2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
19 #include <linux/proc_fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/kmemcheck.h>
22 #include <linux/cpu.h>
23 #include <linux/cpuset.h>
24 #include <linux/mempolicy.h>
25 #include <linux/ctype.h>
26 #include <linux/debugobjects.h>
27 #include <linux/kallsyms.h>
28 #include <linux/memory.h>
29 #include <linux/math64.h>
30 #include <linux/fault-inject.h>
31 #include <linux/stacktrace.h>
33 #include <trace/events/kmem.h>
37 * 1. slub_lock (Global Semaphore)
39 * 3. slab_lock(page) (Only on some arches and for debugging)
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
88 * Overloading of page flags that are otherwise used for LRU management.
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
111 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
114 static inline int kmem_cache_debug(struct kmem_cache *s)
116 #ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
124 * Issues still to be resolved:
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 * - Variable sizing of the per node arrays
131 /* Enable to test recovery from slab corruption on boot */
132 #undef SLUB_RESILIENCY_TEST
134 /* Enable to log cmpxchg failures */
135 #undef SLUB_DEBUG_CMPXCHG
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 #define MIN_PARTIAL 5
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
148 #define MAX_PARTIAL 10
150 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
158 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
161 * Set of flags that will prevent slab merging
163 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
167 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
171 #define OO_MASK ((1 << OO_SHIFT) - 1)
172 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174 /* Internal SLUB flags */
175 #define __OBJECT_POISON 0x80000000UL /* Poison object */
176 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178 static int kmem_size = sizeof(struct kmem_cache);
181 static struct notifier_block slab_notifier;
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
191 /* A list of all slab caches on the system */
192 static DECLARE_RWSEM(slub_lock);
193 static LIST_HEAD(slab_caches);
196 * Tracking user of a slab.
198 #define TRACK_ADDRS_COUNT 16
200 unsigned long addr; /* Called from address */
201 #ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
209 enum track_item { TRACK_ALLOC, TRACK_FREE };
212 static int sysfs_slab_add(struct kmem_cache *);
213 static int sysfs_slab_alias(struct kmem_cache *, const char *);
214 static void sysfs_slab_remove(struct kmem_cache *);
217 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
220 static inline void sysfs_slab_remove(struct kmem_cache *s)
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 int slab_is_available(void)
241 return slab_state >= UP;
244 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246 return s->node[node];
249 /* Verify that a pointer has an address that is valid within a slab page */
250 static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
267 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 return *(void **)(object + s->offset);
272 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
276 #ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
279 p = get_freepointer(s, object);
284 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
286 *(void **)(object + s->offset) = fp;
289 /* Loop over all objects in a slab */
290 #define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
294 /* Determine object index from a given position */
295 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
297 return (p - addr) / s->size;
300 static inline size_t slab_ksize(const struct kmem_cache *s)
302 #ifdef CONFIG_SLUB_DEBUG
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
319 * Else we can use all the padding etc for the allocation
324 static inline int order_objects(int order, unsigned long size, int reserved)
326 return ((PAGE_SIZE << order) - reserved) / size;
329 static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
339 static inline int oo_order(struct kmem_cache_order_objects x)
341 return x.x >> OO_SHIFT;
344 static inline int oo_objects(struct kmem_cache_order_objects x)
346 return x.x & OO_MASK;
350 * Per slab locking using the pagelock
352 static __always_inline void slab_lock(struct page *page)
354 bit_spin_lock(PG_locked, &page->flags);
357 static __always_inline void slab_unlock(struct page *page)
359 __bit_spin_unlock(PG_locked, &page->flags);
362 /* Interrupts must be disabled (for the fallback code to work right) */
363 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
368 VM_BUG_ON(!irqs_disabled());
369 #ifdef CONFIG_CMPXCHG_DOUBLE
370 if (s->flags & __CMPXCHG_DOUBLE) {
371 if (cmpxchg_double(&page->freelist,
372 freelist_old, counters_old,
373 freelist_new, counters_new))
379 if (page->freelist == freelist_old && page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
389 stat(s, CMPXCHG_DOUBLE_FAIL);
391 #ifdef SLUB_DEBUG_CMPXCHG
392 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
398 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
403 #ifdef CONFIG_CMPXCHG_DOUBLE
404 if (s->flags & __CMPXCHG_DOUBLE) {
405 if (cmpxchg_double(&page->freelist,
406 freelist_old, counters_old,
407 freelist_new, counters_new))
414 local_irq_save(flags);
416 if (page->freelist == freelist_old && page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
420 local_irq_restore(flags);
424 local_irq_restore(flags);
428 stat(s, CMPXCHG_DOUBLE_FAIL);
430 #ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
437 #ifdef CONFIG_SLUB_DEBUG
439 * Determine a map of object in use on a page.
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
444 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
447 void *addr = page_address(page);
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
456 #ifdef CONFIG_SLUB_DEBUG_ON
457 static int slub_debug = DEBUG_DEFAULT_FLAGS;
459 static int slub_debug;
462 static char *slub_debug_slabs;
463 static int disable_higher_order_debug;
468 static void print_section(char *text, u8 *addr, unsigned int length)
476 for (i = 0; i < length; i++) {
478 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
481 printk(KERN_CONT " %02x", addr[i]);
483 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
485 printk(KERN_CONT " %s\n", ascii);
492 printk(KERN_CONT " ");
496 printk(KERN_CONT " %s\n", ascii);
500 static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
506 p = object + s->offset + sizeof(void *);
508 p = object + s->inuse;
513 static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
516 struct track *p = get_track(s, object, alloc);
519 #ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
527 save_stack_trace(&trace);
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
542 memset(p, 0, sizeof(struct track));
545 static void init_tracking(struct kmem_cache *s, void *object)
547 if (!(s->flags & SLAB_STORE_USER))
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
554 static void print_track(const char *s, struct track *t)
559 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561 #ifdef CONFIG_STACKTRACE
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
566 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
573 static void print_tracking(struct kmem_cache *s, void *object)
575 if (!(s->flags & SLAB_STORE_USER))
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
582 static void print_page_info(struct page *page)
584 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
589 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
595 vsnprintf(buf, sizeof(buf), fmt, args);
597 printk(KERN_ERR "========================================"
598 "=====================================\n");
599 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
600 printk(KERN_ERR "----------------------------------------"
601 "-------------------------------------\n\n");
604 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
610 vsnprintf(buf, sizeof(buf), fmt, args);
612 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
615 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
617 unsigned int off; /* Offset of last byte */
618 u8 *addr = page_address(page);
620 print_tracking(s, p);
622 print_page_info(page);
624 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
625 p, p - addr, get_freepointer(s, p));
628 print_section("Bytes b4", p - 16, 16);
630 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
632 if (s->flags & SLAB_RED_ZONE)
633 print_section("Redzone", p + s->objsize,
634 s->inuse - s->objsize);
637 off = s->offset + sizeof(void *);
641 if (s->flags & SLAB_STORE_USER)
642 off += 2 * sizeof(struct track);
645 /* Beginning of the filler is the free pointer */
646 print_section("Padding", p + off, s->size - off);
651 static void object_err(struct kmem_cache *s, struct page *page,
652 u8 *object, char *reason)
654 slab_bug(s, "%s", reason);
655 print_trailer(s, page, object);
658 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
664 vsnprintf(buf, sizeof(buf), fmt, args);
666 slab_bug(s, "%s", buf);
667 print_page_info(page);
671 static void init_object(struct kmem_cache *s, void *object, u8 val)
675 if (s->flags & __OBJECT_POISON) {
676 memset(p, POISON_FREE, s->objsize - 1);
677 p[s->objsize - 1] = POISON_END;
680 if (s->flags & SLAB_RED_ZONE)
681 memset(p + s->objsize, val, s->inuse - s->objsize);
684 static u8 *check_bytes8(u8 *start, u8 value, unsigned int bytes)
695 static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
698 unsigned int words, prefix;
701 return check_bytes8(start, value, bytes);
703 value64 = value | value << 8 | value << 16 | value << 24;
704 value64 = (value64 & 0xffffffff) | value64 << 32;
705 prefix = 8 - ((unsigned long)start) % 8;
708 u8 *r = check_bytes8(start, value, prefix);
718 if (*(u64 *)start != value64)
719 return check_bytes8(start, value, 8);
724 return check_bytes8(start, value, bytes % 8);
727 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
728 void *from, void *to)
730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
731 memset(from, data, to - from);
734 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
735 u8 *object, char *what,
736 u8 *start, unsigned int value, unsigned int bytes)
741 fault = check_bytes(start, value, bytes);
746 while (end > fault && end[-1] == value)
749 slab_bug(s, "%s overwritten", what);
750 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
751 fault, end - 1, fault[0], value);
752 print_trailer(s, page, object);
754 restore_bytes(s, what, value, fault, end);
762 * Bytes of the object to be managed.
763 * If the freepointer may overlay the object then the free
764 * pointer is the first word of the object.
766 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
769 * object + s->objsize
770 * Padding to reach word boundary. This is also used for Redzoning.
771 * Padding is extended by another word if Redzoning is enabled and
774 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
775 * 0xcc (RED_ACTIVE) for objects in use.
778 * Meta data starts here.
780 * A. Free pointer (if we cannot overwrite object on free)
781 * B. Tracking data for SLAB_STORE_USER
782 * C. Padding to reach required alignment boundary or at mininum
783 * one word if debugging is on to be able to detect writes
784 * before the word boundary.
786 * Padding is done using 0x5a (POISON_INUSE)
789 * Nothing is used beyond s->size.
791 * If slabcaches are merged then the objsize and inuse boundaries are mostly
792 * ignored. And therefore no slab options that rely on these boundaries
793 * may be used with merged slabcaches.
796 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
798 unsigned long off = s->inuse; /* The end of info */
801 /* Freepointer is placed after the object. */
802 off += sizeof(void *);
804 if (s->flags & SLAB_STORE_USER)
805 /* We also have user information there */
806 off += 2 * sizeof(struct track);
811 return check_bytes_and_report(s, page, p, "Object padding",
812 p + off, POISON_INUSE, s->size - off);
815 /* Check the pad bytes at the end of a slab page */
816 static int slab_pad_check(struct kmem_cache *s, struct page *page)
824 if (!(s->flags & SLAB_POISON))
827 start = page_address(page);
828 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
829 end = start + length;
830 remainder = length % s->size;
834 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
837 while (end > fault && end[-1] == POISON_INUSE)
840 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
841 print_section("Padding", end - remainder, remainder);
843 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
847 static int check_object(struct kmem_cache *s, struct page *page,
848 void *object, u8 val)
851 u8 *endobject = object + s->objsize;
853 if (s->flags & SLAB_RED_ZONE) {
854 if (!check_bytes_and_report(s, page, object, "Redzone",
855 endobject, val, s->inuse - s->objsize))
858 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
859 check_bytes_and_report(s, page, p, "Alignment padding",
860 endobject, POISON_INUSE, s->inuse - s->objsize);
864 if (s->flags & SLAB_POISON) {
865 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
866 (!check_bytes_and_report(s, page, p, "Poison", p,
867 POISON_FREE, s->objsize - 1) ||
868 !check_bytes_and_report(s, page, p, "Poison",
869 p + s->objsize - 1, POISON_END, 1)))
872 * check_pad_bytes cleans up on its own.
874 check_pad_bytes(s, page, p);
877 if (!s->offset && val == SLUB_RED_ACTIVE)
879 * Object and freepointer overlap. Cannot check
880 * freepointer while object is allocated.
884 /* Check free pointer validity */
885 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
886 object_err(s, page, p, "Freepointer corrupt");
888 * No choice but to zap it and thus lose the remainder
889 * of the free objects in this slab. May cause
890 * another error because the object count is now wrong.
892 set_freepointer(s, p, NULL);
898 static int check_slab(struct kmem_cache *s, struct page *page)
902 VM_BUG_ON(!irqs_disabled());
904 if (!PageSlab(page)) {
905 slab_err(s, page, "Not a valid slab page");
909 maxobj = order_objects(compound_order(page), s->size, s->reserved);
910 if (page->objects > maxobj) {
911 slab_err(s, page, "objects %u > max %u",
912 s->name, page->objects, maxobj);
915 if (page->inuse > page->objects) {
916 slab_err(s, page, "inuse %u > max %u",
917 s->name, page->inuse, page->objects);
920 /* Slab_pad_check fixes things up after itself */
921 slab_pad_check(s, page);
926 * Determine if a certain object on a page is on the freelist. Must hold the
927 * slab lock to guarantee that the chains are in a consistent state.
929 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
934 unsigned long max_objects;
937 while (fp && nr <= page->objects) {
940 if (!check_valid_pointer(s, page, fp)) {
942 object_err(s, page, object,
943 "Freechain corrupt");
944 set_freepointer(s, object, NULL);
947 slab_err(s, page, "Freepointer corrupt");
948 page->freelist = NULL;
949 page->inuse = page->objects;
950 slab_fix(s, "Freelist cleared");
956 fp = get_freepointer(s, object);
960 max_objects = order_objects(compound_order(page), s->size, s->reserved);
961 if (max_objects > MAX_OBJS_PER_PAGE)
962 max_objects = MAX_OBJS_PER_PAGE;
964 if (page->objects != max_objects) {
965 slab_err(s, page, "Wrong number of objects. Found %d but "
966 "should be %d", page->objects, max_objects);
967 page->objects = max_objects;
968 slab_fix(s, "Number of objects adjusted.");
970 if (page->inuse != page->objects - nr) {
971 slab_err(s, page, "Wrong object count. Counter is %d but "
972 "counted were %d", page->inuse, page->objects - nr);
973 page->inuse = page->objects - nr;
974 slab_fix(s, "Object count adjusted.");
976 return search == NULL;
979 static void trace(struct kmem_cache *s, struct page *page, void *object,
982 if (s->flags & SLAB_TRACE) {
983 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
985 alloc ? "alloc" : "free",
990 print_section("Object", (void *)object, s->objsize);
997 * Hooks for other subsystems that check memory allocations. In a typical
998 * production configuration these hooks all should produce no code at all.
1000 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1002 flags &= gfp_allowed_mask;
1003 lockdep_trace_alloc(flags);
1004 might_sleep_if(flags & __GFP_WAIT);
1006 return should_failslab(s->objsize, flags, s->flags);
1009 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
1011 flags &= gfp_allowed_mask;
1012 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1013 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
1016 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1018 kmemleak_free_recursive(x, s->flags);
1021 * Trouble is that we may no longer disable interupts in the fast path
1022 * So in order to make the debug calls that expect irqs to be
1023 * disabled we need to disable interrupts temporarily.
1025 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1027 unsigned long flags;
1029 local_irq_save(flags);
1030 kmemcheck_slab_free(s, x, s->objsize);
1031 debug_check_no_locks_freed(x, s->objsize);
1032 local_irq_restore(flags);
1035 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1036 debug_check_no_obj_freed(x, s->objsize);
1040 * Tracking of fully allocated slabs for debugging purposes.
1042 * list_lock must be held.
1044 static void add_full(struct kmem_cache *s,
1045 struct kmem_cache_node *n, struct page *page)
1047 if (!(s->flags & SLAB_STORE_USER))
1050 list_add(&page->lru, &n->full);
1054 * list_lock must be held.
1056 static void remove_full(struct kmem_cache *s, struct page *page)
1058 if (!(s->flags & SLAB_STORE_USER))
1061 list_del(&page->lru);
1064 /* Tracking of the number of slabs for debugging purposes */
1065 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1067 struct kmem_cache_node *n = get_node(s, node);
1069 return atomic_long_read(&n->nr_slabs);
1072 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1074 return atomic_long_read(&n->nr_slabs);
1077 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1079 struct kmem_cache_node *n = get_node(s, node);
1082 * May be called early in order to allocate a slab for the
1083 * kmem_cache_node structure. Solve the chicken-egg
1084 * dilemma by deferring the increment of the count during
1085 * bootstrap (see early_kmem_cache_node_alloc).
1088 atomic_long_inc(&n->nr_slabs);
1089 atomic_long_add(objects, &n->total_objects);
1092 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1094 struct kmem_cache_node *n = get_node(s, node);
1096 atomic_long_dec(&n->nr_slabs);
1097 atomic_long_sub(objects, &n->total_objects);
1100 /* Object debug checks for alloc/free paths */
1101 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1104 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1107 init_object(s, object, SLUB_RED_INACTIVE);
1108 init_tracking(s, object);
1111 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1112 void *object, unsigned long addr)
1114 if (!check_slab(s, page))
1117 if (!check_valid_pointer(s, page, object)) {
1118 object_err(s, page, object, "Freelist Pointer check fails");
1122 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1125 /* Success perform special debug activities for allocs */
1126 if (s->flags & SLAB_STORE_USER)
1127 set_track(s, object, TRACK_ALLOC, addr);
1128 trace(s, page, object, 1);
1129 init_object(s, object, SLUB_RED_ACTIVE);
1133 if (PageSlab(page)) {
1135 * If this is a slab page then lets do the best we can
1136 * to avoid issues in the future. Marking all objects
1137 * as used avoids touching the remaining objects.
1139 slab_fix(s, "Marking all objects used");
1140 page->inuse = page->objects;
1141 page->freelist = NULL;
1146 static noinline int free_debug_processing(struct kmem_cache *s,
1147 struct page *page, void *object, unsigned long addr)
1149 unsigned long flags;
1152 local_irq_save(flags);
1155 if (!check_slab(s, page))
1158 if (!check_valid_pointer(s, page, object)) {
1159 slab_err(s, page, "Invalid object pointer 0x%p", object);
1163 if (on_freelist(s, page, object)) {
1164 object_err(s, page, object, "Object already free");
1168 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1171 if (unlikely(s != page->slab)) {
1172 if (!PageSlab(page)) {
1173 slab_err(s, page, "Attempt to free object(0x%p) "
1174 "outside of slab", object);
1175 } else if (!page->slab) {
1177 "SLUB <none>: no slab for object 0x%p.\n",
1181 object_err(s, page, object,
1182 "page slab pointer corrupt.");
1186 if (s->flags & SLAB_STORE_USER)
1187 set_track(s, object, TRACK_FREE, addr);
1188 trace(s, page, object, 0);
1189 init_object(s, object, SLUB_RED_INACTIVE);
1193 local_irq_restore(flags);
1197 slab_fix(s, "Object at 0x%p not freed", object);
1201 static int __init setup_slub_debug(char *str)
1203 slub_debug = DEBUG_DEFAULT_FLAGS;
1204 if (*str++ != '=' || !*str)
1206 * No options specified. Switch on full debugging.
1212 * No options but restriction on slabs. This means full
1213 * debugging for slabs matching a pattern.
1217 if (tolower(*str) == 'o') {
1219 * Avoid enabling debugging on caches if its minimum order
1220 * would increase as a result.
1222 disable_higher_order_debug = 1;
1229 * Switch off all debugging measures.
1234 * Determine which debug features should be switched on
1236 for (; *str && *str != ','; str++) {
1237 switch (tolower(*str)) {
1239 slub_debug |= SLAB_DEBUG_FREE;
1242 slub_debug |= SLAB_RED_ZONE;
1245 slub_debug |= SLAB_POISON;
1248 slub_debug |= SLAB_STORE_USER;
1251 slub_debug |= SLAB_TRACE;
1254 slub_debug |= SLAB_FAILSLAB;
1257 printk(KERN_ERR "slub_debug option '%c' "
1258 "unknown. skipped\n", *str);
1264 slub_debug_slabs = str + 1;
1269 __setup("slub_debug", setup_slub_debug);
1271 static unsigned long kmem_cache_flags(unsigned long objsize,
1272 unsigned long flags, const char *name,
1273 void (*ctor)(void *))
1276 * Enable debugging if selected on the kernel commandline.
1278 if (slub_debug && (!slub_debug_slabs ||
1279 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1280 flags |= slub_debug;
1285 static inline void setup_object_debug(struct kmem_cache *s,
1286 struct page *page, void *object) {}
1288 static inline int alloc_debug_processing(struct kmem_cache *s,
1289 struct page *page, void *object, unsigned long addr) { return 0; }
1291 static inline int free_debug_processing(struct kmem_cache *s,
1292 struct page *page, void *object, unsigned long addr) { return 0; }
1294 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1296 static inline int check_object(struct kmem_cache *s, struct page *page,
1297 void *object, u8 val) { return 1; }
1298 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1299 struct page *page) {}
1300 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1301 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1302 unsigned long flags, const char *name,
1303 void (*ctor)(void *))
1307 #define slub_debug 0
1309 #define disable_higher_order_debug 0
1311 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1313 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1315 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1317 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1320 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1323 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1326 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1328 #endif /* CONFIG_SLUB_DEBUG */
1331 * Slab allocation and freeing
1333 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1334 struct kmem_cache_order_objects oo)
1336 int order = oo_order(oo);
1338 flags |= __GFP_NOTRACK;
1340 if (node == NUMA_NO_NODE)
1341 return alloc_pages(flags, order);
1343 return alloc_pages_exact_node(node, flags, order);
1346 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1349 struct kmem_cache_order_objects oo = s->oo;
1352 flags &= gfp_allowed_mask;
1354 if (flags & __GFP_WAIT)
1357 flags |= s->allocflags;
1360 * Let the initial higher-order allocation fail under memory pressure
1361 * so we fall-back to the minimum order allocation.
1363 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1365 page = alloc_slab_page(alloc_gfp, node, oo);
1366 if (unlikely(!page)) {
1369 * Allocation may have failed due to fragmentation.
1370 * Try a lower order alloc if possible
1372 page = alloc_slab_page(flags, node, oo);
1375 stat(s, ORDER_FALLBACK);
1378 if (flags & __GFP_WAIT)
1379 local_irq_disable();
1384 if (kmemcheck_enabled
1385 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1386 int pages = 1 << oo_order(oo);
1388 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1391 * Objects from caches that have a constructor don't get
1392 * cleared when they're allocated, so we need to do it here.
1395 kmemcheck_mark_uninitialized_pages(page, pages);
1397 kmemcheck_mark_unallocated_pages(page, pages);
1400 page->objects = oo_objects(oo);
1401 mod_zone_page_state(page_zone(page),
1402 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1403 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1409 static void setup_object(struct kmem_cache *s, struct page *page,
1412 setup_object_debug(s, page, object);
1413 if (unlikely(s->ctor))
1417 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1424 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1426 page = allocate_slab(s,
1427 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1431 inc_slabs_node(s, page_to_nid(page), page->objects);
1433 page->flags |= 1 << PG_slab;
1435 start = page_address(page);
1437 if (unlikely(s->flags & SLAB_POISON))
1438 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1441 for_each_object(p, s, start, page->objects) {
1442 setup_object(s, page, last);
1443 set_freepointer(s, last, p);
1446 setup_object(s, page, last);
1447 set_freepointer(s, last, NULL);
1449 page->freelist = start;
1450 page->inuse = page->objects;
1456 static void __free_slab(struct kmem_cache *s, struct page *page)
1458 int order = compound_order(page);
1459 int pages = 1 << order;
1461 if (kmem_cache_debug(s)) {
1464 slab_pad_check(s, page);
1465 for_each_object(p, s, page_address(page),
1467 check_object(s, page, p, SLUB_RED_INACTIVE);
1470 kmemcheck_free_shadow(page, compound_order(page));
1472 mod_zone_page_state(page_zone(page),
1473 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1474 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1477 __ClearPageSlab(page);
1478 reset_page_mapcount(page);
1479 if (current->reclaim_state)
1480 current->reclaim_state->reclaimed_slab += pages;
1481 __free_pages(page, order);
1484 #define need_reserve_slab_rcu \
1485 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1487 static void rcu_free_slab(struct rcu_head *h)
1491 if (need_reserve_slab_rcu)
1492 page = virt_to_head_page(h);
1494 page = container_of((struct list_head *)h, struct page, lru);
1496 __free_slab(page->slab, page);
1499 static void free_slab(struct kmem_cache *s, struct page *page)
1501 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1502 struct rcu_head *head;
1504 if (need_reserve_slab_rcu) {
1505 int order = compound_order(page);
1506 int offset = (PAGE_SIZE << order) - s->reserved;
1508 VM_BUG_ON(s->reserved != sizeof(*head));
1509 head = page_address(page) + offset;
1512 * RCU free overloads the RCU head over the LRU
1514 head = (void *)&page->lru;
1517 call_rcu(head, rcu_free_slab);
1519 __free_slab(s, page);
1522 static void discard_slab(struct kmem_cache *s, struct page *page)
1524 dec_slabs_node(s, page_to_nid(page), page->objects);
1529 * Management of partially allocated slabs.
1531 * list_lock must be held.
1533 static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1538 list_add_tail(&page->lru, &n->partial);
1540 list_add(&page->lru, &n->partial);
1544 * list_lock must be held.
1546 static inline void remove_partial(struct kmem_cache_node *n,
1549 list_del(&page->lru);
1554 * Lock slab, remove from the partial list and put the object into the
1557 * Returns a list of objects or NULL if it fails.
1559 * Must hold list_lock.
1561 static inline void *acquire_slab(struct kmem_cache *s,
1562 struct kmem_cache_node *n, struct page *page,
1566 unsigned long counters;
1570 * Zap the freelist and set the frozen bit.
1571 * The old freelist is the list of objects for the
1572 * per cpu allocation list.
1575 freelist = page->freelist;
1576 counters = page->counters;
1577 new.counters = counters;
1579 new.inuse = page->objects;
1581 VM_BUG_ON(new.frozen);
1584 } while (!__cmpxchg_double_slab(s, page,
1587 "lock and freeze"));
1589 remove_partial(n, page);
1593 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1596 * Try to allocate a partial slab from a specific node.
1598 static void *get_partial_node(struct kmem_cache *s,
1599 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1601 struct page *page, *page2;
1602 void *object = NULL;
1605 * Racy check. If we mistakenly see no partial slabs then we
1606 * just allocate an empty slab. If we mistakenly try to get a
1607 * partial slab and there is none available then get_partials()
1610 if (!n || !n->nr_partial)
1613 spin_lock(&n->list_lock);
1614 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1615 void *t = acquire_slab(s, n, page, object == NULL);
1623 c->node = page_to_nid(page);
1624 stat(s, ALLOC_FROM_PARTIAL);
1626 available = page->objects - page->inuse;
1629 available = put_cpu_partial(s, page, 0);
1631 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1635 spin_unlock(&n->list_lock);
1640 * Get a page from somewhere. Search in increasing NUMA distances.
1642 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1643 struct kmem_cache_cpu *c)
1646 struct zonelist *zonelist;
1649 enum zone_type high_zoneidx = gfp_zone(flags);
1653 * The defrag ratio allows a configuration of the tradeoffs between
1654 * inter node defragmentation and node local allocations. A lower
1655 * defrag_ratio increases the tendency to do local allocations
1656 * instead of attempting to obtain partial slabs from other nodes.
1658 * If the defrag_ratio is set to 0 then kmalloc() always
1659 * returns node local objects. If the ratio is higher then kmalloc()
1660 * may return off node objects because partial slabs are obtained
1661 * from other nodes and filled up.
1663 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1664 * defrag_ratio = 1000) then every (well almost) allocation will
1665 * first attempt to defrag slab caches on other nodes. This means
1666 * scanning over all nodes to look for partial slabs which may be
1667 * expensive if we do it every time we are trying to find a slab
1668 * with available objects.
1670 if (!s->remote_node_defrag_ratio ||
1671 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1675 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1676 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1677 struct kmem_cache_node *n;
1679 n = get_node(s, zone_to_nid(zone));
1681 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1682 n->nr_partial > s->min_partial) {
1683 object = get_partial_node(s, n, c);
1696 * Get a partial page, lock it and return it.
1698 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1699 struct kmem_cache_cpu *c)
1702 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1704 object = get_partial_node(s, get_node(s, searchnode), c);
1705 if (object || node != NUMA_NO_NODE)
1708 return get_any_partial(s, flags, c);
1711 #ifdef CONFIG_PREEMPT
1713 * Calculate the next globally unique transaction for disambiguiation
1714 * during cmpxchg. The transactions start with the cpu number and are then
1715 * incremented by CONFIG_NR_CPUS.
1717 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1720 * No preemption supported therefore also no need to check for
1726 static inline unsigned long next_tid(unsigned long tid)
1728 return tid + TID_STEP;
1731 static inline unsigned int tid_to_cpu(unsigned long tid)
1733 return tid % TID_STEP;
1736 static inline unsigned long tid_to_event(unsigned long tid)
1738 return tid / TID_STEP;
1741 static inline unsigned int init_tid(int cpu)
1746 static inline void note_cmpxchg_failure(const char *n,
1747 const struct kmem_cache *s, unsigned long tid)
1749 #ifdef SLUB_DEBUG_CMPXCHG
1750 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1752 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1754 #ifdef CONFIG_PREEMPT
1755 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1756 printk("due to cpu change %d -> %d\n",
1757 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1760 if (tid_to_event(tid) != tid_to_event(actual_tid))
1761 printk("due to cpu running other code. Event %ld->%ld\n",
1762 tid_to_event(tid), tid_to_event(actual_tid));
1764 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1765 actual_tid, tid, next_tid(tid));
1767 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1770 void init_kmem_cache_cpus(struct kmem_cache *s)
1774 for_each_possible_cpu(cpu)
1775 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1779 * Remove the cpu slab
1781 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1783 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1784 struct page *page = c->page;
1785 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1787 enum slab_modes l = M_NONE, m = M_NONE;
1794 if (page->freelist) {
1795 stat(s, DEACTIVATE_REMOTE_FREES);
1799 c->tid = next_tid(c->tid);
1801 freelist = c->freelist;
1805 * Stage one: Free all available per cpu objects back
1806 * to the page freelist while it is still frozen. Leave the
1809 * There is no need to take the list->lock because the page
1812 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1814 unsigned long counters;
1817 prior = page->freelist;
1818 counters = page->counters;
1819 set_freepointer(s, freelist, prior);
1820 new.counters = counters;
1822 VM_BUG_ON(!new.frozen);
1824 } while (!__cmpxchg_double_slab(s, page,
1826 freelist, new.counters,
1827 "drain percpu freelist"));
1829 freelist = nextfree;
1833 * Stage two: Ensure that the page is unfrozen while the
1834 * list presence reflects the actual number of objects
1837 * We setup the list membership and then perform a cmpxchg
1838 * with the count. If there is a mismatch then the page
1839 * is not unfrozen but the page is on the wrong list.
1841 * Then we restart the process which may have to remove
1842 * the page from the list that we just put it on again
1843 * because the number of objects in the slab may have
1848 old.freelist = page->freelist;
1849 old.counters = page->counters;
1850 VM_BUG_ON(!old.frozen);
1852 /* Determine target state of the slab */
1853 new.counters = old.counters;
1856 set_freepointer(s, freelist, old.freelist);
1857 new.freelist = freelist;
1859 new.freelist = old.freelist;
1863 if (!new.inuse && n->nr_partial > s->min_partial)
1865 else if (new.freelist) {
1870 * Taking the spinlock removes the possiblity
1871 * that acquire_slab() will see a slab page that
1874 spin_lock(&n->list_lock);
1878 if (kmem_cache_debug(s) && !lock) {
1881 * This also ensures that the scanning of full
1882 * slabs from diagnostic functions will not see
1885 spin_lock(&n->list_lock);
1893 remove_partial(n, page);
1895 else if (l == M_FULL)
1897 remove_full(s, page);
1899 if (m == M_PARTIAL) {
1901 add_partial(n, page, tail);
1902 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1904 } else if (m == M_FULL) {
1906 stat(s, DEACTIVATE_FULL);
1907 add_full(s, n, page);
1913 if (!__cmpxchg_double_slab(s, page,
1914 old.freelist, old.counters,
1915 new.freelist, new.counters,
1920 spin_unlock(&n->list_lock);
1923 stat(s, DEACTIVATE_EMPTY);
1924 discard_slab(s, page);
1929 /* Unfreeze all the cpu partial slabs */
1930 static void unfreeze_partials(struct kmem_cache *s)
1932 struct kmem_cache_node *n = NULL;
1933 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1936 while ((page = c->partial)) {
1937 enum slab_modes { M_PARTIAL, M_FREE };
1938 enum slab_modes l, m;
1942 c->partial = page->next;
1947 old.freelist = page->freelist;
1948 old.counters = page->counters;
1949 VM_BUG_ON(!old.frozen);
1951 new.counters = old.counters;
1952 new.freelist = old.freelist;
1956 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1959 struct kmem_cache_node *n2 = get_node(s,
1965 spin_unlock(&n->list_lock);
1968 spin_lock(&n->list_lock);
1974 remove_partial(n, page);
1976 add_partial(n, page, 1);
1981 } while (!cmpxchg_double_slab(s, page,
1982 old.freelist, old.counters,
1983 new.freelist, new.counters,
1984 "unfreezing slab"));
1987 stat(s, DEACTIVATE_EMPTY);
1988 discard_slab(s, page);
1994 spin_unlock(&n->list_lock);
1998 * Put a page that was just frozen (in __slab_free) into a partial page
1999 * slot if available. This is done without interrupts disabled and without
2000 * preemption disabled. The cmpxchg is racy and may put the partial page
2001 * onto a random cpus partial slot.
2003 * If we did not find a slot then simply move all the partials to the
2004 * per node partial list.
2006 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2008 struct page *oldpage;
2015 oldpage = this_cpu_read(s->cpu_slab->partial);
2018 pobjects = oldpage->pobjects;
2019 pages = oldpage->pages;
2020 if (drain && pobjects > s->cpu_partial) {
2021 unsigned long flags;
2023 * partial array is full. Move the existing
2024 * set to the per node partial list.
2026 local_irq_save(flags);
2027 unfreeze_partials(s);
2028 local_irq_restore(flags);
2035 pobjects += page->objects - page->inuse;
2037 page->pages = pages;
2038 page->pobjects = pobjects;
2039 page->next = oldpage;
2041 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2042 stat(s, CPU_PARTIAL_FREE);
2046 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2048 stat(s, CPUSLAB_FLUSH);
2049 deactivate_slab(s, c);
2055 * Called from IPI handler with interrupts disabled.
2057 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2059 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2065 unfreeze_partials(s);
2069 static void flush_cpu_slab(void *d)
2071 struct kmem_cache *s = d;
2073 __flush_cpu_slab(s, smp_processor_id());
2076 static void flush_all(struct kmem_cache *s)
2078 on_each_cpu(flush_cpu_slab, s, 1);
2082 * Check if the objects in a per cpu structure fit numa
2083 * locality expectations.
2085 static inline int node_match(struct kmem_cache_cpu *c, int node)
2088 if (node != NUMA_NO_NODE && c->node != node)
2094 static int count_free(struct page *page)
2096 return page->objects - page->inuse;
2099 static unsigned long count_partial(struct kmem_cache_node *n,
2100 int (*get_count)(struct page *))
2102 unsigned long flags;
2103 unsigned long x = 0;
2106 spin_lock_irqsave(&n->list_lock, flags);
2107 list_for_each_entry(page, &n->partial, lru)
2108 x += get_count(page);
2109 spin_unlock_irqrestore(&n->list_lock, flags);
2113 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2115 #ifdef CONFIG_SLUB_DEBUG
2116 return atomic_long_read(&n->total_objects);
2122 static noinline void
2123 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2128 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2130 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2131 "default order: %d, min order: %d\n", s->name, s->objsize,
2132 s->size, oo_order(s->oo), oo_order(s->min));
2134 if (oo_order(s->min) > get_order(s->objsize))
2135 printk(KERN_WARNING " %s debugging increased min order, use "
2136 "slub_debug=O to disable.\n", s->name);
2138 for_each_online_node(node) {
2139 struct kmem_cache_node *n = get_node(s, node);
2140 unsigned long nr_slabs;
2141 unsigned long nr_objs;
2142 unsigned long nr_free;
2147 nr_free = count_partial(n, count_free);
2148 nr_slabs = node_nr_slabs(n);
2149 nr_objs = node_nr_objs(n);
2152 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2153 node, nr_slabs, nr_objs, nr_free);
2157 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2158 int node, struct kmem_cache_cpu **pc)
2161 struct kmem_cache_cpu *c;
2162 struct page *page = new_slab(s, flags, node);
2165 c = __this_cpu_ptr(s->cpu_slab);
2170 * No other reference to the page yet so we can
2171 * muck around with it freely without cmpxchg
2173 object = page->freelist;
2174 page->freelist = NULL;
2176 stat(s, ALLOC_SLAB);
2177 c->node = page_to_nid(page);
2187 * Slow path. The lockless freelist is empty or we need to perform
2190 * Processing is still very fast if new objects have been freed to the
2191 * regular freelist. In that case we simply take over the regular freelist
2192 * as the lockless freelist and zap the regular freelist.
2194 * If that is not working then we fall back to the partial lists. We take the
2195 * first element of the freelist as the object to allocate now and move the
2196 * rest of the freelist to the lockless freelist.
2198 * And if we were unable to get a new slab from the partial slab lists then
2199 * we need to allocate a new slab. This is the slowest path since it involves
2200 * a call to the page allocator and the setup of a new slab.
2202 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2203 unsigned long addr, struct kmem_cache_cpu *c)
2206 unsigned long flags;
2208 unsigned long counters;
2210 local_irq_save(flags);
2211 #ifdef CONFIG_PREEMPT
2213 * We may have been preempted and rescheduled on a different
2214 * cpu before disabling interrupts. Need to reload cpu area
2217 c = this_cpu_ptr(s->cpu_slab);
2223 if (unlikely(!node_match(c, node))) {
2224 stat(s, ALLOC_NODE_MISMATCH);
2225 deactivate_slab(s, c);
2229 stat(s, ALLOC_SLOWPATH);
2232 object = c->page->freelist;
2233 counters = c->page->counters;
2234 new.counters = counters;
2235 VM_BUG_ON(!new.frozen);
2238 * If there is no object left then we use this loop to
2239 * deactivate the slab which is simple since no objects
2240 * are left in the slab and therefore we do not need to
2241 * put the page back onto the partial list.
2243 * If there are objects left then we retrieve them
2244 * and use them to refill the per cpu queue.
2247 new.inuse = c->page->objects;
2248 new.frozen = object != NULL;
2250 } while (!__cmpxchg_double_slab(s, c->page,
2257 stat(s, DEACTIVATE_BYPASS);
2261 stat(s, ALLOC_REFILL);
2264 c->freelist = get_freepointer(s, object);
2265 c->tid = next_tid(c->tid);
2266 local_irq_restore(flags);
2272 c->page = c->partial;
2273 c->partial = c->page->next;
2274 c->node = page_to_nid(c->page);
2275 stat(s, CPU_PARTIAL_ALLOC);
2280 /* Then do expensive stuff like retrieving pages from the partial lists */
2281 object = get_partial(s, gfpflags, node, c);
2283 if (unlikely(!object)) {
2285 object = new_slab_objects(s, gfpflags, node, &c);
2287 if (unlikely(!object)) {
2288 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2289 slab_out_of_memory(s, gfpflags, node);
2291 local_irq_restore(flags);
2296 if (likely(!kmem_cache_debug(s)))
2299 /* Only entered in the debug case */
2300 if (!alloc_debug_processing(s, c->page, object, addr))
2301 goto new_slab; /* Slab failed checks. Next slab needed */
2303 c->freelist = get_freepointer(s, object);
2304 deactivate_slab(s, c);
2305 c->node = NUMA_NO_NODE;
2306 local_irq_restore(flags);
2311 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2312 * have the fastpath folded into their functions. So no function call
2313 * overhead for requests that can be satisfied on the fastpath.
2315 * The fastpath works by first checking if the lockless freelist can be used.
2316 * If not then __slab_alloc is called for slow processing.
2318 * Otherwise we can simply pick the next object from the lockless free list.
2320 static __always_inline void *slab_alloc(struct kmem_cache *s,
2321 gfp_t gfpflags, int node, unsigned long addr)
2324 struct kmem_cache_cpu *c;
2327 if (slab_pre_alloc_hook(s, gfpflags))
2333 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2334 * enabled. We may switch back and forth between cpus while
2335 * reading from one cpu area. That does not matter as long
2336 * as we end up on the original cpu again when doing the cmpxchg.
2338 c = __this_cpu_ptr(s->cpu_slab);
2341 * The transaction ids are globally unique per cpu and per operation on
2342 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2343 * occurs on the right processor and that there was no operation on the
2344 * linked list in between.
2349 object = c->freelist;
2350 if (unlikely(!object || !node_match(c, node)))
2352 object = __slab_alloc(s, gfpflags, node, addr, c);
2356 * The cmpxchg will only match if there was no additional
2357 * operation and if we are on the right processor.
2359 * The cmpxchg does the following atomically (without lock semantics!)
2360 * 1. Relocate first pointer to the current per cpu area.
2361 * 2. Verify that tid and freelist have not been changed
2362 * 3. If they were not changed replace tid and freelist
2364 * Since this is without lock semantics the protection is only against
2365 * code executing on this cpu *not* from access by other cpus.
2367 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2368 s->cpu_slab->freelist, s->cpu_slab->tid,
2370 get_freepointer_safe(s, object), next_tid(tid)))) {
2372 note_cmpxchg_failure("slab_alloc", s, tid);
2375 stat(s, ALLOC_FASTPATH);
2378 if (unlikely(gfpflags & __GFP_ZERO) && object)
2379 memset(object, 0, s->objsize);
2381 slab_post_alloc_hook(s, gfpflags, object);
2386 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2388 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2390 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2394 EXPORT_SYMBOL(kmem_cache_alloc);
2396 #ifdef CONFIG_TRACING
2397 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2399 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2400 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2403 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2405 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2407 void *ret = kmalloc_order(size, flags, order);
2408 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2411 EXPORT_SYMBOL(kmalloc_order_trace);
2415 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2417 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2419 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2420 s->objsize, s->size, gfpflags, node);
2424 EXPORT_SYMBOL(kmem_cache_alloc_node);
2426 #ifdef CONFIG_TRACING
2427 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2429 int node, size_t size)
2431 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2433 trace_kmalloc_node(_RET_IP_, ret,
2434 size, s->size, gfpflags, node);
2437 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2442 * Slow patch handling. This may still be called frequently since objects
2443 * have a longer lifetime than the cpu slabs in most processing loads.
2445 * So we still attempt to reduce cache line usage. Just take the slab
2446 * lock and free the item. If there is no additional partial page
2447 * handling required then we can return immediately.
2449 static void __slab_free(struct kmem_cache *s, struct page *page,
2450 void *x, unsigned long addr)
2453 void **object = (void *)x;
2457 unsigned long counters;
2458 struct kmem_cache_node *n = NULL;
2459 unsigned long uninitialized_var(flags);
2461 stat(s, FREE_SLOWPATH);
2463 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2467 prior = page->freelist;
2468 counters = page->counters;
2469 set_freepointer(s, object, prior);
2470 new.counters = counters;
2471 was_frozen = new.frozen;
2473 if ((!new.inuse || !prior) && !was_frozen && !n) {
2475 if (!kmem_cache_debug(s) && !prior)
2478 * Slab was on no list before and will be partially empty
2479 * We can defer the list move and instead freeze it.
2483 else { /* Needs to be taken off a list */
2485 n = get_node(s, page_to_nid(page));
2487 * Speculatively acquire the list_lock.
2488 * If the cmpxchg does not succeed then we may
2489 * drop the list_lock without any processing.
2491 * Otherwise the list_lock will synchronize with
2492 * other processors updating the list of slabs.
2494 spin_lock_irqsave(&n->list_lock, flags);
2500 } while (!cmpxchg_double_slab(s, page,
2502 object, new.counters,
2508 * If we just froze the page then put it onto the
2509 * per cpu partial list.
2511 if (new.frozen && !was_frozen)
2512 put_cpu_partial(s, page, 1);
2515 * The list lock was not taken therefore no list
2516 * activity can be necessary.
2519 stat(s, FREE_FROZEN);
2524 * was_frozen may have been set after we acquired the list_lock in
2525 * an earlier loop. So we need to check it here again.
2528 stat(s, FREE_FROZEN);
2530 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2534 * Objects left in the slab. If it was not on the partial list before
2537 if (unlikely(!prior)) {
2538 remove_full(s, page);
2539 add_partial(n, page, 0);
2540 stat(s, FREE_ADD_PARTIAL);
2543 spin_unlock_irqrestore(&n->list_lock, flags);
2549 * Slab on the partial list.
2551 remove_partial(n, page);
2552 stat(s, FREE_REMOVE_PARTIAL);
2554 /* Slab must be on the full list */
2555 remove_full(s, page);
2557 spin_unlock_irqrestore(&n->list_lock, flags);
2559 discard_slab(s, page);
2563 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2564 * can perform fastpath freeing without additional function calls.
2566 * The fastpath is only possible if we are freeing to the current cpu slab
2567 * of this processor. This typically the case if we have just allocated
2570 * If fastpath is not possible then fall back to __slab_free where we deal
2571 * with all sorts of special processing.
2573 static __always_inline void slab_free(struct kmem_cache *s,
2574 struct page *page, void *x, unsigned long addr)
2576 void **object = (void *)x;
2577 struct kmem_cache_cpu *c;
2580 slab_free_hook(s, x);
2584 * Determine the currently cpus per cpu slab.
2585 * The cpu may change afterward. However that does not matter since
2586 * data is retrieved via this pointer. If we are on the same cpu
2587 * during the cmpxchg then the free will succedd.
2589 c = __this_cpu_ptr(s->cpu_slab);
2594 if (likely(page == c->page)) {
2595 set_freepointer(s, object, c->freelist);
2597 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2598 s->cpu_slab->freelist, s->cpu_slab->tid,
2600 object, next_tid(tid)))) {
2602 note_cmpxchg_failure("slab_free", s, tid);
2605 stat(s, FREE_FASTPATH);
2607 __slab_free(s, page, x, addr);
2611 void kmem_cache_free(struct kmem_cache *s, void *x)
2615 page = virt_to_head_page(x);
2617 slab_free(s, page, x, _RET_IP_);
2619 trace_kmem_cache_free(_RET_IP_, x);
2621 EXPORT_SYMBOL(kmem_cache_free);
2624 * Object placement in a slab is made very easy because we always start at
2625 * offset 0. If we tune the size of the object to the alignment then we can
2626 * get the required alignment by putting one properly sized object after
2629 * Notice that the allocation order determines the sizes of the per cpu
2630 * caches. Each processor has always one slab available for allocations.
2631 * Increasing the allocation order reduces the number of times that slabs
2632 * must be moved on and off the partial lists and is therefore a factor in
2637 * Mininum / Maximum order of slab pages. This influences locking overhead
2638 * and slab fragmentation. A higher order reduces the number of partial slabs
2639 * and increases the number of allocations possible without having to
2640 * take the list_lock.
2642 static int slub_min_order;
2643 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2644 static int slub_min_objects;
2647 * Merge control. If this is set then no merging of slab caches will occur.
2648 * (Could be removed. This was introduced to pacify the merge skeptics.)
2650 static int slub_nomerge;
2653 * Calculate the order of allocation given an slab object size.
2655 * The order of allocation has significant impact on performance and other
2656 * system components. Generally order 0 allocations should be preferred since
2657 * order 0 does not cause fragmentation in the page allocator. Larger objects
2658 * be problematic to put into order 0 slabs because there may be too much
2659 * unused space left. We go to a higher order if more than 1/16th of the slab
2662 * In order to reach satisfactory performance we must ensure that a minimum
2663 * number of objects is in one slab. Otherwise we may generate too much
2664 * activity on the partial lists which requires taking the list_lock. This is
2665 * less a concern for large slabs though which are rarely used.
2667 * slub_max_order specifies the order where we begin to stop considering the
2668 * number of objects in a slab as critical. If we reach slub_max_order then
2669 * we try to keep the page order as low as possible. So we accept more waste
2670 * of space in favor of a small page order.
2672 * Higher order allocations also allow the placement of more objects in a
2673 * slab and thereby reduce object handling overhead. If the user has
2674 * requested a higher mininum order then we start with that one instead of
2675 * the smallest order which will fit the object.
2677 static inline int slab_order(int size, int min_objects,
2678 int max_order, int fract_leftover, int reserved)
2682 int min_order = slub_min_order;
2684 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2685 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2687 for (order = max(min_order,
2688 fls(min_objects * size - 1) - PAGE_SHIFT);
2689 order <= max_order; order++) {
2691 unsigned long slab_size = PAGE_SIZE << order;
2693 if (slab_size < min_objects * size + reserved)
2696 rem = (slab_size - reserved) % size;
2698 if (rem <= slab_size / fract_leftover)
2706 static inline int calculate_order(int size, int reserved)
2714 * Attempt to find best configuration for a slab. This
2715 * works by first attempting to generate a layout with
2716 * the best configuration and backing off gradually.
2718 * First we reduce the acceptable waste in a slab. Then
2719 * we reduce the minimum objects required in a slab.
2721 min_objects = slub_min_objects;
2723 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2724 max_objects = order_objects(slub_max_order, size, reserved);
2725 min_objects = min(min_objects, max_objects);
2727 while (min_objects > 1) {
2729 while (fraction >= 4) {
2730 order = slab_order(size, min_objects,
2731 slub_max_order, fraction, reserved);
2732 if (order <= slub_max_order)
2740 * We were unable to place multiple objects in a slab. Now
2741 * lets see if we can place a single object there.
2743 order = slab_order(size, 1, slub_max_order, 1, reserved);
2744 if (order <= slub_max_order)
2748 * Doh this slab cannot be placed using slub_max_order.
2750 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2751 if (order < MAX_ORDER)
2757 * Figure out what the alignment of the objects will be.
2759 static unsigned long calculate_alignment(unsigned long flags,
2760 unsigned long align, unsigned long size)
2763 * If the user wants hardware cache aligned objects then follow that
2764 * suggestion if the object is sufficiently large.
2766 * The hardware cache alignment cannot override the specified
2767 * alignment though. If that is greater then use it.
2769 if (flags & SLAB_HWCACHE_ALIGN) {
2770 unsigned long ralign = cache_line_size();
2771 while (size <= ralign / 2)
2773 align = max(align, ralign);
2776 if (align < ARCH_SLAB_MINALIGN)
2777 align = ARCH_SLAB_MINALIGN;
2779 return ALIGN(align, sizeof(void *));
2783 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2786 spin_lock_init(&n->list_lock);
2787 INIT_LIST_HEAD(&n->partial);
2788 #ifdef CONFIG_SLUB_DEBUG
2789 atomic_long_set(&n->nr_slabs, 0);
2790 atomic_long_set(&n->total_objects, 0);
2791 INIT_LIST_HEAD(&n->full);
2795 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2797 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2798 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2801 * Must align to double word boundary for the double cmpxchg
2802 * instructions to work; see __pcpu_double_call_return_bool().
2804 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2805 2 * sizeof(void *));
2810 init_kmem_cache_cpus(s);
2815 static struct kmem_cache *kmem_cache_node;
2818 * No kmalloc_node yet so do it by hand. We know that this is the first
2819 * slab on the node for this slabcache. There are no concurrent accesses
2822 * Note that this function only works on the kmalloc_node_cache
2823 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2824 * memory on a fresh node that has no slab structures yet.
2826 static void early_kmem_cache_node_alloc(int node)
2829 struct kmem_cache_node *n;
2831 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2833 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2836 if (page_to_nid(page) != node) {
2837 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2839 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2840 "in order to be able to continue\n");
2845 page->freelist = get_freepointer(kmem_cache_node, n);
2848 kmem_cache_node->node[node] = n;
2849 #ifdef CONFIG_SLUB_DEBUG
2850 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2851 init_tracking(kmem_cache_node, n);
2853 init_kmem_cache_node(n, kmem_cache_node);
2854 inc_slabs_node(kmem_cache_node, node, page->objects);
2856 add_partial(n, page, 0);
2859 static void free_kmem_cache_nodes(struct kmem_cache *s)
2863 for_each_node_state(node, N_NORMAL_MEMORY) {
2864 struct kmem_cache_node *n = s->node[node];
2867 kmem_cache_free(kmem_cache_node, n);
2869 s->node[node] = NULL;
2873 static int init_kmem_cache_nodes(struct kmem_cache *s)
2877 for_each_node_state(node, N_NORMAL_MEMORY) {
2878 struct kmem_cache_node *n;
2880 if (slab_state == DOWN) {
2881 early_kmem_cache_node_alloc(node);
2884 n = kmem_cache_alloc_node(kmem_cache_node,
2888 free_kmem_cache_nodes(s);
2893 init_kmem_cache_node(n, s);
2898 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2900 if (min < MIN_PARTIAL)
2902 else if (min > MAX_PARTIAL)
2904 s->min_partial = min;
2908 * calculate_sizes() determines the order and the distribution of data within
2911 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2913 unsigned long flags = s->flags;
2914 unsigned long size = s->objsize;
2915 unsigned long align = s->align;
2919 * Round up object size to the next word boundary. We can only
2920 * place the free pointer at word boundaries and this determines
2921 * the possible location of the free pointer.
2923 size = ALIGN(size, sizeof(void *));
2925 #ifdef CONFIG_SLUB_DEBUG
2927 * Determine if we can poison the object itself. If the user of
2928 * the slab may touch the object after free or before allocation
2929 * then we should never poison the object itself.
2931 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2933 s->flags |= __OBJECT_POISON;
2935 s->flags &= ~__OBJECT_POISON;
2939 * If we are Redzoning then check if there is some space between the
2940 * end of the object and the free pointer. If not then add an
2941 * additional word to have some bytes to store Redzone information.
2943 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2944 size += sizeof(void *);
2948 * With that we have determined the number of bytes in actual use
2949 * by the object. This is the potential offset to the free pointer.
2953 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2956 * Relocate free pointer after the object if it is not
2957 * permitted to overwrite the first word of the object on
2960 * This is the case if we do RCU, have a constructor or
2961 * destructor or are poisoning the objects.
2964 size += sizeof(void *);
2967 #ifdef CONFIG_SLUB_DEBUG
2968 if (flags & SLAB_STORE_USER)
2970 * Need to store information about allocs and frees after
2973 size += 2 * sizeof(struct track);
2975 if (flags & SLAB_RED_ZONE)
2977 * Add some empty padding so that we can catch
2978 * overwrites from earlier objects rather than let
2979 * tracking information or the free pointer be
2980 * corrupted if a user writes before the start
2983 size += sizeof(void *);
2987 * Determine the alignment based on various parameters that the
2988 * user specified and the dynamic determination of cache line size
2991 align = calculate_alignment(flags, align, s->objsize);
2995 * SLUB stores one object immediately after another beginning from
2996 * offset 0. In order to align the objects we have to simply size
2997 * each object to conform to the alignment.
2999 size = ALIGN(size, align);
3001 if (forced_order >= 0)
3002 order = forced_order;
3004 order = calculate_order(size, s->reserved);
3011 s->allocflags |= __GFP_COMP;
3013 if (s->flags & SLAB_CACHE_DMA)
3014 s->allocflags |= SLUB_DMA;
3016 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3017 s->allocflags |= __GFP_RECLAIMABLE;
3020 * Determine the number of objects per slab
3022 s->oo = oo_make(order, size, s->reserved);
3023 s->min = oo_make(get_order(size), size, s->reserved);
3024 if (oo_objects(s->oo) > oo_objects(s->max))
3027 return !!oo_objects(s->oo);
3031 static int kmem_cache_open(struct kmem_cache *s,
3032 const char *name, size_t size,
3033 size_t align, unsigned long flags,
3034 void (*ctor)(void *))
3036 memset(s, 0, kmem_size);
3041 s->flags = kmem_cache_flags(size, flags, name, ctor);
3044 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3045 s->reserved = sizeof(struct rcu_head);
3047 if (!calculate_sizes(s, -1))
3049 if (disable_higher_order_debug) {
3051 * Disable debugging flags that store metadata if the min slab
3054 if (get_order(s->size) > get_order(s->objsize)) {
3055 s->flags &= ~DEBUG_METADATA_FLAGS;
3057 if (!calculate_sizes(s, -1))
3062 #ifdef CONFIG_CMPXCHG_DOUBLE
3063 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3064 /* Enable fast mode */
3065 s->flags |= __CMPXCHG_DOUBLE;
3069 * The larger the object size is, the more pages we want on the partial
3070 * list to avoid pounding the page allocator excessively.
3072 set_min_partial(s, ilog2(s->size) / 2);
3075 * cpu_partial determined the maximum number of objects kept in the
3076 * per cpu partial lists of a processor.
3078 * Per cpu partial lists mainly contain slabs that just have one
3079 * object freed. If they are used for allocation then they can be
3080 * filled up again with minimal effort. The slab will never hit the
3081 * per node partial lists and therefore no locking will be required.
3083 * This setting also determines
3085 * A) The number of objects from per cpu partial slabs dumped to the
3086 * per node list when we reach the limit.
3087 * B) The number of objects in cpu partial slabs to extract from the
3088 * per node list when we run out of per cpu objects. We only fetch 50%
3089 * to keep some capacity around for frees.
3091 if (s->size >= PAGE_SIZE)
3093 else if (s->size >= 1024)
3095 else if (s->size >= 256)
3096 s->cpu_partial = 13;
3098 s->cpu_partial = 30;
3102 s->remote_node_defrag_ratio = 1000;
3104 if (!init_kmem_cache_nodes(s))
3107 if (alloc_kmem_cache_cpus(s))
3110 free_kmem_cache_nodes(s);
3112 if (flags & SLAB_PANIC)
3113 panic("Cannot create slab %s size=%lu realsize=%u "
3114 "order=%u offset=%u flags=%lx\n",
3115 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3121 * Determine the size of a slab object
3123 unsigned int kmem_cache_size(struct kmem_cache *s)
3127 EXPORT_SYMBOL(kmem_cache_size);
3129 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3132 #ifdef CONFIG_SLUB_DEBUG
3133 void *addr = page_address(page);
3135 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3136 sizeof(long), GFP_ATOMIC);
3139 slab_err(s, page, "%s", text);
3142 get_map(s, page, map);
3143 for_each_object(p, s, addr, page->objects) {
3145 if (!test_bit(slab_index(p, s, addr), map)) {
3146 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3148 print_tracking(s, p);
3157 * Attempt to free all partial slabs on a node.
3158 * This is called from kmem_cache_close(). We must be the last thread
3159 * using the cache and therefore we do not need to lock anymore.
3161 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3163 struct page *page, *h;
3165 list_for_each_entry_safe(page, h, &n->partial, lru) {
3167 remove_partial(n, page);
3168 discard_slab(s, page);
3170 list_slab_objects(s, page,
3171 "Objects remaining on kmem_cache_close()");
3177 * Release all resources used by a slab cache.
3179 static inline int kmem_cache_close(struct kmem_cache *s)
3184 free_percpu(s->cpu_slab);
3185 /* Attempt to free all objects */
3186 for_each_node_state(node, N_NORMAL_MEMORY) {
3187 struct kmem_cache_node *n = get_node(s, node);
3190 if (n->nr_partial || slabs_node(s, node))
3193 free_kmem_cache_nodes(s);
3198 * Close a cache and release the kmem_cache structure
3199 * (must be used for caches created using kmem_cache_create)
3201 void kmem_cache_destroy(struct kmem_cache *s)
3203 down_write(&slub_lock);
3207 up_write(&slub_lock);
3208 if (kmem_cache_close(s)) {
3209 printk(KERN_ERR "SLUB %s: %s called for cache that "
3210 "still has objects.\n", s->name, __func__);
3213 if (s->flags & SLAB_DESTROY_BY_RCU)
3215 sysfs_slab_remove(s);
3217 up_write(&slub_lock);
3219 EXPORT_SYMBOL(kmem_cache_destroy);
3221 /********************************************************************
3223 *******************************************************************/
3225 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3226 EXPORT_SYMBOL(kmalloc_caches);
3228 static struct kmem_cache *kmem_cache;
3230 #ifdef CONFIG_ZONE_DMA
3231 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3234 static int __init setup_slub_min_order(char *str)
3236 get_option(&str, &slub_min_order);
3241 __setup("slub_min_order=", setup_slub_min_order);
3243 static int __init setup_slub_max_order(char *str)
3245 get_option(&str, &slub_max_order);
3246 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3251 __setup("slub_max_order=", setup_slub_max_order);
3253 static int __init setup_slub_min_objects(char *str)
3255 get_option(&str, &slub_min_objects);
3260 __setup("slub_min_objects=", setup_slub_min_objects);
3262 static int __init setup_slub_nomerge(char *str)
3268 __setup("slub_nomerge", setup_slub_nomerge);
3270 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3271 int size, unsigned int flags)
3273 struct kmem_cache *s;
3275 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3278 * This function is called with IRQs disabled during early-boot on
3279 * single CPU so there's no need to take slub_lock here.
3281 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3285 list_add(&s->list, &slab_caches);
3289 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3294 * Conversion table for small slabs sizes / 8 to the index in the
3295 * kmalloc array. This is necessary for slabs < 192 since we have non power
3296 * of two cache sizes there. The size of larger slabs can be determined using
3299 static s8 size_index[24] = {
3326 static inline int size_index_elem(size_t bytes)
3328 return (bytes - 1) / 8;
3331 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3337 return ZERO_SIZE_PTR;
3339 index = size_index[size_index_elem(size)];
3341 index = fls(size - 1);
3343 #ifdef CONFIG_ZONE_DMA
3344 if (unlikely((flags & SLUB_DMA)))
3345 return kmalloc_dma_caches[index];
3348 return kmalloc_caches[index];
3351 void *__kmalloc(size_t size, gfp_t flags)
3353 struct kmem_cache *s;
3356 if (unlikely(size > SLUB_MAX_SIZE))
3357 return kmalloc_large(size, flags);
3359 s = get_slab(size, flags);
3361 if (unlikely(ZERO_OR_NULL_PTR(s)))
3364 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3366 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3370 EXPORT_SYMBOL(__kmalloc);
3373 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3378 flags |= __GFP_COMP | __GFP_NOTRACK;
3379 page = alloc_pages_node(node, flags, get_order(size));
3381 ptr = page_address(page);
3383 kmemleak_alloc(ptr, size, 1, flags);
3387 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3389 struct kmem_cache *s;
3392 if (unlikely(size > SLUB_MAX_SIZE)) {
3393 ret = kmalloc_large_node(size, flags, node);
3395 trace_kmalloc_node(_RET_IP_, ret,
3396 size, PAGE_SIZE << get_order(size),
3402 s = get_slab(size, flags);
3404 if (unlikely(ZERO_OR_NULL_PTR(s)))
3407 ret = slab_alloc(s, flags, node, _RET_IP_);
3409 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3413 EXPORT_SYMBOL(__kmalloc_node);
3416 size_t ksize(const void *object)
3420 if (unlikely(object == ZERO_SIZE_PTR))
3423 page = virt_to_head_page(object);
3425 if (unlikely(!PageSlab(page))) {
3426 WARN_ON(!PageCompound(page));
3427 return PAGE_SIZE << compound_order(page);
3430 return slab_ksize(page->slab);
3432 EXPORT_SYMBOL(ksize);
3434 #ifdef CONFIG_SLUB_DEBUG
3435 bool verify_mem_not_deleted(const void *x)
3438 void *object = (void *)x;
3439 unsigned long flags;
3442 if (unlikely(ZERO_OR_NULL_PTR(x)))
3445 local_irq_save(flags);
3447 page = virt_to_head_page(x);
3448 if (unlikely(!PageSlab(page))) {
3449 /* maybe it was from stack? */
3455 if (on_freelist(page->slab, page, object)) {
3456 object_err(page->slab, page, object, "Object is on free-list");
3464 local_irq_restore(flags);
3467 EXPORT_SYMBOL(verify_mem_not_deleted);
3470 void kfree(const void *x)
3473 void *object = (void *)x;
3475 trace_kfree(_RET_IP_, x);
3477 if (unlikely(ZERO_OR_NULL_PTR(x)))
3480 page = virt_to_head_page(x);
3481 if (unlikely(!PageSlab(page))) {
3482 BUG_ON(!PageCompound(page));
3487 slab_free(page->slab, page, object, _RET_IP_);
3489 EXPORT_SYMBOL(kfree);
3492 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3493 * the remaining slabs by the number of items in use. The slabs with the
3494 * most items in use come first. New allocations will then fill those up
3495 * and thus they can be removed from the partial lists.
3497 * The slabs with the least items are placed last. This results in them
3498 * being allocated from last increasing the chance that the last objects
3499 * are freed in them.
3501 int kmem_cache_shrink(struct kmem_cache *s)
3505 struct kmem_cache_node *n;
3508 int objects = oo_objects(s->max);
3509 struct list_head *slabs_by_inuse =
3510 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3511 unsigned long flags;
3513 if (!slabs_by_inuse)
3517 for_each_node_state(node, N_NORMAL_MEMORY) {
3518 n = get_node(s, node);
3523 for (i = 0; i < objects; i++)
3524 INIT_LIST_HEAD(slabs_by_inuse + i);
3526 spin_lock_irqsave(&n->list_lock, flags);
3529 * Build lists indexed by the items in use in each slab.
3531 * Note that concurrent frees may occur while we hold the
3532 * list_lock. page->inuse here is the upper limit.
3534 list_for_each_entry_safe(page, t, &n->partial, lru) {
3535 list_move(&page->lru, slabs_by_inuse + page->inuse);
3541 * Rebuild the partial list with the slabs filled up most
3542 * first and the least used slabs at the end.
3544 for (i = objects - 1; i > 0; i--)
3545 list_splice(slabs_by_inuse + i, n->partial.prev);
3547 spin_unlock_irqrestore(&n->list_lock, flags);
3549 /* Release empty slabs */
3550 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3551 discard_slab(s, page);
3554 kfree(slabs_by_inuse);
3557 EXPORT_SYMBOL(kmem_cache_shrink);
3559 #if defined(CONFIG_MEMORY_HOTPLUG)
3560 static int slab_mem_going_offline_callback(void *arg)
3562 struct kmem_cache *s;
3564 down_read(&slub_lock);
3565 list_for_each_entry(s, &slab_caches, list)
3566 kmem_cache_shrink(s);
3567 up_read(&slub_lock);
3572 static void slab_mem_offline_callback(void *arg)
3574 struct kmem_cache_node *n;
3575 struct kmem_cache *s;
3576 struct memory_notify *marg = arg;
3579 offline_node = marg->status_change_nid;
3582 * If the node still has available memory. we need kmem_cache_node
3585 if (offline_node < 0)
3588 down_read(&slub_lock);
3589 list_for_each_entry(s, &slab_caches, list) {
3590 n = get_node(s, offline_node);
3593 * if n->nr_slabs > 0, slabs still exist on the node
3594 * that is going down. We were unable to free them,
3595 * and offline_pages() function shouldn't call this
3596 * callback. So, we must fail.
3598 BUG_ON(slabs_node(s, offline_node));
3600 s->node[offline_node] = NULL;
3601 kmem_cache_free(kmem_cache_node, n);
3604 up_read(&slub_lock);
3607 static int slab_mem_going_online_callback(void *arg)
3609 struct kmem_cache_node *n;
3610 struct kmem_cache *s;
3611 struct memory_notify *marg = arg;
3612 int nid = marg->status_change_nid;
3616 * If the node's memory is already available, then kmem_cache_node is
3617 * already created. Nothing to do.
3623 * We are bringing a node online. No memory is available yet. We must
3624 * allocate a kmem_cache_node structure in order to bring the node
3627 down_read(&slub_lock);
3628 list_for_each_entry(s, &slab_caches, list) {
3630 * XXX: kmem_cache_alloc_node will fallback to other nodes
3631 * since memory is not yet available from the node that
3634 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3639 init_kmem_cache_node(n, s);
3643 up_read(&slub_lock);
3647 static int slab_memory_callback(struct notifier_block *self,
3648 unsigned long action, void *arg)
3653 case MEM_GOING_ONLINE:
3654 ret = slab_mem_going_online_callback(arg);
3656 case MEM_GOING_OFFLINE:
3657 ret = slab_mem_going_offline_callback(arg);
3660 case MEM_CANCEL_ONLINE:
3661 slab_mem_offline_callback(arg);
3664 case MEM_CANCEL_OFFLINE:
3668 ret = notifier_from_errno(ret);
3674 #endif /* CONFIG_MEMORY_HOTPLUG */
3676 /********************************************************************
3677 * Basic setup of slabs
3678 *******************************************************************/
3681 * Used for early kmem_cache structures that were allocated using
3682 * the page allocator
3685 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3689 list_add(&s->list, &slab_caches);
3692 for_each_node_state(node, N_NORMAL_MEMORY) {
3693 struct kmem_cache_node *n = get_node(s, node);
3697 list_for_each_entry(p, &n->partial, lru)
3700 #ifdef CONFIG_SLUB_DEBUG
3701 list_for_each_entry(p, &n->full, lru)
3708 void __init kmem_cache_init(void)
3712 struct kmem_cache *temp_kmem_cache;
3714 struct kmem_cache *temp_kmem_cache_node;
3715 unsigned long kmalloc_size;
3717 kmem_size = offsetof(struct kmem_cache, node) +
3718 nr_node_ids * sizeof(struct kmem_cache_node *);
3720 /* Allocate two kmem_caches from the page allocator */
3721 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3722 order = get_order(2 * kmalloc_size);
3723 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3726 * Must first have the slab cache available for the allocations of the
3727 * struct kmem_cache_node's. There is special bootstrap code in
3728 * kmem_cache_open for slab_state == DOWN.
3730 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3732 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3733 sizeof(struct kmem_cache_node),
3734 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3736 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3738 /* Able to allocate the per node structures */
3739 slab_state = PARTIAL;
3741 temp_kmem_cache = kmem_cache;
3742 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3743 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3744 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3745 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3748 * Allocate kmem_cache_node properly from the kmem_cache slab.
3749 * kmem_cache_node is separately allocated so no need to
3750 * update any list pointers.
3752 temp_kmem_cache_node = kmem_cache_node;
3754 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3755 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3757 kmem_cache_bootstrap_fixup(kmem_cache_node);
3760 kmem_cache_bootstrap_fixup(kmem_cache);
3762 /* Free temporary boot structure */
3763 free_pages((unsigned long)temp_kmem_cache, order);
3765 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3768 * Patch up the size_index table if we have strange large alignment
3769 * requirements for the kmalloc array. This is only the case for
3770 * MIPS it seems. The standard arches will not generate any code here.
3772 * Largest permitted alignment is 256 bytes due to the way we
3773 * handle the index determination for the smaller caches.
3775 * Make sure that nothing crazy happens if someone starts tinkering
3776 * around with ARCH_KMALLOC_MINALIGN
3778 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3779 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3781 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3782 int elem = size_index_elem(i);
3783 if (elem >= ARRAY_SIZE(size_index))
3785 size_index[elem] = KMALLOC_SHIFT_LOW;
3788 if (KMALLOC_MIN_SIZE == 64) {
3790 * The 96 byte size cache is not used if the alignment
3793 for (i = 64 + 8; i <= 96; i += 8)
3794 size_index[size_index_elem(i)] = 7;
3795 } else if (KMALLOC_MIN_SIZE == 128) {
3797 * The 192 byte sized cache is not used if the alignment
3798 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3801 for (i = 128 + 8; i <= 192; i += 8)
3802 size_index[size_index_elem(i)] = 8;
3805 /* Caches that are not of the two-to-the-power-of size */
3806 if (KMALLOC_MIN_SIZE <= 32) {
3807 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3811 if (KMALLOC_MIN_SIZE <= 64) {
3812 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3816 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3817 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3823 /* Provide the correct kmalloc names now that the caches are up */
3824 if (KMALLOC_MIN_SIZE <= 32) {
3825 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3826 BUG_ON(!kmalloc_caches[1]->name);
3829 if (KMALLOC_MIN_SIZE <= 64) {
3830 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3831 BUG_ON(!kmalloc_caches[2]->name);
3834 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3835 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3838 kmalloc_caches[i]->name = s;
3842 register_cpu_notifier(&slab_notifier);
3845 #ifdef CONFIG_ZONE_DMA
3846 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3847 struct kmem_cache *s = kmalloc_caches[i];
3850 char *name = kasprintf(GFP_NOWAIT,
3851 "dma-kmalloc-%d", s->objsize);
3854 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3855 s->objsize, SLAB_CACHE_DMA);
3860 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3861 " CPUs=%d, Nodes=%d\n",
3862 caches, cache_line_size(),
3863 slub_min_order, slub_max_order, slub_min_objects,
3864 nr_cpu_ids, nr_node_ids);
3867 void __init kmem_cache_init_late(void)
3872 * Find a mergeable slab cache
3874 static int slab_unmergeable(struct kmem_cache *s)
3876 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3883 * We may have set a slab to be unmergeable during bootstrap.
3885 if (s->refcount < 0)
3891 static struct kmem_cache *find_mergeable(size_t size,
3892 size_t align, unsigned long flags, const char *name,
3893 void (*ctor)(void *))
3895 struct kmem_cache *s;
3897 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3903 size = ALIGN(size, sizeof(void *));
3904 align = calculate_alignment(flags, align, size);
3905 size = ALIGN(size, align);
3906 flags = kmem_cache_flags(size, flags, name, NULL);
3908 list_for_each_entry(s, &slab_caches, list) {
3909 if (slab_unmergeable(s))
3915 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3918 * Check if alignment is compatible.
3919 * Courtesy of Adrian Drzewiecki
3921 if ((s->size & ~(align - 1)) != s->size)
3924 if (s->size - size >= sizeof(void *))
3932 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3933 size_t align, unsigned long flags, void (*ctor)(void *))
3935 struct kmem_cache *s;
3941 down_write(&slub_lock);
3942 s = find_mergeable(size, align, flags, name, ctor);
3946 * Adjust the object sizes so that we clear
3947 * the complete object on kzalloc.
3949 s->objsize = max(s->objsize, (int)size);
3950 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3952 if (sysfs_slab_alias(s, name)) {
3956 up_write(&slub_lock);
3960 n = kstrdup(name, GFP_KERNEL);
3964 s = kmalloc(kmem_size, GFP_KERNEL);
3966 if (kmem_cache_open(s, n,
3967 size, align, flags, ctor)) {
3968 list_add(&s->list, &slab_caches);
3969 if (sysfs_slab_add(s)) {
3975 up_write(&slub_lock);
3982 up_write(&slub_lock);
3984 if (flags & SLAB_PANIC)
3985 panic("Cannot create slabcache %s\n", name);
3990 EXPORT_SYMBOL(kmem_cache_create);
3994 * Use the cpu notifier to insure that the cpu slabs are flushed when
3997 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3998 unsigned long action, void *hcpu)
4000 long cpu = (long)hcpu;
4001 struct kmem_cache *s;
4002 unsigned long flags;
4005 case CPU_UP_CANCELED:
4006 case CPU_UP_CANCELED_FROZEN:
4008 case CPU_DEAD_FROZEN:
4009 down_read(&slub_lock);
4010 list_for_each_entry(s, &slab_caches, list) {
4011 local_irq_save(flags);
4012 __flush_cpu_slab(s, cpu);
4013 local_irq_restore(flags);
4015 up_read(&slub_lock);
4023 static struct notifier_block __cpuinitdata slab_notifier = {
4024 .notifier_call = slab_cpuup_callback
4029 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4031 struct kmem_cache *s;
4034 if (unlikely(size > SLUB_MAX_SIZE))
4035 return kmalloc_large(size, gfpflags);
4037 s = get_slab(size, gfpflags);
4039 if (unlikely(ZERO_OR_NULL_PTR(s)))
4042 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4044 /* Honor the call site pointer we received. */
4045 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4051 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4052 int node, unsigned long caller)
4054 struct kmem_cache *s;
4057 if (unlikely(size > SLUB_MAX_SIZE)) {
4058 ret = kmalloc_large_node(size, gfpflags, node);
4060 trace_kmalloc_node(caller, ret,
4061 size, PAGE_SIZE << get_order(size),
4067 s = get_slab(size, gfpflags);
4069 if (unlikely(ZERO_OR_NULL_PTR(s)))
4072 ret = slab_alloc(s, gfpflags, node, caller);
4074 /* Honor the call site pointer we received. */
4075 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4082 static int count_inuse(struct page *page)
4087 static int count_total(struct page *page)
4089 return page->objects;
4093 #ifdef CONFIG_SLUB_DEBUG
4094 static int validate_slab(struct kmem_cache *s, struct page *page,
4098 void *addr = page_address(page);
4100 if (!check_slab(s, page) ||
4101 !on_freelist(s, page, NULL))
4104 /* Now we know that a valid freelist exists */
4105 bitmap_zero(map, page->objects);
4107 get_map(s, page, map);
4108 for_each_object(p, s, addr, page->objects) {
4109 if (test_bit(slab_index(p, s, addr), map))
4110 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4114 for_each_object(p, s, addr, page->objects)
4115 if (!test_bit(slab_index(p, s, addr), map))
4116 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4121 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4125 validate_slab(s, page, map);
4129 static int validate_slab_node(struct kmem_cache *s,
4130 struct kmem_cache_node *n, unsigned long *map)
4132 unsigned long count = 0;
4134 unsigned long flags;
4136 spin_lock_irqsave(&n->list_lock, flags);
4138 list_for_each_entry(page, &n->partial, lru) {
4139 validate_slab_slab(s, page, map);
4142 if (count != n->nr_partial)
4143 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4144 "counter=%ld\n", s->name, count, n->nr_partial);
4146 if (!(s->flags & SLAB_STORE_USER))
4149 list_for_each_entry(page, &n->full, lru) {
4150 validate_slab_slab(s, page, map);
4153 if (count != atomic_long_read(&n->nr_slabs))
4154 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4155 "counter=%ld\n", s->name, count,
4156 atomic_long_read(&n->nr_slabs));
4159 spin_unlock_irqrestore(&n->list_lock, flags);
4163 static long validate_slab_cache(struct kmem_cache *s)
4166 unsigned long count = 0;
4167 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4168 sizeof(unsigned long), GFP_KERNEL);
4174 for_each_node_state(node, N_NORMAL_MEMORY) {
4175 struct kmem_cache_node *n = get_node(s, node);
4177 count += validate_slab_node(s, n, map);
4183 * Generate lists of code addresses where slabcache objects are allocated
4188 unsigned long count;
4195 DECLARE_BITMAP(cpus, NR_CPUS);
4201 unsigned long count;
4202 struct location *loc;
4205 static void free_loc_track(struct loc_track *t)
4208 free_pages((unsigned long)t->loc,
4209 get_order(sizeof(struct location) * t->max));
4212 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4217 order = get_order(sizeof(struct location) * max);
4219 l = (void *)__get_free_pages(flags, order);
4224 memcpy(l, t->loc, sizeof(struct location) * t->count);
4232 static int add_location(struct loc_track *t, struct kmem_cache *s,
4233 const struct track *track)
4235 long start, end, pos;
4237 unsigned long caddr;
4238 unsigned long age = jiffies - track->when;
4244 pos = start + (end - start + 1) / 2;
4247 * There is nothing at "end". If we end up there
4248 * we need to add something to before end.
4253 caddr = t->loc[pos].addr;
4254 if (track->addr == caddr) {
4260 if (age < l->min_time)
4262 if (age > l->max_time)
4265 if (track->pid < l->min_pid)
4266 l->min_pid = track->pid;
4267 if (track->pid > l->max_pid)
4268 l->max_pid = track->pid;
4270 cpumask_set_cpu(track->cpu,
4271 to_cpumask(l->cpus));
4273 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4277 if (track->addr < caddr)
4284 * Not found. Insert new tracking element.
4286 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4292 (t->count - pos) * sizeof(struct location));
4295 l->addr = track->addr;
4299 l->min_pid = track->pid;
4300 l->max_pid = track->pid;
4301 cpumask_clear(to_cpumask(l->cpus));
4302 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4303 nodes_clear(l->nodes);
4304 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4308 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4309 struct page *page, enum track_item alloc,
4312 void *addr = page_address(page);
4315 bitmap_zero(map, page->objects);
4316 get_map(s, page, map);
4318 for_each_object(p, s, addr, page->objects)
4319 if (!test_bit(slab_index(p, s, addr), map))
4320 add_location(t, s, get_track(s, p, alloc));
4323 static int list_locations(struct kmem_cache *s, char *buf,
4324 enum track_item alloc)
4328 struct loc_track t = { 0, 0, NULL };
4330 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4331 sizeof(unsigned long), GFP_KERNEL);
4333 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4336 return sprintf(buf, "Out of memory\n");
4338 /* Push back cpu slabs */
4341 for_each_node_state(node, N_NORMAL_MEMORY) {
4342 struct kmem_cache_node *n = get_node(s, node);
4343 unsigned long flags;
4346 if (!atomic_long_read(&n->nr_slabs))
4349 spin_lock_irqsave(&n->list_lock, flags);
4350 list_for_each_entry(page, &n->partial, lru)
4351 process_slab(&t, s, page, alloc, map);
4352 list_for_each_entry(page, &n->full, lru)
4353 process_slab(&t, s, page, alloc, map);
4354 spin_unlock_irqrestore(&n->list_lock, flags);
4357 for (i = 0; i < t.count; i++) {
4358 struct location *l = &t.loc[i];
4360 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4362 len += sprintf(buf + len, "%7ld ", l->count);
4365 len += sprintf(buf + len, "%pS", (void *)l->addr);
4367 len += sprintf(buf + len, "<not-available>");
4369 if (l->sum_time != l->min_time) {
4370 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4372 (long)div_u64(l->sum_time, l->count),
4375 len += sprintf(buf + len, " age=%ld",
4378 if (l->min_pid != l->max_pid)
4379 len += sprintf(buf + len, " pid=%ld-%ld",
4380 l->min_pid, l->max_pid);
4382 len += sprintf(buf + len, " pid=%ld",
4385 if (num_online_cpus() > 1 &&
4386 !cpumask_empty(to_cpumask(l->cpus)) &&
4387 len < PAGE_SIZE - 60) {
4388 len += sprintf(buf + len, " cpus=");
4389 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4390 to_cpumask(l->cpus));
4393 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4394 len < PAGE_SIZE - 60) {
4395 len += sprintf(buf + len, " nodes=");
4396 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4400 len += sprintf(buf + len, "\n");
4406 len += sprintf(buf, "No data\n");
4411 #ifdef SLUB_RESILIENCY_TEST
4412 static void resiliency_test(void)
4416 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4418 printk(KERN_ERR "SLUB resiliency testing\n");
4419 printk(KERN_ERR "-----------------------\n");
4420 printk(KERN_ERR "A. Corruption after allocation\n");
4422 p = kzalloc(16, GFP_KERNEL);
4424 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4425 " 0x12->0x%p\n\n", p + 16);
4427 validate_slab_cache(kmalloc_caches[4]);
4429 /* Hmmm... The next two are dangerous */
4430 p = kzalloc(32, GFP_KERNEL);
4431 p[32 + sizeof(void *)] = 0x34;
4432 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4433 " 0x34 -> -0x%p\n", p);
4435 "If allocated object is overwritten then not detectable\n\n");
4437 validate_slab_cache(kmalloc_caches[5]);
4438 p = kzalloc(64, GFP_KERNEL);
4439 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4441 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4444 "If allocated object is overwritten then not detectable\n\n");
4445 validate_slab_cache(kmalloc_caches[6]);
4447 printk(KERN_ERR "\nB. Corruption after free\n");
4448 p = kzalloc(128, GFP_KERNEL);
4451 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4452 validate_slab_cache(kmalloc_caches[7]);
4454 p = kzalloc(256, GFP_KERNEL);
4457 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4459 validate_slab_cache(kmalloc_caches[8]);
4461 p = kzalloc(512, GFP_KERNEL);
4464 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4465 validate_slab_cache(kmalloc_caches[9]);
4469 static void resiliency_test(void) {};
4474 enum slab_stat_type {
4475 SL_ALL, /* All slabs */
4476 SL_PARTIAL, /* Only partially allocated slabs */
4477 SL_CPU, /* Only slabs used for cpu caches */
4478 SL_OBJECTS, /* Determine allocated objects not slabs */
4479 SL_TOTAL /* Determine object capacity not slabs */
4482 #define SO_ALL (1 << SL_ALL)
4483 #define SO_PARTIAL (1 << SL_PARTIAL)
4484 #define SO_CPU (1 << SL_CPU)
4485 #define SO_OBJECTS (1 << SL_OBJECTS)
4486 #define SO_TOTAL (1 << SL_TOTAL)
4488 static ssize_t show_slab_objects(struct kmem_cache *s,
4489 char *buf, unsigned long flags)
4491 unsigned long total = 0;
4494 unsigned long *nodes;
4495 unsigned long *per_cpu;
4497 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4500 per_cpu = nodes + nr_node_ids;
4502 if (flags & SO_CPU) {
4505 for_each_possible_cpu(cpu) {
4506 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4509 if (!c || c->node < 0)
4513 if (flags & SO_TOTAL)
4514 x = c->page->objects;
4515 else if (flags & SO_OBJECTS)
4521 nodes[c->node] += x;
4528 nodes[c->node] += x;
4534 lock_memory_hotplug();
4535 #ifdef CONFIG_SLUB_DEBUG
4536 if (flags & SO_ALL) {
4537 for_each_node_state(node, N_NORMAL_MEMORY) {
4538 struct kmem_cache_node *n = get_node(s, node);
4540 if (flags & SO_TOTAL)
4541 x = atomic_long_read(&n->total_objects);
4542 else if (flags & SO_OBJECTS)
4543 x = atomic_long_read(&n->total_objects) -
4544 count_partial(n, count_free);
4547 x = atomic_long_read(&n->nr_slabs);
4554 if (flags & SO_PARTIAL) {
4555 for_each_node_state(node, N_NORMAL_MEMORY) {
4556 struct kmem_cache_node *n = get_node(s, node);
4558 if (flags & SO_TOTAL)
4559 x = count_partial(n, count_total);
4560 else if (flags & SO_OBJECTS)
4561 x = count_partial(n, count_inuse);
4568 x = sprintf(buf, "%lu", total);
4570 for_each_node_state(node, N_NORMAL_MEMORY)
4572 x += sprintf(buf + x, " N%d=%lu",
4575 unlock_memory_hotplug();
4577 return x + sprintf(buf + x, "\n");
4580 #ifdef CONFIG_SLUB_DEBUG
4581 static int any_slab_objects(struct kmem_cache *s)
4585 for_each_online_node(node) {
4586 struct kmem_cache_node *n = get_node(s, node);
4591 if (atomic_long_read(&n->total_objects))
4598 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4599 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4601 struct slab_attribute {
4602 struct attribute attr;
4603 ssize_t (*show)(struct kmem_cache *s, char *buf);
4604 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4607 #define SLAB_ATTR_RO(_name) \
4608 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4610 #define SLAB_ATTR(_name) \
4611 static struct slab_attribute _name##_attr = \
4612 __ATTR(_name, 0644, _name##_show, _name##_store)
4614 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4616 return sprintf(buf, "%d\n", s->size);
4618 SLAB_ATTR_RO(slab_size);
4620 static ssize_t align_show(struct kmem_cache *s, char *buf)
4622 return sprintf(buf, "%d\n", s->align);
4624 SLAB_ATTR_RO(align);
4626 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4628 return sprintf(buf, "%d\n", s->objsize);
4630 SLAB_ATTR_RO(object_size);
4632 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4634 return sprintf(buf, "%d\n", oo_objects(s->oo));
4636 SLAB_ATTR_RO(objs_per_slab);
4638 static ssize_t order_store(struct kmem_cache *s,
4639 const char *buf, size_t length)
4641 unsigned long order;
4644 err = strict_strtoul(buf, 10, &order);
4648 if (order > slub_max_order || order < slub_min_order)
4651 calculate_sizes(s, order);
4655 static ssize_t order_show(struct kmem_cache *s, char *buf)
4657 return sprintf(buf, "%d\n", oo_order(s->oo));
4661 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4663 return sprintf(buf, "%lu\n", s->min_partial);
4666 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4672 err = strict_strtoul(buf, 10, &min);
4676 set_min_partial(s, min);
4679 SLAB_ATTR(min_partial);
4681 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4683 return sprintf(buf, "%u\n", s->cpu_partial);
4686 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4689 unsigned long objects;
4692 err = strict_strtoul(buf, 10, &objects);
4696 s->cpu_partial = objects;
4700 SLAB_ATTR(cpu_partial);
4702 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4706 return sprintf(buf, "%pS\n", s->ctor);
4710 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4712 return sprintf(buf, "%d\n", s->refcount - 1);
4714 SLAB_ATTR_RO(aliases);
4716 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4718 return show_slab_objects(s, buf, SO_PARTIAL);
4720 SLAB_ATTR_RO(partial);
4722 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4724 return show_slab_objects(s, buf, SO_CPU);
4726 SLAB_ATTR_RO(cpu_slabs);
4728 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4730 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4732 SLAB_ATTR_RO(objects);
4734 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4736 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4738 SLAB_ATTR_RO(objects_partial);
4740 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4747 for_each_online_cpu(cpu) {
4748 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4751 pages += page->pages;
4752 objects += page->pobjects;
4756 len = sprintf(buf, "%d(%d)", objects, pages);
4759 for_each_online_cpu(cpu) {
4760 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4762 if (page && len < PAGE_SIZE - 20)
4763 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4764 page->pobjects, page->pages);
4767 return len + sprintf(buf + len, "\n");
4769 SLAB_ATTR_RO(slabs_cpu_partial);
4771 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4773 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4776 static ssize_t reclaim_account_store(struct kmem_cache *s,
4777 const char *buf, size_t length)
4779 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4781 s->flags |= SLAB_RECLAIM_ACCOUNT;
4784 SLAB_ATTR(reclaim_account);
4786 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4788 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4790 SLAB_ATTR_RO(hwcache_align);
4792 #ifdef CONFIG_ZONE_DMA
4793 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4795 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4797 SLAB_ATTR_RO(cache_dma);
4800 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4802 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4804 SLAB_ATTR_RO(destroy_by_rcu);
4806 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4808 return sprintf(buf, "%d\n", s->reserved);
4810 SLAB_ATTR_RO(reserved);
4812 #ifdef CONFIG_SLUB_DEBUG
4813 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4815 return show_slab_objects(s, buf, SO_ALL);
4817 SLAB_ATTR_RO(slabs);
4819 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4821 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4823 SLAB_ATTR_RO(total_objects);
4825 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4827 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4830 static ssize_t sanity_checks_store(struct kmem_cache *s,
4831 const char *buf, size_t length)
4833 s->flags &= ~SLAB_DEBUG_FREE;
4834 if (buf[0] == '1') {
4835 s->flags &= ~__CMPXCHG_DOUBLE;
4836 s->flags |= SLAB_DEBUG_FREE;
4840 SLAB_ATTR(sanity_checks);
4842 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4847 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4850 s->flags &= ~SLAB_TRACE;
4851 if (buf[0] == '1') {
4852 s->flags &= ~__CMPXCHG_DOUBLE;
4853 s->flags |= SLAB_TRACE;
4859 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4861 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4864 static ssize_t red_zone_store(struct kmem_cache *s,
4865 const char *buf, size_t length)
4867 if (any_slab_objects(s))
4870 s->flags &= ~SLAB_RED_ZONE;
4871 if (buf[0] == '1') {
4872 s->flags &= ~__CMPXCHG_DOUBLE;
4873 s->flags |= SLAB_RED_ZONE;
4875 calculate_sizes(s, -1);
4878 SLAB_ATTR(red_zone);
4880 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4882 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4885 static ssize_t poison_store(struct kmem_cache *s,
4886 const char *buf, size_t length)
4888 if (any_slab_objects(s))
4891 s->flags &= ~SLAB_POISON;
4892 if (buf[0] == '1') {
4893 s->flags &= ~__CMPXCHG_DOUBLE;
4894 s->flags |= SLAB_POISON;
4896 calculate_sizes(s, -1);
4901 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4903 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4906 static ssize_t store_user_store(struct kmem_cache *s,
4907 const char *buf, size_t length)
4909 if (any_slab_objects(s))
4912 s->flags &= ~SLAB_STORE_USER;
4913 if (buf[0] == '1') {
4914 s->flags &= ~__CMPXCHG_DOUBLE;
4915 s->flags |= SLAB_STORE_USER;
4917 calculate_sizes(s, -1);
4920 SLAB_ATTR(store_user);
4922 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4927 static ssize_t validate_store(struct kmem_cache *s,
4928 const char *buf, size_t length)
4932 if (buf[0] == '1') {
4933 ret = validate_slab_cache(s);
4939 SLAB_ATTR(validate);
4941 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4943 if (!(s->flags & SLAB_STORE_USER))
4945 return list_locations(s, buf, TRACK_ALLOC);
4947 SLAB_ATTR_RO(alloc_calls);
4949 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4951 if (!(s->flags & SLAB_STORE_USER))
4953 return list_locations(s, buf, TRACK_FREE);
4955 SLAB_ATTR_RO(free_calls);
4956 #endif /* CONFIG_SLUB_DEBUG */
4958 #ifdef CONFIG_FAILSLAB
4959 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4961 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4964 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4967 s->flags &= ~SLAB_FAILSLAB;
4969 s->flags |= SLAB_FAILSLAB;
4972 SLAB_ATTR(failslab);
4975 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4980 static ssize_t shrink_store(struct kmem_cache *s,
4981 const char *buf, size_t length)
4983 if (buf[0] == '1') {
4984 int rc = kmem_cache_shrink(s);
4995 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4997 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5000 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5001 const char *buf, size_t length)
5003 unsigned long ratio;
5006 err = strict_strtoul(buf, 10, &ratio);
5011 s->remote_node_defrag_ratio = ratio * 10;
5015 SLAB_ATTR(remote_node_defrag_ratio);
5018 #ifdef CONFIG_SLUB_STATS
5019 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5021 unsigned long sum = 0;
5024 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5029 for_each_online_cpu(cpu) {
5030 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5036 len = sprintf(buf, "%lu", sum);
5039 for_each_online_cpu(cpu) {
5040 if (data[cpu] && len < PAGE_SIZE - 20)
5041 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5045 return len + sprintf(buf + len, "\n");
5048 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5052 for_each_online_cpu(cpu)
5053 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5056 #define STAT_ATTR(si, text) \
5057 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5059 return show_stat(s, buf, si); \
5061 static ssize_t text##_store(struct kmem_cache *s, \
5062 const char *buf, size_t length) \
5064 if (buf[0] != '0') \
5066 clear_stat(s, si); \
5071 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5072 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5073 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5074 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5075 STAT_ATTR(FREE_FROZEN, free_frozen);
5076 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5077 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5078 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5079 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5080 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5081 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5082 STAT_ATTR(FREE_SLAB, free_slab);
5083 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5084 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5085 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5086 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5087 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5088 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5089 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5090 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5091 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5092 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5093 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5094 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5097 static struct attribute *slab_attrs[] = {
5098 &slab_size_attr.attr,
5099 &object_size_attr.attr,
5100 &objs_per_slab_attr.attr,
5102 &min_partial_attr.attr,
5103 &cpu_partial_attr.attr,
5105 &objects_partial_attr.attr,
5107 &cpu_slabs_attr.attr,
5111 &hwcache_align_attr.attr,
5112 &reclaim_account_attr.attr,
5113 &destroy_by_rcu_attr.attr,
5115 &reserved_attr.attr,
5116 &slabs_cpu_partial_attr.attr,
5117 #ifdef CONFIG_SLUB_DEBUG
5118 &total_objects_attr.attr,
5120 &sanity_checks_attr.attr,
5122 &red_zone_attr.attr,
5124 &store_user_attr.attr,
5125 &validate_attr.attr,
5126 &alloc_calls_attr.attr,
5127 &free_calls_attr.attr,
5129 #ifdef CONFIG_ZONE_DMA
5130 &cache_dma_attr.attr,
5133 &remote_node_defrag_ratio_attr.attr,
5135 #ifdef CONFIG_SLUB_STATS
5136 &alloc_fastpath_attr.attr,
5137 &alloc_slowpath_attr.attr,
5138 &free_fastpath_attr.attr,
5139 &free_slowpath_attr.attr,
5140 &free_frozen_attr.attr,
5141 &free_add_partial_attr.attr,
5142 &free_remove_partial_attr.attr,
5143 &alloc_from_partial_attr.attr,
5144 &alloc_slab_attr.attr,
5145 &alloc_refill_attr.attr,
5146 &alloc_node_mismatch_attr.attr,
5147 &free_slab_attr.attr,
5148 &cpuslab_flush_attr.attr,
5149 &deactivate_full_attr.attr,
5150 &deactivate_empty_attr.attr,
5151 &deactivate_to_head_attr.attr,
5152 &deactivate_to_tail_attr.attr,
5153 &deactivate_remote_frees_attr.attr,
5154 &deactivate_bypass_attr.attr,
5155 &order_fallback_attr.attr,
5156 &cmpxchg_double_fail_attr.attr,
5157 &cmpxchg_double_cpu_fail_attr.attr,
5158 &cpu_partial_alloc_attr.attr,
5159 &cpu_partial_free_attr.attr,
5161 #ifdef CONFIG_FAILSLAB
5162 &failslab_attr.attr,
5168 static struct attribute_group slab_attr_group = {
5169 .attrs = slab_attrs,
5172 static ssize_t slab_attr_show(struct kobject *kobj,
5173 struct attribute *attr,
5176 struct slab_attribute *attribute;
5177 struct kmem_cache *s;
5180 attribute = to_slab_attr(attr);
5183 if (!attribute->show)
5186 err = attribute->show(s, buf);
5191 static ssize_t slab_attr_store(struct kobject *kobj,
5192 struct attribute *attr,
5193 const char *buf, size_t len)
5195 struct slab_attribute *attribute;
5196 struct kmem_cache *s;
5199 attribute = to_slab_attr(attr);
5202 if (!attribute->store)
5205 err = attribute->store(s, buf, len);
5210 static void kmem_cache_release(struct kobject *kobj)
5212 struct kmem_cache *s = to_slab(kobj);
5218 static const struct sysfs_ops slab_sysfs_ops = {
5219 .show = slab_attr_show,
5220 .store = slab_attr_store,
5223 static struct kobj_type slab_ktype = {
5224 .sysfs_ops = &slab_sysfs_ops,
5225 .release = kmem_cache_release
5228 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5230 struct kobj_type *ktype = get_ktype(kobj);
5232 if (ktype == &slab_ktype)
5237 static const struct kset_uevent_ops slab_uevent_ops = {
5238 .filter = uevent_filter,
5241 static struct kset *slab_kset;
5243 #define ID_STR_LENGTH 64
5245 /* Create a unique string id for a slab cache:
5247 * Format :[flags-]size
5249 static char *create_unique_id(struct kmem_cache *s)
5251 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5258 * First flags affecting slabcache operations. We will only
5259 * get here for aliasable slabs so we do not need to support
5260 * too many flags. The flags here must cover all flags that
5261 * are matched during merging to guarantee that the id is
5264 if (s->flags & SLAB_CACHE_DMA)
5266 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5268 if (s->flags & SLAB_DEBUG_FREE)
5270 if (!(s->flags & SLAB_NOTRACK))
5274 p += sprintf(p, "%07d", s->size);
5275 BUG_ON(p > name + ID_STR_LENGTH - 1);
5279 static int sysfs_slab_add(struct kmem_cache *s)
5285 if (slab_state < SYSFS)
5286 /* Defer until later */
5289 unmergeable = slab_unmergeable(s);
5292 * Slabcache can never be merged so we can use the name proper.
5293 * This is typically the case for debug situations. In that
5294 * case we can catch duplicate names easily.
5296 sysfs_remove_link(&slab_kset->kobj, s->name);
5300 * Create a unique name for the slab as a target
5303 name = create_unique_id(s);
5306 s->kobj.kset = slab_kset;
5307 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5309 kobject_put(&s->kobj);
5313 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5315 kobject_del(&s->kobj);
5316 kobject_put(&s->kobj);
5319 kobject_uevent(&s->kobj, KOBJ_ADD);
5321 /* Setup first alias */
5322 sysfs_slab_alias(s, s->name);
5328 static void sysfs_slab_remove(struct kmem_cache *s)
5330 if (slab_state < SYSFS)
5332 * Sysfs has not been setup yet so no need to remove the
5337 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5338 kobject_del(&s->kobj);
5339 kobject_put(&s->kobj);
5343 * Need to buffer aliases during bootup until sysfs becomes
5344 * available lest we lose that information.
5346 struct saved_alias {
5347 struct kmem_cache *s;
5349 struct saved_alias *next;
5352 static struct saved_alias *alias_list;
5354 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5356 struct saved_alias *al;
5358 if (slab_state == SYSFS) {
5360 * If we have a leftover link then remove it.
5362 sysfs_remove_link(&slab_kset->kobj, name);
5363 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5366 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5372 al->next = alias_list;
5377 static int __init slab_sysfs_init(void)
5379 struct kmem_cache *s;
5382 down_write(&slub_lock);
5384 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5386 up_write(&slub_lock);
5387 printk(KERN_ERR "Cannot register slab subsystem.\n");
5393 list_for_each_entry(s, &slab_caches, list) {
5394 err = sysfs_slab_add(s);
5396 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5397 " to sysfs\n", s->name);
5400 while (alias_list) {
5401 struct saved_alias *al = alias_list;
5403 alias_list = alias_list->next;
5404 err = sysfs_slab_alias(al->s, al->name);
5406 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5407 " %s to sysfs\n", s->name);
5411 up_write(&slub_lock);
5416 __initcall(slab_sysfs_init);
5417 #endif /* CONFIG_SYSFS */
5420 * The /proc/slabinfo ABI
5422 #ifdef CONFIG_SLABINFO
5423 static void print_slabinfo_header(struct seq_file *m)
5425 seq_puts(m, "slabinfo - version: 2.1\n");
5426 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5427 "<objperslab> <pagesperslab>");
5428 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5429 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5433 static void *s_start(struct seq_file *m, loff_t *pos)
5437 down_read(&slub_lock);
5439 print_slabinfo_header(m);
5441 return seq_list_start(&slab_caches, *pos);
5444 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5446 return seq_list_next(p, &slab_caches, pos);
5449 static void s_stop(struct seq_file *m, void *p)
5451 up_read(&slub_lock);
5454 static int s_show(struct seq_file *m, void *p)
5456 unsigned long nr_partials = 0;
5457 unsigned long nr_slabs = 0;
5458 unsigned long nr_inuse = 0;
5459 unsigned long nr_objs = 0;
5460 unsigned long nr_free = 0;
5461 struct kmem_cache *s;
5464 s = list_entry(p, struct kmem_cache, list);
5466 for_each_online_node(node) {
5467 struct kmem_cache_node *n = get_node(s, node);
5472 nr_partials += n->nr_partial;
5473 nr_slabs += atomic_long_read(&n->nr_slabs);
5474 nr_objs += atomic_long_read(&n->total_objects);
5475 nr_free += count_partial(n, count_free);
5478 nr_inuse = nr_objs - nr_free;
5480 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5481 nr_objs, s->size, oo_objects(s->oo),
5482 (1 << oo_order(s->oo)));
5483 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5484 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5490 static const struct seq_operations slabinfo_op = {
5497 static int slabinfo_open(struct inode *inode, struct file *file)
5499 return seq_open(file, &slabinfo_op);
5502 static const struct file_operations proc_slabinfo_operations = {
5503 .open = slabinfo_open,
5505 .llseek = seq_lseek,
5506 .release = seq_release,
5509 static int __init slab_proc_init(void)
5511 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5514 module_init(slab_proc_init);
5515 #endif /* CONFIG_SLABINFO */