2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
31 #include <linux/stacktrace.h>
33 #include <trace/events/kmem.h>
37 * 1. slub_lock (Global Semaphore)
39 * 3. slab_lock(page) (Only on some arches and for debugging)
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
134 /* Enable to log cmpxchg failures */
135 #undef SLUB_DEBUG_CMPXCHG
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 #define MIN_PARTIAL 5
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
148 #define MAX_PARTIAL 10
150 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
158 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
161 * Set of flags that will prevent slab merging
163 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
167 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
171 #define OO_MASK ((1 << OO_SHIFT) - 1)
172 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174 /* Internal SLUB flags */
175 #define __OBJECT_POISON 0x80000000UL /* Poison object */
176 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178 static int kmem_size = sizeof(struct kmem_cache);
181 static struct notifier_block slab_notifier;
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
191 /* A list of all slab caches on the system */
192 static DECLARE_RWSEM(slub_lock);
193 static LIST_HEAD(slab_caches);
196 * Tracking user of a slab.
198 #define TRACK_ADDRS_COUNT 16
200 unsigned long addr; /* Called from address */
201 #ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
209 enum track_item { TRACK_ALLOC, TRACK_FREE };
212 static int sysfs_slab_add(struct kmem_cache *);
213 static int sysfs_slab_alias(struct kmem_cache *, const char *);
214 static void sysfs_slab_remove(struct kmem_cache *);
217 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
220 static inline void sysfs_slab_remove(struct kmem_cache *s)
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 int slab_is_available(void)
241 return slab_state >= UP;
244 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246 return s->node[node];
249 /* Verify that a pointer has an address that is valid within a slab page */
250 static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
267 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 return *(void **)(object + s->offset);
272 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
276 #ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
279 p = get_freepointer(s, object);
284 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
286 *(void **)(object + s->offset) = fp;
289 /* Loop over all objects in a slab */
290 #define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
294 /* Determine object index from a given position */
295 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
297 return (p - addr) / s->size;
300 static inline size_t slab_ksize(const struct kmem_cache *s)
302 #ifdef CONFIG_SLUB_DEBUG
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
319 * Else we can use all the padding etc for the allocation
324 static inline int order_objects(int order, unsigned long size, int reserved)
326 return ((PAGE_SIZE << order) - reserved) / size;
329 static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
339 static inline int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 bit_spin_lock(PG_locked, &page->flags);
357 static __always_inline void slab_unlock(struct page *page)
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
368 VM_BUG_ON(!irqs_disabled());
369 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
370 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
371 if (s->flags & __CMPXCHG_DOUBLE) {
372 if (cmpxchg_double(&page->freelist,
373 freelist_old, counters_old,
374 freelist_new, counters_new))
380 if (page->freelist == freelist_old && page->counters == counters_old) {
381 page->freelist = freelist_new;
382 page->counters = counters_new;
390 stat(s, CMPXCHG_DOUBLE_FAIL);
392 #ifdef SLUB_DEBUG_CMPXCHG
393 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
399 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
400 void *freelist_old, unsigned long counters_old,
401 void *freelist_new, unsigned long counters_new,
404 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
405 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
406 if (s->flags & __CMPXCHG_DOUBLE) {
407 if (cmpxchg_double(&page->freelist,
408 freelist_old, counters_old,
409 freelist_new, counters_new))
416 local_irq_save(flags);
418 if (page->freelist == freelist_old && page->counters == counters_old) {
419 page->freelist = freelist_new;
420 page->counters = counters_new;
422 local_irq_restore(flags);
426 local_irq_restore(flags);
430 stat(s, CMPXCHG_DOUBLE_FAIL);
432 #ifdef SLUB_DEBUG_CMPXCHG
433 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
439 #ifdef CONFIG_SLUB_DEBUG
441 * Determine a map of object in use on a page.
443 * Node listlock must be held to guarantee that the page does
444 * not vanish from under us.
446 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
449 void *addr = page_address(page);
451 for (p = page->freelist; p; p = get_freepointer(s, p))
452 set_bit(slab_index(p, s, addr), map);
458 #ifdef CONFIG_SLUB_DEBUG_ON
459 static int slub_debug = DEBUG_DEFAULT_FLAGS;
461 static int slub_debug;
464 static char *slub_debug_slabs;
465 static int disable_higher_order_debug;
470 static void print_section(char *text, u8 *addr, unsigned int length)
472 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
476 static struct track *get_track(struct kmem_cache *s, void *object,
477 enum track_item alloc)
482 p = object + s->offset + sizeof(void *);
484 p = object + s->inuse;
489 static void set_track(struct kmem_cache *s, void *object,
490 enum track_item alloc, unsigned long addr)
492 struct track *p = get_track(s, object, alloc);
495 #ifdef CONFIG_STACKTRACE
496 struct stack_trace trace;
499 trace.nr_entries = 0;
500 trace.max_entries = TRACK_ADDRS_COUNT;
501 trace.entries = p->addrs;
503 save_stack_trace(&trace);
505 /* See rant in lockdep.c */
506 if (trace.nr_entries != 0 &&
507 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
510 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
514 p->cpu = smp_processor_id();
515 p->pid = current->pid;
518 memset(p, 0, sizeof(struct track));
521 static void init_tracking(struct kmem_cache *s, void *object)
523 if (!(s->flags & SLAB_STORE_USER))
526 set_track(s, object, TRACK_FREE, 0UL);
527 set_track(s, object, TRACK_ALLOC, 0UL);
530 static void print_track(const char *s, struct track *t)
535 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
536 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
537 #ifdef CONFIG_STACKTRACE
540 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
542 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
549 static void print_tracking(struct kmem_cache *s, void *object)
551 if (!(s->flags & SLAB_STORE_USER))
554 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
555 print_track("Freed", get_track(s, object, TRACK_FREE));
558 static void print_page_info(struct page *page)
560 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
561 page, page->objects, page->inuse, page->freelist, page->flags);
565 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
571 vsnprintf(buf, sizeof(buf), fmt, args);
573 printk(KERN_ERR "========================================"
574 "=====================================\n");
575 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
576 printk(KERN_ERR "----------------------------------------"
577 "-------------------------------------\n\n");
580 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
586 vsnprintf(buf, sizeof(buf), fmt, args);
588 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
591 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
593 unsigned int off; /* Offset of last byte */
594 u8 *addr = page_address(page);
596 print_tracking(s, p);
598 print_page_info(page);
600 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
601 p, p - addr, get_freepointer(s, p));
604 print_section("Bytes b4 ", p - 16, 16);
606 print_section("Object ", p, min_t(unsigned long, s->objsize,
608 if (s->flags & SLAB_RED_ZONE)
609 print_section("Redzone ", p + s->objsize,
610 s->inuse - s->objsize);
613 off = s->offset + sizeof(void *);
617 if (s->flags & SLAB_STORE_USER)
618 off += 2 * sizeof(struct track);
621 /* Beginning of the filler is the free pointer */
622 print_section("Padding ", p + off, s->size - off);
627 static void object_err(struct kmem_cache *s, struct page *page,
628 u8 *object, char *reason)
630 slab_bug(s, "%s", reason);
631 print_trailer(s, page, object);
634 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
640 vsnprintf(buf, sizeof(buf), fmt, args);
642 slab_bug(s, "%s", buf);
643 print_page_info(page);
647 static void init_object(struct kmem_cache *s, void *object, u8 val)
651 if (s->flags & __OBJECT_POISON) {
652 memset(p, POISON_FREE, s->objsize - 1);
653 p[s->objsize - 1] = POISON_END;
656 if (s->flags & SLAB_RED_ZONE)
657 memset(p + s->objsize, val, s->inuse - s->objsize);
660 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
661 void *from, void *to)
663 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
664 memset(from, data, to - from);
667 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
668 u8 *object, char *what,
669 u8 *start, unsigned int value, unsigned int bytes)
674 fault = memchr_inv(start, value, bytes);
679 while (end > fault && end[-1] == value)
682 slab_bug(s, "%s overwritten", what);
683 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
684 fault, end - 1, fault[0], value);
685 print_trailer(s, page, object);
687 restore_bytes(s, what, value, fault, end);
695 * Bytes of the object to be managed.
696 * If the freepointer may overlay the object then the free
697 * pointer is the first word of the object.
699 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
702 * object + s->objsize
703 * Padding to reach word boundary. This is also used for Redzoning.
704 * Padding is extended by another word if Redzoning is enabled and
707 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
708 * 0xcc (RED_ACTIVE) for objects in use.
711 * Meta data starts here.
713 * A. Free pointer (if we cannot overwrite object on free)
714 * B. Tracking data for SLAB_STORE_USER
715 * C. Padding to reach required alignment boundary or at mininum
716 * one word if debugging is on to be able to detect writes
717 * before the word boundary.
719 * Padding is done using 0x5a (POISON_INUSE)
722 * Nothing is used beyond s->size.
724 * If slabcaches are merged then the objsize and inuse boundaries are mostly
725 * ignored. And therefore no slab options that rely on these boundaries
726 * may be used with merged slabcaches.
729 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
731 unsigned long off = s->inuse; /* The end of info */
734 /* Freepointer is placed after the object. */
735 off += sizeof(void *);
737 if (s->flags & SLAB_STORE_USER)
738 /* We also have user information there */
739 off += 2 * sizeof(struct track);
744 return check_bytes_and_report(s, page, p, "Object padding",
745 p + off, POISON_INUSE, s->size - off);
748 /* Check the pad bytes at the end of a slab page */
749 static int slab_pad_check(struct kmem_cache *s, struct page *page)
757 if (!(s->flags & SLAB_POISON))
760 start = page_address(page);
761 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
762 end = start + length;
763 remainder = length % s->size;
767 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
770 while (end > fault && end[-1] == POISON_INUSE)
773 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
774 print_section("Padding ", end - remainder, remainder);
776 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
780 static int check_object(struct kmem_cache *s, struct page *page,
781 void *object, u8 val)
784 u8 *endobject = object + s->objsize;
786 if (s->flags & SLAB_RED_ZONE) {
787 if (!check_bytes_and_report(s, page, object, "Redzone",
788 endobject, val, s->inuse - s->objsize))
791 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
792 check_bytes_and_report(s, page, p, "Alignment padding",
793 endobject, POISON_INUSE, s->inuse - s->objsize);
797 if (s->flags & SLAB_POISON) {
798 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
799 (!check_bytes_and_report(s, page, p, "Poison", p,
800 POISON_FREE, s->objsize - 1) ||
801 !check_bytes_and_report(s, page, p, "Poison",
802 p + s->objsize - 1, POISON_END, 1)))
805 * check_pad_bytes cleans up on its own.
807 check_pad_bytes(s, page, p);
810 if (!s->offset && val == SLUB_RED_ACTIVE)
812 * Object and freepointer overlap. Cannot check
813 * freepointer while object is allocated.
817 /* Check free pointer validity */
818 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
819 object_err(s, page, p, "Freepointer corrupt");
821 * No choice but to zap it and thus lose the remainder
822 * of the free objects in this slab. May cause
823 * another error because the object count is now wrong.
825 set_freepointer(s, p, NULL);
831 static int check_slab(struct kmem_cache *s, struct page *page)
835 VM_BUG_ON(!irqs_disabled());
837 if (!PageSlab(page)) {
838 slab_err(s, page, "Not a valid slab page");
842 maxobj = order_objects(compound_order(page), s->size, s->reserved);
843 if (page->objects > maxobj) {
844 slab_err(s, page, "objects %u > max %u",
845 s->name, page->objects, maxobj);
848 if (page->inuse > page->objects) {
849 slab_err(s, page, "inuse %u > max %u",
850 s->name, page->inuse, page->objects);
853 /* Slab_pad_check fixes things up after itself */
854 slab_pad_check(s, page);
859 * Determine if a certain object on a page is on the freelist. Must hold the
860 * slab lock to guarantee that the chains are in a consistent state.
862 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
867 unsigned long max_objects;
870 while (fp && nr <= page->objects) {
873 if (!check_valid_pointer(s, page, fp)) {
875 object_err(s, page, object,
876 "Freechain corrupt");
877 set_freepointer(s, object, NULL);
880 slab_err(s, page, "Freepointer corrupt");
881 page->freelist = NULL;
882 page->inuse = page->objects;
883 slab_fix(s, "Freelist cleared");
889 fp = get_freepointer(s, object);
893 max_objects = order_objects(compound_order(page), s->size, s->reserved);
894 if (max_objects > MAX_OBJS_PER_PAGE)
895 max_objects = MAX_OBJS_PER_PAGE;
897 if (page->objects != max_objects) {
898 slab_err(s, page, "Wrong number of objects. Found %d but "
899 "should be %d", page->objects, max_objects);
900 page->objects = max_objects;
901 slab_fix(s, "Number of objects adjusted.");
903 if (page->inuse != page->objects - nr) {
904 slab_err(s, page, "Wrong object count. Counter is %d but "
905 "counted were %d", page->inuse, page->objects - nr);
906 page->inuse = page->objects - nr;
907 slab_fix(s, "Object count adjusted.");
909 return search == NULL;
912 static void trace(struct kmem_cache *s, struct page *page, void *object,
915 if (s->flags & SLAB_TRACE) {
916 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
918 alloc ? "alloc" : "free",
923 print_section("Object ", (void *)object, s->objsize);
930 * Hooks for other subsystems that check memory allocations. In a typical
931 * production configuration these hooks all should produce no code at all.
933 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
935 flags &= gfp_allowed_mask;
936 lockdep_trace_alloc(flags);
937 might_sleep_if(flags & __GFP_WAIT);
939 return should_failslab(s->objsize, flags, s->flags);
942 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
944 flags &= gfp_allowed_mask;
945 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
946 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
949 static inline void slab_free_hook(struct kmem_cache *s, void *x)
951 kmemleak_free_recursive(x, s->flags);
954 * Trouble is that we may no longer disable interupts in the fast path
955 * So in order to make the debug calls that expect irqs to be
956 * disabled we need to disable interrupts temporarily.
958 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
962 local_irq_save(flags);
963 kmemcheck_slab_free(s, x, s->objsize);
964 debug_check_no_locks_freed(x, s->objsize);
965 local_irq_restore(flags);
968 if (!(s->flags & SLAB_DEBUG_OBJECTS))
969 debug_check_no_obj_freed(x, s->objsize);
973 * Tracking of fully allocated slabs for debugging purposes.
975 * list_lock must be held.
977 static void add_full(struct kmem_cache *s,
978 struct kmem_cache_node *n, struct page *page)
980 if (!(s->flags & SLAB_STORE_USER))
983 list_add(&page->lru, &n->full);
987 * list_lock must be held.
989 static void remove_full(struct kmem_cache *s, struct page *page)
991 if (!(s->flags & SLAB_STORE_USER))
994 list_del(&page->lru);
997 /* Tracking of the number of slabs for debugging purposes */
998 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1000 struct kmem_cache_node *n = get_node(s, node);
1002 return atomic_long_read(&n->nr_slabs);
1005 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1007 return atomic_long_read(&n->nr_slabs);
1010 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1012 struct kmem_cache_node *n = get_node(s, node);
1015 * May be called early in order to allocate a slab for the
1016 * kmem_cache_node structure. Solve the chicken-egg
1017 * dilemma by deferring the increment of the count during
1018 * bootstrap (see early_kmem_cache_node_alloc).
1021 atomic_long_inc(&n->nr_slabs);
1022 atomic_long_add(objects, &n->total_objects);
1025 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1027 struct kmem_cache_node *n = get_node(s, node);
1029 atomic_long_dec(&n->nr_slabs);
1030 atomic_long_sub(objects, &n->total_objects);
1033 /* Object debug checks for alloc/free paths */
1034 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1037 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1040 init_object(s, object, SLUB_RED_INACTIVE);
1041 init_tracking(s, object);
1044 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1045 void *object, unsigned long addr)
1047 if (!check_slab(s, page))
1050 if (!check_valid_pointer(s, page, object)) {
1051 object_err(s, page, object, "Freelist Pointer check fails");
1055 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1058 /* Success perform special debug activities for allocs */
1059 if (s->flags & SLAB_STORE_USER)
1060 set_track(s, object, TRACK_ALLOC, addr);
1061 trace(s, page, object, 1);
1062 init_object(s, object, SLUB_RED_ACTIVE);
1066 if (PageSlab(page)) {
1068 * If this is a slab page then lets do the best we can
1069 * to avoid issues in the future. Marking all objects
1070 * as used avoids touching the remaining objects.
1072 slab_fix(s, "Marking all objects used");
1073 page->inuse = page->objects;
1074 page->freelist = NULL;
1079 static noinline int free_debug_processing(struct kmem_cache *s,
1080 struct page *page, void *object, unsigned long addr)
1082 unsigned long flags;
1085 local_irq_save(flags);
1088 if (!check_slab(s, page))
1091 if (!check_valid_pointer(s, page, object)) {
1092 slab_err(s, page, "Invalid object pointer 0x%p", object);
1096 if (on_freelist(s, page, object)) {
1097 object_err(s, page, object, "Object already free");
1101 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1104 if (unlikely(s != page->slab)) {
1105 if (!PageSlab(page)) {
1106 slab_err(s, page, "Attempt to free object(0x%p) "
1107 "outside of slab", object);
1108 } else if (!page->slab) {
1110 "SLUB <none>: no slab for object 0x%p.\n",
1114 object_err(s, page, object,
1115 "page slab pointer corrupt.");
1119 if (s->flags & SLAB_STORE_USER)
1120 set_track(s, object, TRACK_FREE, addr);
1121 trace(s, page, object, 0);
1122 init_object(s, object, SLUB_RED_INACTIVE);
1126 local_irq_restore(flags);
1130 slab_fix(s, "Object at 0x%p not freed", object);
1134 static int __init setup_slub_debug(char *str)
1136 slub_debug = DEBUG_DEFAULT_FLAGS;
1137 if (*str++ != '=' || !*str)
1139 * No options specified. Switch on full debugging.
1145 * No options but restriction on slabs. This means full
1146 * debugging for slabs matching a pattern.
1150 if (tolower(*str) == 'o') {
1152 * Avoid enabling debugging on caches if its minimum order
1153 * would increase as a result.
1155 disable_higher_order_debug = 1;
1162 * Switch off all debugging measures.
1167 * Determine which debug features should be switched on
1169 for (; *str && *str != ','; str++) {
1170 switch (tolower(*str)) {
1172 slub_debug |= SLAB_DEBUG_FREE;
1175 slub_debug |= SLAB_RED_ZONE;
1178 slub_debug |= SLAB_POISON;
1181 slub_debug |= SLAB_STORE_USER;
1184 slub_debug |= SLAB_TRACE;
1187 slub_debug |= SLAB_FAILSLAB;
1190 printk(KERN_ERR "slub_debug option '%c' "
1191 "unknown. skipped\n", *str);
1197 slub_debug_slabs = str + 1;
1202 __setup("slub_debug", setup_slub_debug);
1204 static unsigned long kmem_cache_flags(unsigned long objsize,
1205 unsigned long flags, const char *name,
1206 void (*ctor)(void *))
1209 * Enable debugging if selected on the kernel commandline.
1211 if (slub_debug && (!slub_debug_slabs ||
1212 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1213 flags |= slub_debug;
1218 static inline void setup_object_debug(struct kmem_cache *s,
1219 struct page *page, void *object) {}
1221 static inline int alloc_debug_processing(struct kmem_cache *s,
1222 struct page *page, void *object, unsigned long addr) { return 0; }
1224 static inline int free_debug_processing(struct kmem_cache *s,
1225 struct page *page, void *object, unsigned long addr) { return 0; }
1227 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1229 static inline int check_object(struct kmem_cache *s, struct page *page,
1230 void *object, u8 val) { return 1; }
1231 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1232 struct page *page) {}
1233 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1234 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1235 unsigned long flags, const char *name,
1236 void (*ctor)(void *))
1240 #define slub_debug 0
1242 #define disable_higher_order_debug 0
1244 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1246 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1248 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1250 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1253 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1256 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1259 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1261 #endif /* CONFIG_SLUB_DEBUG */
1264 * Slab allocation and freeing
1266 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1267 struct kmem_cache_order_objects oo)
1269 int order = oo_order(oo);
1271 flags |= __GFP_NOTRACK;
1273 if (node == NUMA_NO_NODE)
1274 return alloc_pages(flags, order);
1276 return alloc_pages_exact_node(node, flags, order);
1279 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1282 struct kmem_cache_order_objects oo = s->oo;
1285 flags &= gfp_allowed_mask;
1287 if (flags & __GFP_WAIT)
1290 flags |= s->allocflags;
1293 * Let the initial higher-order allocation fail under memory pressure
1294 * so we fall-back to the minimum order allocation.
1296 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1298 page = alloc_slab_page(alloc_gfp, node, oo);
1299 if (unlikely(!page)) {
1302 * Allocation may have failed due to fragmentation.
1303 * Try a lower order alloc if possible
1305 page = alloc_slab_page(flags, node, oo);
1308 stat(s, ORDER_FALLBACK);
1311 if (flags & __GFP_WAIT)
1312 local_irq_disable();
1317 if (kmemcheck_enabled
1318 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1319 int pages = 1 << oo_order(oo);
1321 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1324 * Objects from caches that have a constructor don't get
1325 * cleared when they're allocated, so we need to do it here.
1328 kmemcheck_mark_uninitialized_pages(page, pages);
1330 kmemcheck_mark_unallocated_pages(page, pages);
1333 page->objects = oo_objects(oo);
1334 mod_zone_page_state(page_zone(page),
1335 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1336 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1342 static void setup_object(struct kmem_cache *s, struct page *page,
1345 setup_object_debug(s, page, object);
1346 if (unlikely(s->ctor))
1350 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1357 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1359 page = allocate_slab(s,
1360 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1364 inc_slabs_node(s, page_to_nid(page), page->objects);
1366 page->flags |= 1 << PG_slab;
1368 start = page_address(page);
1370 if (unlikely(s->flags & SLAB_POISON))
1371 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1374 for_each_object(p, s, start, page->objects) {
1375 setup_object(s, page, last);
1376 set_freepointer(s, last, p);
1379 setup_object(s, page, last);
1380 set_freepointer(s, last, NULL);
1382 page->freelist = start;
1383 page->inuse = page->objects;
1389 static void __free_slab(struct kmem_cache *s, struct page *page)
1391 int order = compound_order(page);
1392 int pages = 1 << order;
1394 if (kmem_cache_debug(s)) {
1397 slab_pad_check(s, page);
1398 for_each_object(p, s, page_address(page),
1400 check_object(s, page, p, SLUB_RED_INACTIVE);
1403 kmemcheck_free_shadow(page, compound_order(page));
1405 mod_zone_page_state(page_zone(page),
1406 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1407 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1410 __ClearPageSlab(page);
1411 reset_page_mapcount(page);
1412 if (current->reclaim_state)
1413 current->reclaim_state->reclaimed_slab += pages;
1414 __free_pages(page, order);
1417 #define need_reserve_slab_rcu \
1418 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1420 static void rcu_free_slab(struct rcu_head *h)
1424 if (need_reserve_slab_rcu)
1425 page = virt_to_head_page(h);
1427 page = container_of((struct list_head *)h, struct page, lru);
1429 __free_slab(page->slab, page);
1432 static void free_slab(struct kmem_cache *s, struct page *page)
1434 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1435 struct rcu_head *head;
1437 if (need_reserve_slab_rcu) {
1438 int order = compound_order(page);
1439 int offset = (PAGE_SIZE << order) - s->reserved;
1441 VM_BUG_ON(s->reserved != sizeof(*head));
1442 head = page_address(page) + offset;
1445 * RCU free overloads the RCU head over the LRU
1447 head = (void *)&page->lru;
1450 call_rcu(head, rcu_free_slab);
1452 __free_slab(s, page);
1455 static void discard_slab(struct kmem_cache *s, struct page *page)
1457 dec_slabs_node(s, page_to_nid(page), page->objects);
1462 * Management of partially allocated slabs.
1464 * list_lock must be held.
1466 static inline void add_partial(struct kmem_cache_node *n,
1467 struct page *page, int tail)
1470 if (tail == DEACTIVATE_TO_TAIL)
1471 list_add_tail(&page->lru, &n->partial);
1473 list_add(&page->lru, &n->partial);
1477 * list_lock must be held.
1479 static inline void remove_partial(struct kmem_cache_node *n,
1482 list_del(&page->lru);
1487 * Lock slab, remove from the partial list and put the object into the
1490 * Returns a list of objects or NULL if it fails.
1492 * Must hold list_lock.
1494 static inline void *acquire_slab(struct kmem_cache *s,
1495 struct kmem_cache_node *n, struct page *page,
1499 unsigned long counters;
1503 * Zap the freelist and set the frozen bit.
1504 * The old freelist is the list of objects for the
1505 * per cpu allocation list.
1508 freelist = page->freelist;
1509 counters = page->counters;
1510 new.counters = counters;
1512 new.inuse = page->objects;
1514 VM_BUG_ON(new.frozen);
1517 } while (!__cmpxchg_double_slab(s, page,
1520 "lock and freeze"));
1522 remove_partial(n, page);
1526 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1529 * Try to allocate a partial slab from a specific node.
1531 static void *get_partial_node(struct kmem_cache *s,
1532 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1534 struct page *page, *page2;
1535 void *object = NULL;
1538 * Racy check. If we mistakenly see no partial slabs then we
1539 * just allocate an empty slab. If we mistakenly try to get a
1540 * partial slab and there is none available then get_partials()
1543 if (!n || !n->nr_partial)
1546 spin_lock(&n->list_lock);
1547 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1548 void *t = acquire_slab(s, n, page, object == NULL);
1556 c->node = page_to_nid(page);
1557 stat(s, ALLOC_FROM_PARTIAL);
1559 available = page->objects - page->inuse;
1562 available = put_cpu_partial(s, page, 0);
1564 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1568 spin_unlock(&n->list_lock);
1573 * Get a page from somewhere. Search in increasing NUMA distances.
1575 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1576 struct kmem_cache_cpu *c)
1579 struct zonelist *zonelist;
1582 enum zone_type high_zoneidx = gfp_zone(flags);
1586 * The defrag ratio allows a configuration of the tradeoffs between
1587 * inter node defragmentation and node local allocations. A lower
1588 * defrag_ratio increases the tendency to do local allocations
1589 * instead of attempting to obtain partial slabs from other nodes.
1591 * If the defrag_ratio is set to 0 then kmalloc() always
1592 * returns node local objects. If the ratio is higher then kmalloc()
1593 * may return off node objects because partial slabs are obtained
1594 * from other nodes and filled up.
1596 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1597 * defrag_ratio = 1000) then every (well almost) allocation will
1598 * first attempt to defrag slab caches on other nodes. This means
1599 * scanning over all nodes to look for partial slabs which may be
1600 * expensive if we do it every time we are trying to find a slab
1601 * with available objects.
1603 if (!s->remote_node_defrag_ratio ||
1604 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1608 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1609 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1610 struct kmem_cache_node *n;
1612 n = get_node(s, zone_to_nid(zone));
1614 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1615 n->nr_partial > s->min_partial) {
1616 object = get_partial_node(s, n, c);
1629 * Get a partial page, lock it and return it.
1631 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1632 struct kmem_cache_cpu *c)
1635 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1637 object = get_partial_node(s, get_node(s, searchnode), c);
1638 if (object || node != NUMA_NO_NODE)
1641 return get_any_partial(s, flags, c);
1644 #ifdef CONFIG_PREEMPT
1646 * Calculate the next globally unique transaction for disambiguiation
1647 * during cmpxchg. The transactions start with the cpu number and are then
1648 * incremented by CONFIG_NR_CPUS.
1650 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1653 * No preemption supported therefore also no need to check for
1659 static inline unsigned long next_tid(unsigned long tid)
1661 return tid + TID_STEP;
1664 static inline unsigned int tid_to_cpu(unsigned long tid)
1666 return tid % TID_STEP;
1669 static inline unsigned long tid_to_event(unsigned long tid)
1671 return tid / TID_STEP;
1674 static inline unsigned int init_tid(int cpu)
1679 static inline void note_cmpxchg_failure(const char *n,
1680 const struct kmem_cache *s, unsigned long tid)
1682 #ifdef SLUB_DEBUG_CMPXCHG
1683 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1685 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1687 #ifdef CONFIG_PREEMPT
1688 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1689 printk("due to cpu change %d -> %d\n",
1690 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1693 if (tid_to_event(tid) != tid_to_event(actual_tid))
1694 printk("due to cpu running other code. Event %ld->%ld\n",
1695 tid_to_event(tid), tid_to_event(actual_tid));
1697 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1698 actual_tid, tid, next_tid(tid));
1700 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1703 void init_kmem_cache_cpus(struct kmem_cache *s)
1707 for_each_possible_cpu(cpu)
1708 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1712 * Remove the cpu slab
1714 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1716 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1717 struct page *page = c->page;
1718 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1720 enum slab_modes l = M_NONE, m = M_NONE;
1723 int tail = DEACTIVATE_TO_HEAD;
1727 if (page->freelist) {
1728 stat(s, DEACTIVATE_REMOTE_FREES);
1729 tail = DEACTIVATE_TO_TAIL;
1732 c->tid = next_tid(c->tid);
1734 freelist = c->freelist;
1738 * Stage one: Free all available per cpu objects back
1739 * to the page freelist while it is still frozen. Leave the
1742 * There is no need to take the list->lock because the page
1745 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1747 unsigned long counters;
1750 prior = page->freelist;
1751 counters = page->counters;
1752 set_freepointer(s, freelist, prior);
1753 new.counters = counters;
1755 VM_BUG_ON(!new.frozen);
1757 } while (!__cmpxchg_double_slab(s, page,
1759 freelist, new.counters,
1760 "drain percpu freelist"));
1762 freelist = nextfree;
1766 * Stage two: Ensure that the page is unfrozen while the
1767 * list presence reflects the actual number of objects
1770 * We setup the list membership and then perform a cmpxchg
1771 * with the count. If there is a mismatch then the page
1772 * is not unfrozen but the page is on the wrong list.
1774 * Then we restart the process which may have to remove
1775 * the page from the list that we just put it on again
1776 * because the number of objects in the slab may have
1781 old.freelist = page->freelist;
1782 old.counters = page->counters;
1783 VM_BUG_ON(!old.frozen);
1785 /* Determine target state of the slab */
1786 new.counters = old.counters;
1789 set_freepointer(s, freelist, old.freelist);
1790 new.freelist = freelist;
1792 new.freelist = old.freelist;
1796 if (!new.inuse && n->nr_partial > s->min_partial)
1798 else if (new.freelist) {
1803 * Taking the spinlock removes the possiblity
1804 * that acquire_slab() will see a slab page that
1807 spin_lock(&n->list_lock);
1811 if (kmem_cache_debug(s) && !lock) {
1814 * This also ensures that the scanning of full
1815 * slabs from diagnostic functions will not see
1818 spin_lock(&n->list_lock);
1826 remove_partial(n, page);
1828 else if (l == M_FULL)
1830 remove_full(s, page);
1832 if (m == M_PARTIAL) {
1834 add_partial(n, page, tail);
1837 } else if (m == M_FULL) {
1839 stat(s, DEACTIVATE_FULL);
1840 add_full(s, n, page);
1846 if (!__cmpxchg_double_slab(s, page,
1847 old.freelist, old.counters,
1848 new.freelist, new.counters,
1853 spin_unlock(&n->list_lock);
1856 stat(s, DEACTIVATE_EMPTY);
1857 discard_slab(s, page);
1862 /* Unfreeze all the cpu partial slabs */
1863 static void unfreeze_partials(struct kmem_cache *s)
1865 struct kmem_cache_node *n = NULL;
1866 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1867 struct page *page, *discard_page = NULL;
1869 while ((page = c->partial)) {
1870 enum slab_modes { M_PARTIAL, M_FREE };
1871 enum slab_modes l, m;
1875 c->partial = page->next;
1880 old.freelist = page->freelist;
1881 old.counters = page->counters;
1882 VM_BUG_ON(!old.frozen);
1884 new.counters = old.counters;
1885 new.freelist = old.freelist;
1889 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1892 struct kmem_cache_node *n2 = get_node(s,
1898 spin_unlock(&n->list_lock);
1901 spin_lock(&n->list_lock);
1906 if (l == M_PARTIAL) {
1907 remove_partial(n, page);
1908 stat(s, FREE_REMOVE_PARTIAL);
1910 add_partial(n, page,
1911 DEACTIVATE_TO_TAIL);
1912 stat(s, FREE_ADD_PARTIAL);
1918 } while (!cmpxchg_double_slab(s, page,
1919 old.freelist, old.counters,
1920 new.freelist, new.counters,
1921 "unfreezing slab"));
1924 page->next = discard_page;
1925 discard_page = page;
1930 spin_unlock(&n->list_lock);
1932 while (discard_page) {
1933 page = discard_page;
1934 discard_page = discard_page->next;
1936 stat(s, DEACTIVATE_EMPTY);
1937 discard_slab(s, page);
1943 * Put a page that was just frozen (in __slab_free) into a partial page
1944 * slot if available. This is done without interrupts disabled and without
1945 * preemption disabled. The cmpxchg is racy and may put the partial page
1946 * onto a random cpus partial slot.
1948 * If we did not find a slot then simply move all the partials to the
1949 * per node partial list.
1951 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1953 struct page *oldpage;
1960 oldpage = this_cpu_read(s->cpu_slab->partial);
1963 pobjects = oldpage->pobjects;
1964 pages = oldpage->pages;
1965 if (drain && pobjects > s->cpu_partial) {
1966 unsigned long flags;
1968 * partial array is full. Move the existing
1969 * set to the per node partial list.
1971 local_irq_save(flags);
1972 unfreeze_partials(s);
1973 local_irq_restore(flags);
1980 pobjects += page->objects - page->inuse;
1982 page->pages = pages;
1983 page->pobjects = pobjects;
1984 page->next = oldpage;
1986 } while (irqsafe_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1987 stat(s, CPU_PARTIAL_FREE);
1991 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1993 stat(s, CPUSLAB_FLUSH);
1994 deactivate_slab(s, c);
2000 * Called from IPI handler with interrupts disabled.
2002 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2004 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2010 unfreeze_partials(s);
2014 static void flush_cpu_slab(void *d)
2016 struct kmem_cache *s = d;
2018 __flush_cpu_slab(s, smp_processor_id());
2021 static void flush_all(struct kmem_cache *s)
2023 on_each_cpu(flush_cpu_slab, s, 1);
2027 * Check if the objects in a per cpu structure fit numa
2028 * locality expectations.
2030 static inline int node_match(struct kmem_cache_cpu *c, int node)
2033 if (node != NUMA_NO_NODE && c->node != node)
2039 static int count_free(struct page *page)
2041 return page->objects - page->inuse;
2044 static unsigned long count_partial(struct kmem_cache_node *n,
2045 int (*get_count)(struct page *))
2047 unsigned long flags;
2048 unsigned long x = 0;
2051 spin_lock_irqsave(&n->list_lock, flags);
2052 list_for_each_entry(page, &n->partial, lru)
2053 x += get_count(page);
2054 spin_unlock_irqrestore(&n->list_lock, flags);
2058 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2060 #ifdef CONFIG_SLUB_DEBUG
2061 return atomic_long_read(&n->total_objects);
2067 static noinline void
2068 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2073 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2075 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2076 "default order: %d, min order: %d\n", s->name, s->objsize,
2077 s->size, oo_order(s->oo), oo_order(s->min));
2079 if (oo_order(s->min) > get_order(s->objsize))
2080 printk(KERN_WARNING " %s debugging increased min order, use "
2081 "slub_debug=O to disable.\n", s->name);
2083 for_each_online_node(node) {
2084 struct kmem_cache_node *n = get_node(s, node);
2085 unsigned long nr_slabs;
2086 unsigned long nr_objs;
2087 unsigned long nr_free;
2092 nr_free = count_partial(n, count_free);
2093 nr_slabs = node_nr_slabs(n);
2094 nr_objs = node_nr_objs(n);
2097 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2098 node, nr_slabs, nr_objs, nr_free);
2102 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2103 int node, struct kmem_cache_cpu **pc)
2106 struct kmem_cache_cpu *c;
2107 struct page *page = new_slab(s, flags, node);
2110 c = __this_cpu_ptr(s->cpu_slab);
2115 * No other reference to the page yet so we can
2116 * muck around with it freely without cmpxchg
2118 object = page->freelist;
2119 page->freelist = NULL;
2121 stat(s, ALLOC_SLAB);
2122 c->node = page_to_nid(page);
2132 * Slow path. The lockless freelist is empty or we need to perform
2135 * Processing is still very fast if new objects have been freed to the
2136 * regular freelist. In that case we simply take over the regular freelist
2137 * as the lockless freelist and zap the regular freelist.
2139 * If that is not working then we fall back to the partial lists. We take the
2140 * first element of the freelist as the object to allocate now and move the
2141 * rest of the freelist to the lockless freelist.
2143 * And if we were unable to get a new slab from the partial slab lists then
2144 * we need to allocate a new slab. This is the slowest path since it involves
2145 * a call to the page allocator and the setup of a new slab.
2147 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2148 unsigned long addr, struct kmem_cache_cpu *c)
2151 unsigned long flags;
2153 unsigned long counters;
2155 local_irq_save(flags);
2156 #ifdef CONFIG_PREEMPT
2158 * We may have been preempted and rescheduled on a different
2159 * cpu before disabling interrupts. Need to reload cpu area
2162 c = this_cpu_ptr(s->cpu_slab);
2168 if (unlikely(!node_match(c, node))) {
2169 stat(s, ALLOC_NODE_MISMATCH);
2170 deactivate_slab(s, c);
2174 stat(s, ALLOC_SLOWPATH);
2177 object = c->page->freelist;
2178 counters = c->page->counters;
2179 new.counters = counters;
2180 VM_BUG_ON(!new.frozen);
2183 * If there is no object left then we use this loop to
2184 * deactivate the slab which is simple since no objects
2185 * are left in the slab and therefore we do not need to
2186 * put the page back onto the partial list.
2188 * If there are objects left then we retrieve them
2189 * and use them to refill the per cpu queue.
2192 new.inuse = c->page->objects;
2193 new.frozen = object != NULL;
2195 } while (!__cmpxchg_double_slab(s, c->page,
2202 stat(s, DEACTIVATE_BYPASS);
2206 stat(s, ALLOC_REFILL);
2209 c->freelist = get_freepointer(s, object);
2210 c->tid = next_tid(c->tid);
2211 local_irq_restore(flags);
2217 c->page = c->partial;
2218 c->partial = c->page->next;
2219 c->node = page_to_nid(c->page);
2220 stat(s, CPU_PARTIAL_ALLOC);
2225 /* Then do expensive stuff like retrieving pages from the partial lists */
2226 object = get_partial(s, gfpflags, node, c);
2228 if (unlikely(!object)) {
2230 object = new_slab_objects(s, gfpflags, node, &c);
2232 if (unlikely(!object)) {
2233 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2234 slab_out_of_memory(s, gfpflags, node);
2236 local_irq_restore(flags);
2241 if (likely(!kmem_cache_debug(s)))
2244 /* Only entered in the debug case */
2245 if (!alloc_debug_processing(s, c->page, object, addr))
2246 goto new_slab; /* Slab failed checks. Next slab needed */
2248 c->freelist = get_freepointer(s, object);
2249 deactivate_slab(s, c);
2250 c->node = NUMA_NO_NODE;
2251 local_irq_restore(flags);
2256 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2257 * have the fastpath folded into their functions. So no function call
2258 * overhead for requests that can be satisfied on the fastpath.
2260 * The fastpath works by first checking if the lockless freelist can be used.
2261 * If not then __slab_alloc is called for slow processing.
2263 * Otherwise we can simply pick the next object from the lockless free list.
2265 static __always_inline void *slab_alloc(struct kmem_cache *s,
2266 gfp_t gfpflags, int node, unsigned long addr)
2269 struct kmem_cache_cpu *c;
2272 if (slab_pre_alloc_hook(s, gfpflags))
2278 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2279 * enabled. We may switch back and forth between cpus while
2280 * reading from one cpu area. That does not matter as long
2281 * as we end up on the original cpu again when doing the cmpxchg.
2283 c = __this_cpu_ptr(s->cpu_slab);
2286 * The transaction ids are globally unique per cpu and per operation on
2287 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2288 * occurs on the right processor and that there was no operation on the
2289 * linked list in between.
2294 object = c->freelist;
2295 if (unlikely(!object || !node_match(c, node)))
2297 object = __slab_alloc(s, gfpflags, node, addr, c);
2301 * The cmpxchg will only match if there was no additional
2302 * operation and if we are on the right processor.
2304 * The cmpxchg does the following atomically (without lock semantics!)
2305 * 1. Relocate first pointer to the current per cpu area.
2306 * 2. Verify that tid and freelist have not been changed
2307 * 3. If they were not changed replace tid and freelist
2309 * Since this is without lock semantics the protection is only against
2310 * code executing on this cpu *not* from access by other cpus.
2312 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2313 s->cpu_slab->freelist, s->cpu_slab->tid,
2315 get_freepointer_safe(s, object), next_tid(tid)))) {
2317 note_cmpxchg_failure("slab_alloc", s, tid);
2320 stat(s, ALLOC_FASTPATH);
2323 if (unlikely(gfpflags & __GFP_ZERO) && object)
2324 memset(object, 0, s->objsize);
2326 slab_post_alloc_hook(s, gfpflags, object);
2331 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2333 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2335 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2339 EXPORT_SYMBOL(kmem_cache_alloc);
2341 #ifdef CONFIG_TRACING
2342 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2344 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2345 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2348 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2350 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2352 void *ret = kmalloc_order(size, flags, order);
2353 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2356 EXPORT_SYMBOL(kmalloc_order_trace);
2360 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2362 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2364 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2365 s->objsize, s->size, gfpflags, node);
2369 EXPORT_SYMBOL(kmem_cache_alloc_node);
2371 #ifdef CONFIG_TRACING
2372 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2374 int node, size_t size)
2376 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2378 trace_kmalloc_node(_RET_IP_, ret,
2379 size, s->size, gfpflags, node);
2382 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2387 * Slow patch handling. This may still be called frequently since objects
2388 * have a longer lifetime than the cpu slabs in most processing loads.
2390 * So we still attempt to reduce cache line usage. Just take the slab
2391 * lock and free the item. If there is no additional partial page
2392 * handling required then we can return immediately.
2394 static void __slab_free(struct kmem_cache *s, struct page *page,
2395 void *x, unsigned long addr)
2398 void **object = (void *)x;
2402 unsigned long counters;
2403 struct kmem_cache_node *n = NULL;
2404 unsigned long uninitialized_var(flags);
2406 stat(s, FREE_SLOWPATH);
2408 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2412 prior = page->freelist;
2413 counters = page->counters;
2414 set_freepointer(s, object, prior);
2415 new.counters = counters;
2416 was_frozen = new.frozen;
2418 if ((!new.inuse || !prior) && !was_frozen && !n) {
2420 if (!kmem_cache_debug(s) && !prior)
2423 * Slab was on no list before and will be partially empty
2424 * We can defer the list move and instead freeze it.
2428 else { /* Needs to be taken off a list */
2430 n = get_node(s, page_to_nid(page));
2432 * Speculatively acquire the list_lock.
2433 * If the cmpxchg does not succeed then we may
2434 * drop the list_lock without any processing.
2436 * Otherwise the list_lock will synchronize with
2437 * other processors updating the list of slabs.
2439 spin_lock_irqsave(&n->list_lock, flags);
2445 } while (!cmpxchg_double_slab(s, page,
2447 object, new.counters,
2453 * If we just froze the page then put it onto the
2454 * per cpu partial list.
2456 if (new.frozen && !was_frozen)
2457 put_cpu_partial(s, page, 1);
2460 * The list lock was not taken therefore no list
2461 * activity can be necessary.
2464 stat(s, FREE_FROZEN);
2469 * was_frozen may have been set after we acquired the list_lock in
2470 * an earlier loop. So we need to check it here again.
2473 stat(s, FREE_FROZEN);
2475 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2479 * Objects left in the slab. If it was not on the partial list before
2482 if (unlikely(!prior)) {
2483 remove_full(s, page);
2484 add_partial(n, page, DEACTIVATE_TO_TAIL);
2485 stat(s, FREE_ADD_PARTIAL);
2488 spin_unlock_irqrestore(&n->list_lock, flags);
2494 * Slab on the partial list.
2496 remove_partial(n, page);
2497 stat(s, FREE_REMOVE_PARTIAL);
2499 /* Slab must be on the full list */
2500 remove_full(s, page);
2502 spin_unlock_irqrestore(&n->list_lock, flags);
2504 discard_slab(s, page);
2508 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2509 * can perform fastpath freeing without additional function calls.
2511 * The fastpath is only possible if we are freeing to the current cpu slab
2512 * of this processor. This typically the case if we have just allocated
2515 * If fastpath is not possible then fall back to __slab_free where we deal
2516 * with all sorts of special processing.
2518 static __always_inline void slab_free(struct kmem_cache *s,
2519 struct page *page, void *x, unsigned long addr)
2521 void **object = (void *)x;
2522 struct kmem_cache_cpu *c;
2525 slab_free_hook(s, x);
2529 * Determine the currently cpus per cpu slab.
2530 * The cpu may change afterward. However that does not matter since
2531 * data is retrieved via this pointer. If we are on the same cpu
2532 * during the cmpxchg then the free will succedd.
2534 c = __this_cpu_ptr(s->cpu_slab);
2539 if (likely(page == c->page)) {
2540 set_freepointer(s, object, c->freelist);
2542 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2543 s->cpu_slab->freelist, s->cpu_slab->tid,
2545 object, next_tid(tid)))) {
2547 note_cmpxchg_failure("slab_free", s, tid);
2550 stat(s, FREE_FASTPATH);
2552 __slab_free(s, page, x, addr);
2556 void kmem_cache_free(struct kmem_cache *s, void *x)
2560 page = virt_to_head_page(x);
2562 slab_free(s, page, x, _RET_IP_);
2564 trace_kmem_cache_free(_RET_IP_, x);
2566 EXPORT_SYMBOL(kmem_cache_free);
2569 * Object placement in a slab is made very easy because we always start at
2570 * offset 0. If we tune the size of the object to the alignment then we can
2571 * get the required alignment by putting one properly sized object after
2574 * Notice that the allocation order determines the sizes of the per cpu
2575 * caches. Each processor has always one slab available for allocations.
2576 * Increasing the allocation order reduces the number of times that slabs
2577 * must be moved on and off the partial lists and is therefore a factor in
2582 * Mininum / Maximum order of slab pages. This influences locking overhead
2583 * and slab fragmentation. A higher order reduces the number of partial slabs
2584 * and increases the number of allocations possible without having to
2585 * take the list_lock.
2587 static int slub_min_order;
2588 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2589 static int slub_min_objects;
2592 * Merge control. If this is set then no merging of slab caches will occur.
2593 * (Could be removed. This was introduced to pacify the merge skeptics.)
2595 static int slub_nomerge;
2598 * Calculate the order of allocation given an slab object size.
2600 * The order of allocation has significant impact on performance and other
2601 * system components. Generally order 0 allocations should be preferred since
2602 * order 0 does not cause fragmentation in the page allocator. Larger objects
2603 * be problematic to put into order 0 slabs because there may be too much
2604 * unused space left. We go to a higher order if more than 1/16th of the slab
2607 * In order to reach satisfactory performance we must ensure that a minimum
2608 * number of objects is in one slab. Otherwise we may generate too much
2609 * activity on the partial lists which requires taking the list_lock. This is
2610 * less a concern for large slabs though which are rarely used.
2612 * slub_max_order specifies the order where we begin to stop considering the
2613 * number of objects in a slab as critical. If we reach slub_max_order then
2614 * we try to keep the page order as low as possible. So we accept more waste
2615 * of space in favor of a small page order.
2617 * Higher order allocations also allow the placement of more objects in a
2618 * slab and thereby reduce object handling overhead. If the user has
2619 * requested a higher mininum order then we start with that one instead of
2620 * the smallest order which will fit the object.
2622 static inline int slab_order(int size, int min_objects,
2623 int max_order, int fract_leftover, int reserved)
2627 int min_order = slub_min_order;
2629 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2630 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2632 for (order = max(min_order,
2633 fls(min_objects * size - 1) - PAGE_SHIFT);
2634 order <= max_order; order++) {
2636 unsigned long slab_size = PAGE_SIZE << order;
2638 if (slab_size < min_objects * size + reserved)
2641 rem = (slab_size - reserved) % size;
2643 if (rem <= slab_size / fract_leftover)
2651 static inline int calculate_order(int size, int reserved)
2659 * Attempt to find best configuration for a slab. This
2660 * works by first attempting to generate a layout with
2661 * the best configuration and backing off gradually.
2663 * First we reduce the acceptable waste in a slab. Then
2664 * we reduce the minimum objects required in a slab.
2666 min_objects = slub_min_objects;
2668 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2669 max_objects = order_objects(slub_max_order, size, reserved);
2670 min_objects = min(min_objects, max_objects);
2672 while (min_objects > 1) {
2674 while (fraction >= 4) {
2675 order = slab_order(size, min_objects,
2676 slub_max_order, fraction, reserved);
2677 if (order <= slub_max_order)
2685 * We were unable to place multiple objects in a slab. Now
2686 * lets see if we can place a single object there.
2688 order = slab_order(size, 1, slub_max_order, 1, reserved);
2689 if (order <= slub_max_order)
2693 * Doh this slab cannot be placed using slub_max_order.
2695 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2696 if (order < MAX_ORDER)
2702 * Figure out what the alignment of the objects will be.
2704 static unsigned long calculate_alignment(unsigned long flags,
2705 unsigned long align, unsigned long size)
2708 * If the user wants hardware cache aligned objects then follow that
2709 * suggestion if the object is sufficiently large.
2711 * The hardware cache alignment cannot override the specified
2712 * alignment though. If that is greater then use it.
2714 if (flags & SLAB_HWCACHE_ALIGN) {
2715 unsigned long ralign = cache_line_size();
2716 while (size <= ralign / 2)
2718 align = max(align, ralign);
2721 if (align < ARCH_SLAB_MINALIGN)
2722 align = ARCH_SLAB_MINALIGN;
2724 return ALIGN(align, sizeof(void *));
2728 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2731 spin_lock_init(&n->list_lock);
2732 INIT_LIST_HEAD(&n->partial);
2733 #ifdef CONFIG_SLUB_DEBUG
2734 atomic_long_set(&n->nr_slabs, 0);
2735 atomic_long_set(&n->total_objects, 0);
2736 INIT_LIST_HEAD(&n->full);
2740 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2742 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2743 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2746 * Must align to double word boundary for the double cmpxchg
2747 * instructions to work; see __pcpu_double_call_return_bool().
2749 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2750 2 * sizeof(void *));
2755 init_kmem_cache_cpus(s);
2760 static struct kmem_cache *kmem_cache_node;
2763 * No kmalloc_node yet so do it by hand. We know that this is the first
2764 * slab on the node for this slabcache. There are no concurrent accesses
2767 * Note that this function only works on the kmalloc_node_cache
2768 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2769 * memory on a fresh node that has no slab structures yet.
2771 static void early_kmem_cache_node_alloc(int node)
2774 struct kmem_cache_node *n;
2776 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2778 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2781 if (page_to_nid(page) != node) {
2782 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2784 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2785 "in order to be able to continue\n");
2790 page->freelist = get_freepointer(kmem_cache_node, n);
2793 kmem_cache_node->node[node] = n;
2794 #ifdef CONFIG_SLUB_DEBUG
2795 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2796 init_tracking(kmem_cache_node, n);
2798 init_kmem_cache_node(n, kmem_cache_node);
2799 inc_slabs_node(kmem_cache_node, node, page->objects);
2801 add_partial(n, page, DEACTIVATE_TO_HEAD);
2804 static void free_kmem_cache_nodes(struct kmem_cache *s)
2808 for_each_node_state(node, N_NORMAL_MEMORY) {
2809 struct kmem_cache_node *n = s->node[node];
2812 kmem_cache_free(kmem_cache_node, n);
2814 s->node[node] = NULL;
2818 static int init_kmem_cache_nodes(struct kmem_cache *s)
2822 for_each_node_state(node, N_NORMAL_MEMORY) {
2823 struct kmem_cache_node *n;
2825 if (slab_state == DOWN) {
2826 early_kmem_cache_node_alloc(node);
2829 n = kmem_cache_alloc_node(kmem_cache_node,
2833 free_kmem_cache_nodes(s);
2838 init_kmem_cache_node(n, s);
2843 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2845 if (min < MIN_PARTIAL)
2847 else if (min > MAX_PARTIAL)
2849 s->min_partial = min;
2853 * calculate_sizes() determines the order and the distribution of data within
2856 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2858 unsigned long flags = s->flags;
2859 unsigned long size = s->objsize;
2860 unsigned long align = s->align;
2864 * Round up object size to the next word boundary. We can only
2865 * place the free pointer at word boundaries and this determines
2866 * the possible location of the free pointer.
2868 size = ALIGN(size, sizeof(void *));
2870 #ifdef CONFIG_SLUB_DEBUG
2872 * Determine if we can poison the object itself. If the user of
2873 * the slab may touch the object after free or before allocation
2874 * then we should never poison the object itself.
2876 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2878 s->flags |= __OBJECT_POISON;
2880 s->flags &= ~__OBJECT_POISON;
2884 * If we are Redzoning then check if there is some space between the
2885 * end of the object and the free pointer. If not then add an
2886 * additional word to have some bytes to store Redzone information.
2888 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2889 size += sizeof(void *);
2893 * With that we have determined the number of bytes in actual use
2894 * by the object. This is the potential offset to the free pointer.
2898 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2901 * Relocate free pointer after the object if it is not
2902 * permitted to overwrite the first word of the object on
2905 * This is the case if we do RCU, have a constructor or
2906 * destructor or are poisoning the objects.
2909 size += sizeof(void *);
2912 #ifdef CONFIG_SLUB_DEBUG
2913 if (flags & SLAB_STORE_USER)
2915 * Need to store information about allocs and frees after
2918 size += 2 * sizeof(struct track);
2920 if (flags & SLAB_RED_ZONE)
2922 * Add some empty padding so that we can catch
2923 * overwrites from earlier objects rather than let
2924 * tracking information or the free pointer be
2925 * corrupted if a user writes before the start
2928 size += sizeof(void *);
2932 * Determine the alignment based on various parameters that the
2933 * user specified and the dynamic determination of cache line size
2936 align = calculate_alignment(flags, align, s->objsize);
2940 * SLUB stores one object immediately after another beginning from
2941 * offset 0. In order to align the objects we have to simply size
2942 * each object to conform to the alignment.
2944 size = ALIGN(size, align);
2946 if (forced_order >= 0)
2947 order = forced_order;
2949 order = calculate_order(size, s->reserved);
2956 s->allocflags |= __GFP_COMP;
2958 if (s->flags & SLAB_CACHE_DMA)
2959 s->allocflags |= SLUB_DMA;
2961 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2962 s->allocflags |= __GFP_RECLAIMABLE;
2965 * Determine the number of objects per slab
2967 s->oo = oo_make(order, size, s->reserved);
2968 s->min = oo_make(get_order(size), size, s->reserved);
2969 if (oo_objects(s->oo) > oo_objects(s->max))
2972 return !!oo_objects(s->oo);
2976 static int kmem_cache_open(struct kmem_cache *s,
2977 const char *name, size_t size,
2978 size_t align, unsigned long flags,
2979 void (*ctor)(void *))
2981 memset(s, 0, kmem_size);
2986 s->flags = kmem_cache_flags(size, flags, name, ctor);
2989 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2990 s->reserved = sizeof(struct rcu_head);
2992 if (!calculate_sizes(s, -1))
2994 if (disable_higher_order_debug) {
2996 * Disable debugging flags that store metadata if the min slab
2999 if (get_order(s->size) > get_order(s->objsize)) {
3000 s->flags &= ~DEBUG_METADATA_FLAGS;
3002 if (!calculate_sizes(s, -1))
3007 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3008 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3009 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3010 /* Enable fast mode */
3011 s->flags |= __CMPXCHG_DOUBLE;
3015 * The larger the object size is, the more pages we want on the partial
3016 * list to avoid pounding the page allocator excessively.
3018 set_min_partial(s, ilog2(s->size) / 2);
3021 * cpu_partial determined the maximum number of objects kept in the
3022 * per cpu partial lists of a processor.
3024 * Per cpu partial lists mainly contain slabs that just have one
3025 * object freed. If they are used for allocation then they can be
3026 * filled up again with minimal effort. The slab will never hit the
3027 * per node partial lists and therefore no locking will be required.
3029 * This setting also determines
3031 * A) The number of objects from per cpu partial slabs dumped to the
3032 * per node list when we reach the limit.
3033 * B) The number of objects in cpu partial slabs to extract from the
3034 * per node list when we run out of per cpu objects. We only fetch 50%
3035 * to keep some capacity around for frees.
3037 if (s->size >= PAGE_SIZE)
3039 else if (s->size >= 1024)
3041 else if (s->size >= 256)
3042 s->cpu_partial = 13;
3044 s->cpu_partial = 30;
3048 s->remote_node_defrag_ratio = 1000;
3050 if (!init_kmem_cache_nodes(s))
3053 if (alloc_kmem_cache_cpus(s))
3056 free_kmem_cache_nodes(s);
3058 if (flags & SLAB_PANIC)
3059 panic("Cannot create slab %s size=%lu realsize=%u "
3060 "order=%u offset=%u flags=%lx\n",
3061 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3067 * Determine the size of a slab object
3069 unsigned int kmem_cache_size(struct kmem_cache *s)
3073 EXPORT_SYMBOL(kmem_cache_size);
3075 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3078 #ifdef CONFIG_SLUB_DEBUG
3079 void *addr = page_address(page);
3081 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3082 sizeof(long), GFP_ATOMIC);
3085 slab_err(s, page, "%s", text);
3088 get_map(s, page, map);
3089 for_each_object(p, s, addr, page->objects) {
3091 if (!test_bit(slab_index(p, s, addr), map)) {
3092 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3094 print_tracking(s, p);
3103 * Attempt to free all partial slabs on a node.
3104 * This is called from kmem_cache_close(). We must be the last thread
3105 * using the cache and therefore we do not need to lock anymore.
3107 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3109 struct page *page, *h;
3111 list_for_each_entry_safe(page, h, &n->partial, lru) {
3113 remove_partial(n, page);
3114 discard_slab(s, page);
3116 list_slab_objects(s, page,
3117 "Objects remaining on kmem_cache_close()");
3123 * Release all resources used by a slab cache.
3125 static inline int kmem_cache_close(struct kmem_cache *s)
3130 free_percpu(s->cpu_slab);
3131 /* Attempt to free all objects */
3132 for_each_node_state(node, N_NORMAL_MEMORY) {
3133 struct kmem_cache_node *n = get_node(s, node);
3136 if (n->nr_partial || slabs_node(s, node))
3139 free_kmem_cache_nodes(s);
3144 * Close a cache and release the kmem_cache structure
3145 * (must be used for caches created using kmem_cache_create)
3147 void kmem_cache_destroy(struct kmem_cache *s)
3149 down_write(&slub_lock);
3153 up_write(&slub_lock);
3154 if (kmem_cache_close(s)) {
3155 printk(KERN_ERR "SLUB %s: %s called for cache that "
3156 "still has objects.\n", s->name, __func__);
3159 if (s->flags & SLAB_DESTROY_BY_RCU)
3161 sysfs_slab_remove(s);
3163 up_write(&slub_lock);
3165 EXPORT_SYMBOL(kmem_cache_destroy);
3167 /********************************************************************
3169 *******************************************************************/
3171 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3172 EXPORT_SYMBOL(kmalloc_caches);
3174 static struct kmem_cache *kmem_cache;
3176 #ifdef CONFIG_ZONE_DMA
3177 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3180 static int __init setup_slub_min_order(char *str)
3182 get_option(&str, &slub_min_order);
3187 __setup("slub_min_order=", setup_slub_min_order);
3189 static int __init setup_slub_max_order(char *str)
3191 get_option(&str, &slub_max_order);
3192 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3197 __setup("slub_max_order=", setup_slub_max_order);
3199 static int __init setup_slub_min_objects(char *str)
3201 get_option(&str, &slub_min_objects);
3206 __setup("slub_min_objects=", setup_slub_min_objects);
3208 static int __init setup_slub_nomerge(char *str)
3214 __setup("slub_nomerge", setup_slub_nomerge);
3216 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3217 int size, unsigned int flags)
3219 struct kmem_cache *s;
3221 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3224 * This function is called with IRQs disabled during early-boot on
3225 * single CPU so there's no need to take slub_lock here.
3227 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3231 list_add(&s->list, &slab_caches);
3235 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3240 * Conversion table for small slabs sizes / 8 to the index in the
3241 * kmalloc array. This is necessary for slabs < 192 since we have non power
3242 * of two cache sizes there. The size of larger slabs can be determined using
3245 static s8 size_index[24] = {
3272 static inline int size_index_elem(size_t bytes)
3274 return (bytes - 1) / 8;
3277 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3283 return ZERO_SIZE_PTR;
3285 index = size_index[size_index_elem(size)];
3287 index = fls(size - 1);
3289 #ifdef CONFIG_ZONE_DMA
3290 if (unlikely((flags & SLUB_DMA)))
3291 return kmalloc_dma_caches[index];
3294 return kmalloc_caches[index];
3297 void *__kmalloc(size_t size, gfp_t flags)
3299 struct kmem_cache *s;
3302 if (unlikely(size > SLUB_MAX_SIZE))
3303 return kmalloc_large(size, flags);
3305 s = get_slab(size, flags);
3307 if (unlikely(ZERO_OR_NULL_PTR(s)))
3310 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3312 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3316 EXPORT_SYMBOL(__kmalloc);
3319 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3324 flags |= __GFP_COMP | __GFP_NOTRACK;
3325 page = alloc_pages_node(node, flags, get_order(size));
3327 ptr = page_address(page);
3329 kmemleak_alloc(ptr, size, 1, flags);
3333 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3335 struct kmem_cache *s;
3338 if (unlikely(size > SLUB_MAX_SIZE)) {
3339 ret = kmalloc_large_node(size, flags, node);
3341 trace_kmalloc_node(_RET_IP_, ret,
3342 size, PAGE_SIZE << get_order(size),
3348 s = get_slab(size, flags);
3350 if (unlikely(ZERO_OR_NULL_PTR(s)))
3353 ret = slab_alloc(s, flags, node, _RET_IP_);
3355 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3359 EXPORT_SYMBOL(__kmalloc_node);
3362 size_t ksize(const void *object)
3366 if (unlikely(object == ZERO_SIZE_PTR))
3369 page = virt_to_head_page(object);
3371 if (unlikely(!PageSlab(page))) {
3372 WARN_ON(!PageCompound(page));
3373 return PAGE_SIZE << compound_order(page);
3376 return slab_ksize(page->slab);
3378 EXPORT_SYMBOL(ksize);
3380 #ifdef CONFIG_SLUB_DEBUG
3381 bool verify_mem_not_deleted(const void *x)
3384 void *object = (void *)x;
3385 unsigned long flags;
3388 if (unlikely(ZERO_OR_NULL_PTR(x)))
3391 local_irq_save(flags);
3393 page = virt_to_head_page(x);
3394 if (unlikely(!PageSlab(page))) {
3395 /* maybe it was from stack? */
3401 if (on_freelist(page->slab, page, object)) {
3402 object_err(page->slab, page, object, "Object is on free-list");
3410 local_irq_restore(flags);
3413 EXPORT_SYMBOL(verify_mem_not_deleted);
3416 void kfree(const void *x)
3419 void *object = (void *)x;
3421 trace_kfree(_RET_IP_, x);
3423 if (unlikely(ZERO_OR_NULL_PTR(x)))
3426 page = virt_to_head_page(x);
3427 if (unlikely(!PageSlab(page))) {
3428 BUG_ON(!PageCompound(page));
3433 slab_free(page->slab, page, object, _RET_IP_);
3435 EXPORT_SYMBOL(kfree);
3438 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3439 * the remaining slabs by the number of items in use. The slabs with the
3440 * most items in use come first. New allocations will then fill those up
3441 * and thus they can be removed from the partial lists.
3443 * The slabs with the least items are placed last. This results in them
3444 * being allocated from last increasing the chance that the last objects
3445 * are freed in them.
3447 int kmem_cache_shrink(struct kmem_cache *s)
3451 struct kmem_cache_node *n;
3454 int objects = oo_objects(s->max);
3455 struct list_head *slabs_by_inuse =
3456 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3457 unsigned long flags;
3459 if (!slabs_by_inuse)
3463 for_each_node_state(node, N_NORMAL_MEMORY) {
3464 n = get_node(s, node);
3469 for (i = 0; i < objects; i++)
3470 INIT_LIST_HEAD(slabs_by_inuse + i);
3472 spin_lock_irqsave(&n->list_lock, flags);
3475 * Build lists indexed by the items in use in each slab.
3477 * Note that concurrent frees may occur while we hold the
3478 * list_lock. page->inuse here is the upper limit.
3480 list_for_each_entry_safe(page, t, &n->partial, lru) {
3481 list_move(&page->lru, slabs_by_inuse + page->inuse);
3487 * Rebuild the partial list with the slabs filled up most
3488 * first and the least used slabs at the end.
3490 for (i = objects - 1; i > 0; i--)
3491 list_splice(slabs_by_inuse + i, n->partial.prev);
3493 spin_unlock_irqrestore(&n->list_lock, flags);
3495 /* Release empty slabs */
3496 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3497 discard_slab(s, page);
3500 kfree(slabs_by_inuse);
3503 EXPORT_SYMBOL(kmem_cache_shrink);
3505 #if defined(CONFIG_MEMORY_HOTPLUG)
3506 static int slab_mem_going_offline_callback(void *arg)
3508 struct kmem_cache *s;
3510 down_read(&slub_lock);
3511 list_for_each_entry(s, &slab_caches, list)
3512 kmem_cache_shrink(s);
3513 up_read(&slub_lock);
3518 static void slab_mem_offline_callback(void *arg)
3520 struct kmem_cache_node *n;
3521 struct kmem_cache *s;
3522 struct memory_notify *marg = arg;
3525 offline_node = marg->status_change_nid;
3528 * If the node still has available memory. we need kmem_cache_node
3531 if (offline_node < 0)
3534 down_read(&slub_lock);
3535 list_for_each_entry(s, &slab_caches, list) {
3536 n = get_node(s, offline_node);
3539 * if n->nr_slabs > 0, slabs still exist on the node
3540 * that is going down. We were unable to free them,
3541 * and offline_pages() function shouldn't call this
3542 * callback. So, we must fail.
3544 BUG_ON(slabs_node(s, offline_node));
3546 s->node[offline_node] = NULL;
3547 kmem_cache_free(kmem_cache_node, n);
3550 up_read(&slub_lock);
3553 static int slab_mem_going_online_callback(void *arg)
3555 struct kmem_cache_node *n;
3556 struct kmem_cache *s;
3557 struct memory_notify *marg = arg;
3558 int nid = marg->status_change_nid;
3562 * If the node's memory is already available, then kmem_cache_node is
3563 * already created. Nothing to do.
3569 * We are bringing a node online. No memory is available yet. We must
3570 * allocate a kmem_cache_node structure in order to bring the node
3573 down_read(&slub_lock);
3574 list_for_each_entry(s, &slab_caches, list) {
3576 * XXX: kmem_cache_alloc_node will fallback to other nodes
3577 * since memory is not yet available from the node that
3580 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3585 init_kmem_cache_node(n, s);
3589 up_read(&slub_lock);
3593 static int slab_memory_callback(struct notifier_block *self,
3594 unsigned long action, void *arg)
3599 case MEM_GOING_ONLINE:
3600 ret = slab_mem_going_online_callback(arg);
3602 case MEM_GOING_OFFLINE:
3603 ret = slab_mem_going_offline_callback(arg);
3606 case MEM_CANCEL_ONLINE:
3607 slab_mem_offline_callback(arg);
3610 case MEM_CANCEL_OFFLINE:
3614 ret = notifier_from_errno(ret);
3620 #endif /* CONFIG_MEMORY_HOTPLUG */
3622 /********************************************************************
3623 * Basic setup of slabs
3624 *******************************************************************/
3627 * Used for early kmem_cache structures that were allocated using
3628 * the page allocator
3631 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3635 list_add(&s->list, &slab_caches);
3638 for_each_node_state(node, N_NORMAL_MEMORY) {
3639 struct kmem_cache_node *n = get_node(s, node);
3643 list_for_each_entry(p, &n->partial, lru)
3646 #ifdef CONFIG_SLUB_DEBUG
3647 list_for_each_entry(p, &n->full, lru)
3654 void __init kmem_cache_init(void)
3658 struct kmem_cache *temp_kmem_cache;
3660 struct kmem_cache *temp_kmem_cache_node;
3661 unsigned long kmalloc_size;
3663 if (debug_guardpage_minorder())
3666 kmem_size = offsetof(struct kmem_cache, node) +
3667 nr_node_ids * sizeof(struct kmem_cache_node *);
3669 /* Allocate two kmem_caches from the page allocator */
3670 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3671 order = get_order(2 * kmalloc_size);
3672 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3675 * Must first have the slab cache available for the allocations of the
3676 * struct kmem_cache_node's. There is special bootstrap code in
3677 * kmem_cache_open for slab_state == DOWN.
3679 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3681 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3682 sizeof(struct kmem_cache_node),
3683 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3685 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3687 /* Able to allocate the per node structures */
3688 slab_state = PARTIAL;
3690 temp_kmem_cache = kmem_cache;
3691 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3692 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3693 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3694 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3697 * Allocate kmem_cache_node properly from the kmem_cache slab.
3698 * kmem_cache_node is separately allocated so no need to
3699 * update any list pointers.
3701 temp_kmem_cache_node = kmem_cache_node;
3703 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3704 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3706 kmem_cache_bootstrap_fixup(kmem_cache_node);
3709 kmem_cache_bootstrap_fixup(kmem_cache);
3711 /* Free temporary boot structure */
3712 free_pages((unsigned long)temp_kmem_cache, order);
3714 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3717 * Patch up the size_index table if we have strange large alignment
3718 * requirements for the kmalloc array. This is only the case for
3719 * MIPS it seems. The standard arches will not generate any code here.
3721 * Largest permitted alignment is 256 bytes due to the way we
3722 * handle the index determination for the smaller caches.
3724 * Make sure that nothing crazy happens if someone starts tinkering
3725 * around with ARCH_KMALLOC_MINALIGN
3727 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3728 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3730 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3731 int elem = size_index_elem(i);
3732 if (elem >= ARRAY_SIZE(size_index))
3734 size_index[elem] = KMALLOC_SHIFT_LOW;
3737 if (KMALLOC_MIN_SIZE == 64) {
3739 * The 96 byte size cache is not used if the alignment
3742 for (i = 64 + 8; i <= 96; i += 8)
3743 size_index[size_index_elem(i)] = 7;
3744 } else if (KMALLOC_MIN_SIZE == 128) {
3746 * The 192 byte sized cache is not used if the alignment
3747 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3750 for (i = 128 + 8; i <= 192; i += 8)
3751 size_index[size_index_elem(i)] = 8;
3754 /* Caches that are not of the two-to-the-power-of size */
3755 if (KMALLOC_MIN_SIZE <= 32) {
3756 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3760 if (KMALLOC_MIN_SIZE <= 64) {
3761 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3765 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3766 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3772 /* Provide the correct kmalloc names now that the caches are up */
3773 if (KMALLOC_MIN_SIZE <= 32) {
3774 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3775 BUG_ON(!kmalloc_caches[1]->name);
3778 if (KMALLOC_MIN_SIZE <= 64) {
3779 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3780 BUG_ON(!kmalloc_caches[2]->name);
3783 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3784 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3787 kmalloc_caches[i]->name = s;
3791 register_cpu_notifier(&slab_notifier);
3794 #ifdef CONFIG_ZONE_DMA
3795 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3796 struct kmem_cache *s = kmalloc_caches[i];
3799 char *name = kasprintf(GFP_NOWAIT,
3800 "dma-kmalloc-%d", s->objsize);
3803 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3804 s->objsize, SLAB_CACHE_DMA);
3809 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3810 " CPUs=%d, Nodes=%d\n",
3811 caches, cache_line_size(),
3812 slub_min_order, slub_max_order, slub_min_objects,
3813 nr_cpu_ids, nr_node_ids);
3816 void __init kmem_cache_init_late(void)
3821 * Find a mergeable slab cache
3823 static int slab_unmergeable(struct kmem_cache *s)
3825 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3832 * We may have set a slab to be unmergeable during bootstrap.
3834 if (s->refcount < 0)
3840 static struct kmem_cache *find_mergeable(size_t size,
3841 size_t align, unsigned long flags, const char *name,
3842 void (*ctor)(void *))
3844 struct kmem_cache *s;
3846 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3852 size = ALIGN(size, sizeof(void *));
3853 align = calculate_alignment(flags, align, size);
3854 size = ALIGN(size, align);
3855 flags = kmem_cache_flags(size, flags, name, NULL);
3857 list_for_each_entry(s, &slab_caches, list) {
3858 if (slab_unmergeable(s))
3864 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3867 * Check if alignment is compatible.
3868 * Courtesy of Adrian Drzewiecki
3870 if ((s->size & ~(align - 1)) != s->size)
3873 if (s->size - size >= sizeof(void *))
3881 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3882 size_t align, unsigned long flags, void (*ctor)(void *))
3884 struct kmem_cache *s;
3890 down_write(&slub_lock);
3891 s = find_mergeable(size, align, flags, name, ctor);
3895 * Adjust the object sizes so that we clear
3896 * the complete object on kzalloc.
3898 s->objsize = max(s->objsize, (int)size);
3899 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3901 if (sysfs_slab_alias(s, name)) {
3905 up_write(&slub_lock);
3909 n = kstrdup(name, GFP_KERNEL);
3913 s = kmalloc(kmem_size, GFP_KERNEL);
3915 if (kmem_cache_open(s, n,
3916 size, align, flags, ctor)) {
3917 list_add(&s->list, &slab_caches);
3918 if (sysfs_slab_add(s)) {
3924 up_write(&slub_lock);
3931 up_write(&slub_lock);
3933 if (flags & SLAB_PANIC)
3934 panic("Cannot create slabcache %s\n", name);
3939 EXPORT_SYMBOL(kmem_cache_create);
3943 * Use the cpu notifier to insure that the cpu slabs are flushed when
3946 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3947 unsigned long action, void *hcpu)
3949 long cpu = (long)hcpu;
3950 struct kmem_cache *s;
3951 unsigned long flags;
3954 case CPU_UP_CANCELED:
3955 case CPU_UP_CANCELED_FROZEN:
3957 case CPU_DEAD_FROZEN:
3958 down_read(&slub_lock);
3959 list_for_each_entry(s, &slab_caches, list) {
3960 local_irq_save(flags);
3961 __flush_cpu_slab(s, cpu);
3962 local_irq_restore(flags);
3964 up_read(&slub_lock);
3972 static struct notifier_block __cpuinitdata slab_notifier = {
3973 .notifier_call = slab_cpuup_callback
3978 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3980 struct kmem_cache *s;
3983 if (unlikely(size > SLUB_MAX_SIZE))
3984 return kmalloc_large(size, gfpflags);
3986 s = get_slab(size, gfpflags);
3988 if (unlikely(ZERO_OR_NULL_PTR(s)))
3991 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3993 /* Honor the call site pointer we received. */
3994 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4000 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4001 int node, unsigned long caller)
4003 struct kmem_cache *s;
4006 if (unlikely(size > SLUB_MAX_SIZE)) {
4007 ret = kmalloc_large_node(size, gfpflags, node);
4009 trace_kmalloc_node(caller, ret,
4010 size, PAGE_SIZE << get_order(size),
4016 s = get_slab(size, gfpflags);
4018 if (unlikely(ZERO_OR_NULL_PTR(s)))
4021 ret = slab_alloc(s, gfpflags, node, caller);
4023 /* Honor the call site pointer we received. */
4024 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4031 static int count_inuse(struct page *page)
4036 static int count_total(struct page *page)
4038 return page->objects;
4042 #ifdef CONFIG_SLUB_DEBUG
4043 static int validate_slab(struct kmem_cache *s, struct page *page,
4047 void *addr = page_address(page);
4049 if (!check_slab(s, page) ||
4050 !on_freelist(s, page, NULL))
4053 /* Now we know that a valid freelist exists */
4054 bitmap_zero(map, page->objects);
4056 get_map(s, page, map);
4057 for_each_object(p, s, addr, page->objects) {
4058 if (test_bit(slab_index(p, s, addr), map))
4059 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4063 for_each_object(p, s, addr, page->objects)
4064 if (!test_bit(slab_index(p, s, addr), map))
4065 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4070 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4074 validate_slab(s, page, map);
4078 static int validate_slab_node(struct kmem_cache *s,
4079 struct kmem_cache_node *n, unsigned long *map)
4081 unsigned long count = 0;
4083 unsigned long flags;
4085 spin_lock_irqsave(&n->list_lock, flags);
4087 list_for_each_entry(page, &n->partial, lru) {
4088 validate_slab_slab(s, page, map);
4091 if (count != n->nr_partial)
4092 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4093 "counter=%ld\n", s->name, count, n->nr_partial);
4095 if (!(s->flags & SLAB_STORE_USER))
4098 list_for_each_entry(page, &n->full, lru) {
4099 validate_slab_slab(s, page, map);
4102 if (count != atomic_long_read(&n->nr_slabs))
4103 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4104 "counter=%ld\n", s->name, count,
4105 atomic_long_read(&n->nr_slabs));
4108 spin_unlock_irqrestore(&n->list_lock, flags);
4112 static long validate_slab_cache(struct kmem_cache *s)
4115 unsigned long count = 0;
4116 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4117 sizeof(unsigned long), GFP_KERNEL);
4123 for_each_node_state(node, N_NORMAL_MEMORY) {
4124 struct kmem_cache_node *n = get_node(s, node);
4126 count += validate_slab_node(s, n, map);
4132 * Generate lists of code addresses where slabcache objects are allocated
4137 unsigned long count;
4144 DECLARE_BITMAP(cpus, NR_CPUS);
4150 unsigned long count;
4151 struct location *loc;
4154 static void free_loc_track(struct loc_track *t)
4157 free_pages((unsigned long)t->loc,
4158 get_order(sizeof(struct location) * t->max));
4161 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4166 order = get_order(sizeof(struct location) * max);
4168 l = (void *)__get_free_pages(flags, order);
4173 memcpy(l, t->loc, sizeof(struct location) * t->count);
4181 static int add_location(struct loc_track *t, struct kmem_cache *s,
4182 const struct track *track)
4184 long start, end, pos;
4186 unsigned long caddr;
4187 unsigned long age = jiffies - track->when;
4193 pos = start + (end - start + 1) / 2;
4196 * There is nothing at "end". If we end up there
4197 * we need to add something to before end.
4202 caddr = t->loc[pos].addr;
4203 if (track->addr == caddr) {
4209 if (age < l->min_time)
4211 if (age > l->max_time)
4214 if (track->pid < l->min_pid)
4215 l->min_pid = track->pid;
4216 if (track->pid > l->max_pid)
4217 l->max_pid = track->pid;
4219 cpumask_set_cpu(track->cpu,
4220 to_cpumask(l->cpus));
4222 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4226 if (track->addr < caddr)
4233 * Not found. Insert new tracking element.
4235 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4241 (t->count - pos) * sizeof(struct location));
4244 l->addr = track->addr;
4248 l->min_pid = track->pid;
4249 l->max_pid = track->pid;
4250 cpumask_clear(to_cpumask(l->cpus));
4251 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4252 nodes_clear(l->nodes);
4253 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4257 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4258 struct page *page, enum track_item alloc,
4261 void *addr = page_address(page);
4264 bitmap_zero(map, page->objects);
4265 get_map(s, page, map);
4267 for_each_object(p, s, addr, page->objects)
4268 if (!test_bit(slab_index(p, s, addr), map))
4269 add_location(t, s, get_track(s, p, alloc));
4272 static int list_locations(struct kmem_cache *s, char *buf,
4273 enum track_item alloc)
4277 struct loc_track t = { 0, 0, NULL };
4279 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4280 sizeof(unsigned long), GFP_KERNEL);
4282 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4285 return sprintf(buf, "Out of memory\n");
4287 /* Push back cpu slabs */
4290 for_each_node_state(node, N_NORMAL_MEMORY) {
4291 struct kmem_cache_node *n = get_node(s, node);
4292 unsigned long flags;
4295 if (!atomic_long_read(&n->nr_slabs))
4298 spin_lock_irqsave(&n->list_lock, flags);
4299 list_for_each_entry(page, &n->partial, lru)
4300 process_slab(&t, s, page, alloc, map);
4301 list_for_each_entry(page, &n->full, lru)
4302 process_slab(&t, s, page, alloc, map);
4303 spin_unlock_irqrestore(&n->list_lock, flags);
4306 for (i = 0; i < t.count; i++) {
4307 struct location *l = &t.loc[i];
4309 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4311 len += sprintf(buf + len, "%7ld ", l->count);
4314 len += sprintf(buf + len, "%pS", (void *)l->addr);
4316 len += sprintf(buf + len, "<not-available>");
4318 if (l->sum_time != l->min_time) {
4319 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4321 (long)div_u64(l->sum_time, l->count),
4324 len += sprintf(buf + len, " age=%ld",
4327 if (l->min_pid != l->max_pid)
4328 len += sprintf(buf + len, " pid=%ld-%ld",
4329 l->min_pid, l->max_pid);
4331 len += sprintf(buf + len, " pid=%ld",
4334 if (num_online_cpus() > 1 &&
4335 !cpumask_empty(to_cpumask(l->cpus)) &&
4336 len < PAGE_SIZE - 60) {
4337 len += sprintf(buf + len, " cpus=");
4338 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4339 to_cpumask(l->cpus));
4342 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4343 len < PAGE_SIZE - 60) {
4344 len += sprintf(buf + len, " nodes=");
4345 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4349 len += sprintf(buf + len, "\n");
4355 len += sprintf(buf, "No data\n");
4360 #ifdef SLUB_RESILIENCY_TEST
4361 static void resiliency_test(void)
4365 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4367 printk(KERN_ERR "SLUB resiliency testing\n");
4368 printk(KERN_ERR "-----------------------\n");
4369 printk(KERN_ERR "A. Corruption after allocation\n");
4371 p = kzalloc(16, GFP_KERNEL);
4373 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4374 " 0x12->0x%p\n\n", p + 16);
4376 validate_slab_cache(kmalloc_caches[4]);
4378 /* Hmmm... The next two are dangerous */
4379 p = kzalloc(32, GFP_KERNEL);
4380 p[32 + sizeof(void *)] = 0x34;
4381 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4382 " 0x34 -> -0x%p\n", p);
4384 "If allocated object is overwritten then not detectable\n\n");
4386 validate_slab_cache(kmalloc_caches[5]);
4387 p = kzalloc(64, GFP_KERNEL);
4388 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4390 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4393 "If allocated object is overwritten then not detectable\n\n");
4394 validate_slab_cache(kmalloc_caches[6]);
4396 printk(KERN_ERR "\nB. Corruption after free\n");
4397 p = kzalloc(128, GFP_KERNEL);
4400 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4401 validate_slab_cache(kmalloc_caches[7]);
4403 p = kzalloc(256, GFP_KERNEL);
4406 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4408 validate_slab_cache(kmalloc_caches[8]);
4410 p = kzalloc(512, GFP_KERNEL);
4413 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4414 validate_slab_cache(kmalloc_caches[9]);
4418 static void resiliency_test(void) {};
4423 enum slab_stat_type {
4424 SL_ALL, /* All slabs */
4425 SL_PARTIAL, /* Only partially allocated slabs */
4426 SL_CPU, /* Only slabs used for cpu caches */
4427 SL_OBJECTS, /* Determine allocated objects not slabs */
4428 SL_TOTAL /* Determine object capacity not slabs */
4431 #define SO_ALL (1 << SL_ALL)
4432 #define SO_PARTIAL (1 << SL_PARTIAL)
4433 #define SO_CPU (1 << SL_CPU)
4434 #define SO_OBJECTS (1 << SL_OBJECTS)
4435 #define SO_TOTAL (1 << SL_TOTAL)
4437 static ssize_t show_slab_objects(struct kmem_cache *s,
4438 char *buf, unsigned long flags)
4440 unsigned long total = 0;
4443 unsigned long *nodes;
4444 unsigned long *per_cpu;
4446 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4449 per_cpu = nodes + nr_node_ids;
4451 if (flags & SO_CPU) {
4454 for_each_possible_cpu(cpu) {
4455 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4456 int node = ACCESS_ONCE(c->node);
4461 page = ACCESS_ONCE(c->page);
4463 if (flags & SO_TOTAL)
4465 else if (flags & SO_OBJECTS)
4484 lock_memory_hotplug();
4485 #ifdef CONFIG_SLUB_DEBUG
4486 if (flags & SO_ALL) {
4487 for_each_node_state(node, N_NORMAL_MEMORY) {
4488 struct kmem_cache_node *n = get_node(s, node);
4490 if (flags & SO_TOTAL)
4491 x = atomic_long_read(&n->total_objects);
4492 else if (flags & SO_OBJECTS)
4493 x = atomic_long_read(&n->total_objects) -
4494 count_partial(n, count_free);
4497 x = atomic_long_read(&n->nr_slabs);
4504 if (flags & SO_PARTIAL) {
4505 for_each_node_state(node, N_NORMAL_MEMORY) {
4506 struct kmem_cache_node *n = get_node(s, node);
4508 if (flags & SO_TOTAL)
4509 x = count_partial(n, count_total);
4510 else if (flags & SO_OBJECTS)
4511 x = count_partial(n, count_inuse);
4518 x = sprintf(buf, "%lu", total);
4520 for_each_node_state(node, N_NORMAL_MEMORY)
4522 x += sprintf(buf + x, " N%d=%lu",
4525 unlock_memory_hotplug();
4527 return x + sprintf(buf + x, "\n");
4530 #ifdef CONFIG_SLUB_DEBUG
4531 static int any_slab_objects(struct kmem_cache *s)
4535 for_each_online_node(node) {
4536 struct kmem_cache_node *n = get_node(s, node);
4541 if (atomic_long_read(&n->total_objects))
4548 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4549 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4551 struct slab_attribute {
4552 struct attribute attr;
4553 ssize_t (*show)(struct kmem_cache *s, char *buf);
4554 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4557 #define SLAB_ATTR_RO(_name) \
4558 static struct slab_attribute _name##_attr = \
4559 __ATTR(_name, 0400, _name##_show, NULL)
4561 #define SLAB_ATTR(_name) \
4562 static struct slab_attribute _name##_attr = \
4563 __ATTR(_name, 0600, _name##_show, _name##_store)
4565 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4567 return sprintf(buf, "%d\n", s->size);
4569 SLAB_ATTR_RO(slab_size);
4571 static ssize_t align_show(struct kmem_cache *s, char *buf)
4573 return sprintf(buf, "%d\n", s->align);
4575 SLAB_ATTR_RO(align);
4577 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4579 return sprintf(buf, "%d\n", s->objsize);
4581 SLAB_ATTR_RO(object_size);
4583 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4585 return sprintf(buf, "%d\n", oo_objects(s->oo));
4587 SLAB_ATTR_RO(objs_per_slab);
4589 static ssize_t order_store(struct kmem_cache *s,
4590 const char *buf, size_t length)
4592 unsigned long order;
4595 err = strict_strtoul(buf, 10, &order);
4599 if (order > slub_max_order || order < slub_min_order)
4602 calculate_sizes(s, order);
4606 static ssize_t order_show(struct kmem_cache *s, char *buf)
4608 return sprintf(buf, "%d\n", oo_order(s->oo));
4612 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4614 return sprintf(buf, "%lu\n", s->min_partial);
4617 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4623 err = strict_strtoul(buf, 10, &min);
4627 set_min_partial(s, min);
4630 SLAB_ATTR(min_partial);
4632 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4634 return sprintf(buf, "%u\n", s->cpu_partial);
4637 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4640 unsigned long objects;
4643 err = strict_strtoul(buf, 10, &objects);
4647 s->cpu_partial = objects;
4651 SLAB_ATTR(cpu_partial);
4653 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4657 return sprintf(buf, "%pS\n", s->ctor);
4661 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4663 return sprintf(buf, "%d\n", s->refcount - 1);
4665 SLAB_ATTR_RO(aliases);
4667 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4669 return show_slab_objects(s, buf, SO_PARTIAL);
4671 SLAB_ATTR_RO(partial);
4673 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4675 return show_slab_objects(s, buf, SO_CPU);
4677 SLAB_ATTR_RO(cpu_slabs);
4679 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4681 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4683 SLAB_ATTR_RO(objects);
4685 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4687 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4689 SLAB_ATTR_RO(objects_partial);
4691 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4698 for_each_online_cpu(cpu) {
4699 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4702 pages += page->pages;
4703 objects += page->pobjects;
4707 len = sprintf(buf, "%d(%d)", objects, pages);
4710 for_each_online_cpu(cpu) {
4711 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4713 if (page && len < PAGE_SIZE - 20)
4714 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4715 page->pobjects, page->pages);
4718 return len + sprintf(buf + len, "\n");
4720 SLAB_ATTR_RO(slabs_cpu_partial);
4722 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4724 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4727 static ssize_t reclaim_account_store(struct kmem_cache *s,
4728 const char *buf, size_t length)
4730 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4732 s->flags |= SLAB_RECLAIM_ACCOUNT;
4735 SLAB_ATTR(reclaim_account);
4737 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4739 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4741 SLAB_ATTR_RO(hwcache_align);
4743 #ifdef CONFIG_ZONE_DMA
4744 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4746 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4748 SLAB_ATTR_RO(cache_dma);
4751 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4753 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4755 SLAB_ATTR_RO(destroy_by_rcu);
4757 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4759 return sprintf(buf, "%d\n", s->reserved);
4761 SLAB_ATTR_RO(reserved);
4763 #ifdef CONFIG_SLUB_DEBUG
4764 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4766 return show_slab_objects(s, buf, SO_ALL);
4768 SLAB_ATTR_RO(slabs);
4770 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4772 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4774 SLAB_ATTR_RO(total_objects);
4776 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4778 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4781 static ssize_t sanity_checks_store(struct kmem_cache *s,
4782 const char *buf, size_t length)
4784 s->flags &= ~SLAB_DEBUG_FREE;
4785 if (buf[0] == '1') {
4786 s->flags &= ~__CMPXCHG_DOUBLE;
4787 s->flags |= SLAB_DEBUG_FREE;
4791 SLAB_ATTR(sanity_checks);
4793 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4795 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4798 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4801 s->flags &= ~SLAB_TRACE;
4802 if (buf[0] == '1') {
4803 s->flags &= ~__CMPXCHG_DOUBLE;
4804 s->flags |= SLAB_TRACE;
4810 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4812 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4815 static ssize_t red_zone_store(struct kmem_cache *s,
4816 const char *buf, size_t length)
4818 if (any_slab_objects(s))
4821 s->flags &= ~SLAB_RED_ZONE;
4822 if (buf[0] == '1') {
4823 s->flags &= ~__CMPXCHG_DOUBLE;
4824 s->flags |= SLAB_RED_ZONE;
4826 calculate_sizes(s, -1);
4829 SLAB_ATTR(red_zone);
4831 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4833 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4836 static ssize_t poison_store(struct kmem_cache *s,
4837 const char *buf, size_t length)
4839 if (any_slab_objects(s))
4842 s->flags &= ~SLAB_POISON;
4843 if (buf[0] == '1') {
4844 s->flags &= ~__CMPXCHG_DOUBLE;
4845 s->flags |= SLAB_POISON;
4847 calculate_sizes(s, -1);
4852 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4854 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4857 static ssize_t store_user_store(struct kmem_cache *s,
4858 const char *buf, size_t length)
4860 if (any_slab_objects(s))
4863 s->flags &= ~SLAB_STORE_USER;
4864 if (buf[0] == '1') {
4865 s->flags &= ~__CMPXCHG_DOUBLE;
4866 s->flags |= SLAB_STORE_USER;
4868 calculate_sizes(s, -1);
4871 SLAB_ATTR(store_user);
4873 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4878 static ssize_t validate_store(struct kmem_cache *s,
4879 const char *buf, size_t length)
4883 if (buf[0] == '1') {
4884 ret = validate_slab_cache(s);
4890 SLAB_ATTR(validate);
4892 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4894 if (!(s->flags & SLAB_STORE_USER))
4896 return list_locations(s, buf, TRACK_ALLOC);
4898 SLAB_ATTR_RO(alloc_calls);
4900 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4902 if (!(s->flags & SLAB_STORE_USER))
4904 return list_locations(s, buf, TRACK_FREE);
4906 SLAB_ATTR_RO(free_calls);
4907 #endif /* CONFIG_SLUB_DEBUG */
4909 #ifdef CONFIG_FAILSLAB
4910 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4912 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4915 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4918 s->flags &= ~SLAB_FAILSLAB;
4920 s->flags |= SLAB_FAILSLAB;
4923 SLAB_ATTR(failslab);
4926 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4931 static ssize_t shrink_store(struct kmem_cache *s,
4932 const char *buf, size_t length)
4934 if (buf[0] == '1') {
4935 int rc = kmem_cache_shrink(s);
4946 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4948 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4951 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4952 const char *buf, size_t length)
4954 unsigned long ratio;
4957 err = strict_strtoul(buf, 10, &ratio);
4962 s->remote_node_defrag_ratio = ratio * 10;
4966 SLAB_ATTR(remote_node_defrag_ratio);
4969 #ifdef CONFIG_SLUB_STATS
4970 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4972 unsigned long sum = 0;
4975 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4980 for_each_online_cpu(cpu) {
4981 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4987 len = sprintf(buf, "%lu", sum);
4990 for_each_online_cpu(cpu) {
4991 if (data[cpu] && len < PAGE_SIZE - 20)
4992 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4996 return len + sprintf(buf + len, "\n");
4999 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5003 for_each_online_cpu(cpu)
5004 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5007 #define STAT_ATTR(si, text) \
5008 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5010 return show_stat(s, buf, si); \
5012 static ssize_t text##_store(struct kmem_cache *s, \
5013 const char *buf, size_t length) \
5015 if (buf[0] != '0') \
5017 clear_stat(s, si); \
5022 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5023 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5024 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5025 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5026 STAT_ATTR(FREE_FROZEN, free_frozen);
5027 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5028 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5029 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5030 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5031 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5032 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5033 STAT_ATTR(FREE_SLAB, free_slab);
5034 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5035 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5036 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5037 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5038 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5039 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5040 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5041 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5042 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5043 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5044 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5045 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5048 static struct attribute *slab_attrs[] = {
5049 &slab_size_attr.attr,
5050 &object_size_attr.attr,
5051 &objs_per_slab_attr.attr,
5053 &min_partial_attr.attr,
5054 &cpu_partial_attr.attr,
5056 &objects_partial_attr.attr,
5058 &cpu_slabs_attr.attr,
5062 &hwcache_align_attr.attr,
5063 &reclaim_account_attr.attr,
5064 &destroy_by_rcu_attr.attr,
5066 &reserved_attr.attr,
5067 &slabs_cpu_partial_attr.attr,
5068 #ifdef CONFIG_SLUB_DEBUG
5069 &total_objects_attr.attr,
5071 &sanity_checks_attr.attr,
5073 &red_zone_attr.attr,
5075 &store_user_attr.attr,
5076 &validate_attr.attr,
5077 &alloc_calls_attr.attr,
5078 &free_calls_attr.attr,
5080 #ifdef CONFIG_ZONE_DMA
5081 &cache_dma_attr.attr,
5084 &remote_node_defrag_ratio_attr.attr,
5086 #ifdef CONFIG_SLUB_STATS
5087 &alloc_fastpath_attr.attr,
5088 &alloc_slowpath_attr.attr,
5089 &free_fastpath_attr.attr,
5090 &free_slowpath_attr.attr,
5091 &free_frozen_attr.attr,
5092 &free_add_partial_attr.attr,
5093 &free_remove_partial_attr.attr,
5094 &alloc_from_partial_attr.attr,
5095 &alloc_slab_attr.attr,
5096 &alloc_refill_attr.attr,
5097 &alloc_node_mismatch_attr.attr,
5098 &free_slab_attr.attr,
5099 &cpuslab_flush_attr.attr,
5100 &deactivate_full_attr.attr,
5101 &deactivate_empty_attr.attr,
5102 &deactivate_to_head_attr.attr,
5103 &deactivate_to_tail_attr.attr,
5104 &deactivate_remote_frees_attr.attr,
5105 &deactivate_bypass_attr.attr,
5106 &order_fallback_attr.attr,
5107 &cmpxchg_double_fail_attr.attr,
5108 &cmpxchg_double_cpu_fail_attr.attr,
5109 &cpu_partial_alloc_attr.attr,
5110 &cpu_partial_free_attr.attr,
5112 #ifdef CONFIG_FAILSLAB
5113 &failslab_attr.attr,
5119 static struct attribute_group slab_attr_group = {
5120 .attrs = slab_attrs,
5123 static ssize_t slab_attr_show(struct kobject *kobj,
5124 struct attribute *attr,
5127 struct slab_attribute *attribute;
5128 struct kmem_cache *s;
5131 attribute = to_slab_attr(attr);
5134 if (!attribute->show)
5137 err = attribute->show(s, buf);
5142 static ssize_t slab_attr_store(struct kobject *kobj,
5143 struct attribute *attr,
5144 const char *buf, size_t len)
5146 struct slab_attribute *attribute;
5147 struct kmem_cache *s;
5150 attribute = to_slab_attr(attr);
5153 if (!attribute->store)
5156 err = attribute->store(s, buf, len);
5161 static void kmem_cache_release(struct kobject *kobj)
5163 struct kmem_cache *s = to_slab(kobj);
5169 static const struct sysfs_ops slab_sysfs_ops = {
5170 .show = slab_attr_show,
5171 .store = slab_attr_store,
5174 static struct kobj_type slab_ktype = {
5175 .sysfs_ops = &slab_sysfs_ops,
5176 .release = kmem_cache_release
5179 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5181 struct kobj_type *ktype = get_ktype(kobj);
5183 if (ktype == &slab_ktype)
5188 static const struct kset_uevent_ops slab_uevent_ops = {
5189 .filter = uevent_filter,
5192 static struct kset *slab_kset;
5194 #define ID_STR_LENGTH 64
5196 /* Create a unique string id for a slab cache:
5198 * Format :[flags-]size
5200 static char *create_unique_id(struct kmem_cache *s)
5202 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5209 * First flags affecting slabcache operations. We will only
5210 * get here for aliasable slabs so we do not need to support
5211 * too many flags. The flags here must cover all flags that
5212 * are matched during merging to guarantee that the id is
5215 if (s->flags & SLAB_CACHE_DMA)
5217 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5219 if (s->flags & SLAB_DEBUG_FREE)
5221 if (!(s->flags & SLAB_NOTRACK))
5225 p += sprintf(p, "%07d", s->size);
5226 BUG_ON(p > name + ID_STR_LENGTH - 1);
5230 static int sysfs_slab_add(struct kmem_cache *s)
5236 if (slab_state < SYSFS)
5237 /* Defer until later */
5240 unmergeable = slab_unmergeable(s);
5243 * Slabcache can never be merged so we can use the name proper.
5244 * This is typically the case for debug situations. In that
5245 * case we can catch duplicate names easily.
5247 sysfs_remove_link(&slab_kset->kobj, s->name);
5251 * Create a unique name for the slab as a target
5254 name = create_unique_id(s);
5257 s->kobj.kset = slab_kset;
5258 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5260 kobject_put(&s->kobj);
5264 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5266 kobject_del(&s->kobj);
5267 kobject_put(&s->kobj);
5270 kobject_uevent(&s->kobj, KOBJ_ADD);
5272 /* Setup first alias */
5273 sysfs_slab_alias(s, s->name);
5279 static void sysfs_slab_remove(struct kmem_cache *s)
5281 if (slab_state < SYSFS)
5283 * Sysfs has not been setup yet so no need to remove the
5288 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5289 kobject_del(&s->kobj);
5290 kobject_put(&s->kobj);
5294 * Need to buffer aliases during bootup until sysfs becomes
5295 * available lest we lose that information.
5297 struct saved_alias {
5298 struct kmem_cache *s;
5300 struct saved_alias *next;
5303 static struct saved_alias *alias_list;
5305 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5307 struct saved_alias *al;
5309 if (slab_state == SYSFS) {
5311 * If we have a leftover link then remove it.
5313 sysfs_remove_link(&slab_kset->kobj, name);
5314 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5317 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5323 al->next = alias_list;
5328 static int __init slab_sysfs_init(void)
5330 struct kmem_cache *s;
5333 down_write(&slub_lock);
5335 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5337 up_write(&slub_lock);
5338 printk(KERN_ERR "Cannot register slab subsystem.\n");
5344 list_for_each_entry(s, &slab_caches, list) {
5345 err = sysfs_slab_add(s);
5347 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5348 " to sysfs\n", s->name);
5351 while (alias_list) {
5352 struct saved_alias *al = alias_list;
5354 alias_list = alias_list->next;
5355 err = sysfs_slab_alias(al->s, al->name);
5357 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5358 " %s to sysfs\n", s->name);
5362 up_write(&slub_lock);
5367 __initcall(slab_sysfs_init);
5368 #endif /* CONFIG_SYSFS */
5371 * The /proc/slabinfo ABI
5373 #ifdef CONFIG_SLABINFO
5374 static void print_slabinfo_header(struct seq_file *m)
5376 seq_puts(m, "slabinfo - version: 2.1\n");
5377 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5378 "<objperslab> <pagesperslab>");
5379 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5380 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5384 static void *s_start(struct seq_file *m, loff_t *pos)
5388 down_read(&slub_lock);
5390 print_slabinfo_header(m);
5392 return seq_list_start(&slab_caches, *pos);
5395 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5397 return seq_list_next(p, &slab_caches, pos);
5400 static void s_stop(struct seq_file *m, void *p)
5402 up_read(&slub_lock);
5405 static int s_show(struct seq_file *m, void *p)
5407 unsigned long nr_partials = 0;
5408 unsigned long nr_slabs = 0;
5409 unsigned long nr_inuse = 0;
5410 unsigned long nr_objs = 0;
5411 unsigned long nr_free = 0;
5412 struct kmem_cache *s;
5415 s = list_entry(p, struct kmem_cache, list);
5417 for_each_online_node(node) {
5418 struct kmem_cache_node *n = get_node(s, node);
5423 nr_partials += n->nr_partial;
5424 nr_slabs += atomic_long_read(&n->nr_slabs);
5425 nr_objs += atomic_long_read(&n->total_objects);
5426 nr_free += count_partial(n, count_free);
5429 nr_inuse = nr_objs - nr_free;
5431 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5432 nr_objs, s->size, oo_objects(s->oo),
5433 (1 << oo_order(s->oo)));
5434 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5435 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5441 static const struct seq_operations slabinfo_op = {
5448 static int slabinfo_open(struct inode *inode, struct file *file)
5450 return seq_open(file, &slabinfo_op);
5453 static const struct file_operations proc_slabinfo_operations = {
5454 .open = slabinfo_open,
5456 .llseek = seq_lseek,
5457 .release = seq_release,
5460 static int __init slab_proc_init(void)
5462 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5465 module_init(slab_proc_init);
5466 #endif /* CONFIG_SLABINFO */