&slab_kset->kobj, name);
5313 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5316 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5322 al->next = alias_list;
5327 static int __init slab_sysfs_init(void)
5329 struct kmem_cache *s;
5332 down_write(&slub_lock);
5334 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5336 up_write(&slub_lock);
5337 printk(KERN_ERR "Cannot register slab subsystem.\n");
5343 list_for_each_entry(s, &slab_caches, list) {
5344 err = sysfs_slab_add(s);
5346 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5347 " to sysfs\n", s->name);
5350 while (alias_list) {
5351 struct saved_alias *al = alias_list;
5353 alias_list = alias_list->next;
5354 err = sysfs_slab_alias(al->s, al->name);
5356 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5357 " %s to sysfs\n", s->name);
5361 up_write(&slub_lock);
5366 __initcall(slab_sysfs_init);
5367 #endif /* CONFIG_SYSFS */
5370 * The /proc/slabinfo ABI
5372 #ifdef CONFIG_SLABINFO
5373 static void print_slabinfo_header(struct seq_file *m)
5375 seq_puts(m, "slabinfo - version: 2.1\n");
5376 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5377 "<objperslab> <pagesperslab>");
5378 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5379 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5383 static void *s_start(struct seq_file *m, loff_t *pos)
5387 down_read(&slub_lock);
5389 print_slabinfo_header(m);
5391 return seq_list_start(&slab_caches, *pos);
5394 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5396 return seq_list_next(p, &slab_caches, pos);
5399 static void s_stop(struct seq_file *m, void *p)
5401 up_read(&slub_lock);
5404 static int s_show(struct seq_file *m, void *p)
5406 unsigned long nr_partials = 0;
5407 unsigned long nr_slabs = 0;
5408 unsigned long nr_inuse = 0;
5409 unsigned long nr_objs = 0;
5410 unsigned long nr_free = 0;
5411 struct kmem_cache *s;
5414 s = list_entry(p, struct kmem_cache, list);
5416 for_each_online_node(node) {
5417 struct kmem_cache_node *n = get_node(s, node);
5422 nr_partials += n->nr_partial;
5423 nr_slabs += atomic_long_read(&n->nr_slabs);
5424 nr_objs += atomic_long_read(&n->total_objects);
5425 nr_free += count_partial(n, count_free);
5428 nr_inuse = nr_objs - nr_free;
5430 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5431 nr_objs, s->size, oo_objects(s->oo),
5432 (1 << oo_order(s->oo)));
5433 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5434 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5440 static const struct seq_operations slabinfo_op = {
5447 static int slabinfo_open(struct inode *inode, struct file *file)
5449 return seq_open(file, &slabinfo_op);
5452 static const struct file_operations proc_slabinfo_operations = {
5453 .open = slabinfo_open,
5455 .llseek = seq_lseek,
5456 .release = seq_release,
5459 static int __init slab_proc_init(void)
5461 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5464 module_init(slab_proc_init);
5465 #endif /* CONFIG_SLABINFO */