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 */
180 static struct notifier_block slab_notifier;
184 * Tracking user of a slab.
186 #define TRACK_ADDRS_COUNT 16
188 unsigned long addr; /* Called from address */
189 #ifdef CONFIG_STACKTRACE
190 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
192 int cpu; /* Was running on cpu */
193 int pid; /* Pid context */
194 unsigned long when; /* When did the operation occur */
197 enum track_item { TRACK_ALLOC, TRACK_FREE };
200 static int sysfs_slab_add(struct kmem_cache *);
201 static int sysfs_slab_alias(struct kmem_cache *, const char *);
202 static void sysfs_slab_remove(struct kmem_cache *);
205 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
206 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
208 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
215 __this_cpu_inc(s->cpu_slab->stat[si]);
219 /********************************************************************
220 * Core slab cache functions
221 *******************************************************************/
223 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
225 return s->node[node];
228 /* Verify that a pointer has an address that is valid within a slab page */
229 static inline int check_valid_pointer(struct kmem_cache *s,
230 struct page *page, const void *object)
237 base = page_address(page);
238 if (object < base || object >= base + page->objects * s->size ||
239 (object - base) % s->size) {
246 static inline void *get_freepointer(struct kmem_cache *s, void *object)
248 return *(void **)(object + s->offset);
251 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
253 prefetch(object + s->offset);
256 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
260 #ifdef CONFIG_DEBUG_PAGEALLOC
261 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
263 p = get_freepointer(s, object);
268 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
270 *(void **)(object + s->offset) = fp;
273 /* Loop over all objects in a slab */
274 #define for_each_object(__p, __s, __addr, __objects) \
275 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
278 /* Determine object index from a given position */
279 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
281 return (p - addr) / s->size;
284 static inline size_t slab_ksize(const struct kmem_cache *s)
286 #ifdef CONFIG_SLUB_DEBUG
288 * Debugging requires use of the padding between object
289 * and whatever may come after it.
291 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
292 return s->object_size;
296 * If we have the need to store the freelist pointer
297 * back there or track user information then we can
298 * only use the space before that information.
300 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
303 * Else we can use all the padding etc for the allocation
308 static inline int order_objects(int order, unsigned long size, int reserved)
310 return ((PAGE_SIZE << order) - reserved) / size;
313 static inline struct kmem_cache_order_objects oo_make(int order,
314 unsigned long size, int reserved)
316 struct kmem_cache_order_objects x = {
317 (order << OO_SHIFT) + order_objects(order, size, reserved)
323 static inline int oo_order(struct kmem_cache_order_objects x)
325 return x.x >> OO_SHIFT;
328 static inline int oo_objects(struct kmem_cache_order_objects x)
330 return x.x & OO_MASK;
334 * Per slab locking using the pagelock
336 static __always_inline void slab_lock(struct page *page)
338 bit_spin_lock(PG_locked, &page->flags);
341 static __always_inline void slab_unlock(struct page *page)
343 __bit_spin_unlock(PG_locked, &page->flags);
346 /* Interrupts must be disabled (for the fallback code to work right) */
347 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
348 void *freelist_old, unsigned long counters_old,
349 void *freelist_new, unsigned long counters_new,
352 VM_BUG_ON(!irqs_disabled());
353 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
354 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
355 if (s->flags & __CMPXCHG_DOUBLE) {
356 if (cmpxchg_double(&page->freelist, &page->counters,
357 freelist_old, counters_old,
358 freelist_new, counters_new))
364 if (page->freelist == freelist_old && page->counters == counters_old) {
365 page->freelist = freelist_new;
366 page->counters = counters_new;
374 stat(s, CMPXCHG_DOUBLE_FAIL);
376 #ifdef SLUB_DEBUG_CMPXCHG
377 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
383 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
384 void *freelist_old, unsigned long counters_old,
385 void *freelist_new, unsigned long counters_new,
388 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
389 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
390 if (s->flags & __CMPXCHG_DOUBLE) {
391 if (cmpxchg_double(&page->freelist, &page->counters,
392 freelist_old, counters_old,
393 freelist_new, counters_new))
400 local_irq_save(flags);
402 if (page->freelist == freelist_old && page->counters == counters_old) {
403 page->freelist = freelist_new;
404 page->counters = counters_new;
406 local_irq_restore(flags);
410 local_irq_restore(flags);
414 stat(s, CMPXCHG_DOUBLE_FAIL);
416 #ifdef SLUB_DEBUG_CMPXCHG
417 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
423 #ifdef CONFIG_SLUB_DEBUG
425 * Determine a map of object in use on a page.
427 * Node listlock must be held to guarantee that the page does
428 * not vanish from under us.
430 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
433 void *addr = page_address(page);
435 for (p = page->freelist; p; p = get_freepointer(s, p))
436 set_bit(slab_index(p, s, addr), map);
442 #ifdef CONFIG_SLUB_DEBUG_ON
443 static int slub_debug = DEBUG_DEFAULT_FLAGS;
445 static int slub_debug;
448 static char *slub_debug_slabs;
449 static int disable_higher_order_debug;
454 static void print_section(char *text, u8 *addr, unsigned int length)
456 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
460 static struct track *get_track(struct kmem_cache *s, void *object,
461 enum track_item alloc)
466 p = object + s->offset + sizeof(void *);
468 p = object + s->inuse;
473 static void set_track(struct kmem_cache *s, void *object,
474 enum track_item alloc, unsigned long addr)
476 struct track *p = get_track(s, object, alloc);
479 #ifdef CONFIG_STACKTRACE
480 struct stack_trace trace;
483 trace.nr_entries = 0;
484 trace.max_entries = TRACK_ADDRS_COUNT;
485 trace.entries = p->addrs;
487 save_stack_trace(&trace);
489 /* See rant in lockdep.c */
490 if (trace.nr_entries != 0 &&
491 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
494 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
498 p->cpu = smp_processor_id();
499 p->pid = current->pid;
502 memset(p, 0, sizeof(struct track));
505 static void init_tracking(struct kmem_cache *s, void *object)
507 if (!(s->flags & SLAB_STORE_USER))
510 set_track(s, object, TRACK_FREE, 0UL);
511 set_track(s, object, TRACK_ALLOC, 0UL);
514 static void print_track(const char *s, struct track *t)
519 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
520 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
521 #ifdef CONFIG_STACKTRACE
524 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
526 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
533 static void print_tracking(struct kmem_cache *s, void *object)
535 if (!(s->flags & SLAB_STORE_USER))
538 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
539 print_track("Freed", get_track(s, object, TRACK_FREE));
542 static void print_page_info(struct page *page)
544 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
545 page, page->objects, page->inuse, page->freelist, page->flags);
549 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
555 vsnprintf(buf, sizeof(buf), fmt, args);
557 printk(KERN_ERR "========================================"
558 "=====================================\n");
559 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
560 printk(KERN_ERR "----------------------------------------"
561 "-------------------------------------\n\n");
563 add_taint(TAINT_BAD_PAGE);
566 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
572 vsnprintf(buf, sizeof(buf), fmt, args);
574 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
577 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
579 unsigned int off; /* Offset of last byte */
580 u8 *addr = page_address(page);
582 print_tracking(s, p);
584 print_page_info(page);
586 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
587 p, p - addr, get_freepointer(s, p));
590 print_section("Bytes b4 ", p - 16, 16);
592 print_section("Object ", p, min_t(unsigned long, s->object_size,
594 if (s->flags & SLAB_RED_ZONE)
595 print_section("Redzone ", p + s->object_size,
596 s->inuse - s->object_size);
599 off = s->offset + sizeof(void *);
603 if (s->flags & SLAB_STORE_USER)
604 off += 2 * sizeof(struct track);
607 /* Beginning of the filler is the free pointer */
608 print_section("Padding ", p + off, s->size - off);
613 static void object_err(struct kmem_cache *s, struct page *page,
614 u8 *object, char *reason)
616 slab_bug(s, "%s", reason);
617 print_trailer(s, page, object);
620 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
626 vsnprintf(buf, sizeof(buf), fmt, args);
628 slab_bug(s, "%s", buf);
629 print_page_info(page);
633 static void init_object(struct kmem_cache *s, void *object, u8 val)
637 if (s->flags & __OBJECT_POISON) {
638 memset(p, POISON_FREE, s->object_size - 1);
639 p[s->object_size - 1] = POISON_END;
642 if (s->flags & SLAB_RED_ZONE)
643 memset(p + s->object_size, val, s->inuse - s->object_size);
646 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
647 void *from, void *to)
649 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
650 memset(from, data, to - from);
653 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
654 u8 *object, char *what,
655 u8 *start, unsigned int value, unsigned int bytes)
660 fault = memchr_inv(start, value, bytes);
665 while (end > fault && end[-1] == value)
668 slab_bug(s, "%s overwritten", what);
669 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
670 fault, end - 1, fault[0], value);
671 print_trailer(s, page, object);
673 restore_bytes(s, what, value, fault, end);
681 * Bytes of the object to be managed.
682 * If the freepointer may overlay the object then the free
683 * pointer is the first word of the object.
685 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
688 * object + s->object_size
689 * Padding to reach word boundary. This is also used for Redzoning.
690 * Padding is extended by another word if Redzoning is enabled and
691 * object_size == inuse.
693 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
694 * 0xcc (RED_ACTIVE) for objects in use.
697 * Meta data starts here.
699 * A. Free pointer (if we cannot overwrite object on free)
700 * B. Tracking data for SLAB_STORE_USER
701 * C. Padding to reach required alignment boundary or at mininum
702 * one word if debugging is on to be able to detect writes
703 * before the word boundary.
705 * Padding is done using 0x5a (POISON_INUSE)
708 * Nothing is used beyond s->size.
710 * If slabcaches are merged then the object_size and inuse boundaries are mostly
711 * ignored. And therefore no slab options that rely on these boundaries
712 * may be used with merged slabcaches.
715 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
717 unsigned long off = s->inuse; /* The end of info */
720 /* Freepointer is placed after the object. */
721 off += sizeof(void *);
723 if (s->flags & SLAB_STORE_USER)
724 /* We also have user information there */
725 off += 2 * sizeof(struct track);
730 return check_bytes_and_report(s, page, p, "Object padding",
731 p + off, POISON_INUSE, s->size - off);
734 /* Check the pad bytes at the end of a slab page */
735 static int slab_pad_check(struct kmem_cache *s, struct page *page)
743 if (!(s->flags & SLAB_POISON))
746 start = page_address(page);
747 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
748 end = start + length;
749 remainder = length % s->size;
753 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
756 while (end > fault && end[-1] == POISON_INUSE)
759 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
760 print_section("Padding ", end - remainder, remainder);
762 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
766 static int check_object(struct kmem_cache *s, struct page *page,
767 void *object, u8 val)
770 u8 *endobject = object + s->object_size;
772 if (s->flags & SLAB_RED_ZONE) {
773 if (!check_bytes_and_report(s, page, object, "Redzone",
774 endobject, val, s->inuse - s->object_size))
777 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
778 check_bytes_and_report(s, page, p, "Alignment padding",
779 endobject, POISON_INUSE, s->inuse - s->object_size);
783 if (s->flags & SLAB_POISON) {
784 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
785 (!check_bytes_and_report(s, page, p, "Poison", p,
786 POISON_FREE, s->object_size - 1) ||
787 !check_bytes_and_report(s, page, p, "Poison",
788 p + s->object_size - 1, POISON_END, 1)))
791 * check_pad_bytes cleans up on its own.
793 check_pad_bytes(s, page, p);
796 if (!s->offset && val == SLUB_RED_ACTIVE)
798 * Object and freepointer overlap. Cannot check
799 * freepointer while object is allocated.
803 /* Check free pointer validity */
804 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
805 object_err(s, page, p, "Freepointer corrupt");
807 * No choice but to zap it and thus lose the remainder
808 * of the free objects in this slab. May cause
809 * another error because the object count is now wrong.
811 set_freepointer(s, p, NULL);
817 static int check_slab(struct kmem_cache *s, struct page *page)
821 VM_BUG_ON(!irqs_disabled());
823 if (!PageSlab(page)) {
824 slab_err(s, page, "Not a valid slab page");
828 maxobj = order_objects(compound_order(page), s->size, s->reserved);
829 if (page->objects > maxobj) {
830 slab_err(s, page, "objects %u > max %u",
831 s->name, page->objects, maxobj);
834 if (page->inuse > page->objects) {
835 slab_err(s, page, "inuse %u > max %u",
836 s->name, page->inuse, page->objects);
839 /* Slab_pad_check fixes things up after itself */
840 slab_pad_check(s, page);
845 * Determine if a certain object on a page is on the freelist. Must hold the
846 * slab lock to guarantee that the chains are in a consistent state.
848 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
853 unsigned long max_objects;
856 while (fp && nr <= page->objects) {
859 if (!check_valid_pointer(s, page, fp)) {
861 object_err(s, page, object,
862 "Freechain corrupt");
863 set_freepointer(s, object, NULL);
866 slab_err(s, page, "Freepointer corrupt");
867 page->freelist = NULL;
868 page->inuse = page->objects;
869 slab_fix(s, "Freelist cleared");
875 fp = get_freepointer(s, object);
879 max_objects = order_objects(compound_order(page), s->size, s->reserved);
880 if (max_objects > MAX_OBJS_PER_PAGE)
881 max_objects = MAX_OBJS_PER_PAGE;
883 if (page->objects != max_objects) {
884 slab_err(s, page, "Wrong number of objects. Found %d but "
885 "should be %d", page->objects, max_objects);
886 page->objects = max_objects;
887 slab_fix(s, "Number of objects adjusted.");
889 if (page->inuse != page->objects - nr) {
890 slab_err(s, page, "Wrong object count. Counter is %d but "
891 "counted were %d", page->inuse, page->objects - nr);
892 page->inuse = page->objects - nr;
893 slab_fix(s, "Object count adjusted.");
895 return search == NULL;
898 static void trace(struct kmem_cache *s, struct page *page, void *object,
901 if (s->flags & SLAB_TRACE) {
902 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
904 alloc ? "alloc" : "free",
909 print_section("Object ", (void *)object, s->object_size);
916 * Hooks for other subsystems that check memory allocations. In a typical
917 * production configuration these hooks all should produce no code at all.
919 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
921 flags &= gfp_allowed_mask;
922 lockdep_trace_alloc(flags);
923 might_sleep_if(flags & __GFP_WAIT);
925 return should_failslab(s->object_size, flags, s->flags);
928 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
930 flags &= gfp_allowed_mask;
931 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
932 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
935 static inline void slab_free_hook(struct kmem_cache *s, void *x)
937 kmemleak_free_recursive(x, s->flags);
940 * Trouble is that we may no longer disable interupts in the fast path
941 * So in order to make the debug calls that expect irqs to be
942 * disabled we need to disable interrupts temporarily.
944 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
948 local_irq_save(flags);
949 kmemcheck_slab_free(s, x, s->object_size);
950 debug_check_no_locks_freed(x, s->object_size);
951 local_irq_restore(flags);
954 if (!(s->flags & SLAB_DEBUG_OBJECTS))
955 debug_check_no_obj_freed(x, s->object_size);
959 * Tracking of fully allocated slabs for debugging purposes.
961 * list_lock must be held.
963 static void add_full(struct kmem_cache *s,
964 struct kmem_cache_node *n, struct page *page)
966 if (!(s->flags & SLAB_STORE_USER))
969 list_add(&page->lru, &n->full);
973 * list_lock must be held.
975 static void remove_full(struct kmem_cache *s, struct page *page)
977 if (!(s->flags & SLAB_STORE_USER))
980 list_del(&page->lru);
983 /* Tracking of the number of slabs for debugging purposes */
984 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
986 struct kmem_cache_node *n = get_node(s, node);
988 return atomic_long_read(&n->nr_slabs);
991 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
993 return atomic_long_read(&n->nr_slabs);
996 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
998 struct kmem_cache_node *n = get_node(s, node);
1001 * May be called early in order to allocate a slab for the
1002 * kmem_cache_node structure. Solve the chicken-egg
1003 * dilemma by deferring the increment of the count during
1004 * bootstrap (see early_kmem_cache_node_alloc).
1007 atomic_long_inc(&n->nr_slabs);
1008 atomic_long_add(objects, &n->total_objects);
1011 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1013 struct kmem_cache_node *n = get_node(s, node);
1015 atomic_long_dec(&n->nr_slabs);
1016 atomic_long_sub(objects, &n->total_objects);
1019 /* Object debug checks for alloc/free paths */
1020 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1023 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1026 init_object(s, object, SLUB_RED_INACTIVE);
1027 init_tracking(s, object);
1030 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1031 void *object, unsigned long addr)
1033 if (!check_slab(s, page))
1036 if (!check_valid_pointer(s, page, object)) {
1037 object_err(s, page, object, "Freelist Pointer check fails");
1041 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1044 /* Success perform special debug activities for allocs */
1045 if (s->flags & SLAB_STORE_USER)
1046 set_track(s, object, TRACK_ALLOC, addr);
1047 trace(s, page, object, 1);
1048 init_object(s, object, SLUB_RED_ACTIVE);
1052 if (PageSlab(page)) {
1054 * If this is a slab page then lets do the best we can
1055 * to avoid issues in the future. Marking all objects
1056 * as used avoids touching the remaining objects.
1058 slab_fix(s, "Marking all objects used");
1059 page->inuse = page->objects;
1060 page->freelist = NULL;
1065 static noinline struct kmem_cache_node *free_debug_processing(
1066 struct kmem_cache *s, struct page *page, void *object,
1067 unsigned long addr, unsigned long *flags)
1069 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1071 spin_lock_irqsave(&n->list_lock, *flags);
1074 if (!check_slab(s, page))
1077 if (!check_valid_pointer(s, page, object)) {
1078 slab_err(s, page, "Invalid object pointer 0x%p", object);
1082 if (on_freelist(s, page, object)) {
1083 object_err(s, page, object, "Object already free");
1087 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1090 if (unlikely(s != page->slab_cache)) {
1091 if (!PageSlab(page)) {
1092 slab_err(s, page, "Attempt to free object(0x%p) "
1093 "outside of slab", object);
1094 } else if (!page->slab_cache) {
1096 "SLUB <none>: no slab for object 0x%p.\n",
1100 object_err(s, page, object,
1101 "page slab pointer corrupt.");
1105 if (s->flags & SLAB_STORE_USER)
1106 set_track(s, object, TRACK_FREE, addr);
1107 trace(s, page, object, 0);
1108 init_object(s, object, SLUB_RED_INACTIVE);
1112 * Keep node_lock to preserve integrity
1113 * until the object is actually freed
1119 spin_unlock_irqrestore(&n->list_lock, *flags);
1120 slab_fix(s, "Object at 0x%p not freed", object);
1124 static int __init setup_slub_debug(char *str)
1126 slub_debug = DEBUG_DEFAULT_FLAGS;
1127 if (*str++ != '=' || !*str)
1129 * No options specified. Switch on full debugging.
1135 * No options but restriction on slabs. This means full
1136 * debugging for slabs matching a pattern.
1140 if (tolower(*str) == 'o') {
1142 * Avoid enabling debugging on caches if its minimum order
1143 * would increase as a result.
1145 disable_higher_order_debug = 1;
1152 * Switch off all debugging measures.
1157 * Determine which debug features should be switched on
1159 for (; *str && *str != ','; str++) {
1160 switch (tolower(*str)) {
1162 slub_debug |= SLAB_DEBUG_FREE;
1165 slub_debug |= SLAB_RED_ZONE;
1168 slub_debug |= SLAB_POISON;
1171 slub_debug |= SLAB_STORE_USER;
1174 slub_debug |= SLAB_TRACE;
1177 slub_debug |= SLAB_FAILSLAB;
1180 printk(KERN_ERR "slub_debug option '%c' "
1181 "unknown. skipped\n", *str);
1187 slub_debug_slabs = str + 1;
1192 __setup("slub_debug", setup_slub_debug);
1194 static unsigned long kmem_cache_flags(unsigned long object_size,
1195 unsigned long flags, const char *name,
1196 void (*ctor)(void *))
1199 * Enable debugging if selected on the kernel commandline.
1201 if (slub_debug && (!slub_debug_slabs ||
1202 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1203 flags |= slub_debug;
1208 static inline void setup_object_debug(struct kmem_cache *s,
1209 struct page *page, void *object) {}
1211 static inline int alloc_debug_processing(struct kmem_cache *s,
1212 struct page *page, void *object, unsigned long addr) { return 0; }
1214 static inline struct kmem_cache_node *free_debug_processing(
1215 struct kmem_cache *s, struct page *page, void *object,
1216 unsigned long addr, unsigned long *flags) { return NULL; }
1218 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1220 static inline int check_object(struct kmem_cache *s, struct page *page,
1221 void *object, u8 val) { return 1; }
1222 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1223 struct page *page) {}
1224 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1225 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1226 unsigned long flags, const char *name,
1227 void (*ctor)(void *))
1231 #define slub_debug 0
1233 #define disable_higher_order_debug 0
1235 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1237 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1239 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1241 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1244 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1247 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1250 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1252 #endif /* CONFIG_SLUB_DEBUG */
1255 * Slab allocation and freeing
1257 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1258 struct kmem_cache_order_objects oo)
1260 int order = oo_order(oo);
1262 flags |= __GFP_NOTRACK;
1264 if (node == NUMA_NO_NODE)
1265 return alloc_pages(flags, order);
1267 return alloc_pages_exact_node(node, flags, order);
1270 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1273 struct kmem_cache_order_objects oo = s->oo;
1276 flags &= gfp_allowed_mask;
1278 if (flags & __GFP_WAIT)
1281 flags |= s->allocflags;
1284 * Let the initial higher-order allocation fail under memory pressure
1285 * so we fall-back to the minimum order allocation.
1287 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1289 page = alloc_slab_page(alloc_gfp, node, oo);
1290 if (unlikely(!page)) {
1293 * Allocation may have failed due to fragmentation.
1294 * Try a lower order alloc if possible
1296 page = alloc_slab_page(flags, node, oo);
1299 stat(s, ORDER_FALLBACK);
1302 if (kmemcheck_enabled && page
1303 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1304 int pages = 1 << oo_order(oo);
1306 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1309 * Objects from caches that have a constructor don't get
1310 * cleared when they're allocated, so we need to do it here.
1313 kmemcheck_mark_uninitialized_pages(page, pages);
1315 kmemcheck_mark_unallocated_pages(page, pages);
1318 if (flags & __GFP_WAIT)
1319 local_irq_disable();
1323 page->objects = oo_objects(oo);
1324 mod_zone_page_state(page_zone(page),
1325 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1326 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1332 static void setup_object(struct kmem_cache *s, struct page *page,
1335 setup_object_debug(s, page, object);
1336 if (unlikely(s->ctor))
1340 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1347 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1349 page = allocate_slab(s,
1350 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1354 inc_slabs_node(s, page_to_nid(page), page->objects);
1355 page->slab_cache = s;
1356 __SetPageSlab(page);
1357 if (page->pfmemalloc)
1358 SetPageSlabPfmemalloc(page);
1360 start = page_address(page);
1362 if (unlikely(s->flags & SLAB_POISON))
1363 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1366 for_each_object(p, s, start, page->objects) {
1367 setup_object(s, page, last);
1368 set_freepointer(s, last, p);
1371 setup_object(s, page, last);
1372 set_freepointer(s, last, NULL);
1374 page->freelist = start;
1375 page->inuse = page->objects;
1381 static void __free_slab(struct kmem_cache *s, struct page *page)
1383 int order = compound_order(page);
1384 int pages = 1 << order;
1386 if (kmem_cache_debug(s)) {
1389 slab_pad_check(s, page);
1390 for_each_object(p, s, page_address(page),
1392 check_object(s, page, p, SLUB_RED_INACTIVE);
1395 kmemcheck_free_shadow(page, compound_order(page));
1397 mod_zone_page_state(page_zone(page),
1398 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1399 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1402 __ClearPageSlabPfmemalloc(page);
1403 __ClearPageSlab(page);
1404 reset_page_mapcount(page);
1405 if (current->reclaim_state)
1406 current->reclaim_state->reclaimed_slab += pages;
1407 __free_pages(page, order);
1410 #define need_reserve_slab_rcu \
1411 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1413 static void rcu_free_slab(struct rcu_head *h)
1417 if (need_reserve_slab_rcu)
1418 page = virt_to_head_page(h);
1420 page = container_of((struct list_head *)h, struct page, lru);
1422 __free_slab(page->slab_cache, page);
1425 static void free_slab(struct kmem_cache *s, struct page *page)
1427 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1428 struct rcu_head *head;
1430 if (need_reserve_slab_rcu) {
1431 int order = compound_order(page);
1432 int offset = (PAGE_SIZE << order) - s->reserved;
1434 VM_BUG_ON(s->reserved != sizeof(*head));
1435 head = page_address(page) + offset;
1438 * RCU free overloads the RCU head over the LRU
1440 head = (void *)&page->lru;
1443 call_rcu(head, rcu_free_slab);
1445 __free_slab(s, page);
1448 static void discard_slab(struct kmem_cache *s, struct page *page)
1450 dec_slabs_node(s, page_to_nid(page), page->objects);
1455 * Management of partially allocated slabs.
1457 * list_lock must be held.
1459 static inline void add_partial(struct kmem_cache_node *n,
1460 struct page *page, int tail)
1463 if (tail == DEACTIVATE_TO_TAIL)
1464 list_add_tail(&page->lru, &n->partial);
1466 list_add(&page->lru, &n->partial);
1470 * list_lock must be held.
1472 static inline void remove_partial(struct kmem_cache_node *n,
1475 list_del(&page->lru);
1480 * Remove slab from the partial list, freeze it and
1481 * return the pointer to the freelist.
1483 * Returns a list of objects or NULL if it fails.
1485 * Must hold list_lock since we modify the partial list.
1487 static inline void *acquire_slab(struct kmem_cache *s,
1488 struct kmem_cache_node *n, struct page *page,
1492 unsigned long counters;
1496 * Zap the freelist and set the frozen bit.
1497 * The old freelist is the list of objects for the
1498 * per cpu allocation list.
1500 freelist = page->freelist;
1501 counters = page->counters;
1502 new.counters = counters;
1504 new.inuse = page->objects;
1505 new.freelist = NULL;
1507 new.freelist = freelist;
1510 VM_BUG_ON(new.frozen);
1513 if (!__cmpxchg_double_slab(s, page,
1515 new.freelist, new.counters,
1519 remove_partial(n, page);
1524 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1525 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1528 * Try to allocate a partial slab from a specific node.
1530 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1531 struct kmem_cache_cpu *c, gfp_t flags)
1533 struct page *page, *page2;
1534 void *object = NULL;
1537 * Racy check. If we mistakenly see no partial slabs then we
1538 * just allocate an empty slab. If we mistakenly try to get a
1539 * partial slab and there is none available then get_partials()
1542 if (!n || !n->nr_partial)
1545 spin_lock(&n->list_lock);
1546 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1550 if (!pfmemalloc_match(page, flags))
1553 t = acquire_slab(s, n, page, object == NULL);
1559 stat(s, ALLOC_FROM_PARTIAL);
1561 available = page->objects - page->inuse;
1563 available = put_cpu_partial(s, page, 0);
1564 stat(s, CPU_PARTIAL_NODE);
1566 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1570 spin_unlock(&n->list_lock);
1575 * Get a page from somewhere. Search in increasing NUMA distances.
1577 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1578 struct kmem_cache_cpu *c)
1581 struct zonelist *zonelist;
1584 enum zone_type high_zoneidx = gfp_zone(flags);
1586 unsigned int cpuset_mems_cookie;
1589 * The defrag ratio allows a configuration of the tradeoffs between
1590 * inter node defragmentation and node local allocations. A lower
1591 * defrag_ratio increases the tendency to do local allocations
1592 * instead of attempting to obtain partial slabs from other nodes.
1594 * If the defrag_ratio is set to 0 then kmalloc() always
1595 * returns node local objects. If the ratio is higher then kmalloc()
1596 * may return off node objects because partial slabs are obtained
1597 * from other nodes and filled up.
1599 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1600 * defrag_ratio = 1000) then every (well almost) allocation will
1601 * first attempt to defrag slab caches on other nodes. This means
1602 * scanning over all nodes to look for partial slabs which may be
1603 * expensive if we do it every time we are trying to find a slab
1604 * with available objects.
1606 if (!s->remote_node_defrag_ratio ||
1607 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1611 cpuset_mems_cookie = get_mems_allowed();
1612 zonelist = node_zonelist(slab_node(), flags);
1613 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1614 struct kmem_cache_node *n;
1616 n = get_node(s, zone_to_nid(zone));
1618 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1619 n->nr_partial > s->min_partial) {
1620 object = get_partial_node(s, n, c, flags);
1623 * Return the object even if
1624 * put_mems_allowed indicated that
1625 * the cpuset mems_allowed was
1626 * updated in parallel. It's a
1627 * harmless race between the alloc
1628 * and the cpuset update.
1630 put_mems_allowed(cpuset_mems_cookie);
1635 } while (!put_mems_allowed(cpuset_mems_cookie));
1641 * Get a partial page, lock it and return it.
1643 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1644 struct kmem_cache_cpu *c)
1647 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1649 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1650 if (object || node != NUMA_NO_NODE)
1653 return get_any_partial(s, flags, c);
1656 #ifdef CONFIG_PREEMPT
1658 * Calculate the next globally unique transaction for disambiguiation
1659 * during cmpxchg. The transactions start with the cpu number and are then
1660 * incremented by CONFIG_NR_CPUS.
1662 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1665 * No preemption supported therefore also no need to check for
1671 static inline unsigned long next_tid(unsigned long tid)
1673 return tid + TID_STEP;
1676 static inline unsigned int tid_to_cpu(unsigned long tid)
1678 return tid % TID_STEP;
1681 static inline unsigned long tid_to_event(unsigned long tid)
1683 return tid / TID_STEP;
1686 static inline unsigned int init_tid(int cpu)
1691 static inline void note_cmpxchg_failure(const char *n,
1692 const struct kmem_cache *s, unsigned long tid)
1694 #ifdef SLUB_DEBUG_CMPXCHG
1695 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1697 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1699 #ifdef CONFIG_PREEMPT
1700 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1701 printk("due to cpu change %d -> %d\n",
1702 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1705 if (tid_to_event(tid) != tid_to_event(actual_tid))
1706 printk("due to cpu running other code. Event %ld->%ld\n",
1707 tid_to_event(tid), tid_to_event(actual_tid));
1709 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1710 actual_tid, tid, next_tid(tid));
1712 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1715 static void init_kmem_cache_cpus(struct kmem_cache *s)
1719 for_each_possible_cpu(cpu)
1720 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1724 * Remove the cpu slab
1726 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1728 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1729 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1731 enum slab_modes l = M_NONE, m = M_NONE;
1733 int tail = DEACTIVATE_TO_HEAD;
1737 if (page->freelist) {
1738 stat(s, DEACTIVATE_REMOTE_FREES);
1739 tail = DEACTIVATE_TO_TAIL;
1743 * Stage one: Free all available per cpu objects back
1744 * to the page freelist while it is still frozen. Leave the
1747 * There is no need to take the list->lock because the page
1750 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1752 unsigned long counters;
1755 prior = page->freelist;
1756 counters = page->counters;
1757 set_freepointer(s, freelist, prior);
1758 new.counters = counters;
1760 VM_BUG_ON(!new.frozen);
1762 } while (!__cmpxchg_double_slab(s, page,
1764 freelist, new.counters,
1765 "drain percpu freelist"));
1767 freelist = nextfree;
1771 * Stage two: Ensure that the page is unfrozen while the
1772 * list presence reflects the actual number of objects
1775 * We setup the list membership and then perform a cmpxchg
1776 * with the count. If there is a mismatch then the page
1777 * is not unfrozen but the page is on the wrong list.
1779 * Then we restart the process which may have to remove
1780 * the page from the list that we just put it on again
1781 * because the number of objects in the slab may have
1786 old.freelist = page->freelist;
1787 old.counters = page->counters;
1788 VM_BUG_ON(!old.frozen);
1790 /* Determine target state of the slab */
1791 new.counters = old.counters;
1794 set_freepointer(s, freelist, old.freelist);
1795 new.freelist = freelist;
1797 new.freelist = old.freelist;
1801 if (!new.inuse && n->nr_partial > s->min_partial)
1803 else if (new.freelist) {
1808 * Taking the spinlock removes the possiblity
1809 * that acquire_slab() will see a slab page that
1812 spin_lock(&n->list_lock);
1816 if (kmem_cache_debug(s) && !lock) {
1819 * This also ensures that the scanning of full
1820 * slabs from diagnostic functions will not see
1823 spin_lock(&n->list_lock);
1831 remove_partial(n, page);
1833 else if (l == M_FULL)
1835 remove_full(s, page);
1837 if (m == M_PARTIAL) {
1839 add_partial(n, page, tail);
1842 } else if (m == M_FULL) {
1844 stat(s, DEACTIVATE_FULL);
1845 add_full(s, n, page);
1851 if (!__cmpxchg_double_slab(s, page,
1852 old.freelist, old.counters,
1853 new.freelist, new.counters,
1858 spin_unlock(&n->list_lock);
1861 stat(s, DEACTIVATE_EMPTY);
1862 discard_slab(s, page);
1868 * Unfreeze all the cpu partial slabs.
1870 * This function must be called with interrupts disabled
1871 * for the cpu using c (or some other guarantee must be there
1872 * to guarantee no concurrent accesses).
1874 static void unfreeze_partials(struct kmem_cache *s,
1875 struct kmem_cache_cpu *c)
1877 struct kmem_cache_node *n = NULL, *n2 = NULL;
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, this_cpu_ptr(s->cpu_slab));
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, c);
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 init_kmem_cache_node(struct kmem_cache_node *n)
2767 spin_lock_init(&n->list_lock);
2768 INIT_LIST_HEAD(&n->partial);
2769 #ifdef CONFIG_SLUB_DEBUG
2770 atomic_long_set(&n->nr_slabs, 0);
2771 atomic_long_set(&n->total_objects, 0);
2772 INIT_LIST_HEAD(&n->full);
2776 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2778 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2779 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2782 * Must align to double word boundary for the double cmpxchg
2783 * instructions to work; see __pcpu_double_call_return_bool().
2785 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2786 2 * sizeof(void *));
2791 init_kmem_cache_cpus(s);
2796 static struct kmem_cache *kmem_cache_node;
2799 * No kmalloc_node yet so do it by hand. We know that this is the first
2800 * slab on the node for this slabcache. There are no concurrent accesses
2803 * Note that this function only works on the kmalloc_node_cache
2804 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2805 * memory on a fresh node that has no slab structures yet.
2807 static void early_kmem_cache_node_alloc(int node)
2810 struct kmem_cache_node *n;
2812 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2814 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2817 if (page_to_nid(page) != node) {
2818 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2820 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2821 "in order to be able to continue\n");
2826 page->freelist = get_freepointer(kmem_cache_node, n);
2829 kmem_cache_node->node[node] = n;
2830 #ifdef CONFIG_SLUB_DEBUG
2831 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2832 init_tracking(kmem_cache_node, n);
2834 init_kmem_cache_node(n);
2835 inc_slabs_node(kmem_cache_node, node, page->objects);
2837 add_partial(n, page, DEACTIVATE_TO_HEAD);
2840 static void free_kmem_cache_nodes(struct kmem_cache *s)
2844 for_each_node_state(node, N_NORMAL_MEMORY) {
2845 struct kmem_cache_node *n = s->node[node];
2848 kmem_cache_free(kmem_cache_node, n);
2850 s->node[node] = NULL;
2854 static int init_kmem_cache_nodes(struct kmem_cache *s)
2858 for_each_node_state(node, N_NORMAL_MEMORY) {
2859 struct kmem_cache_node *n;
2861 if (slab_state == DOWN) {
2862 early_kmem_cache_node_alloc(node);
2865 n = kmem_cache_alloc_node(kmem_cache_node,
2869 free_kmem_cache_nodes(s);
2874 init_kmem_cache_node(n);
2879 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2881 if (min < MIN_PARTIAL)
2883 else if (min > MAX_PARTIAL)
2885 s->min_partial = min;
2889 * calculate_sizes() determines the order and the distribution of data within
2892 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2894 unsigned long flags = s->flags;
2895 unsigned long size = s->object_size;
2899 * Round up object size to the next word boundary. We can only
2900 * place the free pointer at word boundaries and this determines
2901 * the possible location of the free pointer.
2903 size = ALIGN(size, sizeof(void *));
2905 #ifdef CONFIG_SLUB_DEBUG
2907 * Determine if we can poison the object itself. If the user of
2908 * the slab may touch the object after free or before allocation
2909 * then we should never poison the object itself.
2911 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2913 s->flags |= __OBJECT_POISON;
2915 s->flags &= ~__OBJECT_POISON;
2919 * If we are Redzoning then check if there is some space between the
2920 * end of the object and the free pointer. If not then add an
2921 * additional word to have some bytes to store Redzone information.
2923 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2924 size += sizeof(void *);
2928 * With that we have determined the number of bytes in actual use
2929 * by the object. This is the potential offset to the free pointer.
2933 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2936 * Relocate free pointer after the object if it is not
2937 * permitted to overwrite the first word of the object on
2940 * This is the case if we do RCU, have a constructor or
2941 * destructor or are poisoning the objects.
2944 size += sizeof(void *);
2947 #ifdef CONFIG_SLUB_DEBUG
2948 if (flags & SLAB_STORE_USER)
2950 * Need to store information about allocs and frees after
2953 size += 2 * sizeof(struct track);
2955 if (flags & SLAB_RED_ZONE)
2957 * Add some empty padding so that we can catch
2958 * overwrites from earlier objects rather than let
2959 * tracking information or the free pointer be
2960 * corrupted if a user writes before the start
2963 size += sizeof(void *);
2967 * SLUB stores one object immediately after another beginning from
2968 * offset 0. In order to align the objects we have to simply size
2969 * each object to conform to the alignment.
2971 size = ALIGN(size, s->align);
2973 if (forced_order >= 0)
2974 order = forced_order;
2976 order = calculate_order(size, s->reserved);
2983 s->allocflags |= __GFP_COMP;
2985 if (s->flags & SLAB_CACHE_DMA)
2986 s->allocflags |= SLUB_DMA;
2988 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2989 s->allocflags |= __GFP_RECLAIMABLE;
2992 * Determine the number of objects per slab
2994 s->oo = oo_make(order, size, s->reserved);
2995 s->min = oo_make(get_order(size), size, s->reserved);
2996 if (oo_objects(s->oo) > oo_objects(s->max))
2999 return !!oo_objects(s->oo);
3002 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3004 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3007 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3008 s->reserved = sizeof(struct rcu_head);
3010 if (!calculate_sizes(s, -1))
3012 if (disable_higher_order_debug) {
3014 * Disable debugging flags that store metadata if the min slab
3017 if (get_order(s->size) > get_order(s->object_size)) {
3018 s->flags &= ~DEBUG_METADATA_FLAGS;
3020 if (!calculate_sizes(s, -1))
3025 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3026 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3027 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3028 /* Enable fast mode */
3029 s->flags |= __CMPXCHG_DOUBLE;
3033 * The larger the object size is, the more pages we want on the partial
3034 * list to avoid pounding the page allocator excessively.
3036 set_min_partial(s, ilog2(s->size) / 2);
3039 * cpu_partial determined the maximum number of objects kept in the
3040 * per cpu partial lists of a processor.
3042 * Per cpu partial lists mainly contain slabs that just have one
3043 * object freed. If they are used for allocation then they can be
3044 * filled up again with minimal effort. The slab will never hit the
3045 * per node partial lists and therefore no locking will be required.
3047 * This setting also determines
3049 * A) The number of objects from per cpu partial slabs dumped to the
3050 * per node list when we reach the limit.
3051 * B) The number of objects in cpu partial slabs to extract from the
3052 * per node list when we run out of per cpu objects. We only fetch 50%
3053 * to keep some capacity around for frees.
3055 if (kmem_cache_debug(s))
3057 else if (s->size >= PAGE_SIZE)
3059 else if (s->size >= 1024)
3061 else if (s->size >= 256)
3062 s->cpu_partial = 13;
3064 s->cpu_partial = 30;
3067 s->remote_node_defrag_ratio = 1000;
3069 if (!init_kmem_cache_nodes(s))
3072 if (alloc_kmem_cache_cpus(s))
3075 free_kmem_cache_nodes(s);
3077 if (flags & SLAB_PANIC)
3078 panic("Cannot create slab %s size=%lu realsize=%u "
3079 "order=%u offset=%u flags=%lx\n",
3080 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3085 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3088 #ifdef CONFIG_SLUB_DEBUG
3089 void *addr = page_address(page);
3091 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3092 sizeof(long), GFP_ATOMIC);
3095 slab_err(s, page, text, s->name);
3098 get_map(s, page, map);
3099 for_each_object(p, s, addr, page->objects) {
3101 if (!test_bit(slab_index(p, s, addr), map)) {
3102 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3104 print_tracking(s, p);
3113 * Attempt to free all partial slabs on a node.
3114 * This is called from kmem_cache_close(). We must be the last thread
3115 * using the cache and therefore we do not need to lock anymore.
3117 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3119 struct page *page, *h;
3121 list_for_each_entry_safe(page, h, &n->partial, lru) {
3123 remove_partial(n, page);
3124 discard_slab(s, page);
3126 list_slab_objects(s, page,
3127 "Objects remaining in %s on kmem_cache_close()");
3133 * Release all resources used by a slab cache.
3135 static inline int kmem_cache_close(struct kmem_cache *s)
3140 /* Attempt to free all objects */
3141 for_each_node_state(node, N_NORMAL_MEMORY) {
3142 struct kmem_cache_node *n = get_node(s, node);
3145 if (n->nr_partial || slabs_node(s, node))
3148 free_percpu(s->cpu_slab);
3149 free_kmem_cache_nodes(s);
3153 int __kmem_cache_shutdown(struct kmem_cache *s)
3155 int rc = kmem_cache_close(s);
3158 sysfs_slab_remove(s);
3163 /********************************************************************
3165 *******************************************************************/
3167 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3168 EXPORT_SYMBOL(kmalloc_caches);
3170 #ifdef CONFIG_ZONE_DMA
3171 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3174 static int __init setup_slub_min_order(char *str)
3176 get_option(&str, &slub_min_order);
3181 __setup("slub_min_order=", setup_slub_min_order);
3183 static int __init setup_slub_max_order(char *str)
3185 get_option(&str, &slub_max_order);
3186 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3191 __setup("slub_max_order=", setup_slub_max_order);
3193 static int __init setup_slub_min_objects(char *str)
3195 get_option(&str, &slub_min_objects);
3200 __setup("slub_min_objects=", setup_slub_min_objects);
3202 static int __init setup_slub_nomerge(char *str)
3208 __setup("slub_nomerge", setup_slub_nomerge);
3211 * Conversion table for small slabs sizes / 8 to the index in the
3212 * kmalloc array. This is necessary for slabs < 192 since we have non power
3213 * of two cache sizes there. The size of larger slabs can be determined using
3216 static s8 size_index[24] = {
3243 static inline int size_index_elem(size_t bytes)
3245 return (bytes - 1) / 8;
3248 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3254 return ZERO_SIZE_PTR;
3256 index = size_index[size_index_elem(size)];
3258 index = fls(size - 1);
3260 #ifdef CONFIG_ZONE_DMA
3261 if (unlikely((flags & SLUB_DMA)))
3262 return kmalloc_dma_caches[index];
3265 return kmalloc_caches[index];
3268 void *__kmalloc(size_t size, gfp_t flags)
3270 struct kmem_cache *s;
3273 if (unlikely(size > SLUB_MAX_SIZE))
3274 return kmalloc_large(size, flags);
3276 s = get_slab(size, flags);
3278 if (unlikely(ZERO_OR_NULL_PTR(s)))
3281 ret = slab_alloc(s, flags, _RET_IP_);
3283 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3287 EXPORT_SYMBOL(__kmalloc);
3290 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3295 flags |= __GFP_COMP | __GFP_NOTRACK;
3296 page = alloc_pages_node(node, flags, get_order(size));
3298 ptr = page_address(page);
3300 kmemleak_alloc(ptr, size, 1, flags);
3304 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3306 struct kmem_cache *s;
3309 if (unlikely(size > SLUB_MAX_SIZE)) {
3310 ret = kmalloc_large_node(size, flags, node);
3312 trace_kmalloc_node(_RET_IP_, ret,
3313 size, PAGE_SIZE << get_order(size),
3319 s = get_slab(size, flags);
3321 if (unlikely(ZERO_OR_NULL_PTR(s)))
3324 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3326 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3330 EXPORT_SYMBOL(__kmalloc_node);
3333 size_t ksize(const void *object)
3337 if (unlikely(object == ZERO_SIZE_PTR))
3340 page = virt_to_head_page(object);
3342 if (unlikely(!PageSlab(page))) {
3343 WARN_ON(!PageCompound(page));
3344 return PAGE_SIZE << compound_order(page);
3347 return slab_ksize(page->slab_cache);
3349 EXPORT_SYMBOL(ksize);
3351 #ifdef CONFIG_SLUB_DEBUG
3352 bool verify_mem_not_deleted(const void *x)
3355 void *object = (void *)x;
3356 unsigned long flags;
3359 if (unlikely(ZERO_OR_NULL_PTR(x)))
3362 local_irq_save(flags);
3364 page = virt_to_head_page(x);
3365 if (unlikely(!PageSlab(page))) {
3366 /* maybe it was from stack? */
3372 if (on_freelist(page->slab_cache, page, object)) {
3373 object_err(page->slab_cache, page, object, "Object is on free-list");
3381 local_irq_restore(flags);
3384 EXPORT_SYMBOL(verify_mem_not_deleted);
3387 void kfree(const void *x)
3390 void *object = (void *)x;
3392 trace_kfree(_RET_IP_, x);
3394 if (unlikely(ZERO_OR_NULL_PTR(x)))
3397 page = virt_to_head_page(x);
3398 if (unlikely(!PageSlab(page))) {
3399 BUG_ON(!PageCompound(page));
3401 __free_pages(page, compound_order(page));
3404 slab_free(page->slab_cache, page, object, _RET_IP_);
3406 EXPORT_SYMBOL(kfree);
3409 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3410 * the remaining slabs by the number of items in use. The slabs with the
3411 * most items in use come first. New allocations will then fill those up
3412 * and thus they can be removed from the partial lists.
3414 * The slabs with the least items are placed last. This results in them
3415 * being allocated from last increasing the chance that the last objects
3416 * are freed in them.
3418 int kmem_cache_shrink(struct kmem_cache *s)
3422 struct kmem_cache_node *n;
3425 int objects = oo_objects(s->max);
3426 struct list_head *slabs_by_inuse =
3427 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3428 unsigned long flags;
3430 if (!slabs_by_inuse)
3434 for_each_node_state(node, N_NORMAL_MEMORY) {
3435 n = get_node(s, node);
3440 for (i = 0; i < objects; i++)
3441 INIT_LIST_HEAD(slabs_by_inuse + i);
3443 spin_lock_irqsave(&n->list_lock, flags);
3446 * Build lists indexed by the items in use in each slab.
3448 * Note that concurrent frees may occur while we hold the
3449 * list_lock. page->inuse here is the upper limit.
3451 list_for_each_entry_safe(page, t, &n->partial, lru) {
3452 list_move(&page->lru, slabs_by_inuse + page->inuse);
3458 * Rebuild the partial list with the slabs filled up most
3459 * first and the least used slabs at the end.
3461 for (i = objects - 1; i > 0; i--)
3462 list_splice(slabs_by_inuse + i, n->partial.prev);
3464 spin_unlock_irqrestore(&n->list_lock, flags);
3466 /* Release empty slabs */
3467 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3468 discard_slab(s, page);
3471 kfree(slabs_by_inuse);
3474 EXPORT_SYMBOL(kmem_cache_shrink);
3476 #if defined(CONFIG_MEMORY_HOTPLUG)
3477 static int slab_mem_going_offline_callback(void *arg)
3479 struct kmem_cache *s;
3481 mutex_lock(&slab_mutex);
3482 list_for_each_entry(s, &slab_caches, list)
3483 kmem_cache_shrink(s);
3484 mutex_unlock(&slab_mutex);
3489 static void slab_mem_offline_callback(void *arg)
3491 struct kmem_cache_node *n;
3492 struct kmem_cache *s;
3493 struct memory_notify *marg = arg;
3496 offline_node = marg->status_change_nid;
3499 * If the node still has available memory. we need kmem_cache_node
3502 if (offline_node < 0)
3505 mutex_lock(&slab_mutex);
3506 list_for_each_entry(s, &slab_caches, list) {
3507 n = get_node(s, offline_node);
3510 * if n->nr_slabs > 0, slabs still exist on the node
3511 * that is going down. We were unable to free them,
3512 * and offline_pages() function shouldn't call this
3513 * callback. So, we must fail.
3515 BUG_ON(slabs_node(s, offline_node));
3517 s->node[offline_node] = NULL;
3518 kmem_cache_free(kmem_cache_node, n);
3521 mutex_unlock(&slab_mutex);
3524 static int slab_mem_going_online_callback(void *arg)
3526 struct kmem_cache_node *n;
3527 struct kmem_cache *s;
3528 struct memory_notify *marg = arg;
3529 int nid = marg->status_change_nid;
3533 * If the node's memory is already available, then kmem_cache_node is
3534 * already created. Nothing to do.
3540 * We are bringing a node online. No memory is available yet. We must
3541 * allocate a kmem_cache_node structure in order to bring the node
3544 mutex_lock(&slab_mutex);
3545 list_for_each_entry(s, &slab_caches, list) {
3547 * XXX: kmem_cache_alloc_node will fallback to other nodes
3548 * since memory is not yet available from the node that
3551 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3556 init_kmem_cache_node(n);
3560 mutex_unlock(&slab_mutex);
3564 static int slab_memory_callback(struct notifier_block *self,
3565 unsigned long action, void *arg)
3570 case MEM_GOING_ONLINE:
3571 ret = slab_mem_going_online_callback(arg);
3573 case MEM_GOING_OFFLINE:
3574 ret = slab_mem_going_offline_callback(arg);
3577 case MEM_CANCEL_ONLINE:
3578 slab_mem_offline_callback(arg);
3581 case MEM_CANCEL_OFFLINE:
3585 ret = notifier_from_errno(ret);
3591 #endif /* CONFIG_MEMORY_HOTPLUG */
3593 /********************************************************************
3594 * Basic setup of slabs
3595 *******************************************************************/
3598 * Used for early kmem_cache structures that were allocated using
3599 * the page allocator. Allocate them properly then fix up the pointers
3600 * that may be pointing to the wrong kmem_cache structure.
3603 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3606 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3608 memcpy(s, static_cache, kmem_cache->object_size);
3610 for_each_node_state(node, N_NORMAL_MEMORY) {
3611 struct kmem_cache_node *n = get_node(s, node);
3615 list_for_each_entry(p, &n->partial, lru)
3618 #ifdef CONFIG_SLUB_DEBUG
3619 list_for_each_entry(p, &n->full, lru)
3624 list_add(&s->list, &slab_caches);
3628 void __init kmem_cache_init(void)
3630 static __initdata struct kmem_cache boot_kmem_cache,
3631 boot_kmem_cache_node;
3635 if (debug_guardpage_minorder())
3638 kmem_cache_node = &boot_kmem_cache_node;
3639 kmem_cache = &boot_kmem_cache;
3641 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3642 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3644 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3646 /* Able to allocate the per node structures */
3647 slab_state = PARTIAL;
3649 create_boot_cache(kmem_cache, "kmem_cache",
3650 offsetof(struct kmem_cache, node) +
3651 nr_node_ids * sizeof(struct kmem_cache_node *),
3652 SLAB_HWCACHE_ALIGN);
3654 kmem_cache = bootstrap(&boot_kmem_cache);
3657 * Allocate kmem_cache_node properly from the kmem_cache slab.
3658 * kmem_cache_node is separately allocated so no need to
3659 * update any list pointers.
3661 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3663 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3666 * Patch up the size_index table if we have strange large alignment
3667 * requirements for the kmalloc array. This is only the case for
3668 * MIPS it seems. The standard arches will not generate any code here.
3670 * Largest permitted alignment is 256 bytes due to the way we
3671 * handle the index determination for the smaller caches.
3673 * Make sure that nothing crazy happens if someone starts tinkering
3674 * around with ARCH_KMALLOC_MINALIGN
3676 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3677 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3679 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3680 int elem = size_index_elem(i);
3681 if (elem >= ARRAY_SIZE(size_index))
3683 size_index[elem] = KMALLOC_SHIFT_LOW;
3686 if (KMALLOC_MIN_SIZE == 64) {
3688 * The 96 byte size cache is not used if the alignment
3691 for (i = 64 + 8; i <= 96; i += 8)
3692 size_index[size_index_elem(i)] = 7;
3693 } else if (KMALLOC_MIN_SIZE == 128) {
3695 * The 192 byte sized cache is not used if the alignment
3696 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3699 for (i = 128 + 8; i <= 192; i += 8)
3700 size_index[size_index_elem(i)] = 8;
3703 /* Caches that are not of the two-to-the-power-of size */
3704 if (KMALLOC_MIN_SIZE <= 32) {
3705 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3709 if (KMALLOC_MIN_SIZE <= 64) {
3710 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3714 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3715 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3721 /* Provide the correct kmalloc names now that the caches are up */
3722 if (KMALLOC_MIN_SIZE <= 32) {
3723 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3724 BUG_ON(!kmalloc_caches[1]->name);
3727 if (KMALLOC_MIN_SIZE <= 64) {
3728 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3729 BUG_ON(!kmalloc_caches[2]->name);
3732 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3733 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3736 kmalloc_caches[i]->name = s;
3740 register_cpu_notifier(&slab_notifier);
3743 #ifdef CONFIG_ZONE_DMA
3744 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3745 struct kmem_cache *s = kmalloc_caches[i];
3748 char *name = kasprintf(GFP_NOWAIT,
3749 "dma-kmalloc-%d", s->object_size);
3752 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3753 s->object_size, SLAB_CACHE_DMA);
3758 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3759 " CPUs=%d, Nodes=%d\n",
3760 caches, cache_line_size(),
3761 slub_min_order, slub_max_order, slub_min_objects,
3762 nr_cpu_ids, nr_node_ids);
3765 void __init kmem_cache_init_late(void)
3770 * Find a mergeable slab cache
3772 static int slab_unmergeable(struct kmem_cache *s)
3774 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3781 * We may have set a slab to be unmergeable during bootstrap.
3783 if (s->refcount < 0)
3789 static struct kmem_cache *find_mergeable(size_t size,
3790 size_t align, unsigned long flags, const char *name,
3791 void (*ctor)(void *))
3793 struct kmem_cache *s;
3795 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3801 size = ALIGN(size, sizeof(void *));
3802 align = calculate_alignment(flags, align, size);
3803 size = ALIGN(size, align);
3804 flags = kmem_cache_flags(size, flags, name, NULL);
3806 list_for_each_entry(s, &slab_caches, list) {
3807 if (slab_unmergeable(s))
3813 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3816 * Check if alignment is compatible.
3817 * Courtesy of Adrian Drzewiecki
3819 if ((s->size & ~(align - 1)) != s->size)
3822 if (s->size - size >= sizeof(void *))
3830 struct kmem_cache *__kmem_cache_alias(const char *name, size_t size,
3831 size_t align, unsigned long flags, void (*ctor)(void *))
3833 struct kmem_cache *s;
3835 s = find_mergeable(size, align, flags, name, ctor);
3839 * Adjust the object sizes so that we clear
3840 * the complete object on kzalloc.
3842 s->object_size = max(s->object_size, (int)size);
3843 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3845 if (sysfs_slab_alias(s, name)) {
3854 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3858 err = kmem_cache_open(s, flags);
3862 /* Mutex is not taken during early boot */
3863 if (slab_state <= UP)
3866 mutex_unlock(&slab_mutex);
3867 err = sysfs_slab_add(s);
3868 mutex_lock(&slab_mutex);
3871 kmem_cache_close(s);
3878 * Use the cpu notifier to insure that the cpu slabs are flushed when
3881 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3882 unsigned long action, void *hcpu)
3884 long cpu = (long)hcpu;
3885 struct kmem_cache *s;
3886 unsigned long flags;
3889 case CPU_UP_CANCELED:
3890 case CPU_UP_CANCELED_FROZEN:
3892 case CPU_DEAD_FROZEN:
3893 mutex_lock(&slab_mutex);
3894 list_for_each_entry(s, &slab_caches, list) {
3895 local_irq_save(flags);
3896 __flush_cpu_slab(s, cpu);
3897 local_irq_restore(flags);
3899 mutex_unlock(&slab_mutex);
3907 static struct notifier_block __cpuinitdata slab_notifier = {
3908 .notifier_call = slab_cpuup_callback
3913 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3915 struct kmem_cache *s;
3918 if (unlikely(size > SLUB_MAX_SIZE))
3919 return kmalloc_large(size, gfpflags);
3921 s = get_slab(size, gfpflags);
3923 if (unlikely(ZERO_OR_NULL_PTR(s)))
3926 ret = slab_alloc(s, gfpflags, caller);
3928 /* Honor the call site pointer we received. */
3929 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3935 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3936 int node, unsigned long caller)
3938 struct kmem_cache *s;
3941 if (unlikely(size > SLUB_MAX_SIZE)) {
3942 ret = kmalloc_large_node(size, gfpflags, node);
3944 trace_kmalloc_node(caller, ret,
3945 size, PAGE_SIZE << get_order(size),
3951 s = get_slab(size, gfpflags);
3953 if (unlikely(ZERO_OR_NULL_PTR(s)))
3956 ret = slab_alloc_node(s, gfpflags, node, caller);
3958 /* Honor the call site pointer we received. */
3959 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3966 static int count_inuse(struct page *page)
3971 static int count_total(struct page *page)
3973 return page->objects;
3977 #ifdef CONFIG_SLUB_DEBUG
3978 static int validate_slab(struct kmem_cache *s, struct page *page,
3982 void *addr = page_address(page);
3984 if (!check_slab(s, page) ||
3985 !on_freelist(s, page, NULL))
3988 /* Now we know that a valid freelist exists */
3989 bitmap_zero(map, page->objects);
3991 get_map(s, page, map);
3992 for_each_object(p, s, addr, page->objects) {
3993 if (test_bit(slab_index(p, s, addr), map))
3994 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3998 for_each_object(p, s, addr, page->objects)
3999 if (!test_bit(slab_index(p, s, addr), map))
4000 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4005 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4009 validate_slab(s, page, map);
4013 static int validate_slab_node(struct kmem_cache *s,
4014 struct kmem_cache_node *n, unsigned long *map)
4016 unsigned long count = 0;
4018 unsigned long flags;
4020 spin_lock_irqsave(&n->list_lock, flags);
4022 list_for_each_entry(page, &n->partial, lru) {
4023 validate_slab_slab(s, page, map);
4026 if (count != n->nr_partial)
4027 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4028 "counter=%ld\n", s->name, count, n->nr_partial);
4030 if (!(s->flags & SLAB_STORE_USER))
4033 list_for_each_entry(page, &n->full, lru) {
4034 validate_slab_slab(s, page, map);
4037 if (count != atomic_long_read(&n->nr_slabs))
4038 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4039 "counter=%ld\n", s->name, count,
4040 atomic_long_read(&n->nr_slabs));
4043 spin_unlock_irqrestore(&n->list_lock, flags);
4047 static long validate_slab_cache(struct kmem_cache *s)
4050 unsigned long count = 0;
4051 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4052 sizeof(unsigned long), GFP_KERNEL);
4058 for_each_node_state(node, N_NORMAL_MEMORY) {
4059 struct kmem_cache_node *n = get_node(s, node);
4061 count += validate_slab_node(s, n, map);
4067 * Generate lists of code addresses where slabcache objects are allocated
4072 unsigned long count;
4079 DECLARE_BITMAP(cpus, NR_CPUS);
4085 unsigned long count;
4086 struct location *loc;
4089 static void free_loc_track(struct loc_track *t)
4092 free_pages((unsigned long)t->loc,
4093 get_order(sizeof(struct location) * t->max));
4096 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4101 order = get_order(sizeof(struct location) * max);
4103 l = (void *)__get_free_pages(flags, order);
4108 memcpy(l, t->loc, sizeof(struct location) * t->count);
4116 static int add_location(struct loc_track *t, struct kmem_cache *s,
4117 const struct track *track)
4119 long start, end, pos;
4121 unsigned long caddr;
4122 unsigned long age = jiffies - track->when;
4128 pos = start + (end - start + 1) / 2;
4131 * There is nothing at "end". If we end up there
4132 * we need to add something to before end.
4137 caddr = t->loc[pos].addr;
4138 if (track->addr == caddr) {
4144 if (age < l->min_time)
4146 if (age > l->max_time)
4149 if (track->pid < l->min_pid)
4150 l->min_pid = track->pid;
4151 if (track->pid > l->max_pid)
4152 l->max_pid = track->pid;
4154 cpumask_set_cpu(track->cpu,
4155 to_cpumask(l->cpus));
4157 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4161 if (track->addr < caddr)
4168 * Not found. Insert new tracking element.
4170 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4176 (t->count - pos) * sizeof(struct location));
4179 l->addr = track->addr;
4183 l->min_pid = track->pid;
4184 l->max_pid = track->pid;
4185 cpumask_clear(to_cpumask(l->cpus));
4186 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4187 nodes_clear(l->nodes);
4188 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4192 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4193 struct page *page, enum track_item alloc,
4196 void *addr = page_address(page);
4199 bitmap_zero(map, page->objects);
4200 get_map(s, page, map);
4202 for_each_object(p, s, addr, page->objects)
4203 if (!test_bit(slab_index(p, s, addr), map))
4204 add_location(t, s, get_track(s, p, alloc));
4207 static int list_locations(struct kmem_cache *s, char *buf,
4208 enum track_item alloc)
4212 struct loc_track t = { 0, 0, NULL };
4214 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4215 sizeof(unsigned long), GFP_KERNEL);
4217 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4220 return sprintf(buf, "Out of memory\n");
4222 /* Push back cpu slabs */
4225 for_each_node_state(node, N_NORMAL_MEMORY) {
4226 struct kmem_cache_node *n = get_node(s, node);
4227 unsigned long flags;
4230 if (!atomic_long_read(&n->nr_slabs))
4233 spin_lock_irqsave(&n->list_lock, flags);
4234 list_for_each_entry(page, &n->partial, lru)
4235 process_slab(&t, s, page, alloc, map);
4236 list_for_each_entry(page, &n->full, lru)
4237 process_slab(&t, s, page, alloc, map);
4238 spin_unlock_irqrestore(&n->list_lock, flags);
4241 for (i = 0; i < t.count; i++) {
4242 struct location *l = &t.loc[i];
4244 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4246 len += sprintf(buf + len, "%7ld ", l->count);
4249 len += sprintf(buf + len, "%pS", (void *)l->addr);
4251 len += sprintf(buf + len, "<not-available>");
4253 if (l->sum_time != l->min_time) {
4254 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4256 (long)div_u64(l->sum_time, l->count),
4259 len += sprintf(buf + len, " age=%ld",
4262 if (l->min_pid != l->max_pid)
4263 len += sprintf(buf + len, " pid=%ld-%ld",
4264 l->min_pid, l->max_pid);
4266 len += sprintf(buf + len, " pid=%ld",
4269 if (num_online_cpus() > 1 &&
4270 !cpumask_empty(to_cpumask(l->cpus)) &&
4271 len < PAGE_SIZE - 60) {
4272 len += sprintf(buf + len, " cpus=");
4273 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4274 to_cpumask(l->cpus));
4277 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4278 len < PAGE_SIZE - 60) {
4279 len += sprintf(buf + len, " nodes=");
4280 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4284 len += sprintf(buf + len, "\n");
4290 len += sprintf(buf, "No data\n");
4295 #ifdef SLUB_RESILIENCY_TEST
4296 static void resiliency_test(void)
4300 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4302 printk(KERN_ERR "SLUB resiliency testing\n");
4303 printk(KERN_ERR "-----------------------\n");
4304 printk(KERN_ERR "A. Corruption after allocation\n");
4306 p = kzalloc(16, GFP_KERNEL);
4308 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4309 " 0x12->0x%p\n\n", p + 16);
4311 validate_slab_cache(kmalloc_caches[4]);
4313 /* Hmmm... The next two are dangerous */
4314 p = kzalloc(32, GFP_KERNEL);
4315 p[32 + sizeof(void *)] = 0x34;
4316 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4317 " 0x34 -> -0x%p\n", p);
4319 "If allocated object is overwritten then not detectable\n\n");
4321 validate_slab_cache(kmalloc_caches[5]);
4322 p = kzalloc(64, GFP_KERNEL);
4323 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4325 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4328 "If allocated object is overwritten then not detectable\n\n");
4329 validate_slab_cache(kmalloc_caches[6]);
4331 printk(KERN_ERR "\nB. Corruption after free\n");
4332 p = kzalloc(128, GFP_KERNEL);
4335 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4336 validate_slab_cache(kmalloc_caches[7]);
4338 p = kzalloc(256, GFP_KERNEL);
4341 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4343 validate_slab_cache(kmalloc_caches[8]);
4345 p = kzalloc(512, GFP_KERNEL);
4348 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4349 validate_slab_cache(kmalloc_caches[9]);
4353 static void resiliency_test(void) {};
4358 enum slab_stat_type {
4359 SL_ALL, /* All slabs */
4360 SL_PARTIAL, /* Only partially allocated slabs */
4361 SL_CPU, /* Only slabs used for cpu caches */
4362 SL_OBJECTS, /* Determine allocated objects not slabs */
4363 SL_TOTAL /* Determine object capacity not slabs */
4366 #define SO_ALL (1 << SL_ALL)
4367 #define SO_PARTIAL (1 << SL_PARTIAL)
4368 #define SO_CPU (1 << SL_CPU)
4369 #define SO_OBJECTS (1 << SL_OBJECTS)
4370 #define SO_TOTAL (1 << SL_TOTAL)
4372 static ssize_t show_slab_objects(struct kmem_cache *s,
4373 char *buf, unsigned long flags)
4375 unsigned long total = 0;
4378 unsigned long *nodes;
4379 unsigned long *per_cpu;
4381 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4384 per_cpu = nodes + nr_node_ids;
4386 if (flags & SO_CPU) {
4389 for_each_possible_cpu(cpu) {
4390 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4394 page = ACCESS_ONCE(c->page);
4398 node = page_to_nid(page);
4399 if (flags & SO_TOTAL)
4401 else if (flags & SO_OBJECTS)
4409 page = ACCESS_ONCE(c->partial);
4420 lock_memory_hotplug();
4421 #ifdef CONFIG_SLUB_DEBUG
4422 if (flags & SO_ALL) {
4423 for_each_node_state(node, N_NORMAL_MEMORY) {
4424 struct kmem_cache_node *n = get_node(s, node);
4426 if (flags & SO_TOTAL)
4427 x = atomic_long_read(&n->total_objects);
4428 else if (flags & SO_OBJECTS)
4429 x = atomic_long_read(&n->total_objects) -
4430 count_partial(n, count_free);
4433 x = atomic_long_read(&n->nr_slabs);
4440 if (flags & SO_PARTIAL) {
4441 for_each_node_state(node, N_NORMAL_MEMORY) {
4442 struct kmem_cache_node *n = get_node(s, node);
4444 if (flags & SO_TOTAL)
4445 x = count_partial(n, count_total);
4446 else if (flags & SO_OBJECTS)
4447 x = count_partial(n, count_inuse);
4454 x = sprintf(buf, "%lu", total);
4456 for_each_node_state(node, N_NORMAL_MEMORY)
4458 x += sprintf(buf + x, " N%d=%lu",
4461 unlock_memory_hotplug();
4463 return x + sprintf(buf + x, "\n");
4466 #ifdef CONFIG_SLUB_DEBUG
4467 static int any_slab_objects(struct kmem_cache *s)
4471 for_each_online_node(node) {
4472 struct kmem_cache_node *n = get_node(s, node);
4477 if (atomic_long_read(&n->total_objects))
4484 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4485 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4487 struct slab_attribute {
4488 struct attribute attr;
4489 ssize_t (*show)(struct kmem_cache *s, char *buf);
4490 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4493 #define SLAB_ATTR_RO(_name) \
4494 static struct slab_attribute _name##_attr = \
4495 __ATTR(_name, 0400, _name##_show, NULL)
4497 #define SLAB_ATTR(_name) \
4498 static struct slab_attribute _name##_attr = \
4499 __ATTR(_name, 0600, _name##_show, _name##_store)
4501 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4503 return sprintf(buf, "%d\n", s->size);
4505 SLAB_ATTR_RO(slab_size);
4507 static ssize_t align_show(struct kmem_cache *s, char *buf)
4509 return sprintf(buf, "%d\n", s->align);
4511 SLAB_ATTR_RO(align);
4513 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4515 return sprintf(buf, "%d\n", s->object_size);
4517 SLAB_ATTR_RO(object_size);
4519 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4521 return sprintf(buf, "%d\n", oo_objects(s->oo));
4523 SLAB_ATTR_RO(objs_per_slab);
4525 static ssize_t order_store(struct kmem_cache *s,
4526 const char *buf, size_t length)
4528 unsigned long order;
4531 err = strict_strtoul(buf, 10, &order);
4535 if (order > slub_max_order || order < slub_min_order)
4538 calculate_sizes(s, order);
4542 static ssize_t order_show(struct kmem_cache *s, char *buf)
4544 return sprintf(buf, "%d\n", oo_order(s->oo));
4548 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4550 return sprintf(buf, "%lu\n", s->min_partial);
4553 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4559 err = strict_strtoul(buf, 10, &min);
4563 set_min_partial(s, min);
4566 SLAB_ATTR(min_partial);
4568 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4570 return sprintf(buf, "%u\n", s->cpu_partial);
4573 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4576 unsigned long objects;
4579 err = strict_strtoul(buf, 10, &objects);
4582 if (objects && kmem_cache_debug(s))
4585 s->cpu_partial = objects;
4589 SLAB_ATTR(cpu_partial);
4591 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4595 return sprintf(buf, "%pS\n", s->ctor);
4599 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4601 return sprintf(buf, "%d\n", s->refcount - 1);
4603 SLAB_ATTR_RO(aliases);
4605 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4607 return show_slab_objects(s, buf, SO_PARTIAL);
4609 SLAB_ATTR_RO(partial);
4611 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4613 return show_slab_objects(s, buf, SO_CPU);
4615 SLAB_ATTR_RO(cpu_slabs);
4617 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4619 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4621 SLAB_ATTR_RO(objects);
4623 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4625 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4627 SLAB_ATTR_RO(objects_partial);
4629 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4636 for_each_online_cpu(cpu) {
4637 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4640 pages += page->pages;
4641 objects += page->pobjects;
4645 len = sprintf(buf, "%d(%d)", objects, pages);
4648 for_each_online_cpu(cpu) {
4649 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4651 if (page && len < PAGE_SIZE - 20)
4652 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4653 page->pobjects, page->pages);
4656 return len + sprintf(buf + len, "\n");
4658 SLAB_ATTR_RO(slabs_cpu_partial);
4660 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4662 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4665 static ssize_t reclaim_account_store(struct kmem_cache *s,
4666 const char *buf, size_t length)
4668 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4670 s->flags |= SLAB_RECLAIM_ACCOUNT;
4673 SLAB_ATTR(reclaim_account);
4675 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4677 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4679 SLAB_ATTR_RO(hwcache_align);
4681 #ifdef CONFIG_ZONE_DMA
4682 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4684 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4686 SLAB_ATTR_RO(cache_dma);
4689 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4691 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4693 SLAB_ATTR_RO(destroy_by_rcu);
4695 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4697 return sprintf(buf, "%d\n", s->reserved);
4699 SLAB_ATTR_RO(reserved);
4701 #ifdef CONFIG_SLUB_DEBUG
4702 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4704 return show_slab_objects(s, buf, SO_ALL);
4706 SLAB_ATTR_RO(slabs);
4708 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4710 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4712 SLAB_ATTR_RO(total_objects);
4714 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4716 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4719 static ssize_t sanity_checks_store(struct kmem_cache *s,
4720 const char *buf, size_t length)
4722 s->flags &= ~SLAB_DEBUG_FREE;
4723 if (buf[0] == '1') {
4724 s->flags &= ~__CMPXCHG_DOUBLE;
4725 s->flags |= SLAB_DEBUG_FREE;
4729 SLAB_ATTR(sanity_checks);
4731 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4733 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4736 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4739 s->flags &= ~SLAB_TRACE;
4740 if (buf[0] == '1') {
4741 s->flags &= ~__CMPXCHG_DOUBLE;
4742 s->flags |= SLAB_TRACE;
4748 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4750 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4753 static ssize_t red_zone_store(struct kmem_cache *s,
4754 const char *buf, size_t length)
4756 if (any_slab_objects(s))
4759 s->flags &= ~SLAB_RED_ZONE;
4760 if (buf[0] == '1') {
4761 s->flags &= ~__CMPXCHG_DOUBLE;
4762 s->flags |= SLAB_RED_ZONE;
4764 calculate_sizes(s, -1);
4767 SLAB_ATTR(red_zone);
4769 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4771 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4774 static ssize_t poison_store(struct kmem_cache *s,
4775 const char *buf, size_t length)
4777 if (any_slab_objects(s))
4780 s->flags &= ~SLAB_POISON;
4781 if (buf[0] == '1') {
4782 s->flags &= ~__CMPXCHG_DOUBLE;
4783 s->flags |= SLAB_POISON;
4785 calculate_sizes(s, -1);
4790 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4792 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4795 static ssize_t store_user_store(struct kmem_cache *s,
4796 const char *buf, size_t length)
4798 if (any_slab_objects(s))
4801 s->flags &= ~SLAB_STORE_USER;
4802 if (buf[0] == '1') {
4803 s->flags &= ~__CMPXCHG_DOUBLE;
4804 s->flags |= SLAB_STORE_USER;
4806 calculate_sizes(s, -1);
4809 SLAB_ATTR(store_user);
4811 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4816 static ssize_t validate_store(struct kmem_cache *s,
4817 const char *buf, size_t length)
4821 if (buf[0] == '1') {
4822 ret = validate_slab_cache(s);
4828 SLAB_ATTR(validate);
4830 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4832 if (!(s->flags & SLAB_STORE_USER))
4834 return list_locations(s, buf, TRACK_ALLOC);
4836 SLAB_ATTR_RO(alloc_calls);
4838 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4840 if (!(s->flags & SLAB_STORE_USER))
4842 return list_locations(s, buf, TRACK_FREE);
4844 SLAB_ATTR_RO(free_calls);
4845 #endif /* CONFIG_SLUB_DEBUG */
4847 #ifdef CONFIG_FAILSLAB
4848 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4850 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4853 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4856 s->flags &= ~SLAB_FAILSLAB;
4858 s->flags |= SLAB_FAILSLAB;
4861 SLAB_ATTR(failslab);
4864 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4869 static ssize_t shrink_store(struct kmem_cache *s,
4870 const char *buf, size_t length)
4872 if (buf[0] == '1') {
4873 int rc = kmem_cache_shrink(s);
4884 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4886 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4889 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4890 const char *buf, size_t length)
4892 unsigned long ratio;
4895 err = strict_strtoul(buf, 10, &ratio);
4900 s->remote_node_defrag_ratio = ratio * 10;
4904 SLAB_ATTR(remote_node_defrag_ratio);
4907 #ifdef CONFIG_SLUB_STATS
4908 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4910 unsigned long sum = 0;
4913 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4918 for_each_online_cpu(cpu) {
4919 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4925 len = sprintf(buf, "%lu", sum);
4928 for_each_online_cpu(cpu) {
4929 if (data[cpu] && len < PAGE_SIZE - 20)
4930 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4934 return len + sprintf(buf + len, "\n");
4937 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4941 for_each_online_cpu(cpu)
4942 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4945 #define STAT_ATTR(si, text) \
4946 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4948 return show_stat(s, buf, si); \
4950 static ssize_t text##_store(struct kmem_cache *s, \
4951 const char *buf, size_t length) \
4953 if (buf[0] != '0') \
4955 clear_stat(s, si); \
4960 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4961 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4962 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4963 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4964 STAT_ATTR(FREE_FROZEN, free_frozen);
4965 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4966 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4967 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4968 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4969 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4970 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4971 STAT_ATTR(FREE_SLAB, free_slab);
4972 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4973 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4974 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4975 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4976 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4977 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4978 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4979 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4980 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4981 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4982 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4983 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4984 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4985 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4988 static struct attribute *slab_attrs[] = {
4989 &slab_size_attr.attr,
4990 &object_size_attr.attr,
4991 &objs_per_slab_attr.attr,
4993 &min_partial_attr.attr,
4994 &cpu_partial_attr.attr,
4996 &objects_partial_attr.attr,
4998 &cpu_slabs_attr.attr,
5002 &hwcache_align_attr.attr,
5003 &reclaim_account_attr.attr,
5004 &destroy_by_rcu_attr.attr,
5006 &reserved_attr.attr,
5007 &slabs_cpu_partial_attr.attr,
5008 #ifdef CONFIG_SLUB_DEBUG
5009 &total_objects_attr.attr,
5011 &sanity_checks_attr.attr,
5013 &red_zone_attr.attr,
5015 &store_user_attr.attr,
5016 &validate_attr.attr,
5017 &alloc_calls_attr.attr,
5018 &free_calls_attr.attr,
5020 #ifdef CONFIG_ZONE_DMA
5021 &cache_dma_attr.attr,
5024 &remote_node_defrag_ratio_attr.attr,
5026 #ifdef CONFIG_SLUB_STATS
5027 &alloc_fastpath_attr.attr,
5028 &alloc_slowpath_attr.attr,
5029 &free_fastpath_attr.attr,
5030 &free_slowpath_attr.attr,
5031 &free_frozen_attr.attr,
5032 &free_add_partial_attr.attr,
5033 &free_remove_partial_attr.attr,
5034 &alloc_from_partial_attr.attr,
5035 &alloc_slab_attr.attr,
5036 &alloc_refill_attr.attr,
5037 &alloc_node_mismatch_attr.attr,
5038 &free_slab_attr.attr,
5039 &cpuslab_flush_attr.attr,
5040 &deactivate_full_attr.attr,
5041 &deactivate_empty_attr.attr,
5042 &deactivate_to_head_attr.attr,
5043 &deactivate_to_tail_attr.attr,
5044 &deactivate_remote_frees_attr.attr,
5045 &deactivate_bypass_attr.attr,
5046 &order_fallback_attr.attr,
5047 &cmpxchg_double_fail_attr.attr,
5048 &cmpxchg_double_cpu_fail_attr.attr,
5049 &cpu_partial_alloc_attr.attr,
5050 &cpu_partial_free_attr.attr,
5051 &cpu_partial_node_attr.attr,
5052 &cpu_partial_drain_attr.attr,
5054 #ifdef CONFIG_FAILSLAB
5055 &failslab_attr.attr,
5061 static struct attribute_group slab_attr_group = {
5062 .attrs = slab_attrs,
5065 static ssize_t slab_attr_show(struct kobject *kobj,
5066 struct attribute *attr,
5069 struct slab_attribute *attribute;
5070 struct kmem_cache *s;
5073 attribute = to_slab_attr(attr);
5076 if (!attribute->show)
5079 err = attribute->show(s, buf);
5084 static ssize_t slab_attr_store(struct kobject *kobj,
5085 struct attribute *attr,
5086 const char *buf, size_t len)
5088 struct slab_attribute *attribute;
5089 struct kmem_cache *s;
5092 attribute = to_slab_attr(attr);
5095 if (!attribute->store)
5098 err = attribute->store(s, buf, len);
5103 static const struct sysfs_ops slab_sysfs_ops = {
5104 .show = slab_attr_show,
5105 .store = slab_attr_store,
5108 static struct kobj_type slab_ktype = {
5109 .sysfs_ops = &slab_sysfs_ops,
5112 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5114 struct kobj_type *ktype = get_ktype(kobj);
5116 if (ktype == &slab_ktype)
5121 static const struct kset_uevent_ops slab_uevent_ops = {
5122 .filter = uevent_filter,
5125 static struct kset *slab_kset;
5127 #define ID_STR_LENGTH 64
5129 /* Create a unique string id for a slab cache:
5131 * Format :[flags-]size
5133 static char *create_unique_id(struct kmem_cache *s)
5135 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5142 * First flags affecting slabcache operations. We will only
5143 * get here for aliasable slabs so we do not need to support
5144 * too many flags. The flags here must cover all flags that
5145 * are matched during merging to guarantee that the id is
5148 if (s->flags & SLAB_CACHE_DMA)
5150 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5152 if (s->flags & SLAB_DEBUG_FREE)
5154 if (!(s->flags & SLAB_NOTRACK))
5158 p += sprintf(p, "%07d", s->size);
5159 BUG_ON(p > name + ID_STR_LENGTH - 1);
5163 static int sysfs_slab_add(struct kmem_cache *s)
5167 int unmergeable = slab_unmergeable(s);
5171 * Slabcache can never be merged so we can use the name proper.
5172 * This is typically the case for debug situations. In that
5173 * case we can catch duplicate names easily.
5175 sysfs_remove_link(&slab_kset->kobj, s->name);
5179 * Create a unique name for the slab as a target
5182 name = create_unique_id(s);
5185 s->kobj.kset = slab_kset;
5186 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5188 kobject_put(&s->kobj);
5192 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5194 kobject_del(&s->kobj);
5195 kobject_put(&s->kobj);
5198 kobject_uevent(&s->kobj, KOBJ_ADD);
5200 /* Setup first alias */
5201 sysfs_slab_alias(s, s->name);
5207 static void sysfs_slab_remove(struct kmem_cache *s)
5209 if (slab_state < FULL)
5211 * Sysfs has not been setup yet so no need to remove the
5216 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5217 kobject_del(&s->kobj);
5218 kobject_put(&s->kobj);
5222 * Need to buffer aliases during bootup until sysfs becomes
5223 * available lest we lose that information.
5225 struct saved_alias {
5226 struct kmem_cache *s;
5228 struct saved_alias *next;
5231 static struct saved_alias *alias_list;
5233 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5235 struct saved_alias *al;
5237 if (slab_state == FULL) {
5239 * If we have a leftover link then remove it.
5241 sysfs_remove_link(&slab_kset->kobj, name);
5242 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5245 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5251 al->next = alias_list;
5256 static int __init slab_sysfs_init(void)
5258 struct kmem_cache *s;
5261 mutex_lock(&slab_mutex);
5263 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5265 mutex_unlock(&slab_mutex);
5266 printk(KERN_ERR "Cannot register slab subsystem.\n");
5272 list_for_each_entry(s, &slab_caches, list) {
5273 err = sysfs_slab_add(s);
5275 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5276 " to sysfs\n", s->name);
5279 while (alias_list) {
5280 struct saved_alias *al = alias_list;
5282 alias_list = alias_list->next;
5283 err = sysfs_slab_alias(al->s, al->name);
5285 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5286 " %s to sysfs\n", al->name);
5290 mutex_unlock(&slab_mutex);
5295 __initcall(slab_sysfs_init);
5296 #endif /* CONFIG_SYSFS */
5299 * The /proc/slabinfo ABI
5301 #ifdef CONFIG_SLABINFO
5302 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5304 unsigned long nr_partials = 0;
5305 unsigned long nr_slabs = 0;
5306 unsigned long nr_objs = 0;
5307 unsigned long nr_free = 0;
5310 for_each_online_node(node) {
5311 struct kmem_cache_node *n = get_node(s, node);
5316 nr_partials += n->nr_partial;
5317 nr_slabs += atomic_long_read(&n->nr_slabs);
5318 nr_objs += atomic_long_read(&n->total_objects);
5319 nr_free += count_partial(n, count_free);
5322 sinfo->active_objs = nr_objs - nr_free;
5323 sinfo->num_objs = nr_objs;
5324 sinfo->active_slabs = nr_slabs;
5325 sinfo->num_slabs = nr_slabs;
5326 sinfo->objects_per_slab = oo_objects(s->oo);
5327 sinfo->cache_order = oo_order(s->oo);
5330 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5334 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5335 size_t count, loff_t *ppos)
5339 #endif /* CONFIG_SLABINFO */