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>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
35 #include <trace/events/kmem.h>
41 * 1. slab_mutex (Global Mutex)
43 * 3. slab_lock(page) (Only on some arches and for debugging)
47 * The role of the slab_mutex is to protect the list of all the slabs
48 * and to synchronize major metadata changes to slab cache structures.
50 * The slab_lock is only used for debugging and on arches that do not
51 * have the ability to do a cmpxchg_double. It only protects the second
52 * double word in the page struct. Meaning
53 * A. page->freelist -> List of object free in a page
54 * B. page->counters -> Counters of objects
55 * C. page->frozen -> frozen state
57 * If a slab is frozen then it is exempt from list management. It is not
58 * on any list. The processor that froze the slab is the one who can
59 * perform list operations on the page. Other processors may put objects
60 * onto the freelist but the processor that froze the slab is the only
61 * one that can retrieve the objects from the page's freelist.
63 * The list_lock protects the partial and full list on each node and
64 * the partial slab counter. If taken then no new slabs may be added or
65 * removed from the lists nor make the number of partial slabs be modified.
66 * (Note that the total number of slabs is an atomic value that may be
67 * modified without taking the list lock).
69 * The list_lock is a centralized lock and thus we avoid taking it as
70 * much as possible. As long as SLUB does not have to handle partial
71 * slabs, operations can continue without any centralized lock. F.e.
72 * allocating a long series of objects that fill up slabs does not require
74 * Interrupts are disabled during allocation and deallocation in order to
75 * make the slab allocator safe to use in the context of an irq. In addition
76 * interrupts are disabled to ensure that the processor does not change
77 * while handling per_cpu slabs, due to kernel preemption.
79 * SLUB assigns one slab for allocation to each processor.
80 * Allocations only occur from these slabs called cpu slabs.
82 * Slabs with free elements are kept on a partial list and during regular
83 * operations no list for full slabs is used. If an object in a full slab is
84 * freed then the slab will show up again on the partial lists.
85 * We track full slabs for debugging purposes though because otherwise we
86 * cannot scan all objects.
88 * Slabs are freed when they become empty. Teardown and setup is
89 * minimal so we rely on the page allocators per cpu caches for
90 * fast frees and allocs.
92 * Overloading of page flags that are otherwise used for LRU management.
94 * PageActive The slab is frozen and exempt from list processing.
95 * This means that the slab is dedicated to a purpose
96 * such as satisfying allocations for a specific
97 * processor. Objects may be freed in the slab while
98 * it is frozen but slab_free will then skip the usual
99 * list operations. It is up to the processor holding
100 * the slab to integrate the slab into the slab lists
101 * when the slab is no longer needed.
103 * One use of this flag is to mark slabs that are
104 * used for allocations. Then such a slab becomes a cpu
105 * slab. The cpu slab may be equipped with an additional
106 * freelist that allows lockless access to
107 * free objects in addition to the regular freelist
108 * that requires the slab lock.
110 * PageError Slab requires special handling due to debug
111 * options set. This moves slab handling out of
112 * the fast path and disables lockless freelists.
115 static inline int kmem_cache_debug(struct kmem_cache *s)
117 #ifdef CONFIG_SLUB_DEBUG
118 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
125 * Issues still to be resolved:
127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
129 * - Variable sizing of the per node arrays
132 /* Enable to test recovery from slab corruption on boot */
133 #undef SLUB_RESILIENCY_TEST
135 /* Enable to log cmpxchg failures */
136 #undef SLUB_DEBUG_CMPXCHG
139 * Mininum number of partial slabs. These will be left on the partial
140 * lists even if they are empty. kmem_cache_shrink may reclaim them.
142 #define MIN_PARTIAL 5
145 * Maximum number of desirable partial slabs.
146 * The existence of more partial slabs makes kmem_cache_shrink
147 * sort the partial list by the number of objects in the.
149 #define MAX_PARTIAL 10
151 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
152 SLAB_POISON | SLAB_STORE_USER)
155 * Debugging flags that require metadata to be stored in the slab. These get
156 * disabled when slub_debug=O is used and a cache's min order increases with
159 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
162 * Set of flags that will prevent slab merging
164 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
165 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
168 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
169 SLAB_CACHE_DMA | SLAB_NOTRACK)
172 #define OO_MASK ((1 << OO_SHIFT) - 1)
173 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
175 /* Internal SLUB flags */
176 #define __OBJECT_POISON 0x80000000UL /* Poison object */
177 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
179 static int kmem_size = sizeof(struct kmem_cache);
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void sysfs_slab_remove(struct kmem_cache *);
207 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
208 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
210 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
214 static inline void stat(const struct kmem_cache *s, enum stat_item si)
216 #ifdef CONFIG_SLUB_STATS
217 __this_cpu_inc(s->cpu_slab->stat[si]);
221 /********************************************************************
222 * Core slab cache functions
223 *******************************************************************/
225 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
227 return s->node[node];
230 /* Verify that a pointer has an address that is valid within a slab page */
231 static inline int check_valid_pointer(struct kmem_cache *s,
232 struct page *page, const void *object)
239 base = page_address(page);
240 if (object < base || object >= base + page->objects * s->size ||
241 (object - base) % s->size) {
248 static inline void *get_freepointer(struct kmem_cache *s, void *object)
250 return *(void **)(object + s->offset);
253 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
255 prefetch(object + s->offset);
258 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
262 #ifdef CONFIG_DEBUG_PAGEALLOC
263 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
265 p = get_freepointer(s, object);
270 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
272 *(void **)(object + s->offset) = fp;
275 /* Loop over all objects in a slab */
276 #define for_each_object(__p, __s, __addr, __objects) \
277 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
283 return (p - addr) / s->size;
286 static inline size_t slab_ksize(const struct kmem_cache *s)
288 #ifdef CONFIG_SLUB_DEBUG
290 * Debugging requires use of the padding between object
291 * and whatever may come after it.
293 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
294 return s->object_size;
298 * If we have the need to store the freelist pointer
299 * back there or track user information then we can
300 * only use the space before that information.
302 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
305 * Else we can use all the padding etc for the allocation
310 static inline int order_objects(int order, unsigned long size, int reserved)
312 return ((PAGE_SIZE << order) - reserved) / size;
315 static inline struct kmem_cache_order_objects oo_make(int order,
316 unsigned long size, int reserved)
318 struct kmem_cache_order_objects x = {
319 (order << OO_SHIFT) + order_objects(order, size, reserved)
325 static inline int oo_order(struct kmem_cache_order_objects x)
327 return x.x >> OO_SHIFT;
330 static inline int oo_objects(struct kmem_cache_order_objects x)
332 return x.x & OO_MASK;
336 * Per slab locking using the pagelock
338 static __always_inline void slab_lock(struct page *page)
340 bit_spin_lock(PG_locked, &page->flags);
343 static __always_inline void slab_unlock(struct page *page)
345 __bit_spin_unlock(PG_locked, &page->flags);
348 /* Interrupts must be disabled (for the fallback code to work right) */
349 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
350 void *freelist_old, unsigned long counters_old,
351 void *freelist_new, unsigned long counters_new,
354 VM_BUG_ON(!irqs_disabled());
355 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
356 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
357 if (s->flags & __CMPXCHG_DOUBLE) {
358 if (cmpxchg_double(&page->freelist, &page->counters,
359 freelist_old, counters_old,
360 freelist_new, counters_new))
366 if (page->freelist == freelist_old && page->counters == counters_old) {
367 page->freelist = freelist_new;
368 page->counters = counters_new;
376 stat(s, CMPXCHG_DOUBLE_FAIL);
378 #ifdef SLUB_DEBUG_CMPXCHG
379 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
385 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
386 void *freelist_old, unsigned long counters_old,
387 void *freelist_new, unsigned long counters_new,
390 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
391 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
392 if (s->flags & __CMPXCHG_DOUBLE) {
393 if (cmpxchg_double(&page->freelist, &page->counters,
394 freelist_old, counters_old,
395 freelist_new, counters_new))
402 local_irq_save(flags);
404 if (page->freelist == freelist_old && page->counters == counters_old) {
405 page->freelist = freelist_new;
406 page->counters = counters_new;
408 local_irq_restore(flags);
412 local_irq_restore(flags);
416 stat(s, CMPXCHG_DOUBLE_FAIL);
418 #ifdef SLUB_DEBUG_CMPXCHG
419 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
425 #ifdef CONFIG_SLUB_DEBUG
427 * Determine a map of object in use on a page.
429 * Node listlock must be held to guarantee that the page does
430 * not vanish from under us.
432 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
435 void *addr = page_address(page);
437 for (p = page->freelist; p; p = get_freepointer(s, p))
438 set_bit(slab_index(p, s, addr), map);
444 #ifdef CONFIG_SLUB_DEBUG_ON
445 static int slub_debug = DEBUG_DEFAULT_FLAGS;
447 static int slub_debug;
450 static char *slub_debug_slabs;
451 static int disable_higher_order_debug;
456 static void print_section(char *text, u8 *addr, unsigned int length)
458 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
462 static struct track *get_track(struct kmem_cache *s, void *object,
463 enum track_item alloc)
468 p = object + s->offset + sizeof(void *);
470 p = object + s->inuse;
475 static void set_track(struct kmem_cache *s, void *object,
476 enum track_item alloc, unsigned long addr)
478 struct track *p = get_track(s, object, alloc);
481 #ifdef CONFIG_STACKTRACE
482 struct stack_trace trace;
485 trace.nr_entries = 0;
486 trace.max_entries = TRACK_ADDRS_COUNT;
487 trace.entries = p->addrs;
489 save_stack_trace(&trace);
491 /* See rant in lockdep.c */
492 if (trace.nr_entries != 0 &&
493 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
496 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
500 p->cpu = smp_processor_id();
501 p->pid = current->pid;
504 memset(p, 0, sizeof(struct track));
507 static void init_tracking(struct kmem_cache *s, void *object)
509 if (!(s->flags & SLAB_STORE_USER))
512 set_track(s, object, TRACK_FREE, 0UL);
513 set_track(s, object, TRACK_ALLOC, 0UL);
516 static void print_track(const char *s, struct track *t)
521 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
522 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
523 #ifdef CONFIG_STACKTRACE
526 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
528 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
535 static void print_tracking(struct kmem_cache *s, void *object)
537 if (!(s->flags & SLAB_STORE_USER))
540 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
541 print_track("Freed", get_track(s, object, TRACK_FREE));
544 static void print_page_info(struct page *page)
546 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
547 page, page->objects, page->inuse, page->freelist, page->flags);
551 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
557 vsnprintf(buf, sizeof(buf), fmt, args);
559 printk(KERN_ERR "========================================"
560 "=====================================\n");
561 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
562 printk(KERN_ERR "----------------------------------------"
563 "-------------------------------------\n\n");
565 add_taint(TAINT_BAD_PAGE);
568 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
574 vsnprintf(buf, sizeof(buf), fmt, args);
576 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
579 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
581 unsigned int off; /* Offset of last byte */
582 u8 *addr = page_address(page);
584 print_tracking(s, p);
586 print_page_info(page);
588 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
589 p, p - addr, get_freepointer(s, p));
592 print_section("Bytes b4 ", p - 16, 16);
594 print_section("Object ", p, min_t(unsigned long, s->object_size,
596 if (s->flags & SLAB_RED_ZONE)
597 print_section("Redzone ", p + s->object_size,
598 s->inuse - s->object_size);
601 off = s->offset + sizeof(void *);
605 if (s->flags & SLAB_STORE_USER)
606 off += 2 * sizeof(struct track);
609 /* Beginning of the filler is the free pointer */
610 print_section("Padding ", p + off, s->size - off);
615 static void object_err(struct kmem_cache *s, struct page *page,
616 u8 *object, char *reason)
618 slab_bug(s, "%s", reason);
619 print_trailer(s, page, object);
622 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
628 vsnprintf(buf, sizeof(buf), fmt, args);
630 slab_bug(s, "%s", buf);
631 print_page_info(page);
635 static void init_object(struct kmem_cache *s, void *object, u8 val)
639 if (s->flags & __OBJECT_POISON) {
640 memset(p, POISON_FREE, s->object_size - 1);
641 p[s->object_size - 1] = POISON_END;
644 if (s->flags & SLAB_RED_ZONE)
645 memset(p + s->object_size, val, s->inuse - s->object_size);
648 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
649 void *from, void *to)
651 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
652 memset(from, data, to - from);
655 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
656 u8 *object, char *what,
657 u8 *start, unsigned int value, unsigned int bytes)
662 fault = memchr_inv(start, value, bytes);
667 while (end > fault && end[-1] == value)
670 slab_bug(s, "%s overwritten", what);
671 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
672 fault, end - 1, fault[0], value);
673 print_trailer(s, page, object);
675 restore_bytes(s, what, value, fault, end);
683 * Bytes of the object to be managed.
684 * If the freepointer may overlay the object then the free
685 * pointer is the first word of the object.
687 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
690 * object + s->object_size
691 * Padding to reach word boundary. This is also used for Redzoning.
692 * Padding is extended by another word if Redzoning is enabled and
693 * object_size == inuse.
695 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
696 * 0xcc (RED_ACTIVE) for objects in use.
699 * Meta data starts here.
701 * A. Free pointer (if we cannot overwrite object on free)
702 * B. Tracking data for SLAB_STORE_USER
703 * C. Padding to reach required alignment boundary or at mininum
704 * one word if debugging is on to be able to detect writes
705 * before the word boundary.
707 * Padding is done using 0x5a (POISON_INUSE)
710 * Nothing is used beyond s->size.
712 * If slabcaches are merged then the object_size and inuse boundaries are mostly
713 * ignored. And therefore no slab options that rely on these boundaries
714 * may be used with merged slabcaches.
717 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
719 unsigned long off = s->inuse; /* The end of info */
722 /* Freepointer is placed after the object. */
723 off += sizeof(void *);
725 if (s->flags & SLAB_STORE_USER)
726 /* We also have user information there */
727 off += 2 * sizeof(struct track);
732 return check_bytes_and_report(s, page, p, "Object padding",
733 p + off, POISON_INUSE, s->size - off);
736 /* Check the pad bytes at the end of a slab page */
737 static int slab_pad_check(struct kmem_cache *s, struct page *page)
745 if (!(s->flags & SLAB_POISON))
748 start = page_address(page);
749 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
750 end = start + length;
751 remainder = length % s->size;
755 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
758 while (end > fault && end[-1] == POISON_INUSE)
761 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
762 print_section("Padding ", end - remainder, remainder);
764 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
768 static int check_object(struct kmem_cache *s, struct page *page,
769 void *object, u8 val)
772 u8 *endobject = object + s->object_size;
774 if (s->flags & SLAB_RED_ZONE) {
775 if (!check_bytes_and_report(s, page, object, "Redzone",
776 endobject, val, s->inuse - s->object_size))
779 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
780 check_bytes_and_report(s, page, p, "Alignment padding",
781 endobject, POISON_INUSE, s->inuse - s->object_size);
785 if (s->flags & SLAB_POISON) {
786 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
787 (!check_bytes_and_report(s, page, p, "Poison", p,
788 POISON_FREE, s->object_size - 1) ||
789 !check_bytes_and_report(s, page, p, "Poison",
790 p + s->object_size - 1, POISON_END, 1)))
793 * check_pad_bytes cleans up on its own.
795 check_pad_bytes(s, page, p);
798 if (!s->offset && val == SLUB_RED_ACTIVE)
800 * Object and freepointer overlap. Cannot check
801 * freepointer while object is allocated.
805 /* Check free pointer validity */
806 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
807 object_err(s, page, p, "Freepointer corrupt");
809 * No choice but to zap it and thus lose the remainder
810 * of the free objects in this slab. May cause
811 * another error because the object count is now wrong.
813 set_freepointer(s, p, NULL);
819 static int check_slab(struct kmem_cache *s, struct page *page)
823 VM_BUG_ON(!irqs_disabled());
825 if (!PageSlab(page)) {
826 slab_err(s, page, "Not a valid slab page");
830 maxobj = order_objects(compound_order(page), s->size, s->reserved);
831 if (page->objects > maxobj) {
832 slab_err(s, page, "objects %u > max %u",
833 s->name, page->objects, maxobj);
836 if (page->inuse > page->objects) {
837 slab_err(s, page, "inuse %u > max %u",
838 s->name, page->inuse, page->objects);
841 /* Slab_pad_check fixes things up after itself */
842 slab_pad_check(s, page);
847 * Determine if a certain object on a page is on the freelist. Must hold the
848 * slab lock to guarantee that the chains are in a consistent state.
850 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
855 unsigned long max_objects;
858 while (fp && nr <= page->objects) {
861 if (!check_valid_pointer(s, page, fp)) {
863 object_err(s, page, object,
864 "Freechain corrupt");
865 set_freepointer(s, object, NULL);
868 slab_err(s, page, "Freepointer corrupt");
869 page->freelist = NULL;
870 page->inuse = page->objects;
871 slab_fix(s, "Freelist cleared");
877 fp = get_freepointer(s, object);
881 max_objects = order_objects(compound_order(page), s->size, s->reserved);
882 if (max_objects > MAX_OBJS_PER_PAGE)
883 max_objects = MAX_OBJS_PER_PAGE;
885 if (page->objects != max_objects) {
886 slab_err(s, page, "Wrong number of objects. Found %d but "
887 "should be %d", page->objects, max_objects);
888 page->objects = max_objects;
889 slab_fix(s, "Number of objects adjusted.");
891 if (page->inuse != page->objects - nr) {
892 slab_err(s, page, "Wrong object count. Counter is %d but "
893 "counted were %d", page->inuse, page->objects - nr);
894 page->inuse = page->objects - nr;
895 slab_fix(s, "Object count adjusted.");
897 return search == NULL;
900 static void trace(struct kmem_cache *s, struct page *page, void *object,
903 if (s->flags & SLAB_TRACE) {
904 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
906 alloc ? "alloc" : "free",
911 print_section("Object ", (void *)object, s->object_size);
918 * Hooks for other subsystems that check memory allocations. In a typical
919 * production configuration these hooks all should produce no code at all.
921 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
923 flags &= gfp_allowed_mask;
924 lockdep_trace_alloc(flags);
925 might_sleep_if(flags & __GFP_WAIT);
927 return should_failslab(s->object_size, flags, s->flags);
930 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
932 flags &= gfp_allowed_mask;
933 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
934 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
937 static inline void slab_free_hook(struct kmem_cache *s, void *x)
939 kmemleak_free_recursive(x, s->flags);
942 * Trouble is that we may no longer disable interupts in the fast path
943 * So in order to make the debug calls that expect irqs to be
944 * disabled we need to disable interrupts temporarily.
946 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
950 local_irq_save(flags);
951 kmemcheck_slab_free(s, x, s->object_size);
952 debug_check_no_locks_freed(x, s->object_size);
953 local_irq_restore(flags);
956 if (!(s->flags & SLAB_DEBUG_OBJECTS))
957 debug_check_no_obj_freed(x, s->object_size);
961 * Tracking of fully allocated slabs for debugging purposes.
963 * list_lock must be held.
965 static void add_full(struct kmem_cache *s,
966 struct kmem_cache_node *n, struct page *page)
968 if (!(s->flags & SLAB_STORE_USER))
971 list_add(&page->lru, &n->full);
975 * list_lock must be held.
977 static void remove_full(struct kmem_cache *s, struct page *page)
979 if (!(s->flags & SLAB_STORE_USER))
982 list_del(&page->lru);
985 /* Tracking of the number of slabs for debugging purposes */
986 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
988 struct kmem_cache_node *n = get_node(s, node);
990 return atomic_long_read(&n->nr_slabs);
993 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
995 return atomic_long_read(&n->nr_slabs);
998 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1000 struct kmem_cache_node *n = get_node(s, node);
1003 * May be called early in order to allocate a slab for the
1004 * kmem_cache_node structure. Solve the chicken-egg
1005 * dilemma by deferring the increment of the count during
1006 * bootstrap (see early_kmem_cache_node_alloc).
1009 atomic_long_inc(&n->nr_slabs);
1010 atomic_long_add(objects, &n->total_objects);
1013 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1015 struct kmem_cache_node *n = get_node(s, node);
1017 atomic_long_dec(&n->nr_slabs);
1018 atomic_long_sub(objects, &n->total_objects);
1021 /* Object debug checks for alloc/free paths */
1022 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1025 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1028 init_object(s, object, SLUB_RED_INACTIVE);
1029 init_tracking(s, object);
1032 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1033 void *object, unsigned long addr)
1035 if (!check_slab(s, page))
1038 if (!check_valid_pointer(s, page, object)) {
1039 object_err(s, page, object, "Freelist Pointer check fails");
1043 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1046 /* Success perform special debug activities for allocs */
1047 if (s->flags & SLAB_STORE_USER)
1048 set_track(s, object, TRACK_ALLOC, addr);
1049 trace(s, page, object, 1);
1050 init_object(s, object, SLUB_RED_ACTIVE);
1054 if (PageSlab(page)) {
1056 * If this is a slab page then lets do the best we can
1057 * to avoid issues in the future. Marking all objects
1058 * as used avoids touching the remaining objects.
1060 slab_fix(s, "Marking all objects used");
1061 page->inuse = page->objects;
1062 page->freelist = NULL;
1067 static noinline struct kmem_cache_node *free_debug_processing(
1068 struct kmem_cache *s, struct page *page, void *object,
1069 unsigned long addr, unsigned long *flags)
1071 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1073 spin_lock_irqsave(&n->list_lock, *flags);
1076 if (!check_slab(s, page))
1079 if (!check_valid_pointer(s, page, object)) {
1080 slab_err(s, page, "Invalid object pointer 0x%p", object);
1084 if (on_freelist(s, page, object)) {
1085 object_err(s, page, object, "Object already free");
1089 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1092 if (unlikely(s != page->slab_cache)) {
1093 if (!PageSlab(page)) {
1094 slab_err(s, page, "Attempt to free object(0x%p) "
1095 "outside of slab", object);
1096 } else if (!page->slab_cache) {
1098 "SLUB <none>: no slab for object 0x%p.\n",
1102 object_err(s, page, object,
1103 "page slab pointer corrupt.");
1107 if (s->flags & SLAB_STORE_USER)
1108 set_track(s, object, TRACK_FREE, addr);
1109 trace(s, page, object, 0);
1110 init_object(s, object, SLUB_RED_INACTIVE);
1114 * Keep node_lock to preserve integrity
1115 * until the object is actually freed
1121 spin_unlock_irqrestore(&n->list_lock, *flags);
1122 slab_fix(s, "Object at 0x%p not freed", object);
1126 static int __init setup_slub_debug(char *str)
1128 slub_debug = DEBUG_DEFAULT_FLAGS;
1129 if (*str++ != '=' || !*str)
1131 * No options specified. Switch on full debugging.
1137 * No options but restriction on slabs. This means full
1138 * debugging for slabs matching a pattern.
1142 if (tolower(*str) == 'o') {
1144 * Avoid enabling debugging on caches if its minimum order
1145 * would increase as a result.
1147 disable_higher_order_debug = 1;
1154 * Switch off all debugging measures.
1159 * Determine which debug features should be switched on
1161 for (; *str && *str != ','; str++) {
1162 switch (tolower(*str)) {
1164 slub_debug |= SLAB_DEBUG_FREE;
1167 slub_debug |= SLAB_RED_ZONE;
1170 slub_debug |= SLAB_POISON;
1173 slub_debug |= SLAB_STORE_USER;
1176 slub_debug |= SLAB_TRACE;
1179 slub_debug |= SLAB_FAILSLAB;
1182 printk(KERN_ERR "slub_debug option '%c' "
1183 "unknown. skipped\n", *str);
1189 slub_debug_slabs = str + 1;
1194 __setup("slub_debug", setup_slub_debug);
1196 static unsigned long kmem_cache_flags(unsigned long object_size,
1197 unsigned long flags, const char *name,
1198 void (*ctor)(void *))
1201 * Enable debugging if selected on the kernel commandline.
1203 if (slub_debug && (!slub_debug_slabs ||
1204 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1205 flags |= slub_debug;
1210 static inline void setup_object_debug(struct kmem_cache *s,
1211 struct page *page, void *object) {}
1213 static inline int alloc_debug_processing(struct kmem_cache *s,
1214 struct page *page, void *object, unsigned long addr) { return 0; }
1216 static inline struct kmem_cache_node *free_debug_processing(
1217 struct kmem_cache *s, struct page *page, void *object,
1218 unsigned long addr, unsigned long *flags) { return NULL; }
1220 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1222 static inline int check_object(struct kmem_cache *s, struct page *page,
1223 void *object, u8 val) { return 1; }
1224 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1225 struct page *page) {}
1226 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1227 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1228 unsigned long flags, const char *name,
1229 void (*ctor)(void *))
1233 #define slub_debug 0
1235 #define disable_higher_order_debug 0
1237 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1239 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1241 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1243 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1246 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1249 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1252 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1254 #endif /* CONFIG_SLUB_DEBUG */
1257 * Slab allocation and freeing
1259 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1260 struct kmem_cache_order_objects oo)
1262 int order = oo_order(oo);
1264 flags |= __GFP_NOTRACK;
1266 if (node == NUMA_NO_NODE)
1267 return alloc_pages(flags, order);
1269 return alloc_pages_exact_node(node, flags, order);
1272 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1275 struct kmem_cache_order_objects oo = s->oo;
1278 flags &= gfp_allowed_mask;
1280 if (flags & __GFP_WAIT)
1283 flags |= s->allocflags;
1286 * Let the initial higher-order allocation fail under memory pressure
1287 * so we fall-back to the minimum order allocation.
1289 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1291 page = alloc_slab_page(alloc_gfp, node, oo);
1292 if (unlikely(!page)) {
1295 * Allocation may have failed due to fragmentation.
1296 * Try a lower order alloc if possible
1298 page = alloc_slab_page(flags, node, oo);
1301 stat(s, ORDER_FALLBACK);
1304 if (kmemcheck_enabled && page
1305 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1306 int pages = 1 << oo_order(oo);
1308 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1311 * Objects from caches that have a constructor don't get
1312 * cleared when they're allocated, so we need to do it here.
1315 kmemcheck_mark_uninitialized_pages(page, pages);
1317 kmemcheck_mark_unallocated_pages(page, pages);
1320 if (flags & __GFP_WAIT)
1321 local_irq_disable();
1325 page->objects = oo_objects(oo);
1326 mod_zone_page_state(page_zone(page),
1327 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1328 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1334 static void setup_object(struct kmem_cache *s, struct page *page,
1337 setup_object_debug(s, page, object);
1338 if (unlikely(s->ctor))
1342 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1349 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1351 page = allocate_slab(s,
1352 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1356 inc_slabs_node(s, page_to_nid(page), page->objects);
1357 page->slab_cache = s;
1358 __SetPageSlab(page);
1359 if (page->pfmemalloc)
1360 SetPageSlabPfmemalloc(page);
1362 start = page_address(page);
1364 if (unlikely(s->flags & SLAB_POISON))
1365 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1368 for_each_object(p, s, start, page->objects) {
1369 setup_object(s, page, last);
1370 set_freepointer(s, last, p);
1373 setup_object(s, page, last);
1374 set_freepointer(s, last, NULL);
1376 page->freelist = start;
1377 page->inuse = page->objects;
1383 static void __free_slab(struct kmem_cache *s, struct page *page)
1385 int order = compound_order(page);
1386 int pages = 1 << order;
1388 if (kmem_cache_debug(s)) {
1391 slab_pad_check(s, page);
1392 for_each_object(p, s, page_address(page),
1394 check_object(s, page, p, SLUB_RED_INACTIVE);
1397 kmemcheck_free_shadow(page, compound_order(page));
1399 mod_zone_page_state(page_zone(page),
1400 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1401 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1404 __ClearPageSlabPfmemalloc(page);
1405 __ClearPageSlab(page);
1406 reset_page_mapcount(page);
1407 if (current->reclaim_state)
1408 current->reclaim_state->reclaimed_slab += pages;
1409 __free_pages(page, order);
1412 #define need_reserve_slab_rcu \
1413 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1415 static void rcu_free_slab(struct rcu_head *h)
1419 if (need_reserve_slab_rcu)
1420 page = virt_to_head_page(h);
1422 page = container_of((struct list_head *)h, struct page, lru);
1424 __free_slab(page->slab_cache, page);
1427 static void free_slab(struct kmem_cache *s, struct page *page)
1429 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1430 struct rcu_head *head;
1432 if (need_reserve_slab_rcu) {
1433 int order = compound_order(page);
1434 int offset = (PAGE_SIZE << order) - s->reserved;
1436 VM_BUG_ON(s->reserved != sizeof(*head));
1437 head = page_address(page) + offset;
1440 * RCU free overloads the RCU head over the LRU
1442 head = (void *)&page->lru;
1445 call_rcu(head, rcu_free_slab);
1447 __free_slab(s, page);
1450 static void discard_slab(struct kmem_cache *s, struct page *page)
1452 dec_slabs_node(s, page_to_nid(page), page->objects);
1457 * Management of partially allocated slabs.
1459 * list_lock must be held.
1461 static inline void add_partial(struct kmem_cache_node *n,
1462 struct page *page, int tail)
1465 if (tail == DEACTIVATE_TO_TAIL)
1466 list_add_tail(&page->lru, &n->partial);
1468 list_add(&page->lru, &n->partial);
1472 * list_lock must be held.
1474 static inline void remove_partial(struct kmem_cache_node *n,
1477 list_del(&page->lru);
1482 * Remove slab from the partial list, freeze it and
1483 * return the pointer to the freelist.
1485 * Returns a list of objects or NULL if it fails.
1487 * Must hold list_lock since we modify the partial list.
1489 static inline void *acquire_slab(struct kmem_cache *s,
1490 struct kmem_cache_node *n, struct page *page,
1494 unsigned long counters;
1498 * Zap the freelist and set the frozen bit.
1499 * The old freelist is the list of objects for the
1500 * per cpu allocation list.
1502 freelist = page->freelist;
1503 counters = page->counters;
1504 new.counters = counters;
1506 new.inuse = page->objects;
1507 new.freelist = NULL;
1509 new.freelist = freelist;
1512 VM_BUG_ON(new.frozen);
1515 if (!__cmpxchg_double_slab(s, page,
1517 new.freelist, new.counters,
1521 remove_partial(n, page);
1526 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1527 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1530 * Try to allocate a partial slab from a specific node.
1532 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1533 struct kmem_cache_cpu *c, gfp_t flags)
1535 struct page *page, *page2;
1536 void *object = NULL;
1539 * Racy check. If we mistakenly see no partial slabs then we
1540 * just allocate an empty slab. If we mistakenly try to get a
1541 * partial slab and there is none available then get_partials()
1544 if (!n || !n->nr_partial)
1547 spin_lock(&n->list_lock);
1548 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1552 if (!pfmemalloc_match(page, flags))
1555 t = acquire_slab(s, n, page, object == NULL);
1561 stat(s, ALLOC_FROM_PARTIAL);
1563 available = page->objects - page->inuse;
1565 available = put_cpu_partial(s, page, 0);
1566 stat(s, CPU_PARTIAL_NODE);
1568 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1572 spin_unlock(&n->list_lock);
1577 * Get a page from somewhere. Search in increasing NUMA distances.
1579 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1580 struct kmem_cache_cpu *c)
1583 struct zonelist *zonelist;
1586 enum zone_type high_zoneidx = gfp_zone(flags);
1588 unsigned int cpuset_mems_cookie;
1591 * The defrag ratio allows a configuration of the tradeoffs between
1592 * inter node defragmentation and node local allocations. A lower
1593 * defrag_ratio increases the tendency to do local allocations
1594 * instead of attempting to obtain partial slabs from other nodes.
1596 * If the defrag_ratio is set to 0 then kmalloc() always
1597 * returns node local objects. If the ratio is higher then kmalloc()
1598 * may return off node objects because partial slabs are obtained
1599 * from other nodes and filled up.
1601 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1602 * defrag_ratio = 1000) then every (well almost) allocation will
1603 * first attempt to defrag slab caches on other nodes. This means
1604 * scanning over all nodes to look for partial slabs which may be
1605 * expensive if we do it every time we are trying to find a slab
1606 * with available objects.
1608 if (!s->remote_node_defrag_ratio ||
1609 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1613 cpuset_mems_cookie = get_mems_allowed();
1614 zonelist = node_zonelist(slab_node(), flags);
1615 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1616 struct kmem_cache_node *n;
1618 n = get_node(s, zone_to_nid(zone));
1620 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1621 n->nr_partial > s->min_partial) {
1622 object = get_partial_node(s, n, c, flags);
1625 * Return the object even if
1626 * put_mems_allowed indicated that
1627 * the cpuset mems_allowed was
1628 * updated in parallel. It's a
1629 * harmless race between the alloc
1630 * and the cpuset update.
1632 put_mems_allowed(cpuset_mems_cookie);
1637 } while (!put_mems_allowed(cpuset_mems_cookie));
1643 * Get a partial page, lock it and return it.
1645 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1646 struct kmem_cache_cpu *c)
1649 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1651 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1652 if (object || node != NUMA_NO_NODE)
1655 return get_any_partial(s, flags, c);
1658 #ifdef CONFIG_PREEMPT
1660 * Calculate the next globally unique transaction for disambiguiation
1661 * during cmpxchg. The transactions start with the cpu number and are then
1662 * incremented by CONFIG_NR_CPUS.
1664 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1667 * No preemption supported therefore also no need to check for
1673 static inline unsigned long next_tid(unsigned long tid)
1675 return tid + TID_STEP;
1678 static inline unsigned int tid_to_cpu(unsigned long tid)
1680 return tid % TID_STEP;
1683 static inline unsigned long tid_to_event(unsigned long tid)
1685 return tid / TID_STEP;
1688 static inline unsigned int init_tid(int cpu)
1693 static inline void note_cmpxchg_failure(const char *n,
1694 const struct kmem_cache *s, unsigned long tid)
1696 #ifdef SLUB_DEBUG_CMPXCHG
1697 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1699 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1701 #ifdef CONFIG_PREEMPT
1702 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1703 printk("due to cpu change %d -> %d\n",
1704 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1707 if (tid_to_event(tid) != tid_to_event(actual_tid))
1708 printk("due to cpu running other code. Event %ld->%ld\n",
1709 tid_to_event(tid), tid_to_event(actual_tid));
1711 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1712 actual_tid, tid, next_tid(tid));
1714 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1717 static void init_kmem_cache_cpus(struct kmem_cache *s)
1721 for_each_possible_cpu(cpu)
1722 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1726 * Remove the cpu slab
1728 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1730 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1731 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1733 enum slab_modes l = M_NONE, m = M_NONE;
1735 int tail = DEACTIVATE_TO_HEAD;
1739 if (page->freelist) {
1740 stat(s, DEACTIVATE_REMOTE_FREES);
1741 tail = DEACTIVATE_TO_TAIL;
1745 * Stage one: Free all available per cpu objects back
1746 * to the page freelist while it is still frozen. Leave the
1749 * There is no need to take the list->lock because the page
1752 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1754 unsigned long counters;
1757 prior = page->freelist;
1758 counters = page->counters;
1759 set_freepointer(s, freelist, prior);
1760 new.counters = counters;
1762 VM_BUG_ON(!new.frozen);
1764 } while (!__cmpxchg_double_slab(s, page,
1766 freelist, new.counters,
1767 "drain percpu freelist"));
1769 freelist = nextfree;
1773 * Stage two: Ensure that the page is unfrozen while the
1774 * list presence reflects the actual number of objects
1777 * We setup the list membership and then perform a cmpxchg
1778 * with the count. If there is a mismatch then the page
1779 * is not unfrozen but the page is on the wrong list.
1781 * Then we restart the process which may have to remove
1782 * the page from the list that we just put it on again
1783 * because the number of objects in the slab may have
1788 old.freelist = page->freelist;
1789 old.counters = page->counters;
1790 VM_BUG_ON(!old.frozen);
1792 /* Determine target state of the slab */
1793 new.counters = old.counters;
1796 set_freepointer(s, freelist, old.freelist);
1797 new.freelist = freelist;
1799 new.freelist = old.freelist;
1803 if (!new.inuse && n->nr_partial > s->min_partial)
1805 else if (new.freelist) {
1810 * Taking the spinlock removes the possiblity
1811 * that acquire_slab() will see a slab page that
1814 spin_lock(&n->list_lock);
1818 if (kmem_cache_debug(s) && !lock) {
1821 * This also ensures that the scanning of full
1822 * slabs from diagnostic functions will not see
1825 spin_lock(&n->list_lock);
1833 remove_partial(n, page);
1835 else if (l == M_FULL)
1837 remove_full(s, page);
1839 if (m == M_PARTIAL) {
1841 add_partial(n, page, tail);
1844 } else if (m == M_FULL) {
1846 stat(s, DEACTIVATE_FULL);
1847 add_full(s, n, page);
1853 if (!__cmpxchg_double_slab(s, page,
1854 old.freelist, old.counters,
1855 new.freelist, new.counters,
1860 spin_unlock(&n->list_lock);
1863 stat(s, DEACTIVATE_EMPTY);
1864 discard_slab(s, page);
1870 * Unfreeze all the cpu partial slabs.
1872 * This function must be called with interrupt disabled.
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 static 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);
1968 stat(s, CPU_PARTIAL_DRAIN);
1973 pobjects += page->objects - page->inuse;
1975 page->pages = pages;
1976 page->pobjects = pobjects;
1977 page->next = oldpage;
1979 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1983 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1985 stat(s, CPUSLAB_FLUSH);
1986 deactivate_slab(s, c->page, c->freelist);
1988 c->tid = next_tid(c->tid);
1996 * Called from IPI handler with interrupts disabled.
1998 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2000 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2006 unfreeze_partials(s);
2010 static void flush_cpu_slab(void *d)
2012 struct kmem_cache *s = d;
2014 __flush_cpu_slab(s, smp_processor_id());
2017 static bool has_cpu_slab(int cpu, void *info)
2019 struct kmem_cache *s = info;
2020 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2022 return c->page || c->partial;
2025 static void flush_all(struct kmem_cache *s)
2027 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2031 * Check if the objects in a per cpu structure fit numa
2032 * locality expectations.
2034 static inline int node_match(struct page *page, int node)
2037 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2043 static int count_free(struct page *page)
2045 return page->objects - page->inuse;
2048 static unsigned long count_partial(struct kmem_cache_node *n,
2049 int (*get_count)(struct page *))
2051 unsigned long flags;
2052 unsigned long x = 0;
2055 spin_lock_irqsave(&n->list_lock, flags);
2056 list_for_each_entry(page, &n->partial, lru)
2057 x += get_count(page);
2058 spin_unlock_irqrestore(&n->list_lock, flags);
2062 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2064 #ifdef CONFIG_SLUB_DEBUG
2065 return atomic_long_read(&n->total_objects);
2071 static noinline void
2072 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2077 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2079 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2080 "default order: %d, min order: %d\n", s->name, s->object_size,
2081 s->size, oo_order(s->oo), oo_order(s->min));
2083 if (oo_order(s->min) > get_order(s->object_size))
2084 printk(KERN_WARNING " %s debugging increased min order, use "
2085 "slub_debug=O to disable.\n", s->name);
2087 for_each_online_node(node) {
2088 struct kmem_cache_node *n = get_node(s, node);
2089 unsigned long nr_slabs;
2090 unsigned long nr_objs;
2091 unsigned long nr_free;
2096 nr_free = count_partial(n, count_free);
2097 nr_slabs = node_nr_slabs(n);
2098 nr_objs = node_nr_objs(n);
2101 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2102 node, nr_slabs, nr_objs, nr_free);
2106 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2107 int node, struct kmem_cache_cpu **pc)
2110 struct kmem_cache_cpu *c = *pc;
2113 freelist = get_partial(s, flags, node, c);
2118 page = new_slab(s, flags, node);
2120 c = __this_cpu_ptr(s->cpu_slab);
2125 * No other reference to the page yet so we can
2126 * muck around with it freely without cmpxchg
2128 freelist = page->freelist;
2129 page->freelist = NULL;
2131 stat(s, ALLOC_SLAB);
2140 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2142 if (unlikely(PageSlabPfmemalloc(page)))
2143 return gfp_pfmemalloc_allowed(gfpflags);
2149 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2150 * or deactivate the page.
2152 * The page is still frozen if the return value is not NULL.
2154 * If this function returns NULL then the page has been unfrozen.
2156 * This function must be called with interrupt disabled.
2158 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2161 unsigned long counters;
2165 freelist = page->freelist;
2166 counters = page->counters;
2168 new.counters = counters;
2169 VM_BUG_ON(!new.frozen);
2171 new.inuse = page->objects;
2172 new.frozen = freelist != NULL;
2174 } while (!__cmpxchg_double_slab(s, page,
2183 * Slow path. The lockless freelist is empty or we need to perform
2186 * Processing is still very fast if new objects have been freed to the
2187 * regular freelist. In that case we simply take over the regular freelist
2188 * as the lockless freelist and zap the regular freelist.
2190 * If that is not working then we fall back to the partial lists. We take the
2191 * first element of the freelist as the object to allocate now and move the
2192 * rest of the freelist to the lockless freelist.
2194 * And if we were unable to get a new slab from the partial slab lists then
2195 * we need to allocate a new slab. This is the slowest path since it involves
2196 * a call to the page allocator and the setup of a new slab.
2198 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2199 unsigned long addr, struct kmem_cache_cpu *c)
2203 unsigned long flags;
2205 local_irq_save(flags);
2206 #ifdef CONFIG_PREEMPT
2208 * We may have been preempted and rescheduled on a different
2209 * cpu before disabling interrupts. Need to reload cpu area
2212 c = this_cpu_ptr(s->cpu_slab);
2220 if (unlikely(!node_match(page, node))) {
2221 stat(s, ALLOC_NODE_MISMATCH);
2222 deactivate_slab(s, page, c->freelist);
2229 * By rights, we should be searching for a slab page that was
2230 * PFMEMALLOC but right now, we are losing the pfmemalloc
2231 * information when the page leaves the per-cpu allocator
2233 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2234 deactivate_slab(s, page, c->freelist);
2240 /* must check again c->freelist in case of cpu migration or IRQ */
2241 freelist = c->freelist;
2245 stat(s, ALLOC_SLOWPATH);
2247 freelist = get_freelist(s, page);
2251 stat(s, DEACTIVATE_BYPASS);
2255 stat(s, ALLOC_REFILL);
2259 * freelist is pointing to the list of objects to be used.
2260 * page is pointing to the page from which the objects are obtained.
2261 * That page must be frozen for per cpu allocations to work.
2263 VM_BUG_ON(!c->page->frozen);
2264 c->freelist = get_freepointer(s, freelist);
2265 c->tid = next_tid(c->tid);
2266 local_irq_restore(flags);
2272 page = c->page = c->partial;
2273 c->partial = page->next;
2274 stat(s, CPU_PARTIAL_ALLOC);
2279 freelist = new_slab_objects(s, gfpflags, node, &c);
2281 if (unlikely(!freelist)) {
2282 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2283 slab_out_of_memory(s, gfpflags, node);
2285 local_irq_restore(flags);
2290 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2293 /* Only entered in the debug case */
2294 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2295 goto new_slab; /* Slab failed checks. Next slab needed */
2297 deactivate_slab(s, page, get_freepointer(s, freelist));
2300 local_irq_restore(flags);
2305 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2306 * have the fastpath folded into their functions. So no function call
2307 * overhead for requests that can be satisfied on the fastpath.
2309 * The fastpath works by first checking if the lockless freelist can be used.
2310 * If not then __slab_alloc is called for slow processing.
2312 * Otherwise we can simply pick the next object from the lockless free list.
2314 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2315 gfp_t gfpflags, int node, unsigned long addr)
2318 struct kmem_cache_cpu *c;
2322 if (slab_pre_alloc_hook(s, gfpflags))
2328 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2329 * enabled. We may switch back and forth between cpus while
2330 * reading from one cpu area. That does not matter as long
2331 * as we end up on the original cpu again when doing the cmpxchg.
2333 c = __this_cpu_ptr(s->cpu_slab);
2336 * The transaction ids are globally unique per cpu and per operation on
2337 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2338 * occurs on the right processor and that there was no operation on the
2339 * linked list in between.
2344 object = c->freelist;
2346 if (unlikely(!object || !node_match(page, node)))
2347 object = __slab_alloc(s, gfpflags, node, addr, c);
2350 void *next_object = get_freepointer_safe(s, object);
2353 * The cmpxchg will only match if there was no additional
2354 * operation and if we are on the right processor.
2356 * The cmpxchg does the following atomically (without lock semantics!)
2357 * 1. Relocate first pointer to the current per cpu area.
2358 * 2. Verify that tid and freelist have not been changed
2359 * 3. If they were not changed replace tid and freelist
2361 * Since this is without lock semantics the protection is only against
2362 * code executing on this cpu *not* from access by other cpus.
2364 if (unlikely(!this_cpu_cmpxchg_double(
2365 s->cpu_slab->freelist, s->cpu_slab->tid,
2367 next_object, next_tid(tid)))) {
2369 note_cmpxchg_failure("slab_alloc", s, tid);
2372 prefetch_freepointer(s, next_object);
2373 stat(s, ALLOC_FASTPATH);
2376 if (unlikely(gfpflags & __GFP_ZERO) && object)
2377 memset(object, 0, s->object_size);
2379 slab_post_alloc_hook(s, gfpflags, object);
2384 static __always_inline void *slab_alloc(struct kmem_cache *s,
2385 gfp_t gfpflags, unsigned long addr)
2387 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2390 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2392 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2394 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2398 EXPORT_SYMBOL(kmem_cache_alloc);
2400 #ifdef CONFIG_TRACING
2401 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2403 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2404 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2407 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2409 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2411 void *ret = kmalloc_order(size, flags, order);
2412 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2415 EXPORT_SYMBOL(kmalloc_order_trace);
2419 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2421 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2423 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2424 s->object_size, s->size, gfpflags, node);
2428 EXPORT_SYMBOL(kmem_cache_alloc_node);
2430 #ifdef CONFIG_TRACING
2431 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2433 int node, size_t size)
2435 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2437 trace_kmalloc_node(_RET_IP_, ret,
2438 size, s->size, gfpflags, node);
2441 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2446 * Slow patch handling. This may still be called frequently since objects
2447 * have a longer lifetime than the cpu slabs in most processing loads.
2449 * So we still attempt to reduce cache line usage. Just take the slab
2450 * lock and free the item. If there is no additional partial page
2451 * handling required then we can return immediately.
2453 static void __slab_free(struct kmem_cache *s, struct page *page,
2454 void *x, unsigned long addr)
2457 void **object = (void *)x;
2460 unsigned long counters;
2461 struct kmem_cache_node *n = NULL;
2462 unsigned long uninitialized_var(flags);
2464 stat(s, FREE_SLOWPATH);
2466 if (kmem_cache_debug(s) &&
2467 !(n = free_debug_processing(s, page, x, addr, &flags)))
2472 spin_unlock_irqrestore(&n->list_lock, flags);
2475 prior = page->freelist;
2476 counters = page->counters;
2477 set_freepointer(s, object, prior);
2478 new.counters = counters;
2479 was_frozen = new.frozen;
2481 if ((!new.inuse || !prior) && !was_frozen) {
2483 if (!kmem_cache_debug(s) && !prior)
2486 * Slab was on no list before and will be partially empty
2487 * We can defer the list move and instead freeze it.
2491 else { /* Needs to be taken off a list */
2493 n = get_node(s, page_to_nid(page));
2495 * Speculatively acquire the list_lock.
2496 * If the cmpxchg does not succeed then we may
2497 * drop the list_lock without any processing.
2499 * Otherwise the list_lock will synchronize with
2500 * other processors updating the list of slabs.
2502 spin_lock_irqsave(&n->list_lock, flags);
2507 } while (!cmpxchg_double_slab(s, page,
2509 object, new.counters,
2515 * If we just froze the page then put it onto the
2516 * per cpu partial list.
2518 if (new.frozen && !was_frozen) {
2519 put_cpu_partial(s, page, 1);
2520 stat(s, CPU_PARTIAL_FREE);
2523 * The list lock was not taken therefore no list
2524 * activity can be necessary.
2527 stat(s, FREE_FROZEN);
2531 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2535 * Objects left in the slab. If it was not on the partial list before
2538 if (kmem_cache_debug(s) && unlikely(!prior)) {
2539 remove_full(s, page);
2540 add_partial(n, page, DEACTIVATE_TO_TAIL);
2541 stat(s, FREE_ADD_PARTIAL);
2543 spin_unlock_irqrestore(&n->list_lock, flags);
2549 * Slab on the partial list.
2551 remove_partial(n, page);
2552 stat(s, FREE_REMOVE_PARTIAL);
2554 /* Slab must be on the full list */
2555 remove_full(s, page);
2557 spin_unlock_irqrestore(&n->list_lock, flags);
2559 discard_slab(s, page);
2563 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2564 * can perform fastpath freeing without additional function calls.
2566 * The fastpath is only possible if we are freeing to the current cpu slab
2567 * of this processor. This typically the case if we have just allocated
2570 * If fastpath is not possible then fall back to __slab_free where we deal
2571 * with all sorts of special processing.
2573 static __always_inline void slab_free(struct kmem_cache *s,
2574 struct page *page, void *x, unsigned long addr)
2576 void **object = (void *)x;
2577 struct kmem_cache_cpu *c;
2580 slab_free_hook(s, x);
2584 * Determine the currently cpus per cpu slab.
2585 * The cpu may change afterward. However that does not matter since
2586 * data is retrieved via this pointer. If we are on the same cpu
2587 * during the cmpxchg then the free will succedd.
2589 c = __this_cpu_ptr(s->cpu_slab);
2594 if (likely(page == c->page)) {
2595 set_freepointer(s, object, c->freelist);
2597 if (unlikely(!this_cpu_cmpxchg_double(
2598 s->cpu_slab->freelist, s->cpu_slab->tid,
2600 object, next_tid(tid)))) {
2602 note_cmpxchg_failure("slab_free", s, tid);
2605 stat(s, FREE_FASTPATH);
2607 __slab_free(s, page, x, addr);
2611 void kmem_cache_free(struct kmem_cache *s, void *x)
2615 page = virt_to_head_page(x);
2617 if (kmem_cache_debug(s) && page->slab_cache != s) {
2618 pr_err("kmem_cache_free: Wrong slab cache. %s but object"
2619 " is from %s\n", page->slab_cache->name, s->name);
2624 slab_free(s, page, x, _RET_IP_);
2626 trace_kmem_cache_free(_RET_IP_, x);
2628 EXPORT_SYMBOL(kmem_cache_free);
2631 * Object placement in a slab is made very easy because we always start at
2632 * offset 0. If we tune the size of the object to the alignment then we can
2633 * get the required alignment by putting one properly sized object after
2636 * Notice that the allocation order determines the sizes of the per cpu
2637 * caches. Each processor has always one slab available for allocations.
2638 * Increasing the allocation order reduces the number of times that slabs
2639 * must be moved on and off the partial lists and is therefore a factor in
2644 * Mininum / Maximum order of slab pages. This influences locking overhead
2645 * and slab fragmentation. A higher order reduces the number of partial slabs
2646 * and increases the number of allocations possible without having to
2647 * take the list_lock.
2649 static int slub_min_order;
2650 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2651 static int slub_min_objects;
2654 * Merge control. If this is set then no merging of slab caches will occur.
2655 * (Could be removed. This was introduced to pacify the merge skeptics.)
2657 static int slub_nomerge;
2660 * Calculate the order of allocation given an slab object size.
2662 * The order of allocation has significant impact on performance and other
2663 * system components. Generally order 0 allocations should be preferred since
2664 * order 0 does not cause fragmentation in the page allocator. Larger objects
2665 * be problematic to put into order 0 slabs because there may be too much
2666 * unused space left. We go to a higher order if more than 1/16th of the slab
2669 * In order to reach satisfactory performance we must ensure that a minimum
2670 * number of objects is in one slab. Otherwise we may generate too much
2671 * activity on the partial lists which requires taking the list_lock. This is
2672 * less a concern for large slabs though which are rarely used.
2674 * slub_max_order specifies the order where we begin to stop considering the
2675 * number of objects in a slab as critical. If we reach slub_max_order then
2676 * we try to keep the page order as low as possible. So we accept more waste
2677 * of space in favor of a small page order.
2679 * Higher order allocations also allow the placement of more objects in a
2680 * slab and thereby reduce object handling overhead. If the user has
2681 * requested a higher mininum order then we start with that one instead of
2682 * the smallest order which will fit the object.
2684 static inline int slab_order(int size, int min_objects,
2685 int max_order, int fract_leftover, int reserved)
2689 int min_order = slub_min_order;
2691 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2692 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2694 for (order = max(min_order,
2695 fls(min_objects * size - 1) - PAGE_SHIFT);
2696 order <= max_order; order++) {
2698 unsigned long slab_size = PAGE_SIZE << order;
2700 if (slab_size < min_objects * size + reserved)
2703 rem = (slab_size - reserved) % size;
2705 if (rem <= slab_size / fract_leftover)
2713 static inline int calculate_order(int size, int reserved)
2721 * Attempt to find best configuration for a slab. This
2722 * works by first attempting to generate a layout with
2723 * the best configuration and backing off gradually.
2725 * First we reduce the acceptable waste in a slab. Then
2726 * we reduce the minimum objects required in a slab.
2728 min_objects = slub_min_objects;
2730 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2731 max_objects = order_objects(slub_max_order, size, reserved);
2732 min_objects = min(min_objects, max_objects);
2734 while (min_objects > 1) {
2736 while (fraction >= 4) {
2737 order = slab_order(size, min_objects,
2738 slub_max_order, fraction, reserved);
2739 if (order <= slub_max_order)
2747 * We were unable to place multiple objects in a slab. Now
2748 * lets see if we can place a single object there.
2750 order = slab_order(size, 1, slub_max_order, 1, reserved);
2751 if (order <= slub_max_order)
2755 * Doh this slab cannot be placed using slub_max_order.
2757 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2758 if (order < MAX_ORDER)
2764 * Figure out what the alignment of the objects will be.
2766 static unsigned long calculate_alignment(unsigned long flags,
2767 unsigned long align, unsigned long size)
2770 * If the user wants hardware cache aligned objects then follow that
2771 * suggestion if the object is sufficiently large.
2773 * The hardware cache alignment cannot override the specified
2774 * alignment though. If that is greater then use it.
2776 if (flags & SLAB_HWCACHE_ALIGN) {
2777 unsigned long ralign = cache_line_size();
2778 while (size <= ralign / 2)
2780 align = max(align, ralign);
2783 if (align < ARCH_SLAB_MINALIGN)
2784 align = ARCH_SLAB_MINALIGN;
2786 return ALIGN(align, sizeof(void *));
2790 init_kmem_cache_node(struct kmem_cache_node *n)
2793 spin_lock_init(&n->list_lock);
2794 INIT_LIST_HEAD(&n->partial);
2795 #ifdef CONFIG_SLUB_DEBUG
2796 atomic_long_set(&n->nr_slabs, 0);
2797 atomic_long_set(&n->total_objects, 0);
2798 INIT_LIST_HEAD(&n->full);
2802 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2804 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2805 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2808 * Must align to double word boundary for the double cmpxchg
2809 * instructions to work; see __pcpu_double_call_return_bool().
2811 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2812 2 * sizeof(void *));
2817 init_kmem_cache_cpus(s);
2822 static struct kmem_cache *kmem_cache_node;
2825 * No kmalloc_node yet so do it by hand. We know that this is the first
2826 * slab on the node for this slabcache. There are no concurrent accesses
2829 * Note that this function only works on the kmalloc_node_cache
2830 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2831 * memory on a fresh node that has no slab structures yet.
2833 static void early_kmem_cache_node_alloc(int node)
2836 struct kmem_cache_node *n;
2838 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2840 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2843 if (page_to_nid(page) != node) {
2844 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2846 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2847 "in order to be able to continue\n");
2852 page->freelist = get_freepointer(kmem_cache_node, n);
2855 kmem_cache_node->node[node] = n;
2856 #ifdef CONFIG_SLUB_DEBUG
2857 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2858 init_tracking(kmem_cache_node, n);
2860 init_kmem_cache_node(n);
2861 inc_slabs_node(kmem_cache_node, node, page->objects);
2863 add_partial(n, page, DEACTIVATE_TO_HEAD);
2866 static void free_kmem_cache_nodes(struct kmem_cache *s)
2870 for_each_node_state(node, N_NORMAL_MEMORY) {
2871 struct kmem_cache_node *n = s->node[node];
2874 kmem_cache_free(kmem_cache_node, n);
2876 s->node[node] = NULL;
2880 static int init_kmem_cache_nodes(struct kmem_cache *s)
2884 for_each_node_state(node, N_NORMAL_MEMORY) {
2885 struct kmem_cache_node *n;
2887 if (slab_state == DOWN) {
2888 early_kmem_cache_node_alloc(node);
2891 n = kmem_cache_alloc_node(kmem_cache_node,
2895 free_kmem_cache_nodes(s);
2900 init_kmem_cache_node(n);
2905 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2907 if (min < MIN_PARTIAL)
2909 else if (min > MAX_PARTIAL)
2911 s->min_partial = min;
2915 * calculate_sizes() determines the order and the distribution of data within
2918 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2920 unsigned long flags = s->flags;
2921 unsigned long size = s->object_size;
2922 unsigned long align = s->align;
2926 * Round up object size to the next word boundary. We can only
2927 * place the free pointer at word boundaries and this determines
2928 * the possible location of the free pointer.
2930 size = ALIGN(size, sizeof(void *));
2932 #ifdef CONFIG_SLUB_DEBUG
2934 * Determine if we can poison the object itself. If the user of
2935 * the slab may touch the object after free or before allocation
2936 * then we should never poison the object itself.
2938 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2940 s->flags |= __OBJECT_POISON;
2942 s->flags &= ~__OBJECT_POISON;
2946 * If we are Redzoning then check if there is some space between the
2947 * end of the object and the free pointer. If not then add an
2948 * additional word to have some bytes to store Redzone information.
2950 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2951 size += sizeof(void *);
2955 * With that we have determined the number of bytes in actual use
2956 * by the object. This is the potential offset to the free pointer.
2960 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2963 * Relocate free pointer after the object if it is not
2964 * permitted to overwrite the first word of the object on
2967 * This is the case if we do RCU, have a constructor or
2968 * destructor or are poisoning the objects.
2971 size += sizeof(void *);
2974 #ifdef CONFIG_SLUB_DEBUG
2975 if (flags & SLAB_STORE_USER)
2977 * Need to store information about allocs and frees after
2980 size += 2 * sizeof(struct track);
2982 if (flags & SLAB_RED_ZONE)
2984 * Add some empty padding so that we can catch
2985 * overwrites from earlier objects rather than let
2986 * tracking information or the free pointer be
2987 * corrupted if a user writes before the start
2990 size += sizeof(void *);
2994 * Determine the alignment based on various parameters that the
2995 * user specified and the dynamic determination of cache line size
2998 align = calculate_alignment(flags, align, s->object_size);
3002 * SLUB stores one object immediately after another beginning from
3003 * offset 0. In order to align the objects we have to simply size
3004 * each object to conform to the alignment.
3006 size = ALIGN(size, align);
3008 if (forced_order >= 0)
3009 order = forced_order;
3011 order = calculate_order(size, s->reserved);
3018 s->allocflags |= __GFP_COMP;
3020 if (s->flags & SLAB_CACHE_DMA)
3021 s->allocflags |= SLUB_DMA;
3023 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3024 s->allocflags |= __GFP_RECLAIMABLE;
3027 * Determine the number of objects per slab
3029 s->oo = oo_make(order, size, s->reserved);
3030 s->min = oo_make(get_order(size), size, s->reserved);
3031 if (oo_objects(s->oo) > oo_objects(s->max))
3034 return !!oo_objects(s->oo);
3038 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3040 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3043 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3044 s->reserved = sizeof(struct rcu_head);
3046 if (!calculate_sizes(s, -1))
3048 if (disable_higher_order_debug) {
3050 * Disable debugging flags that store metadata if the min slab
3053 if (get_order(s->size) > get_order(s->object_size)) {
3054 s->flags &= ~DEBUG_METADATA_FLAGS;
3056 if (!calculate_sizes(s, -1))
3061 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3062 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3063 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3064 /* Enable fast mode */
3065 s->flags |= __CMPXCHG_DOUBLE;
3069 * The larger the object size is, the more pages we want on the partial
3070 * list to avoid pounding the page allocator excessively.
3072 set_min_partial(s, ilog2(s->size) / 2);
3075 * cpu_partial determined the maximum number of objects kept in the
3076 * per cpu partial lists of a processor.
3078 * Per cpu partial lists mainly contain slabs that just have one
3079 * object freed. If they are used for allocation then they can be
3080 * filled up again with minimal effort. The slab will never hit the
3081 * per node partial lists and therefore no locking will be required.
3083 * This setting also determines
3085 * A) The number of objects from per cpu partial slabs dumped to the
3086 * per node list when we reach the limit.
3087 * B) The number of objects in cpu partial slabs to extract from the
3088 * per node list when we run out of per cpu objects. We only fetch 50%
3089 * to keep some capacity around for frees.
3091 if (kmem_cache_debug(s))
3093 else if (s->size >= PAGE_SIZE)
3095 else if (s->size >= 1024)
3097 else if (s->size >= 256)
3098 s->cpu_partial = 13;
3100 s->cpu_partial = 30;
3103 s->remote_node_defrag_ratio = 1000;
3105 if (!init_kmem_cache_nodes(s))
3108 if (alloc_kmem_cache_cpus(s))
3111 free_kmem_cache_nodes(s);
3113 if (flags & SLAB_PANIC)
3114 panic("Cannot create slab %s size=%lu realsize=%u "
3115 "order=%u offset=%u flags=%lx\n",
3116 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3121 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3124 #ifdef CONFIG_SLUB_DEBUG
3125 void *addr = page_address(page);
3127 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3128 sizeof(long), GFP_ATOMIC);
3131 slab_err(s, page, text, s->name);
3134 get_map(s, page, map);
3135 for_each_object(p, s, addr, page->objects) {
3137 if (!test_bit(slab_index(p, s, addr), map)) {
3138 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3140 print_tracking(s, p);
3149 * Attempt to free all partial slabs on a node.
3150 * This is called from kmem_cache_close(). We must be the last thread
3151 * using the cache and therefore we do not need to lock anymore.
3153 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3155 struct page *page, *h;
3157 list_for_each_entry_safe(page, h, &n->partial, lru) {
3159 remove_partial(n, page);
3160 discard_slab(s, page);
3162 list_slab_objects(s, page,
3163 "Objects remaining in %s on kmem_cache_close()");
3169 * Release all resources used by a slab cache.
3171 static inline int kmem_cache_close(struct kmem_cache *s)
3176 /* Attempt to free all objects */
3177 for_each_node_state(node, N_NORMAL_MEMORY) {
3178 struct kmem_cache_node *n = get_node(s, node);
3181 if (n->nr_partial || slabs_node(s, node))
3184 free_percpu(s->cpu_slab);
3185 free_kmem_cache_nodes(s);
3189 int __kmem_cache_shutdown(struct kmem_cache *s)
3191 int rc = kmem_cache_close(s);
3194 sysfs_slab_remove(s);
3199 /********************************************************************
3201 *******************************************************************/
3203 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3204 EXPORT_SYMBOL(kmalloc_caches);
3206 #ifdef CONFIG_ZONE_DMA
3207 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3210 static int __init setup_slub_min_order(char *str)
3212 get_option(&str, &slub_min_order);
3217 __setup("slub_min_order=", setup_slub_min_order);
3219 static int __init setup_slub_max_order(char *str)
3221 get_option(&str, &slub_max_order);
3222 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3227 __setup("slub_max_order=", setup_slub_max_order);
3229 static int __init setup_slub_min_objects(char *str)
3231 get_option(&str, &slub_min_objects);
3236 __setup("slub_min_objects=", setup_slub_min_objects);
3238 static int __init setup_slub_nomerge(char *str)
3244 __setup("slub_nomerge", setup_slub_nomerge);
3246 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3247 int size, unsigned int flags)
3249 struct kmem_cache *s;
3251 s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3254 s->size = s->object_size = size;
3255 s->align = ARCH_KMALLOC_MINALIGN;
3258 * This function is called with IRQs disabled during early-boot on
3259 * single CPU so there's no need to take slab_mutex here.
3261 if (kmem_cache_open(s, flags))
3264 list_add(&s->list, &slab_caches);
3268 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3273 * Conversion table for small slabs sizes / 8 to the index in the
3274 * kmalloc array. This is necessary for slabs < 192 since we have non power
3275 * of two cache sizes there. The size of larger slabs can be determined using
3278 static s8 size_index[24] = {
3305 static inline int size_index_elem(size_t bytes)
3307 return (bytes - 1) / 8;
3310 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3316 return ZERO_SIZE_PTR;
3318 index = size_index[size_index_elem(size)];
3320 index = fls(size - 1);
3322 #ifdef CONFIG_ZONE_DMA
3323 if (unlikely((flags & SLUB_DMA)))
3324 return kmalloc_dma_caches[index];
3327 return kmalloc_caches[index];
3330 void *__kmalloc(size_t size, gfp_t flags)
3332 struct kmem_cache *s;
3335 if (unlikely(size > SLUB_MAX_SIZE))
3336 return kmalloc_large(size, flags);
3338 s = get_slab(size, flags);
3340 if (unlikely(ZERO_OR_NULL_PTR(s)))
3343 ret = slab_alloc(s, flags, _RET_IP_);
3345 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3349 EXPORT_SYMBOL(__kmalloc);
3352 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3357 flags |= __GFP_COMP | __GFP_NOTRACK;
3358 page = alloc_pages_node(node, flags, get_order(size));
3360 ptr = page_address(page);
3362 kmemleak_alloc(ptr, size, 1, flags);
3366 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3368 struct kmem_cache *s;
3371 if (unlikely(size > SLUB_MAX_SIZE)) {
3372 ret = kmalloc_large_node(size, flags, node);
3374 trace_kmalloc_node(_RET_IP_, ret,
3375 size, PAGE_SIZE << get_order(size),
3381 s = get_slab(size, flags);
3383 if (unlikely(ZERO_OR_NULL_PTR(s)))
3386 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3388 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3392 EXPORT_SYMBOL(__kmalloc_node);
3395 size_t ksize(const void *object)
3399 if (unlikely(object == ZERO_SIZE_PTR))
3402 page = virt_to_head_page(object);
3404 if (unlikely(!PageSlab(page))) {
3405 WARN_ON(!PageCompound(page));
3406 return PAGE_SIZE << compound_order(page);
3409 return slab_ksize(page->slab_cache);
3411 EXPORT_SYMBOL(ksize);
3413 #ifdef CONFIG_SLUB_DEBUG
3414 bool verify_mem_not_deleted(const void *x)
3417 void *object = (void *)x;
3418 unsigned long flags;
3421 if (unlikely(ZERO_OR_NULL_PTR(x)))
3424 local_irq_save(flags);
3426 page = virt_to_head_page(x);
3427 if (unlikely(!PageSlab(page))) {
3428 /* maybe it was from stack? */
3434 if (on_freelist(page->slab_cache, page, object)) {
3435 object_err(page->slab_cache, page, object, "Object is on free-list");
3443 local_irq_restore(flags);
3446 EXPORT_SYMBOL(verify_mem_not_deleted);
3449 void kfree(const void *x)
3452 void *object = (void *)x;
3454 trace_kfree(_RET_IP_, x);
3456 if (unlikely(ZERO_OR_NULL_PTR(x)))
3459 page = virt_to_head_page(x);
3460 if (unlikely(!PageSlab(page))) {
3461 BUG_ON(!PageCompound(page));
3463 __free_pages(page, compound_order(page));
3466 slab_free(page->slab_cache, page, object, _RET_IP_);
3468 EXPORT_SYMBOL(kfree);
3471 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3472 * the remaining slabs by the number of items in use. The slabs with the
3473 * most items in use come first. New allocations will then fill those up
3474 * and thus they can be removed from the partial lists.
3476 * The slabs with the least items are placed last. This results in them
3477 * being allocated from last increasing the chance that the last objects
3478 * are freed in them.
3480 int kmem_cache_shrink(struct kmem_cache *s)
3484 struct kmem_cache_node *n;
3487 int objects = oo_objects(s->max);
3488 struct list_head *slabs_by_inuse =
3489 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3490 unsigned long flags;
3492 if (!slabs_by_inuse)
3496 for_each_node_state(node, N_NORMAL_MEMORY) {
3497 n = get_node(s, node);
3502 for (i = 0; i < objects; i++)
3503 INIT_LIST_HEAD(slabs_by_inuse + i);
3505 spin_lock_irqsave(&n->list_lock, flags);
3508 * Build lists indexed by the items in use in each slab.
3510 * Note that concurrent frees may occur while we hold the
3511 * list_lock. page->inuse here is the upper limit.
3513 list_for_each_entry_safe(page, t, &n->partial, lru) {
3514 list_move(&page->lru, slabs_by_inuse + page->inuse);
3520 * Rebuild the partial list with the slabs filled up most
3521 * first and the least used slabs at the end.
3523 for (i = objects - 1; i > 0; i--)
3524 list_splice(slabs_by_inuse + i, n->partial.prev);
3526 spin_unlock_irqrestore(&n->list_lock, flags);
3528 /* Release empty slabs */
3529 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3530 discard_slab(s, page);
3533 kfree(slabs_by_inuse);
3536 EXPORT_SYMBOL(kmem_cache_shrink);
3538 #if defined(CONFIG_MEMORY_HOTPLUG)
3539 static int slab_mem_going_offline_callback(void *arg)
3541 struct kmem_cache *s;
3543 mutex_lock(&slab_mutex);
3544 list_for_each_entry(s, &slab_caches, list)
3545 kmem_cache_shrink(s);
3546 mutex_unlock(&slab_mutex);
3551 static void slab_mem_offline_callback(void *arg)
3553 struct kmem_cache_node *n;
3554 struct kmem_cache *s;
3555 struct memory_notify *marg = arg;
3558 offline_node = marg->status_change_nid;
3561 * If the node still has available memory. we need kmem_cache_node
3564 if (offline_node < 0)
3567 mutex_lock(&slab_mutex);
3568 list_for_each_entry(s, &slab_caches, list) {
3569 n = get_node(s, offline_node);
3572 * if n->nr_slabs > 0, slabs still exist on the node
3573 * that is going down. We were unable to free them,
3574 * and offline_pages() function shouldn't call this
3575 * callback. So, we must fail.
3577 BUG_ON(slabs_node(s, offline_node));
3579 s->node[offline_node] = NULL;
3580 kmem_cache_free(kmem_cache_node, n);
3583 mutex_unlock(&slab_mutex);
3586 static int slab_mem_going_online_callback(void *arg)
3588 struct kmem_cache_node *n;
3589 struct kmem_cache *s;
3590 struct memory_notify *marg = arg;
3591 int nid = marg->status_change_nid;
3595 * If the node's memory is already available, then kmem_cache_node is
3596 * already created. Nothing to do.
3602 * We are bringing a node online. No memory is available yet. We must
3603 * allocate a kmem_cache_node structure in order to bring the node
3606 mutex_lock(&slab_mutex);
3607 list_for_each_entry(s, &slab_caches, list) {
3609 * XXX: kmem_cache_alloc_node will fallback to other nodes
3610 * since memory is not yet available from the node that
3613 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3618 init_kmem_cache_node(n);
3622 mutex_unlock(&slab_mutex);
3626 static int slab_memory_callback(struct notifier_block *self,
3627 unsigned long action, void *arg)
3632 case MEM_GOING_ONLINE:
3633 ret = slab_mem_going_online_callback(arg);
3635 case MEM_GOING_OFFLINE:
3636 ret = slab_mem_going_offline_callback(arg);
3639 case MEM_CANCEL_ONLINE:
3640 slab_mem_offline_callback(arg);
3643 case MEM_CANCEL_OFFLINE:
3647 ret = notifier_from_errno(ret);
3653 #endif /* CONFIG_MEMORY_HOTPLUG */
3655 /********************************************************************
3656 * Basic setup of slabs
3657 *******************************************************************/
3660 * Used for early kmem_cache structures that were allocated using
3661 * the page allocator
3664 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3668 list_add(&s->list, &slab_caches);
3671 for_each_node_state(node, N_NORMAL_MEMORY) {
3672 struct kmem_cache_node *n = get_node(s, node);
3676 list_for_each_entry(p, &n->partial, lru)
3679 #ifdef CONFIG_SLUB_DEBUG
3680 list_for_each_entry(p, &n->full, lru)
3687 void __init kmem_cache_init(void)
3691 struct kmem_cache *temp_kmem_cache;
3693 struct kmem_cache *temp_kmem_cache_node;
3694 unsigned long kmalloc_size;
3696 if (debug_guardpage_minorder())
3699 kmem_size = offsetof(struct kmem_cache, node) +
3700 nr_node_ids * sizeof(struct kmem_cache_node *);
3702 /* Allocate two kmem_caches from the page allocator */
3703 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3704 order = get_order(2 * kmalloc_size);
3705 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT | __GFP_ZERO, order);
3708 * Must first have the slab cache available for the allocations of the
3709 * struct kmem_cache_node's. There is special bootstrap code in
3710 * kmem_cache_open for slab_state == DOWN.
3712 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3714 kmem_cache_node->name = "kmem_cache_node";
3715 kmem_cache_node->size = kmem_cache_node->object_size =
3716 sizeof(struct kmem_cache_node);
3717 kmem_cache_open(kmem_cache_node, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3719 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3721 /* Able to allocate the per node structures */
3722 slab_state = PARTIAL;
3724 temp_kmem_cache = kmem_cache;
3725 kmem_cache->name = "kmem_cache";
3726 kmem_cache->size = kmem_cache->object_size = kmem_size;
3727 kmem_cache_open(kmem_cache, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
3729 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3730 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3733 * Allocate kmem_cache_node properly from the kmem_cache slab.
3734 * kmem_cache_node is separately allocated so no need to
3735 * update any list pointers.
3737 temp_kmem_cache_node = kmem_cache_node;
3739 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3740 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3742 kmem_cache_bootstrap_fixup(kmem_cache_node);
3745 kmem_cache_bootstrap_fixup(kmem_cache);
3747 /* Free temporary boot structure */
3748 free_pages((unsigned long)temp_kmem_cache, order);
3750 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3753 * Patch up the size_index table if we have strange large alignment
3754 * requirements for the kmalloc array. This is only the case for
3755 * MIPS it seems. The standard arches will not generate any code here.
3757 * Largest permitted alignment is 256 bytes due to the way we
3758 * handle the index determination for the smaller caches.
3760 * Make sure that nothing crazy happens if someone starts tinkering
3761 * around with ARCH_KMALLOC_MINALIGN
3763 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3764 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3766 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3767 int elem = size_index_elem(i);
3768 if (elem >= ARRAY_SIZE(size_index))
3770 size_index[elem] = KMALLOC_SHIFT_LOW;
3773 if (KMALLOC_MIN_SIZE == 64) {
3775 * The 96 byte size cache is not used if the alignment
3778 for (i = 64 + 8; i <= 96; i += 8)
3779 size_index[size_index_elem(i)] = 7;
3780 } else if (KMALLOC_MIN_SIZE == 128) {
3782 * The 192 byte sized cache is not used if the alignment
3783 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3786 for (i = 128 + 8; i <= 192; i += 8)
3787 size_index[size_index_elem(i)] = 8;
3790 /* Caches that are not of the two-to-the-power-of size */
3791 if (KMALLOC_MIN_SIZE <= 32) {
3792 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3796 if (KMALLOC_MIN_SIZE <= 64) {
3797 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3801 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3802 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3808 /* Provide the correct kmalloc names now that the caches are up */
3809 if (KMALLOC_MIN_SIZE <= 32) {
3810 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3811 BUG_ON(!kmalloc_caches[1]->name);
3814 if (KMALLOC_MIN_SIZE <= 64) {
3815 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3816 BUG_ON(!kmalloc_caches[2]->name);
3819 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3820 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3823 kmalloc_caches[i]->name = s;
3827 register_cpu_notifier(&slab_notifier);
3830 #ifdef CONFIG_ZONE_DMA
3831 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3832 struct kmem_cache *s = kmalloc_caches[i];
3835 char *name = kasprintf(GFP_NOWAIT,
3836 "dma-kmalloc-%d", s->object_size);
3839 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3840 s->object_size, SLAB_CACHE_DMA);
3845 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3846 " CPUs=%d, Nodes=%d\n",
3847 caches, cache_line_size(),
3848 slub_min_order, slub_max_order, slub_min_objects,
3849 nr_cpu_ids, nr_node_ids);
3852 void __init kmem_cache_init_late(void)
3857 * Find a mergeable slab cache
3859 static int slab_unmergeable(struct kmem_cache *s)
3861 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3868 * We may have set a slab to be unmergeable during bootstrap.
3870 if (s->refcount < 0)
3876 static struct kmem_cache *find_mergeable(size_t size,
3877 size_t align, unsigned long flags, const char *name,
3878 void (*ctor)(void *))
3880 struct kmem_cache *s;
3882 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3888 size = ALIGN(size, sizeof(void *));
3889 align = calculate_alignment(flags, align, size);
3890 size = ALIGN(size, align);
3891 flags = kmem_cache_flags(size, flags, name, NULL);
3893 list_for_each_entry(s, &slab_caches, list) {
3894 if (slab_unmergeable(s))
3900 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3903 * Check if alignment is compatible.
3904 * Courtesy of Adrian Drzewiecki
3906 if ((s->size & ~(align - 1)) != s->size)
3909 if (s->size - size >= sizeof(void *))
3917 struct kmem_cache *__kmem_cache_alias(const char *name, size_t size,
3918 size_t align, unsigned long flags, void (*ctor)(void *))
3920 struct kmem_cache *s;
3922 s = find_mergeable(size, align, flags, name, ctor);
3926 * Adjust the object sizes so that we clear
3927 * the complete object on kzalloc.
3929 s->object_size = max(s->object_size, (int)size);
3930 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3932 if (sysfs_slab_alias(s, name)) {
3941 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3945 err = kmem_cache_open(s, flags);
3949 mutex_unlock(&slab_mutex);
3950 err = sysfs_slab_add(s);
3951 mutex_lock(&slab_mutex);
3954 kmem_cache_close(s);
3961 * Use the cpu notifier to insure that the cpu slabs are flushed when
3964 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3965 unsigned long action, void *hcpu)
3967 long cpu = (long)hcpu;
3968 struct kmem_cache *s;
3969 unsigned long flags;
3972 case CPU_UP_CANCELED:
3973 case CPU_UP_CANCELED_FROZEN:
3975 case CPU_DEAD_FROZEN:
3976 mutex_lock(&slab_mutex);
3977 list_for_each_entry(s, &slab_caches, list) {
3978 local_irq_save(flags);
3979 __flush_cpu_slab(s, cpu);
3980 local_irq_restore(flags);
3982 mutex_unlock(&slab_mutex);
3990 static struct notifier_block __cpuinitdata slab_notifier = {
3991 .notifier_call = slab_cpuup_callback
3996 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3998 struct kmem_cache *s;
4001 if (unlikely(size > SLUB_MAX_SIZE))
4002 return kmalloc_large(size, gfpflags);
4004 s = get_slab(size, gfpflags);
4006 if (unlikely(ZERO_OR_NULL_PTR(s)))
4009 ret = slab_alloc(s, gfpflags, caller);
4011 /* Honor the call site pointer we received. */
4012 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4018 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4019 int node, unsigned long caller)
4021 struct kmem_cache *s;
4024 if (unlikely(size > SLUB_MAX_SIZE)) {
4025 ret = kmalloc_large_node(size, gfpflags, node);
4027 trace_kmalloc_node(caller, ret,
4028 size, PAGE_SIZE << get_order(size),
4034 s = get_slab(size, gfpflags);
4036 if (unlikely(ZERO_OR_NULL_PTR(s)))
4039 ret = slab_alloc_node(s, gfpflags, node, caller);
4041 /* Honor the call site pointer we received. */
4042 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4049 static int count_inuse(struct page *page)
4054 static int count_total(struct page *page)
4056 return page->objects;
4060 #ifdef CONFIG_SLUB_DEBUG
4061 static int validate_slab(struct kmem_cache *s, struct page *page,
4065 void *addr = page_address(page);
4067 if (!check_slab(s, page) ||
4068 !on_freelist(s, page, NULL))
4071 /* Now we know that a valid freelist exists */
4072 bitmap_zero(map, page->objects);
4074 get_map(s, page, map);
4075 for_each_object(p, s, addr, page->objects) {
4076 if (test_bit(slab_index(p, s, addr), map))
4077 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4081 for_each_object(p, s, addr, page->objects)
4082 if (!test_bit(slab_index(p, s, addr), map))
4083 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4088 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4092 validate_slab(s, page, map);
4096 static int validate_slab_node(struct kmem_cache *s,
4097 struct kmem_cache_node *n, unsigned long *map)
4099 unsigned long count = 0;
4101 unsigned long flags;
4103 spin_lock_irqsave(&n->list_lock, flags);
4105 list_for_each_entry(page, &n->partial, lru) {
4106 validate_slab_slab(s, page, map);
4109 if (count != n->nr_partial)
4110 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4111 "counter=%ld\n", s->name, count, n->nr_partial);
4113 if (!(s->flags & SLAB_STORE_USER))
4116 list_for_each_entry(page, &n->full, lru) {
4117 validate_slab_slab(s, page, map);
4120 if (count != atomic_long_read(&n->nr_slabs))
4121 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4122 "counter=%ld\n", s->name, count,
4123 atomic_long_read(&n->nr_slabs));
4126 spin_unlock_irqrestore(&n->list_lock, flags);
4130 static long validate_slab_cache(struct kmem_cache *s)
4133 unsigned long count = 0;
4134 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4135 sizeof(unsigned long), GFP_KERNEL);
4141 for_each_node_state(node, N_NORMAL_MEMORY) {
4142 struct kmem_cache_node *n = get_node(s, node);
4144 count += validate_slab_node(s, n, map);
4150 * Generate lists of code addresses where slabcache objects are allocated
4155 unsigned long count;
4162 DECLARE_BITMAP(cpus, NR_CPUS);
4168 unsigned long count;
4169 struct location *loc;
4172 static void free_loc_track(struct loc_track *t)
4175 free_pages((unsigned long)t->loc,
4176 get_order(sizeof(struct location) * t->max));
4179 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4184 order = get_order(sizeof(struct location) * max);
4186 l = (void *)__get_free_pages(flags, order);
4191 memcpy(l, t->loc, sizeof(struct location) * t->count);
4199 static int add_location(struct loc_track *t, struct kmem_cache *s,
4200 const struct track *track)
4202 long start, end, pos;
4204 unsigned long caddr;
4205 unsigned long age = jiffies - track->when;
4211 pos = start + (end - start + 1) / 2;
4214 * There is nothing at "end". If we end up there
4215 * we need to add something to before end.
4220 caddr = t->loc[pos].addr;
4221 if (track->addr == caddr) {
4227 if (age < l->min_time)
4229 if (age > l->max_time)
4232 if (track->pid < l->min_pid)
4233 l->min_pid = track->pid;
4234 if (track->pid > l->max_pid)
4235 l->max_pid = track->pid;
4237 cpumask_set_cpu(track->cpu,
4238 to_cpumask(l->cpus));
4240 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4244 if (track->addr < caddr)
4251 * Not found. Insert new tracking element.
4253 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4259 (t->count - pos) * sizeof(struct location));
4262 l->addr = track->addr;
4266 l->min_pid = track->pid;
4267 l->max_pid = track->pid;
4268 cpumask_clear(to_cpumask(l->cpus));
4269 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4270 nodes_clear(l->nodes);
4271 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4275 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4276 struct page *page, enum track_item alloc,
4279 void *addr = page_address(page);
4282 bitmap_zero(map, page->objects);
4283 get_map(s, page, map);
4285 for_each_object(p, s, addr, page->objects)
4286 if (!test_bit(slab_index(p, s, addr), map))
4287 add_location(t, s, get_track(s, p, alloc));
4290 static int list_locations(struct kmem_cache *s, char *buf,
4291 enum track_item alloc)
4295 struct loc_track t = { 0, 0, NULL };
4297 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4298 sizeof(unsigned long), GFP_KERNEL);
4300 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4303 return sprintf(buf, "Out of memory\n");
4305 /* Push back cpu slabs */
4308 for_each_node_state(node, N_NORMAL_MEMORY) {
4309 struct kmem_cache_node *n = get_node(s, node);
4310 unsigned long flags;
4313 if (!atomic_long_read(&n->nr_slabs))
4316 spin_lock_irqsave(&n->list_lock, flags);
4317 list_for_each_entry(page, &n->partial, lru)
4318 process_slab(&t, s, page, alloc, map);
4319 list_for_each_entry(page, &n->full, lru)
4320 process_slab(&t, s, page, alloc, map);
4321 spin_unlock_irqrestore(&n->list_lock, flags);
4324 for (i = 0; i < t.count; i++) {
4325 struct location *l = &t.loc[i];
4327 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4329 len += sprintf(buf + len, "%7ld ", l->count);
4332 len += sprintf(buf + len, "%pS", (void *)l->addr);
4334 len += sprintf(buf + len, "<not-available>");
4336 if (l->sum_time != l->min_time) {
4337 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4339 (long)div_u64(l->sum_time, l->count),
4342 len += sprintf(buf + len, " age=%ld",
4345 if (l->min_pid != l->max_pid)
4346 len += sprintf(buf + len, " pid=%ld-%ld",
4347 l->min_pid, l->max_pid);
4349 len += sprintf(buf + len, " pid=%ld",
4352 if (num_online_cpus() > 1 &&
4353 !cpumask_empty(to_cpumask(l->cpus)) &&
4354 len < PAGE_SIZE - 60) {
4355 len += sprintf(buf + len, " cpus=");
4356 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4357 to_cpumask(l->cpus));
4360 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4361 len < PAGE_SIZE - 60) {
4362 len += sprintf(buf + len, " nodes=");
4363 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4367 len += sprintf(buf + len, "\n");
4373 len += sprintf(buf, "No data\n");
4378 #ifdef SLUB_RESILIENCY_TEST
4379 static void resiliency_test(void)
4383 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4385 printk(KERN_ERR "SLUB resiliency testing\n");
4386 printk(KERN_ERR "-----------------------\n");
4387 printk(KERN_ERR "A. Corruption after allocation\n");
4389 p = kzalloc(16, GFP_KERNEL);
4391 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4392 " 0x12->0x%p\n\n", p + 16);
4394 validate_slab_cache(kmalloc_caches[4]);
4396 /* Hmmm... The next two are dangerous */
4397 p = kzalloc(32, GFP_KERNEL);
4398 p[32 + sizeof(void *)] = 0x34;
4399 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4400 " 0x34 -> -0x%p\n", p);
4402 "If allocated object is overwritten then not detectable\n\n");
4404 validate_slab_cache(kmalloc_caches[5]);
4405 p = kzalloc(64, GFP_KERNEL);
4406 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4408 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4411 "If allocated object is overwritten then not detectable\n\n");
4412 validate_slab_cache(kmalloc_caches[6]);
4414 printk(KERN_ERR "\nB. Corruption after free\n");
4415 p = kzalloc(128, GFP_KERNEL);
4418 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4419 validate_slab_cache(kmalloc_caches[7]);
4421 p = kzalloc(256, GFP_KERNEL);
4424 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4426 validate_slab_cache(kmalloc_caches[8]);
4428 p = kzalloc(512, GFP_KERNEL);
4431 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4432 validate_slab_cache(kmalloc_caches[9]);
4436 static void resiliency_test(void) {};
4441 enum slab_stat_type {
4442 SL_ALL, /* All slabs */
4443 SL_PARTIAL, /* Only partially allocated slabs */
4444 SL_CPU, /* Only slabs used for cpu caches */
4445 SL_OBJECTS, /* Determine allocated objects not slabs */
4446 SL_TOTAL /* Determine object capacity not slabs */
4449 #define SO_ALL (1 << SL_ALL)
4450 #define SO_PARTIAL (1 << SL_PARTIAL)
4451 #define SO_CPU (1 << SL_CPU)
4452 #define SO_OBJECTS (1 << SL_OBJECTS)
4453 #define SO_TOTAL (1 << SL_TOTAL)
4455 static ssize_t show_slab_objects(struct kmem_cache *s,
4456 char *buf, unsigned long flags)
4458 unsigned long total = 0;
4461 unsigned long *nodes;
4462 unsigned long *per_cpu;
4464 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4467 per_cpu = nodes + nr_node_ids;
4469 if (flags & SO_CPU) {
4472 for_each_possible_cpu(cpu) {
4473 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4477 page = ACCESS_ONCE(c->page);
4481 node = page_to_nid(page);
4482 if (flags & SO_TOTAL)
4484 else if (flags & SO_OBJECTS)
4492 page = ACCESS_ONCE(c->partial);
4503 lock_memory_hotplug();
4504 #ifdef CONFIG_SLUB_DEBUG
4505 if (flags & SO_ALL) {
4506 for_each_node_state(node, N_NORMAL_MEMORY) {
4507 struct kmem_cache_node *n = get_node(s, node);
4509 if (flags & SO_TOTAL)
4510 x = atomic_long_read(&n->total_objects);
4511 else if (flags & SO_OBJECTS)
4512 x = atomic_long_read(&n->total_objects) -
4513 count_partial(n, count_free);
4516 x = atomic_long_read(&n->nr_slabs);
4523 if (flags & SO_PARTIAL) {
4524 for_each_node_state(node, N_NORMAL_MEMORY) {
4525 struct kmem_cache_node *n = get_node(s, node);
4527 if (flags & SO_TOTAL)
4528 x = count_partial(n, count_total);
4529 else if (flags & SO_OBJECTS)
4530 x = count_partial(n, count_inuse);
4537 x = sprintf(buf, "%lu", total);
4539 for_each_node_state(node, N_NORMAL_MEMORY)
4541 x += sprintf(buf + x, " N%d=%lu",
4544 unlock_memory_hotplug();
4546 return x + sprintf(buf + x, "\n");
4549 #ifdef CONFIG_SLUB_DEBUG
4550 static int any_slab_objects(struct kmem_cache *s)
4554 for_each_online_node(node) {
4555 struct kmem_cache_node *n = get_node(s, node);
4560 if (atomic_long_read(&n->total_objects))
4567 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4568 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4570 struct slab_attribute {
4571 struct attribute attr;
4572 ssize_t (*show)(struct kmem_cache *s, char *buf);
4573 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4576 #define SLAB_ATTR_RO(_name) \
4577 static struct slab_attribute _name##_attr = \
4578 __ATTR(_name, 0400, _name##_show, NULL)
4580 #define SLAB_ATTR(_name) \
4581 static struct slab_attribute _name##_attr = \
4582 __ATTR(_name, 0600, _name##_show, _name##_store)
4584 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4586 return sprintf(buf, "%d\n", s->size);
4588 SLAB_ATTR_RO(slab_size);
4590 static ssize_t align_show(struct kmem_cache *s, char *buf)
4592 return sprintf(buf, "%d\n", s->align);
4594 SLAB_ATTR_RO(align);
4596 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4598 return sprintf(buf, "%d\n", s->object_size);
4600 SLAB_ATTR_RO(object_size);
4602 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4604 return sprintf(buf, "%d\n", oo_objects(s->oo));
4606 SLAB_ATTR_RO(objs_per_slab);
4608 static ssize_t order_store(struct kmem_cache *s,
4609 const char *buf, size_t length)
4611 unsigned long order;
4614 err = strict_strtoul(buf, 10, &order);
4618 if (order > slub_max_order || order < slub_min_order)
4621 calculate_sizes(s, order);
4625 static ssize_t order_show(struct kmem_cache *s, char *buf)
4627 return sprintf(buf, "%d\n", oo_order(s->oo));
4631 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4633 return sprintf(buf, "%lu\n", s->min_partial);
4636 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4642 err = strict_strtoul(buf, 10, &min);
4646 set_min_partial(s, min);
4649 SLAB_ATTR(min_partial);
4651 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4653 return sprintf(buf, "%u\n", s->cpu_partial);
4656 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4659 unsigned long objects;
4662 err = strict_strtoul(buf, 10, &objects);
4665 if (objects && kmem_cache_debug(s))
4668 s->cpu_partial = objects;
4672 SLAB_ATTR(cpu_partial);
4674 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4678 return sprintf(buf, "%pS\n", s->ctor);
4682 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4684 return sprintf(buf, "%d\n", s->refcount - 1);
4686 SLAB_ATTR_RO(aliases);
4688 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4690 return show_slab_objects(s, buf, SO_PARTIAL);
4692 SLAB_ATTR_RO(partial);
4694 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4696 return show_slab_objects(s, buf, SO_CPU);
4698 SLAB_ATTR_RO(cpu_slabs);
4700 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4702 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4704 SLAB_ATTR_RO(objects);
4706 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4708 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4710 SLAB_ATTR_RO(objects_partial);
4712 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4719 for_each_online_cpu(cpu) {
4720 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4723 pages += page->pages;
4724 objects += page->pobjects;
4728 len = sprintf(buf, "%d(%d)", objects, pages);
4731 for_each_online_cpu(cpu) {
4732 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4734 if (page && len < PAGE_SIZE - 20)
4735 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4736 page->pobjects, page->pages);
4739 return len + sprintf(buf + len, "\n");
4741 SLAB_ATTR_RO(slabs_cpu_partial);
4743 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4745 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4748 static ssize_t reclaim_account_store(struct kmem_cache *s,
4749 const char *buf, size_t length)
4751 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4753 s->flags |= SLAB_RECLAIM_ACCOUNT;
4756 SLAB_ATTR(reclaim_account);
4758 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4760 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4762 SLAB_ATTR_RO(hwcache_align);
4764 #ifdef CONFIG_ZONE_DMA
4765 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4767 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4769 SLAB_ATTR_RO(cache_dma);
4772 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4774 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4776 SLAB_ATTR_RO(destroy_by_rcu);
4778 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4780 return sprintf(buf, "%d\n", s->reserved);
4782 SLAB_ATTR_RO(reserved);
4784 #ifdef CONFIG_SLUB_DEBUG
4785 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4787 return show_slab_objects(s, buf, SO_ALL);
4789 SLAB_ATTR_RO(slabs);
4791 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4793 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4795 SLAB_ATTR_RO(total_objects);
4797 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4799 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4802 static ssize_t sanity_checks_store(struct kmem_cache *s,
4803 const char *buf, size_t length)
4805 s->flags &= ~SLAB_DEBUG_FREE;
4806 if (buf[0] == '1') {
4807 s->flags &= ~__CMPXCHG_DOUBLE;
4808 s->flags |= SLAB_DEBUG_FREE;
4812 SLAB_ATTR(sanity_checks);
4814 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4816 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4819 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4822 s->flags &= ~SLAB_TRACE;
4823 if (buf[0] == '1') {
4824 s->flags &= ~__CMPXCHG_DOUBLE;
4825 s->flags |= SLAB_TRACE;
4831 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4833 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4836 static ssize_t red_zone_store(struct kmem_cache *s,
4837 const char *buf, size_t length)
4839 if (any_slab_objects(s))
4842 s->flags &= ~SLAB_RED_ZONE;
4843 if (buf[0] == '1') {
4844 s->flags &= ~__CMPXCHG_DOUBLE;
4845 s->flags |= SLAB_RED_ZONE;
4847 calculate_sizes(s, -1);
4850 SLAB_ATTR(red_zone);
4852 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4854 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4857 static ssize_t poison_store(struct kmem_cache *s,
4858 const char *buf, size_t length)
4860 if (any_slab_objects(s))
4863 s->flags &= ~SLAB_POISON;
4864 if (buf[0] == '1') {
4865 s->flags &= ~__CMPXCHG_DOUBLE;
4866 s->flags |= SLAB_POISON;
4868 calculate_sizes(s, -1);
4873 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4875 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4878 static ssize_t store_user_store(struct kmem_cache *s,
4879 const char *buf, size_t length)
4881 if (any_slab_objects(s))
4884 s->flags &= ~SLAB_STORE_USER;
4885 if (buf[0] == '1') {
4886 s->flags &= ~__CMPXCHG_DOUBLE;
4887 s->flags |= SLAB_STORE_USER;
4889 calculate_sizes(s, -1);
4892 SLAB_ATTR(store_user);
4894 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4899 static ssize_t validate_store(struct kmem_cache *s,
4900 const char *buf, size_t length)
4904 if (buf[0] == '1') {
4905 ret = validate_slab_cache(s);
4911 SLAB_ATTR(validate);
4913 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4915 if (!(s->flags & SLAB_STORE_USER))
4917 return list_locations(s, buf, TRACK_ALLOC);
4919 SLAB_ATTR_RO(alloc_calls);
4921 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4923 if (!(s->flags & SLAB_STORE_USER))
4925 return list_locations(s, buf, TRACK_FREE);
4927 SLAB_ATTR_RO(free_calls);
4928 #endif /* CONFIG_SLUB_DEBUG */
4930 #ifdef CONFIG_FAILSLAB
4931 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4933 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4936 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4939 s->flags &= ~SLAB_FAILSLAB;
4941 s->flags |= SLAB_FAILSLAB;
4944 SLAB_ATTR(failslab);
4947 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4952 static ssize_t shrink_store(struct kmem_cache *s,
4953 const char *buf, size_t length)
4955 if (buf[0] == '1') {
4956 int rc = kmem_cache_shrink(s);
4967 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4969 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4972 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4973 const char *buf, size_t length)
4975 unsigned long ratio;
4978 err = strict_strtoul(buf, 10, &ratio);
4983 s->remote_node_defrag_ratio = ratio * 10;
4987 SLAB_ATTR(remote_node_defrag_ratio);
4990 #ifdef CONFIG_SLUB_STATS
4991 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4993 unsigned long sum = 0;
4996 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5001 for_each_online_cpu(cpu) {
5002 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5008 len = sprintf(buf, "%lu", sum);
5011 for_each_online_cpu(cpu) {
5012 if (data[cpu] && len < PAGE_SIZE - 20)
5013 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5017 return len + sprintf(buf + len, "\n");
5020 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5024 for_each_online_cpu(cpu)
5025 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5028 #define STAT_ATTR(si, text) \
5029 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5031 return show_stat(s, buf, si); \
5033 static ssize_t text##_store(struct kmem_cache *s, \
5034 const char *buf, size_t length) \
5036 if (buf[0] != '0') \
5038 clear_stat(s, si); \
5043 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5044 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5045 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5046 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5047 STAT_ATTR(FREE_FROZEN, free_frozen);
5048 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5049 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5050 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5051 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5052 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5053 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5054 STAT_ATTR(FREE_SLAB, free_slab);
5055 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5056 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5057 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5058 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5059 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5060 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5061 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5062 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5063 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5064 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5065 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5066 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5067 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5068 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5071 static struct attribute *slab_attrs[] = {
5072 &slab_size_attr.attr,
5073 &object_size_attr.attr,
5074 &objs_per_slab_attr.attr,
5076 &min_partial_attr.attr,
5077 &cpu_partial_attr.attr,
5079 &objects_partial_attr.attr,
5081 &cpu_slabs_attr.attr,
5085 &hwcache_align_attr.attr,
5086 &reclaim_account_attr.attr,
5087 &destroy_by_rcu_attr.attr,
5089 &reserved_attr.attr,
5090 &slabs_cpu_partial_attr.attr,
5091 #ifdef CONFIG_SLUB_DEBUG
5092 &total_objects_attr.attr,
5094 &sanity_checks_attr.attr,
5096 &red_zone_attr.attr,
5098 &store_user_attr.attr,
5099 &validate_attr.attr,
5100 &alloc_calls_attr.attr,
5101 &free_calls_attr.attr,
5103 #ifdef CONFIG_ZONE_DMA
5104 &cache_dma_attr.attr,
5107 &remote_node_defrag_ratio_attr.attr,
5109 #ifdef CONFIG_SLUB_STATS
5110 &alloc_fastpath_attr.attr,
5111 &alloc_slowpath_attr.attr,
5112 &free_fastpath_attr.attr,
5113 &free_slowpath_attr.attr,
5114 &free_frozen_attr.attr,
5115 &free_add_partial_attr.attr,
5116 &free_remove_partial_attr.attr,
5117 &alloc_from_partial_attr.attr,
5118 &alloc_slab_attr.attr,
5119 &alloc_refill_attr.attr,
5120 &alloc_node_mismatch_attr.attr,
5121 &free_slab_attr.attr,
5122 &cpuslab_flush_attr.attr,
5123 &deactivate_full_attr.attr,
5124 &deactivate_empty_attr.attr,
5125 &deactivate_to_head_attr.attr,
5126 &deactivate_to_tail_attr.attr,
5127 &deactivate_remote_frees_attr.attr,
5128 &deactivate_bypass_attr.attr,
5129 &order_fallback_attr.attr,
5130 &cmpxchg_double_fail_attr.attr,
5131 &cmpxchg_double_cpu_fail_attr.attr,
5132 &cpu_partial_alloc_attr.attr,
5133 &cpu_partial_free_attr.attr,
5134 &cpu_partial_node_attr.attr,
5135 &cpu_partial_drain_attr.attr,
5137 #ifdef CONFIG_FAILSLAB
5138 &failslab_attr.attr,
5144 static struct attribute_group slab_attr_group = {
5145 .attrs = slab_attrs,
5148 static ssize_t slab_attr_show(struct kobject *kobj,
5149 struct attribute *attr,
5152 struct slab_attribute *attribute;
5153 struct kmem_cache *s;
5156 attribute = to_slab_attr(attr);
5159 if (!attribute->show)
5162 err = attribute->show(s, buf);
5167 static ssize_t slab_attr_store(struct kobject *kobj,
5168 struct attribute *attr,
5169 const char *buf, size_t len)
5171 struct slab_attribute *attribute;
5172 struct kmem_cache *s;
5175 attribute = to_slab_attr(attr);
5178 if (!attribute->store)
5181 err = attribute->store(s, buf, len);
5186 static const struct sysfs_ops slab_sysfs_ops = {
5187 .show = slab_attr_show,
5188 .store = slab_attr_store,
5191 static struct kobj_type slab_ktype = {
5192 .sysfs_ops = &slab_sysfs_ops,
5195 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5197 struct kobj_type *ktype = get_ktype(kobj);
5199 if (ktype == &slab_ktype)
5204 static const struct kset_uevent_ops slab_uevent_ops = {
5205 .filter = uevent_filter,
5208 static struct kset *slab_kset;
5210 #define ID_STR_LENGTH 64
5212 /* Create a unique string id for a slab cache:
5214 * Format :[flags-]size
5216 static char *create_unique_id(struct kmem_cache *s)
5218 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5225 * First flags affecting slabcache operations. We will only
5226 * get here for aliasable slabs so we do not need to support
5227 * too many flags. The flags here must cover all flags that
5228 * are matched during merging to guarantee that the id is
5231 if (s->flags & SLAB_CACHE_DMA)
5233 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5235 if (s->flags & SLAB_DEBUG_FREE)
5237 if (!(s->flags & SLAB_NOTRACK))
5241 p += sprintf(p, "%07d", s->size);
5242 BUG_ON(p > name + ID_STR_LENGTH - 1);
5246 static int sysfs_slab_add(struct kmem_cache *s)
5252 if (slab_state < FULL)
5253 /* Defer until later */
5256 unmergeable = slab_unmergeable(s);
5259 * Slabcache can never be merged so we can use the name proper.
5260 * This is typically the case for debug situations. In that
5261 * case we can catch duplicate names easily.
5263 sysfs_remove_link(&slab_kset->kobj, s->name);
5267 * Create a unique name for the slab as a target
5270 name = create_unique_id(s);
5273 s->kobj.kset = slab_kset;
5274 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5276 kobject_put(&s->kobj);
5280 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5282 kobject_del(&s->kobj);
5283 kobject_put(&s->kobj);
5286 kobject_uevent(&s->kobj, KOBJ_ADD);
5288 /* Setup first alias */
5289 sysfs_slab_alias(s, s->name);
5295 static void sysfs_slab_remove(struct kmem_cache *s)
5297 if (slab_state < FULL)
5299 * Sysfs has not been setup yet so no need to remove the
5304 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5305 kobject_del(&s->kobj);
5306 kobject_put(&s->kobj);
5310 * Need to buffer aliases during bootup until sysfs becomes
5311 * available lest we lose that information.
5313 struct saved_alias {
5314 struct kmem_cache *s;
5316 struct saved_alias *next;
5319 static struct saved_alias *alias_list;
5321 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5323 struct saved_alias *al;
5325 if (slab_state == FULL) {
5327 * If we have a leftover link then remove it.
5329 sysfs_remove_link(&slab_kset->kobj, name);
5330 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5333 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5339 al->next = alias_list;
5344 static int __init slab_sysfs_init(void)
5346 struct kmem_cache *s;
5349 mutex_lock(&slab_mutex);
5351 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5353 mutex_unlock(&slab_mutex);
5354 printk(KERN_ERR "Cannot register slab subsystem.\n");
5360 list_for_each_entry(s, &slab_caches, list) {
5361 err = sysfs_slab_add(s);
5363 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5364 " to sysfs\n", s->name);
5367 while (alias_list) {
5368 struct saved_alias *al = alias_list;
5370 alias_list = alias_list->next;
5371 err = sysfs_slab_alias(al->s, al->name);
5373 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5374 " %s to sysfs\n", al->name);
5378 mutex_unlock(&slab_mutex);
5383 __initcall(slab_sysfs_init);
5384 #endif /* CONFIG_SYSFS */
5387 * The /proc/slabinfo ABI
5389 #ifdef CONFIG_SLABINFO
5390 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5392 unsigned long nr_partials = 0;
5393 unsigned long nr_slabs = 0;
5394 unsigned long nr_objs = 0;
5395 unsigned long nr_free = 0;
5398 for_each_online_node(node) {
5399 struct kmem_cache_node *n = get_node(s, node);
5404 nr_partials += n->nr_partial;
5405 nr_slabs += atomic_long_read(&n->nr_slabs);
5406 nr_objs += atomic_long_read(&n->total_objects);
5407 nr_free += count_partial(n, count_free);
5410 sinfo->active_objs = nr_objs - nr_free;
5411 sinfo->num_objs = nr_objs;
5412 sinfo->active_slabs = nr_slabs;
5413 sinfo->num_slabs = nr_slabs;
5414 sinfo->objects_per_slab = oo_objects(s->oo);
5415 sinfo->cache_order = oo_order(s->oo);
5418 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5422 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5423 size_t count, loff_t *ppos)
5427 #endif /* CONFIG_SLABINFO */