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 void prefetch_freepointer(const struct kmem_cache *s, void *object)
274 prefetch(object + s->offset);
277 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
281 #ifdef CONFIG_DEBUG_PAGEALLOC
282 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
284 p = get_freepointer(s, object);
289 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291 *(void **)(object + s->offset) = fp;
294 /* Loop over all objects in a slab */
295 #define for_each_object(__p, __s, __addr, __objects) \
296 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
299 /* Determine object index from a given position */
300 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
302 return (p - addr) / s->size;
305 static inline size_t slab_ksize(const struct kmem_cache *s)
307 #ifdef CONFIG_SLUB_DEBUG
309 * Debugging requires use of the padding between object
310 * and whatever may come after it.
312 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
317 * If we have the need to store the freelist pointer
318 * back there or track user information then we can
319 * only use the space before that information.
321 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
324 * Else we can use all the padding etc for the allocation
329 static inline int order_objects(int order, unsigned long size, int reserved)
331 return ((PAGE_SIZE << order) - reserved) / size;
334 static inline struct kmem_cache_order_objects oo_make(int order,
335 unsigned long size, int reserved)
337 struct kmem_cache_order_objects x = {
338 (order << OO_SHIFT) + order_objects(order, size, reserved)
344 static inline int oo_order(struct kmem_cache_order_objects x)
346 return x.x >> OO_SHIFT;
349 static inline int oo_objects(struct kmem_cache_order_objects x)
351 return x.x & OO_MASK;
355 * Per slab locking using the pagelock
357 static __always_inline void slab_lock(struct page *page)
359 bit_spin_lock(PG_locked, &page->flags);
362 static __always_inline void slab_unlock(struct page *page)
364 __bit_spin_unlock(PG_locked, &page->flags);
367 /* Interrupts must be disabled (for the fallback code to work right) */
368 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
369 void *freelist_old, unsigned long counters_old,
370 void *freelist_new, unsigned long counters_new,
373 VM_BUG_ON(!irqs_disabled());
374 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
375 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
376 if (s->flags & __CMPXCHG_DOUBLE) {
377 if (cmpxchg_double(&page->freelist, &page->counters,
378 freelist_old, counters_old,
379 freelist_new, counters_new))
385 if (page->freelist == freelist_old && page->counters == counters_old) {
386 page->freelist = freelist_new;
387 page->counters = counters_new;
395 stat(s, CMPXCHG_DOUBLE_FAIL);
397 #ifdef SLUB_DEBUG_CMPXCHG
398 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
404 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
405 void *freelist_old, unsigned long counters_old,
406 void *freelist_new, unsigned long counters_new,
409 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
410 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
411 if (s->flags & __CMPXCHG_DOUBLE) {
412 if (cmpxchg_double(&page->freelist, &page->counters,
413 freelist_old, counters_old,
414 freelist_new, counters_new))
421 local_irq_save(flags);
423 if (page->freelist == freelist_old && page->counters == counters_old) {
424 page->freelist = freelist_new;
425 page->counters = counters_new;
427 local_irq_restore(flags);
431 local_irq_restore(flags);
435 stat(s, CMPXCHG_DOUBLE_FAIL);
437 #ifdef SLUB_DEBUG_CMPXCHG
438 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
444 #ifdef CONFIG_SLUB_DEBUG
446 * Determine a map of object in use on a page.
448 * Node listlock must be held to guarantee that the page does
449 * not vanish from under us.
451 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
454 void *addr = page_address(page);
456 for (p = page->freelist; p; p = get_freepointer(s, p))
457 set_bit(slab_index(p, s, addr), map);
463 #ifdef CONFIG_SLUB_DEBUG_ON
464 static int slub_debug = DEBUG_DEFAULT_FLAGS;
466 static int slub_debug;
469 static char *slub_debug_slabs;
470 static int disable_higher_order_debug;
475 static void print_section(char *text, u8 *addr, unsigned int length)
477 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
481 static struct track *get_track(struct kmem_cache *s, void *object,
482 enum track_item alloc)
487 p = object + s->offset + sizeof(void *);
489 p = object + s->inuse;
494 static void set_track(struct kmem_cache *s, void *object,
495 enum track_item alloc, unsigned long addr)
497 struct track *p = get_track(s, object, alloc);
500 #ifdef CONFIG_STACKTRACE
501 struct stack_trace trace;
504 trace.nr_entries = 0;
505 trace.max_entries = TRACK_ADDRS_COUNT;
506 trace.entries = p->addrs;
508 save_stack_trace(&trace);
510 /* See rant in lockdep.c */
511 if (trace.nr_entries != 0 &&
512 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
515 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
519 p->cpu = smp_processor_id();
520 p->pid = current->pid;
523 memset(p, 0, sizeof(struct track));
526 static void init_tracking(struct kmem_cache *s, void *object)
528 if (!(s->flags & SLAB_STORE_USER))
531 set_track(s, object, TRACK_FREE, 0UL);
532 set_track(s, object, TRACK_ALLOC, 0UL);
535 static void print_track(const char *s, struct track *t)
540 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
541 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
542 #ifdef CONFIG_STACKTRACE
545 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
547 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
554 static void print_tracking(struct kmem_cache *s, void *object)
556 if (!(s->flags & SLAB_STORE_USER))
559 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
560 print_track("Freed", get_track(s, object, TRACK_FREE));
563 static void print_page_info(struct page *page)
565 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
566 page, page->objects, page->inuse, page->freelist, page->flags);
570 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
576 vsnprintf(buf, sizeof(buf), fmt, args);
578 printk(KERN_ERR "========================================"
579 "=====================================\n");
580 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
581 printk(KERN_ERR "----------------------------------------"
582 "-------------------------------------\n\n");
585 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
591 vsnprintf(buf, sizeof(buf), fmt, args);
593 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
596 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
598 unsigned int off; /* Offset of last byte */
599 u8 *addr = page_address(page);
601 print_tracking(s, p);
603 print_page_info(page);
605 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
606 p, p - addr, get_freepointer(s, p));
609 print_section("Bytes b4 ", p - 16, 16);
611 print_section("Object ", p, min_t(unsigned long, s->objsize,
613 if (s->flags & SLAB_RED_ZONE)
614 print_section("Redzone ", p + s->objsize,
615 s->inuse - s->objsize);
618 off = s->offset + sizeof(void *);
622 if (s->flags & SLAB_STORE_USER)
623 off += 2 * sizeof(struct track);
626 /* Beginning of the filler is the free pointer */
627 print_section("Padding ", p + off, s->size - off);
632 static void object_err(struct kmem_cache *s, struct page *page,
633 u8 *object, char *reason)
635 slab_bug(s, "%s", reason);
636 print_trailer(s, page, object);
639 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
645 vsnprintf(buf, sizeof(buf), fmt, args);
647 slab_bug(s, "%s", buf);
648 print_page_info(page);
652 static void init_object(struct kmem_cache *s, void *object, u8 val)
656 if (s->flags & __OBJECT_POISON) {
657 memset(p, POISON_FREE, s->objsize - 1);
658 p[s->objsize - 1] = POISON_END;
661 if (s->flags & SLAB_RED_ZONE)
662 memset(p + s->objsize, val, s->inuse - s->objsize);
665 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
666 void *from, void *to)
668 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
669 memset(from, data, to - from);
672 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
673 u8 *object, char *what,
674 u8 *start, unsigned int value, unsigned int bytes)
679 fault = memchr_inv(start, value, bytes);
684 while (end > fault && end[-1] == value)
687 slab_bug(s, "%s overwritten", what);
688 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
689 fault, end - 1, fault[0], value);
690 print_trailer(s, page, object);
692 restore_bytes(s, what, value, fault, end);
700 * Bytes of the object to be managed.
701 * If the freepointer may overlay the object then the free
702 * pointer is the first word of the object.
704 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
707 * object + s->objsize
708 * Padding to reach word boundary. This is also used for Redzoning.
709 * Padding is extended by another word if Redzoning is enabled and
712 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
713 * 0xcc (RED_ACTIVE) for objects in use.
716 * Meta data starts here.
718 * A. Free pointer (if we cannot overwrite object on free)
719 * B. Tracking data for SLAB_STORE_USER
720 * C. Padding to reach required alignment boundary or at mininum
721 * one word if debugging is on to be able to detect writes
722 * before the word boundary.
724 * Padding is done using 0x5a (POISON_INUSE)
727 * Nothing is used beyond s->size.
729 * If slabcaches are merged then the objsize and inuse boundaries are mostly
730 * ignored. And therefore no slab options that rely on these boundaries
731 * may be used with merged slabcaches.
734 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
736 unsigned long off = s->inuse; /* The end of info */
739 /* Freepointer is placed after the object. */
740 off += sizeof(void *);
742 if (s->flags & SLAB_STORE_USER)
743 /* We also have user information there */
744 off += 2 * sizeof(struct track);
749 return check_bytes_and_report(s, page, p, "Object padding",
750 p + off, POISON_INUSE, s->size - off);
753 /* Check the pad bytes at the end of a slab page */
754 static int slab_pad_check(struct kmem_cache *s, struct page *page)
762 if (!(s->flags & SLAB_POISON))
765 start = page_address(page);
766 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
767 end = start + length;
768 remainder = length % s->size;
772 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
775 while (end > fault && end[-1] == POISON_INUSE)
778 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
779 print_section("Padding ", end - remainder, remainder);
781 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
785 static int check_object(struct kmem_cache *s, struct page *page,
786 void *object, u8 val)
789 u8 *endobject = object + s->objsize;
791 if (s->flags & SLAB_RED_ZONE) {
792 if (!check_bytes_and_report(s, page, object, "Redzone",
793 endobject, val, s->inuse - s->objsize))
796 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
797 check_bytes_and_report(s, page, p, "Alignment padding",
798 endobject, POISON_INUSE, s->inuse - s->objsize);
802 if (s->flags & SLAB_POISON) {
803 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
804 (!check_bytes_and_report(s, page, p, "Poison", p,
805 POISON_FREE, s->objsize - 1) ||
806 !check_bytes_and_report(s, page, p, "Poison",
807 p + s->objsize - 1, POISON_END, 1)))
810 * check_pad_bytes cleans up on its own.
812 check_pad_bytes(s, page, p);
815 if (!s->offset && val == SLUB_RED_ACTIVE)
817 * Object and freepointer overlap. Cannot check
818 * freepointer while object is allocated.
822 /* Check free pointer validity */
823 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
824 object_err(s, page, p, "Freepointer corrupt");
826 * No choice but to zap it and thus lose the remainder
827 * of the free objects in this slab. May cause
828 * another error because the object count is now wrong.
830 set_freepointer(s, p, NULL);
836 static int check_slab(struct kmem_cache *s, struct page *page)
840 VM_BUG_ON(!irqs_disabled());
842 if (!PageSlab(page)) {
843 slab_err(s, page, "Not a valid slab page");
847 maxobj = order_objects(compound_order(page), s->size, s->reserved);
848 if (page->objects > maxobj) {
849 slab_err(s, page, "objects %u > max %u",
850 s->name, page->objects, maxobj);
853 if (page->inuse > page->objects) {
854 slab_err(s, page, "inuse %u > max %u",
855 s->name, page->inuse, page->objects);
858 /* Slab_pad_check fixes things up after itself */
859 slab_pad_check(s, page);
864 * Determine if a certain object on a page is on the freelist. Must hold the
865 * slab lock to guarantee that the chains are in a consistent state.
867 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
872 unsigned long max_objects;
875 while (fp && nr <= page->objects) {
878 if (!check_valid_pointer(s, page, fp)) {
880 object_err(s, page, object,
881 "Freechain corrupt");
882 set_freepointer(s, object, NULL);
885 slab_err(s, page, "Freepointer corrupt");
886 page->freelist = NULL;
887 page->inuse = page->objects;
888 slab_fix(s, "Freelist cleared");
894 fp = get_freepointer(s, object);
898 max_objects = order_objects(compound_order(page), s->size, s->reserved);
899 if (max_objects > MAX_OBJS_PER_PAGE)
900 max_objects = MAX_OBJS_PER_PAGE;
902 if (page->objects != max_objects) {
903 slab_err(s, page, "Wrong number of objects. Found %d but "
904 "should be %d", page->objects, max_objects);
905 page->objects = max_objects;
906 slab_fix(s, "Number of objects adjusted.");
908 if (page->inuse != page->objects - nr) {
909 slab_err(s, page, "Wrong object count. Counter is %d but "
910 "counted were %d", page->inuse, page->objects - nr);
911 page->inuse = page->objects - nr;
912 slab_fix(s, "Object count adjusted.");
914 return search == NULL;
917 static void trace(struct kmem_cache *s, struct page *page, void *object,
920 if (s->flags & SLAB_TRACE) {
921 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
923 alloc ? "alloc" : "free",
928 print_section("Object ", (void *)object, s->objsize);
935 * Hooks for other subsystems that check memory allocations. In a typical
936 * production configuration these hooks all should produce no code at all.
938 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
940 flags &= gfp_allowed_mask;
941 lockdep_trace_alloc(flags);
942 might_sleep_if(flags & __GFP_WAIT);
944 return should_failslab(s->objsize, flags, s->flags);
947 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
949 flags &= gfp_allowed_mask;
950 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
951 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
954 static inline void slab_free_hook(struct kmem_cache *s, void *x)
956 kmemleak_free_recursive(x, s->flags);
959 * Trouble is that we may no longer disable interupts in the fast path
960 * So in order to make the debug calls that expect irqs to be
961 * disabled we need to disable interrupts temporarily.
963 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
967 local_irq_save(flags);
968 kmemcheck_slab_free(s, x, s->objsize);
969 debug_check_no_locks_freed(x, s->objsize);
970 local_irq_restore(flags);
973 if (!(s->flags & SLAB_DEBUG_OBJECTS))
974 debug_check_no_obj_freed(x, s->objsize);
978 * Tracking of fully allocated slabs for debugging purposes.
980 * list_lock must be held.
982 static void add_full(struct kmem_cache *s,
983 struct kmem_cache_node *n, struct page *page)
985 if (!(s->flags & SLAB_STORE_USER))
988 list_add(&page->lru, &n->full);
992 * list_lock must be held.
994 static void remove_full(struct kmem_cache *s, struct page *page)
996 if (!(s->flags & SLAB_STORE_USER))
999 list_del(&page->lru);
1002 /* Tracking of the number of slabs for debugging purposes */
1003 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1005 struct kmem_cache_node *n = get_node(s, node);
1007 return atomic_long_read(&n->nr_slabs);
1010 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1012 return atomic_long_read(&n->nr_slabs);
1015 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1017 struct kmem_cache_node *n = get_node(s, node);
1020 * May be called early in order to allocate a slab for the
1021 * kmem_cache_node structure. Solve the chicken-egg
1022 * dilemma by deferring the increment of the count during
1023 * bootstrap (see early_kmem_cache_node_alloc).
1026 atomic_long_inc(&n->nr_slabs);
1027 atomic_long_add(objects, &n->total_objects);
1030 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1032 struct kmem_cache_node *n = get_node(s, node);
1034 atomic_long_dec(&n->nr_slabs);
1035 atomic_long_sub(objects, &n->total_objects);
1038 /* Object debug checks for alloc/free paths */
1039 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1042 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1045 init_object(s, object, SLUB_RED_INACTIVE);
1046 init_tracking(s, object);
1049 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1050 void *object, unsigned long addr)
1052 if (!check_slab(s, page))
1055 if (!check_valid_pointer(s, page, object)) {
1056 object_err(s, page, object, "Freelist Pointer check fails");
1060 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1063 /* Success perform special debug activities for allocs */
1064 if (s->flags & SLAB_STORE_USER)
1065 set_track(s, object, TRACK_ALLOC, addr);
1066 trace(s, page, object, 1);
1067 init_object(s, object, SLUB_RED_ACTIVE);
1071 if (PageSlab(page)) {
1073 * If this is a slab page then lets do the best we can
1074 * to avoid issues in the future. Marking all objects
1075 * as used avoids touching the remaining objects.
1077 slab_fix(s, "Marking all objects used");
1078 page->inuse = page->objects;
1079 page->freelist = NULL;
1084 static noinline int free_debug_processing(struct kmem_cache *s,
1085 struct page *page, void *object, unsigned long addr)
1087 unsigned long flags;
1090 local_irq_save(flags);
1093 if (!check_slab(s, page))
1096 if (!check_valid_pointer(s, page, object)) {
1097 slab_err(s, page, "Invalid object pointer 0x%p", object);
1101 if (on_freelist(s, page, object)) {
1102 object_err(s, page, object, "Object already free");
1106 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1109 if (unlikely(s != page->slab)) {
1110 if (!PageSlab(page)) {
1111 slab_err(s, page, "Attempt to free object(0x%p) "
1112 "outside of slab", object);
1113 } else if (!page->slab) {
1115 "SLUB <none>: no slab for object 0x%p.\n",
1119 object_err(s, page, object,
1120 "page slab pointer corrupt.");
1124 if (s->flags & SLAB_STORE_USER)
1125 set_track(s, object, TRACK_FREE, addr);
1126 trace(s, page, object, 0);
1127 init_object(s, object, SLUB_RED_INACTIVE);
1131 local_irq_restore(flags);
1135 slab_fix(s, "Object at 0x%p not freed", object);
1139 static int __init setup_slub_debug(char *str)
1141 slub_debug = DEBUG_DEFAULT_FLAGS;
1142 if (*str++ != '=' || !*str)
1144 * No options specified. Switch on full debugging.
1150 * No options but restriction on slabs. This means full
1151 * debugging for slabs matching a pattern.
1155 if (tolower(*str) == 'o') {
1157 * Avoid enabling debugging on caches if its minimum order
1158 * would increase as a result.
1160 disable_higher_order_debug = 1;
1167 * Switch off all debugging measures.
1172 * Determine which debug features should be switched on
1174 for (; *str && *str != ','; str++) {
1175 switch (tolower(*str)) {
1177 slub_debug |= SLAB_DEBUG_FREE;
1180 slub_debug |= SLAB_RED_ZONE;
1183 slub_debug |= SLAB_POISON;
1186 slub_debug |= SLAB_STORE_USER;
1189 slub_debug |= SLAB_TRACE;
1192 slub_debug |= SLAB_FAILSLAB;
1195 printk(KERN_ERR "slub_debug option '%c' "
1196 "unknown. skipped\n", *str);
1202 slub_debug_slabs = str + 1;
1207 __setup("slub_debug", setup_slub_debug);
1209 static unsigned long kmem_cache_flags(unsigned long objsize,
1210 unsigned long flags, const char *name,
1211 void (*ctor)(void *))
1214 * Enable debugging if selected on the kernel commandline.
1216 if (slub_debug && (!slub_debug_slabs ||
1217 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1218 flags |= slub_debug;
1223 static inline void setup_object_debug(struct kmem_cache *s,
1224 struct page *page, void *object) {}
1226 static inline int alloc_debug_processing(struct kmem_cache *s,
1227 struct page *page, void *object, unsigned long addr) { return 0; }
1229 static inline int free_debug_processing(struct kmem_cache *s,
1230 struct page *page, void *object, unsigned long addr) { return 0; }
1232 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1234 static inline int check_object(struct kmem_cache *s, struct page *page,
1235 void *object, u8 val) { return 1; }
1236 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1237 struct page *page) {}
1238 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1239 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1240 unsigned long flags, const char *name,
1241 void (*ctor)(void *))
1245 #define slub_debug 0
1247 #define disable_higher_order_debug 0
1249 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1251 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1253 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1255 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1258 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1261 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1264 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1266 #endif /* CONFIG_SLUB_DEBUG */
1269 * Slab allocation and freeing
1271 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1272 struct kmem_cache_order_objects oo)
1274 int order = oo_order(oo);
1276 flags |= __GFP_NOTRACK;
1278 if (node == NUMA_NO_NODE)
1279 return alloc_pages(flags, order);
1281 return alloc_pages_exact_node(node, flags, order);
1284 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1287 struct kmem_cache_order_objects oo = s->oo;
1290 flags &= gfp_allowed_mask;
1292 if (flags & __GFP_WAIT)
1295 flags |= s->allocflags;
1298 * Let the initial higher-order allocation fail under memory pressure
1299 * so we fall-back to the minimum order allocation.
1301 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1303 page = alloc_slab_page(alloc_gfp, node, oo);
1304 if (unlikely(!page)) {
1307 * Allocation may have failed due to fragmentation.
1308 * Try a lower order alloc if possible
1310 page = alloc_slab_page(flags, node, oo);
1313 stat(s, ORDER_FALLBACK);
1316 if (flags & __GFP_WAIT)
1317 local_irq_disable();
1322 if (kmemcheck_enabled
1323 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1324 int pages = 1 << oo_order(oo);
1326 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1329 * Objects from caches that have a constructor don't get
1330 * cleared when they're allocated, so we need to do it here.
1333 kmemcheck_mark_uninitialized_pages(page, pages);
1335 kmemcheck_mark_unallocated_pages(page, pages);
1338 page->objects = oo_objects(oo);
1339 mod_zone_page_state(page_zone(page),
1340 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1341 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1347 static void setup_object(struct kmem_cache *s, struct page *page,
1350 setup_object_debug(s, page, object);
1351 if (unlikely(s->ctor))
1355 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1362 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1364 page = allocate_slab(s,
1365 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1369 inc_slabs_node(s, page_to_nid(page), page->objects);
1371 page->flags |= 1 << PG_slab;
1373 start = page_address(page);
1375 if (unlikely(s->flags & SLAB_POISON))
1376 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1379 for_each_object(p, s, start, page->objects) {
1380 setup_object(s, page, last);
1381 set_freepointer(s, last, p);
1384 setup_object(s, page, last);
1385 set_freepointer(s, last, NULL);
1387 page->freelist = start;
1388 page->inuse = page->objects;
1394 static void __free_slab(struct kmem_cache *s, struct page *page)
1396 int order = compound_order(page);
1397 int pages = 1 << order;
1399 if (kmem_cache_debug(s)) {
1402 slab_pad_check(s, page);
1403 for_each_object(p, s, page_address(page),
1405 check_object(s, page, p, SLUB_RED_INACTIVE);
1408 kmemcheck_free_shadow(page, compound_order(page));
1410 mod_zone_page_state(page_zone(page),
1411 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1412 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1415 __ClearPageSlab(page);
1416 reset_page_mapcount(page);
1417 if (current->reclaim_state)
1418 current->reclaim_state->reclaimed_slab += pages;
1419 __free_pages(page, order);
1422 #define need_reserve_slab_rcu \
1423 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1425 static void rcu_free_slab(struct rcu_head *h)
1429 if (need_reserve_slab_rcu)
1430 page = virt_to_head_page(h);
1432 page = container_of((struct list_head *)h, struct page, lru);
1434 __free_slab(page->slab, page);
1437 static void free_slab(struct kmem_cache *s, struct page *page)
1439 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1440 struct rcu_head *head;
1442 if (need_reserve_slab_rcu) {
1443 int order = compound_order(page);
1444 int offset = (PAGE_SIZE << order) - s->reserved;
1446 VM_BUG_ON(s->reserved != sizeof(*head));
1447 head = page_address(page) + offset;
1450 * RCU free overloads the RCU head over the LRU
1452 head = (void *)&page->lru;
1455 call_rcu(head, rcu_free_slab);
1457 __free_slab(s, page);
1460 static void discard_slab(struct kmem_cache *s, struct page *page)
1462 dec_slabs_node(s, page_to_nid(page), page->objects);
1467 * Management of partially allocated slabs.
1469 * list_lock must be held.
1471 static inline void add_partial(struct kmem_cache_node *n,
1472 struct page *page, int tail)
1475 if (tail == DEACTIVATE_TO_TAIL)
1476 list_add_tail(&page->lru, &n->partial);
1478 list_add(&page->lru, &n->partial);
1482 * list_lock must be held.
1484 static inline void remove_partial(struct kmem_cache_node *n,
1487 list_del(&page->lru);
1492 * Lock slab, remove from the partial list and put the object into the
1495 * Returns a list of objects or NULL if it fails.
1497 * Must hold list_lock.
1499 static inline void *acquire_slab(struct kmem_cache *s,
1500 struct kmem_cache_node *n, struct page *page,
1504 unsigned long counters;
1508 * Zap the freelist and set the frozen bit.
1509 * The old freelist is the list of objects for the
1510 * per cpu allocation list.
1513 freelist = page->freelist;
1514 counters = page->counters;
1515 new.counters = counters;
1517 new.inuse = page->objects;
1519 VM_BUG_ON(new.frozen);
1522 } while (!__cmpxchg_double_slab(s, page,
1525 "lock and freeze"));
1527 remove_partial(n, page);
1531 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1534 * Try to allocate a partial slab from a specific node.
1536 static void *get_partial_node(struct kmem_cache *s,
1537 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1539 struct page *page, *page2;
1540 void *object = NULL;
1543 * Racy check. If we mistakenly see no partial slabs then we
1544 * just allocate an empty slab. If we mistakenly try to get a
1545 * partial slab and there is none available then get_partials()
1548 if (!n || !n->nr_partial)
1551 spin_lock(&n->list_lock);
1552 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1553 void *t = acquire_slab(s, n, page, object == NULL);
1561 c->node = page_to_nid(page);
1562 stat(s, ALLOC_FROM_PARTIAL);
1564 available = page->objects - page->inuse;
1567 available = put_cpu_partial(s, page, 0);
1569 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1573 spin_unlock(&n->list_lock);
1578 * Get a page from somewhere. Search in increasing NUMA distances.
1580 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1581 struct kmem_cache_cpu *c)
1584 struct zonelist *zonelist;
1587 enum zone_type high_zoneidx = gfp_zone(flags);
1591 * The defrag ratio allows a configuration of the tradeoffs between
1592 * inter node defragmentation and node local allocations. A lower
1593 * defrag_ratio increases the tendency to do local allocations
1594 * instead of attempting to obtain partial slabs from other nodes.
1596 * If the defrag_ratio is set to 0 then kmalloc() always
1597 * returns node local objects. If the ratio is higher then kmalloc()
1598 * may return off node objects because partial slabs are obtained
1599 * from other nodes and filled up.
1601 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1602 * defrag_ratio = 1000) then every (well almost) allocation will
1603 * first attempt to defrag slab caches on other nodes. This means
1604 * scanning over all nodes to look for partial slabs which may be
1605 * expensive if we do it every time we are trying to find a slab
1606 * with available objects.
1608 if (!s->remote_node_defrag_ratio ||
1609 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1613 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1614 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1615 struct kmem_cache_node *n;
1617 n = get_node(s, zone_to_nid(zone));
1619 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1620 n->nr_partial > s->min_partial) {
1621 object = get_partial_node(s, n, c);
1634 * Get a partial page, lock it and return it.
1636 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1637 struct kmem_cache_cpu *c)
1640 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1642 object = get_partial_node(s, get_node(s, searchnode), c);
1643 if (object || node != NUMA_NO_NODE)
1646 return get_any_partial(s, flags, c);
1649 #ifdef CONFIG_PREEMPT
1651 * Calculate the next globally unique transaction for disambiguiation
1652 * during cmpxchg. The transactions start with the cpu number and are then
1653 * incremented by CONFIG_NR_CPUS.
1655 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1658 * No preemption supported therefore also no need to check for
1664 static inline unsigned long next_tid(unsigned long tid)
1666 return tid + TID_STEP;
1669 static inline unsigned int tid_to_cpu(unsigned long tid)
1671 return tid % TID_STEP;
1674 static inline unsigned long tid_to_event(unsigned long tid)
1676 return tid / TID_STEP;
1679 static inline unsigned int init_tid(int cpu)
1684 static inline void note_cmpxchg_failure(const char *n,
1685 const struct kmem_cache *s, unsigned long tid)
1687 #ifdef SLUB_DEBUG_CMPXCHG
1688 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1690 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1692 #ifdef CONFIG_PREEMPT
1693 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1694 printk("due to cpu change %d -> %d\n",
1695 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1698 if (tid_to_event(tid) != tid_to_event(actual_tid))
1699 printk("due to cpu running other code. Event %ld->%ld\n",
1700 tid_to_event(tid), tid_to_event(actual_tid));
1702 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1703 actual_tid, tid, next_tid(tid));
1705 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1708 void init_kmem_cache_cpus(struct kmem_cache *s)
1712 for_each_possible_cpu(cpu)
1713 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1717 * Remove the cpu slab
1719 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1721 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1722 struct page *page = c->page;
1723 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1725 enum slab_modes l = M_NONE, m = M_NONE;
1728 int tail = DEACTIVATE_TO_HEAD;
1732 if (page->freelist) {
1733 stat(s, DEACTIVATE_REMOTE_FREES);
1734 tail = DEACTIVATE_TO_TAIL;
1737 c->tid = next_tid(c->tid);
1739 freelist = c->freelist;
1743 * Stage one: Free all available per cpu objects back
1744 * to the page freelist while it is still frozen. Leave the
1747 * There is no need to take the list->lock because the page
1750 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1752 unsigned long counters;
1755 prior = page->freelist;
1756 counters = page->counters;
1757 set_freepointer(s, freelist, prior);
1758 new.counters = counters;
1760 VM_BUG_ON(!new.frozen);
1762 } while (!__cmpxchg_double_slab(s, page,
1764 freelist, new.counters,
1765 "drain percpu freelist"));
1767 freelist = nextfree;
1771 * Stage two: Ensure that the page is unfrozen while the
1772 * list presence reflects the actual number of objects
1775 * We setup the list membership and then perform a cmpxchg
1776 * with the count. If there is a mismatch then the page
1777 * is not unfrozen but the page is on the wrong list.
1779 * Then we restart the process which may have to remove
1780 * the page from the list that we just put it on again
1781 * because the number of objects in the slab may have
1786 old.freelist = page->freelist;
1787 old.counters = page->counters;
1788 VM_BUG_ON(!old.frozen);
1790 /* Determine target state of the slab */
1791 new.counters = old.counters;
1794 set_freepointer(s, freelist, old.freelist);
1795 new.freelist = freelist;
1797 new.freelist = old.freelist;
1801 if (!new.inuse && n->nr_partial > s->min_partial)
1803 else if (new.freelist) {
1808 * Taking the spinlock removes the possiblity
1809 * that acquire_slab() will see a slab page that
1812 spin_lock(&n->list_lock);
1816 if (kmem_cache_debug(s) && !lock) {
1819 * This also ensures that the scanning of full
1820 * slabs from diagnostic functions will not see
1823 spin_lock(&n->list_lock);
1831 remove_partial(n, page);
1833 else if (l == M_FULL)
1835 remove_full(s, page);
1837 if (m == M_PARTIAL) {
1839 add_partial(n, page, tail);
1842 } else if (m == M_FULL) {
1844 stat(s, DEACTIVATE_FULL);
1845 add_full(s, n, page);
1851 if (!__cmpxchg_double_slab(s, page,
1852 old.freelist, old.counters,
1853 new.freelist, new.counters,
1858 spin_unlock(&n->list_lock);
1861 stat(s, DEACTIVATE_EMPTY);
1862 discard_slab(s, page);
1867 /* Unfreeze all the cpu partial slabs */
1868 static void unfreeze_partials(struct kmem_cache *s)
1870 struct kmem_cache_node *n = NULL;
1871 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1872 struct page *page, *discard_page = NULL;
1874 while ((page = c->partial)) {
1875 enum slab_modes { M_PARTIAL, M_FREE };
1876 enum slab_modes l, m;
1880 c->partial = page->next;
1885 old.freelist = page->freelist;
1886 old.counters = page->counters;
1887 VM_BUG_ON(!old.frozen);
1889 new.counters = old.counters;
1890 new.freelist = old.freelist;
1894 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1897 struct kmem_cache_node *n2 = get_node(s,
1903 spin_unlock(&n->list_lock);
1906 spin_lock(&n->list_lock);
1911 if (l == M_PARTIAL) {
1912 remove_partial(n, page);
1913 stat(s, FREE_REMOVE_PARTIAL);
1915 add_partial(n, page,
1916 DEACTIVATE_TO_TAIL);
1917 stat(s, FREE_ADD_PARTIAL);
1923 } while (!cmpxchg_double_slab(s, page,
1924 old.freelist, old.counters,
1925 new.freelist, new.counters,
1926 "unfreezing slab"));
1929 page->next = discard_page;
1930 discard_page = page;
1935 spin_unlock(&n->list_lock);
1937 while (discard_page) {
1938 page = discard_page;
1939 discard_page = discard_page->next;
1941 stat(s, DEACTIVATE_EMPTY);
1942 discard_slab(s, page);
1948 * Put a page that was just frozen (in __slab_free) into a partial page
1949 * slot if available. This is done without interrupts disabled and without
1950 * preemption disabled. The cmpxchg is racy and may put the partial page
1951 * onto a random cpus partial slot.
1953 * If we did not find a slot then simply move all the partials to the
1954 * per node partial list.
1956 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1958 struct page *oldpage;
1965 oldpage = this_cpu_read(s->cpu_slab->partial);
1968 pobjects = oldpage->pobjects;
1969 pages = oldpage->pages;
1970 if (drain && pobjects > s->cpu_partial) {
1971 unsigned long flags;
1973 * partial array is full. Move the existing
1974 * set to the per node partial list.
1976 local_irq_save(flags);
1977 unfreeze_partials(s);
1978 local_irq_restore(flags);
1985 pobjects += page->objects - page->inuse;
1987 page->pages = pages;
1988 page->pobjects = pobjects;
1989 page->next = oldpage;
1991 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1992 stat(s, CPU_PARTIAL_FREE);
1996 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1998 stat(s, CPUSLAB_FLUSH);
1999 deactivate_slab(s, c);
2005 * Called from IPI handler with interrupts disabled.
2007 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2009 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2015 unfreeze_partials(s);
2019 static void flush_cpu_slab(void *d)
2021 struct kmem_cache *s = d;
2023 __flush_cpu_slab(s, smp_processor_id());
2026 static void flush_all(struct kmem_cache *s)
2028 on_each_cpu(flush_cpu_slab, s, 1);
2032 * Check if the objects in a per cpu structure fit numa
2033 * locality expectations.
2035 static inline int node_match(struct kmem_cache_cpu *c, int node)
2038 if (node != NUMA_NO_NODE && c->node != node)
2044 static int count_free(struct page *page)
2046 return page->objects - page->inuse;
2049 static unsigned long count_partial(struct kmem_cache_node *n,
2050 int (*get_count)(struct page *))
2052 unsigned long flags;
2053 unsigned long x = 0;
2056 spin_lock_irqsave(&n->list_lock, flags);
2057 list_for_each_entry(page, &n->partial, lru)
2058 x += get_count(page);
2059 spin_unlock_irqrestore(&n->list_lock, flags);
2063 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2065 #ifdef CONFIG_SLUB_DEBUG
2066 return atomic_long_read(&n->total_objects);
2072 static noinline void
2073 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2078 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2080 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2081 "default order: %d, min order: %d\n", s->name, s->objsize,
2082 s->size, oo_order(s->oo), oo_order(s->min));
2084 if (oo_order(s->min) > get_order(s->objsize))
2085 printk(KERN_WARNING " %s debugging increased min order, use "
2086 "slub_debug=O to disable.\n", s->name);
2088 for_each_online_node(node) {
2089 struct kmem_cache_node *n = get_node(s, node);
2090 unsigned long nr_slabs;
2091 unsigned long nr_objs;
2092 unsigned long nr_free;
2097 nr_free = count_partial(n, count_free);
2098 nr_slabs = node_nr_slabs(n);
2099 nr_objs = node_nr_objs(n);
2102 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2103 node, nr_slabs, nr_objs, nr_free);
2107 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2108 int node, struct kmem_cache_cpu **pc)
2111 struct kmem_cache_cpu *c;
2112 struct page *page = new_slab(s, flags, node);
2115 c = __this_cpu_ptr(s->cpu_slab);
2120 * No other reference to the page yet so we can
2121 * muck around with it freely without cmpxchg
2123 object = page->freelist;
2124 page->freelist = NULL;
2126 stat(s, ALLOC_SLAB);
2127 c->node = page_to_nid(page);
2137 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2138 * or deactivate the page.
2140 * The page is still frozen if the return value is not NULL.
2142 * If this function returns NULL then the page has been unfrozen.
2144 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2147 unsigned long counters;
2151 freelist = page->freelist;
2152 counters = page->counters;
2153 new.counters = counters;
2154 VM_BUG_ON(!new.frozen);
2156 new.inuse = page->objects;
2157 new.frozen = freelist != NULL;
2159 } while (!cmpxchg_double_slab(s, page,
2168 * Slow path. The lockless freelist is empty or we need to perform
2171 * Processing is still very fast if new objects have been freed to the
2172 * regular freelist. In that case we simply take over the regular freelist
2173 * as the lockless freelist and zap the regular freelist.
2175 * If that is not working then we fall back to the partial lists. We take the
2176 * first element of the freelist as the object to allocate now and move the
2177 * rest of the freelist to the lockless freelist.
2179 * And if we were unable to get a new slab from the partial slab lists then
2180 * we need to allocate a new slab. This is the slowest path since it involves
2181 * a call to the page allocator and the setup of a new slab.
2183 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2184 unsigned long addr, struct kmem_cache_cpu *c)
2187 unsigned long flags;
2189 local_irq_save(flags);
2190 #ifdef CONFIG_PREEMPT
2192 * We may have been preempted and rescheduled on a different
2193 * cpu before disabling interrupts. Need to reload cpu area
2196 c = this_cpu_ptr(s->cpu_slab);
2202 if (unlikely(!node_match(c, node))) {
2203 stat(s, ALLOC_NODE_MISMATCH);
2204 deactivate_slab(s, c);
2208 /* must check again c->freelist in case of cpu migration or IRQ */
2209 object = c->freelist;
2213 stat(s, ALLOC_SLOWPATH);
2215 object = get_freelist(s, c->page);
2219 stat(s, DEACTIVATE_BYPASS);
2223 stat(s, ALLOC_REFILL);
2226 c->freelist = get_freepointer(s, object);
2227 c->tid = next_tid(c->tid);
2228 local_irq_restore(flags);
2234 c->page = c->partial;
2235 c->partial = c->page->next;
2236 c->node = page_to_nid(c->page);
2237 stat(s, CPU_PARTIAL_ALLOC);
2242 /* Then do expensive stuff like retrieving pages from the partial lists */
2243 object = get_partial(s, gfpflags, node, c);
2245 if (unlikely(!object)) {
2247 object = new_slab_objects(s, gfpflags, node, &c);
2249 if (unlikely(!object)) {
2250 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2251 slab_out_of_memory(s, gfpflags, node);
2253 local_irq_restore(flags);
2258 if (likely(!kmem_cache_debug(s)))
2261 /* Only entered in the debug case */
2262 if (!alloc_debug_processing(s, c->page, object, addr))
2263 goto new_slab; /* Slab failed checks. Next slab needed */
2265 c->freelist = get_freepointer(s, object);
2266 deactivate_slab(s, c);
2267 c->node = NUMA_NO_NODE;
2268 local_irq_restore(flags);
2273 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2274 * have the fastpath folded into their functions. So no function call
2275 * overhead for requests that can be satisfied on the fastpath.
2277 * The fastpath works by first checking if the lockless freelist can be used.
2278 * If not then __slab_alloc is called for slow processing.
2280 * Otherwise we can simply pick the next object from the lockless free list.
2282 static __always_inline void *slab_alloc(struct kmem_cache *s,
2283 gfp_t gfpflags, int node, unsigned long addr)
2286 struct kmem_cache_cpu *c;
2289 if (slab_pre_alloc_hook(s, gfpflags))
2295 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2296 * enabled. We may switch back and forth between cpus while
2297 * reading from one cpu area. That does not matter as long
2298 * as we end up on the original cpu again when doing the cmpxchg.
2300 c = __this_cpu_ptr(s->cpu_slab);
2303 * The transaction ids are globally unique per cpu and per operation on
2304 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2305 * occurs on the right processor and that there was no operation on the
2306 * linked list in between.
2311 object = c->freelist;
2312 if (unlikely(!object || !node_match(c, node)))
2314 object = __slab_alloc(s, gfpflags, node, addr, c);
2317 void *next_object = get_freepointer_safe(s, object);
2320 * The cmpxchg will only match if there was no additional
2321 * operation and if we are on the right processor.
2323 * The cmpxchg does the following atomically (without lock semantics!)
2324 * 1. Relocate first pointer to the current per cpu area.
2325 * 2. Verify that tid and freelist have not been changed
2326 * 3. If they were not changed replace tid and freelist
2328 * Since this is without lock semantics the protection is only against
2329 * code executing on this cpu *not* from access by other cpus.
2331 if (unlikely(!this_cpu_cmpxchg_double(
2332 s->cpu_slab->freelist, s->cpu_slab->tid,
2334 next_object, next_tid(tid)))) {
2336 note_cmpxchg_failure("slab_alloc", s, tid);
2339 prefetch_freepointer(s, next_object);
2340 stat(s, ALLOC_FASTPATH);
2343 if (unlikely(gfpflags & __GFP_ZERO) && object)
2344 memset(object, 0, s->objsize);
2346 slab_post_alloc_hook(s, gfpflags, object);
2351 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2353 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2355 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2359 EXPORT_SYMBOL(kmem_cache_alloc);
2361 #ifdef CONFIG_TRACING
2362 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2364 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2365 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2368 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2370 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2372 void *ret = kmalloc_order(size, flags, order);
2373 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2376 EXPORT_SYMBOL(kmalloc_order_trace);
2380 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2382 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2384 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2385 s->objsize, s->size, gfpflags, node);
2389 EXPORT_SYMBOL(kmem_cache_alloc_node);
2391 #ifdef CONFIG_TRACING
2392 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2394 int node, size_t size)
2396 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2398 trace_kmalloc_node(_RET_IP_, ret,
2399 size, s->size, gfpflags, node);
2402 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2407 * Slow patch handling. This may still be called frequently since objects
2408 * have a longer lifetime than the cpu slabs in most processing loads.
2410 * So we still attempt to reduce cache line usage. Just take the slab
2411 * lock and free the item. If there is no additional partial page
2412 * handling required then we can return immediately.
2414 static void __slab_free(struct kmem_cache *s, struct page *page,
2415 void *x, unsigned long addr)
2418 void **object = (void *)x;
2422 unsigned long counters;
2423 struct kmem_cache_node *n = NULL;
2424 unsigned long uninitialized_var(flags);
2426 stat(s, FREE_SLOWPATH);
2428 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2432 prior = page->freelist;
2433 counters = page->counters;
2434 set_freepointer(s, object, prior);
2435 new.counters = counters;
2436 was_frozen = new.frozen;
2438 if ((!new.inuse || !prior) && !was_frozen && !n) {
2440 if (!kmem_cache_debug(s) && !prior)
2443 * Slab was on no list before and will be partially empty
2444 * We can defer the list move and instead freeze it.
2448 else { /* Needs to be taken off a list */
2450 n = get_node(s, page_to_nid(page));
2452 * Speculatively acquire the list_lock.
2453 * If the cmpxchg does not succeed then we may
2454 * drop the list_lock without any processing.
2456 * Otherwise the list_lock will synchronize with
2457 * other processors updating the list of slabs.
2459 spin_lock_irqsave(&n->list_lock, flags);
2465 } while (!cmpxchg_double_slab(s, page,
2467 object, new.counters,
2473 * If we just froze the page then put it onto the
2474 * per cpu partial list.
2476 if (new.frozen && !was_frozen)
2477 put_cpu_partial(s, page, 1);
2480 * The list lock was not taken therefore no list
2481 * activity can be necessary.
2484 stat(s, FREE_FROZEN);
2489 * was_frozen may have been set after we acquired the list_lock in
2490 * an earlier loop. So we need to check it here again.
2493 stat(s, FREE_FROZEN);
2495 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2499 * Objects left in the slab. If it was not on the partial list before
2502 if (unlikely(!prior)) {
2503 remove_full(s, page);
2504 add_partial(n, page, DEACTIVATE_TO_TAIL);
2505 stat(s, FREE_ADD_PARTIAL);
2508 spin_unlock_irqrestore(&n->list_lock, flags);
2514 * Slab on the partial list.
2516 remove_partial(n, page);
2517 stat(s, FREE_REMOVE_PARTIAL);
2519 /* Slab must be on the full list */
2520 remove_full(s, page);
2522 spin_unlock_irqrestore(&n->list_lock, flags);
2524 discard_slab(s, page);
2528 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2529 * can perform fastpath freeing without additional function calls.
2531 * The fastpath is only possible if we are freeing to the current cpu slab
2532 * of this processor. This typically the case if we have just allocated
2535 * If fastpath is not possible then fall back to __slab_free where we deal
2536 * with all sorts of special processing.
2538 static __always_inline void slab_free(struct kmem_cache *s,
2539 struct page *page, void *x, unsigned long addr)
2541 void **object = (void *)x;
2542 struct kmem_cache_cpu *c;
2545 slab_free_hook(s, x);
2549 * Determine the currently cpus per cpu slab.
2550 * The cpu may change afterward. However that does not matter since
2551 * data is retrieved via this pointer. If we are on the same cpu
2552 * during the cmpxchg then the free will succedd.
2554 c = __this_cpu_ptr(s->cpu_slab);
2559 if (likely(page == c->page)) {
2560 set_freepointer(s, object, c->freelist);
2562 if (unlikely(!this_cpu_cmpxchg_double(
2563 s->cpu_slab->freelist, s->cpu_slab->tid,
2565 object, next_tid(tid)))) {
2567 note_cmpxchg_failure("slab_free", s, tid);
2570 stat(s, FREE_FASTPATH);
2572 __slab_free(s, page, x, addr);
2576 void kmem_cache_free(struct kmem_cache *s, void *x)
2580 page = virt_to_head_page(x);
2582 slab_free(s, page, x, _RET_IP_);
2584 trace_kmem_cache_free(_RET_IP_, x);
2586 EXPORT_SYMBOL(kmem_cache_free);
2589 * Object placement in a slab is made very easy because we always start at
2590 * offset 0. If we tune the size of the object to the alignment then we can
2591 * get the required alignment by putting one properly sized object after
2594 * Notice that the allocation order determines the sizes of the per cpu
2595 * caches. Each processor has always one slab available for allocations.
2596 * Increasing the allocation order reduces the number of times that slabs
2597 * must be moved on and off the partial lists and is therefore a factor in
2602 * Mininum / Maximum order of slab pages. This influences locking overhead
2603 * and slab fragmentation. A higher order reduces the number of partial slabs
2604 * and increases the number of allocations possible without having to
2605 * take the list_lock.
2607 static int slub_min_order;
2608 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2609 static int slub_min_objects;
2612 * Merge control. If this is set then no merging of slab caches will occur.
2613 * (Could be removed. This was introduced to pacify the merge skeptics.)
2615 static int slub_nomerge;
2618 * Calculate the order of allocation given an slab object size.
2620 * The order of allocation has significant impact on performance and other
2621 * system components. Generally order 0 allocations should be preferred since
2622 * order 0 does not cause fragmentation in the page allocator. Larger objects
2623 * be problematic to put into order 0 slabs because there may be too much
2624 * unused space left. We go to a higher order if more than 1/16th of the slab
2627 * In order to reach satisfactory performance we must ensure that a minimum
2628 * number of objects is in one slab. Otherwise we may generate too much
2629 * activity on the partial lists which requires taking the list_lock. This is
2630 * less a concern for large slabs though which are rarely used.
2632 * slub_max_order specifies the order where we begin to stop considering the
2633 * number of objects in a slab as critical. If we reach slub_max_order then
2634 * we try to keep the page order as low as possible. So we accept more waste
2635 * of space in favor of a small page order.
2637 * Higher order allocations also allow the placement of more objects in a
2638 * slab and thereby reduce object handling overhead. If the user has
2639 * requested a higher mininum order then we start with that one instead of
2640 * the smallest order which will fit the object.
2642 static inline int slab_order(int size, int min_objects,
2643 int max_order, int fract_leftover, int reserved)
2647 int min_order = slub_min_order;
2649 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2650 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2652 for (order = max(min_order,
2653 fls(min_objects * size - 1) - PAGE_SHIFT);
2654 order <= max_order; order++) {
2656 unsigned long slab_size = PAGE_SIZE << order;
2658 if (slab_size < min_objects * size + reserved)
2661 rem = (slab_size - reserved) % size;
2663 if (rem <= slab_size / fract_leftover)
2671 static inline int calculate_order(int size, int reserved)
2679 * Attempt to find best configuration for a slab. This
2680 * works by first attempting to generate a layout with
2681 * the best configuration and backing off gradually.
2683 * First we reduce the acceptable waste in a slab. Then
2684 * we reduce the minimum objects required in a slab.
2686 min_objects = slub_min_objects;
2688 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2689 max_objects = order_objects(slub_max_order, size, reserved);
2690 min_objects = min(min_objects, max_objects);
2692 while (min_objects > 1) {
2694 while (fraction >= 4) {
2695 order = slab_order(size, min_objects,
2696 slub_max_order, fraction, reserved);
2697 if (order <= slub_max_order)
2705 * We were unable to place multiple objects in a slab. Now
2706 * lets see if we can place a single object there.
2708 order = slab_order(size, 1, slub_max_order, 1, reserved);
2709 if (order <= slub_max_order)
2713 * Doh this slab cannot be placed using slub_max_order.
2715 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2716 if (order < MAX_ORDER)
2722 * Figure out what the alignment of the objects will be.
2724 static unsigned long calculate_alignment(unsigned long flags,
2725 unsigned long align, unsigned long size)
2728 * If the user wants hardware cache aligned objects then follow that
2729 * suggestion if the object is sufficiently large.
2731 * The hardware cache alignment cannot override the specified
2732 * alignment though. If that is greater then use it.
2734 if (flags & SLAB_HWCACHE_ALIGN) {
2735 unsigned long ralign = cache_line_size();
2736 while (size <= ralign / 2)
2738 align = max(align, ralign);
2741 if (align < ARCH_SLAB_MINALIGN)
2742 align = ARCH_SLAB_MINALIGN;
2744 return ALIGN(align, sizeof(void *));
2748 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2751 spin_lock_init(&n->list_lock);
2752 INIT_LIST_HEAD(&n->partial);
2753 #ifdef CONFIG_SLUB_DEBUG
2754 atomic_long_set(&n->nr_slabs, 0);
2755 atomic_long_set(&n->total_objects, 0);
2756 INIT_LIST_HEAD(&n->full);
2760 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2762 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2763 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2766 * Must align to double word boundary for the double cmpxchg
2767 * instructions to work; see __pcpu_double_call_return_bool().
2769 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2770 2 * sizeof(void *));
2775 init_kmem_cache_cpus(s);
2780 static struct kmem_cache *kmem_cache_node;
2783 * No kmalloc_node yet so do it by hand. We know that this is the first
2784 * slab on the node for this slabcache. There are no concurrent accesses
2787 * Note that this function only works on the kmalloc_node_cache
2788 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2789 * memory on a fresh node that has no slab structures yet.
2791 static void early_kmem_cache_node_alloc(int node)
2794 struct kmem_cache_node *n;
2796 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2798 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2801 if (page_to_nid(page) != node) {
2802 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2804 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2805 "in order to be able to continue\n");
2810 page->freelist = get_freepointer(kmem_cache_node, n);
2813 kmem_cache_node->node[node] = n;
2814 #ifdef CONFIG_SLUB_DEBUG
2815 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2816 init_tracking(kmem_cache_node, n);
2818 init_kmem_cache_node(n, kmem_cache_node);
2819 inc_slabs_node(kmem_cache_node, node, page->objects);
2821 add_partial(n, page, DEACTIVATE_TO_HEAD);
2824 static void free_kmem_cache_nodes(struct kmem_cache *s)
2828 for_each_node_state(node, N_NORMAL_MEMORY) {
2829 struct kmem_cache_node *n = s->node[node];
2832 kmem_cache_free(kmem_cache_node, n);
2834 s->node[node] = NULL;
2838 static int init_kmem_cache_nodes(struct kmem_cache *s)
2842 for_each_node_state(node, N_NORMAL_MEMORY) {
2843 struct kmem_cache_node *n;
2845 if (slab_state == DOWN) {
2846 early_kmem_cache_node_alloc(node);
2849 n = kmem_cache_alloc_node(kmem_cache_node,
2853 free_kmem_cache_nodes(s);
2858 init_kmem_cache_node(n, s);
2863 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2865 if (min < MIN_PARTIAL)
2867 else if (min > MAX_PARTIAL)
2869 s->min_partial = min;
2873 * calculate_sizes() determines the order and the distribution of data within
2876 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2878 unsigned long flags = s->flags;
2879 unsigned long size = s->objsize;
2880 unsigned long align = s->align;
2884 * Round up object size to the next word boundary. We can only
2885 * place the free pointer at word boundaries and this determines
2886 * the possible location of the free pointer.
2888 size = ALIGN(size, sizeof(void *));
2890 #ifdef CONFIG_SLUB_DEBUG
2892 * Determine if we can poison the object itself. If the user of
2893 * the slab may touch the object after free or before allocation
2894 * then we should never poison the object itself.
2896 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2898 s->flags |= __OBJECT_POISON;
2900 s->flags &= ~__OBJECT_POISON;
2904 * If we are Redzoning then check if there is some space between the
2905 * end of the object and the free pointer. If not then add an
2906 * additional word to have some bytes to store Redzone information.
2908 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2909 size += sizeof(void *);
2913 * With that we have determined the number of bytes in actual use
2914 * by the object. This is the potential offset to the free pointer.
2918 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2921 * Relocate free pointer after the object if it is not
2922 * permitted to overwrite the first word of the object on
2925 * This is the case if we do RCU, have a constructor or
2926 * destructor or are poisoning the objects.
2929 size += sizeof(void *);
2932 #ifdef CONFIG_SLUB_DEBUG
2933 if (flags & SLAB_STORE_USER)
2935 * Need to store information about allocs and frees after
2938 size += 2 * sizeof(struct track);
2940 if (flags & SLAB_RED_ZONE)
2942 * Add some empty padding so that we can catch
2943 * overwrites from earlier objects rather than let
2944 * tracking information or the free pointer be
2945 * corrupted if a user writes before the start
2948 size += sizeof(void *);
2952 * Determine the alignment based on various parameters that the
2953 * user specified and the dynamic determination of cache line size
2956 align = calculate_alignment(flags, align, s->objsize);
2960 * SLUB stores one object immediately after another beginning from
2961 * offset 0. In order to align the objects we have to simply size
2962 * each object to conform to the alignment.
2964 size = ALIGN(size, align);
2966 if (forced_order >= 0)
2967 order = forced_order;
2969 order = calculate_order(size, s->reserved);
2976 s->allocflags |= __GFP_COMP;
2978 if (s->flags & SLAB_CACHE_DMA)
2979 s->allocflags |= SLUB_DMA;
2981 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2982 s->allocflags |= __GFP_RECLAIMABLE;
2985 * Determine the number of objects per slab
2987 s->oo = oo_make(order, size, s->reserved);
2988 s->min = oo_make(get_order(size), size, s->reserved);
2989 if (oo_objects(s->oo) > oo_objects(s->max))
2992 return !!oo_objects(s->oo);
2996 static int kmem_cache_open(struct kmem_cache *s,
2997 const char *name, size_t size,
2998 size_t align, unsigned long flags,
2999 void (*ctor)(void *))
3001 memset(s, 0, kmem_size);
3006 s->flags = kmem_cache_flags(size, flags, name, ctor);
3009 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3010 s->reserved = sizeof(struct rcu_head);
3012 if (!calculate_sizes(s, -1))
3014 if (disable_higher_order_debug) {
3016 * Disable debugging flags that store metadata if the min slab
3019 if (get_order(s->size) > get_order(s->objsize)) {
3020 s->flags &= ~DEBUG_METADATA_FLAGS;
3022 if (!calculate_sizes(s, -1))
3027 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3028 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3029 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3030 /* Enable fast mode */
3031 s->flags |= __CMPXCHG_DOUBLE;
3035 * The larger the object size is, the more pages we want on the partial
3036 * list to avoid pounding the page allocator excessively.
3038 set_min_partial(s, ilog2(s->size) / 2);
3041 * cpu_partial determined the maximum number of objects kept in the
3042 * per cpu partial lists of a processor.
3044 * Per cpu partial lists mainly contain slabs that just have one
3045 * object freed. If they are used for allocation then they can be
3046 * filled up again with minimal effort. The slab will never hit the
3047 * per node partial lists and therefore no locking will be required.
3049 * This setting also determines
3051 * A) The number of objects from per cpu partial slabs dumped to the
3052 * per node list when we reach the limit.
3053 * B) The number of objects in cpu partial slabs to extract from the
3054 * per node list when we run out of per cpu objects. We only fetch 50%
3055 * to keep some capacity around for frees.
3057 if (kmem_cache_debug(s))
3059 else if (s->size >= PAGE_SIZE)
3061 else if (s->size >= 1024)
3063 else if (s->size >= 256)
3064 s->cpu_partial = 13;
3066 s->cpu_partial = 30;
3070 s->remote_node_defrag_ratio = 1000;
3072 if (!init_kmem_cache_nodes(s))
3075 if (alloc_kmem_cache_cpus(s))
3078 free_kmem_cache_nodes(s);
3080 if (flags & SLAB_PANIC)
3081 panic("Cannot create slab %s size=%lu realsize=%u "
3082 "order=%u offset=%u flags=%lx\n",
3083 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3089 * Determine the size of a slab object
3091 unsigned int kmem_cache_size(struct kmem_cache *s)
3095 EXPORT_SYMBOL(kmem_cache_size);
3097 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3100 #ifdef CONFIG_SLUB_DEBUG
3101 void *addr = page_address(page);
3103 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3104 sizeof(long), GFP_ATOMIC);
3107 slab_err(s, page, "%s", text);
3110 get_map(s, page, map);
3111 for_each_object(p, s, addr, page->objects) {
3113 if (!test_bit(slab_index(p, s, addr), map)) {
3114 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3116 print_tracking(s, p);
3125 * Attempt to free all partial slabs on a node.
3126 * This is called from kmem_cache_close(). We must be the last thread
3127 * using the cache and therefore we do not need to lock anymore.
3129 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3131 struct page *page, *h;
3133 list_for_each_entry_safe(page, h, &n->partial, lru) {
3135 remove_partial(n, page);
3136 discard_slab(s, page);
3138 list_slab_objects(s, page,
3139 "Objects remaining on kmem_cache_close()");
3145 * Release all resources used by a slab cache.
3147 static inline int kmem_cache_close(struct kmem_cache *s)
3152 free_percpu(s->cpu_slab);
3153 /* Attempt to free all objects */
3154 for_each_node_state(node, N_NORMAL_MEMORY) {
3155 struct kmem_cache_node *n = get_node(s, node);
3158 if (n->nr_partial || slabs_node(s, node))
3161 free_kmem_cache_nodes(s);
3166 * Close a cache and release the kmem_cache structure
3167 * (must be used for caches created using kmem_cache_create)
3169 void kmem_cache_destroy(struct kmem_cache *s)
3171 down_write(&slub_lock);
3175 up_write(&slub_lock);
3176 if (kmem_cache_close(s)) {
3177 printk(KERN_ERR "SLUB %s: %s called for cache that "
3178 "still has objects.\n", s->name, __func__);
3181 if (s->flags & SLAB_DESTROY_BY_RCU)
3183 sysfs_slab_remove(s);
3185 up_write(&slub_lock);
3187 EXPORT_SYMBOL(kmem_cache_destroy);
3189 /********************************************************************
3191 *******************************************************************/
3193 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3194 EXPORT_SYMBOL(kmalloc_caches);
3196 static struct kmem_cache *kmem_cache;
3198 #ifdef CONFIG_ZONE_DMA
3199 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3202 static int __init setup_slub_min_order(char *str)
3204 get_option(&str, &slub_min_order);
3209 __setup("slub_min_order=", setup_slub_min_order);
3211 static int __init setup_slub_max_order(char *str)
3213 get_option(&str, &slub_max_order);
3214 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3219 __setup("slub_max_order=", setup_slub_max_order);
3221 static int __init setup_slub_min_objects(char *str)
3223 get_option(&str, &slub_min_objects);
3228 __setup("slub_min_objects=", setup_slub_min_objects);
3230 static int __init setup_slub_nomerge(char *str)
3236 __setup("slub_nomerge", setup_slub_nomerge);
3238 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3239 int size, unsigned int flags)
3241 struct kmem_cache *s;
3243 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3246 * This function is called with IRQs disabled during early-boot on
3247 * single CPU so there's no need to take slub_lock here.
3249 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3253 list_add(&s->list, &slab_caches);
3257 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3262 * Conversion table for small slabs sizes / 8 to the index in the
3263 * kmalloc array. This is necessary for slabs < 192 since we have non power
3264 * of two cache sizes there. The size of larger slabs can be determined using
3267 static s8 size_index[24] = {
3294 static inline int size_index_elem(size_t bytes)
3296 return (bytes - 1) / 8;
3299 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3305 return ZERO_SIZE_PTR;
3307 index = size_index[size_index_elem(size)];
3309 index = fls(size - 1);
3311 #ifdef CONFIG_ZONE_DMA
3312 if (unlikely((flags & SLUB_DMA)))
3313 return kmalloc_dma_caches[index];
3316 return kmalloc_caches[index];
3319 void *__kmalloc(size_t size, gfp_t flags)
3321 struct kmem_cache *s;
3324 if (unlikely(size > SLUB_MAX_SIZE))
3325 return kmalloc_large(size, flags);
3327 s = get_slab(size, flags);
3329 if (unlikely(ZERO_OR_NULL_PTR(s)))
3332 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3334 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3338 EXPORT_SYMBOL(__kmalloc);
3341 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3346 flags |= __GFP_COMP | __GFP_NOTRACK;
3347 page = alloc_pages_node(node, flags, get_order(size));
3349 ptr = page_address(page);
3351 kmemleak_alloc(ptr, size, 1, flags);
3355 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3357 struct kmem_cache *s;
3360 if (unlikely(size > SLUB_MAX_SIZE)) {
3361 ret = kmalloc_large_node(size, flags, node);
3363 trace_kmalloc_node(_RET_IP_, ret,
3364 size, PAGE_SIZE << get_order(size),
3370 s = get_slab(size, flags);
3372 if (unlikely(ZERO_OR_NULL_PTR(s)))
3375 ret = slab_alloc(s, flags, node, _RET_IP_);
3377 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3381 EXPORT_SYMBOL(__kmalloc_node);
3384 size_t ksize(const void *object)
3388 if (unlikely(object == ZERO_SIZE_PTR))
3391 page = virt_to_head_page(object);
3393 if (unlikely(!PageSlab(page))) {
3394 WARN_ON(!PageCompound(page));
3395 return PAGE_SIZE << compound_order(page);
3398 return slab_ksize(page->slab);
3400 EXPORT_SYMBOL(ksize);
3402 #ifdef CONFIG_SLUB_DEBUG
3403 bool verify_mem_not_deleted(const void *x)
3406 void *object = (void *)x;
3407 unsigned long flags;
3410 if (unlikely(ZERO_OR_NULL_PTR(x)))
3413 local_irq_save(flags);
3415 page = virt_to_head_page(x);
3416 if (unlikely(!PageSlab(page))) {
3417 /* maybe it was from stack? */
3423 if (on_freelist(page->slab, page, object)) {
3424 object_err(page->slab, page, object, "Object is on free-list");
3432 local_irq_restore(flags);
3435 EXPORT_SYMBOL(verify_mem_not_deleted);
3438 void kfree(const void *x)
3441 void *object = (void *)x;
3443 trace_kfree(_RET_IP_, x);
3445 if (unlikely(ZERO_OR_NULL_PTR(x)))
3448 page = virt_to_head_page(x);
3449 if (unlikely(!PageSlab(page))) {
3450 BUG_ON(!PageCompound(page));
3455 slab_free(page->slab, page, object, _RET_IP_);
3457 EXPORT_SYMBOL(kfree);
3460 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3461 * the remaining slabs by the number of items in use. The slabs with the
3462 * most items in use come first. New allocations will then fill those up
3463 * and thus they can be removed from the partial lists.
3465 * The slabs with the least items are placed last. This results in them
3466 * being allocated from last increasing the chance that the last objects
3467 * are freed in them.
3469 int kmem_cache_shrink(struct kmem_cache *s)
3473 struct kmem_cache_node *n;
3476 int objects = oo_objects(s->max);
3477 struct list_head *slabs_by_inuse =
3478 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3479 unsigned long flags;
3481 if (!slabs_by_inuse)
3485 for_each_node_state(node, N_NORMAL_MEMORY) {
3486 n = get_node(s, node);
3491 for (i = 0; i < objects; i++)
3492 INIT_LIST_HEAD(slabs_by_inuse + i);
3494 spin_lock_irqsave(&n->list_lock, flags);
3497 * Build lists indexed by the items in use in each slab.
3499 * Note that concurrent frees may occur while we hold the
3500 * list_lock. page->inuse here is the upper limit.
3502 list_for_each_entry_safe(page, t, &n->partial, lru) {
3503 list_move(&page->lru, slabs_by_inuse + page->inuse);
3509 * Rebuild the partial list with the slabs filled up most
3510 * first and the least used slabs at the end.
3512 for (i = objects - 1; i > 0; i--)
3513 list_splice(slabs_by_inuse + i, n->partial.prev);
3515 spin_unlock_irqrestore(&n->list_lock, flags);
3517 /* Release empty slabs */
3518 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3519 discard_slab(s, page);
3522 kfree(slabs_by_inuse);
3525 EXPORT_SYMBOL(kmem_cache_shrink);
3527 #if defined(CONFIG_MEMORY_HOTPLUG)
3528 static int slab_mem_going_offline_callback(void *arg)
3530 struct kmem_cache *s;
3532 down_read(&slub_lock);
3533 list_for_each_entry(s, &slab_caches, list)
3534 kmem_cache_shrink(s);
3535 up_read(&slub_lock);
3540 static void slab_mem_offline_callback(void *arg)
3542 struct kmem_cache_node *n;
3543 struct kmem_cache *s;
3544 struct memory_notify *marg = arg;
3547 offline_node = marg->status_change_nid;
3550 * If the node still has available memory. we need kmem_cache_node
3553 if (offline_node < 0)
3556 down_read(&slub_lock);
3557 list_for_each_entry(s, &slab_caches, list) {
3558 n = get_node(s, offline_node);
3561 * if n->nr_slabs > 0, slabs still exist on the node
3562 * that is going down. We were unable to free them,
3563 * and offline_pages() function shouldn't call this
3564 * callback. So, we must fail.
3566 BUG_ON(slabs_node(s, offline_node));
3568 s->node[offline_node] = NULL;
3569 kmem_cache_free(kmem_cache_node, n);
3572 up_read(&slub_lock);
3575 static int slab_mem_going_online_callback(void *arg)
3577 struct kmem_cache_node *n;
3578 struct kmem_cache *s;
3579 struct memory_notify *marg = arg;
3580 int nid = marg->status_change_nid;
3584 * If the node's memory is already available, then kmem_cache_node is
3585 * already created. Nothing to do.
3591 * We are bringing a node online. No memory is available yet. We must
3592 * allocate a kmem_cache_node structure in order to bring the node
3595 down_read(&slub_lock);
3596 list_for_each_entry(s, &slab_caches, list) {
3598 * XXX: kmem_cache_alloc_node will fallback to other nodes
3599 * since memory is not yet available from the node that
3602 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3607 init_kmem_cache_node(n, s);
3611 up_read(&slub_lock);
3615 static int slab_memory_callback(struct notifier_block *self,
3616 unsigned long action, void *arg)
3621 case MEM_GOING_ONLINE:
3622 ret = slab_mem_going_online_callback(arg);
3624 case MEM_GOING_OFFLINE:
3625 ret = slab_mem_going_offline_callback(arg);
3628 case MEM_CANCEL_ONLINE:
3629 slab_mem_offline_callback(arg);
3632 case MEM_CANCEL_OFFLINE:
3636 ret = notifier_from_errno(ret);
3642 #endif /* CONFIG_MEMORY_HOTPLUG */
3644 /********************************************************************
3645 * Basic setup of slabs
3646 *******************************************************************/
3649 * Used for early kmem_cache structures that were allocated using
3650 * the page allocator
3653 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3657 list_add(&s->list, &slab_caches);
3660 for_each_node_state(node, N_NORMAL_MEMORY) {
3661 struct kmem_cache_node *n = get_node(s, node);
3665 list_for_each_entry(p, &n->partial, lru)
3668 #ifdef CONFIG_SLUB_DEBUG
3669 list_for_each_entry(p, &n->full, lru)
3676 void __init kmem_cache_init(void)
3680 struct kmem_cache *temp_kmem_cache;
3682 struct kmem_cache *temp_kmem_cache_node;
3683 unsigned long kmalloc_size;
3685 if (debug_guardpage_minorder())
3688 kmem_size = offsetof(struct kmem_cache, node) +
3689 nr_node_ids * sizeof(struct kmem_cache_node *);
3691 /* Allocate two kmem_caches from the page allocator */
3692 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3693 order = get_order(2 * kmalloc_size);
3694 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3697 * Must first have the slab cache available for the allocations of the
3698 * struct kmem_cache_node's. There is special bootstrap code in
3699 * kmem_cache_open for slab_state == DOWN.
3701 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3703 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3704 sizeof(struct kmem_cache_node),
3705 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3707 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3709 /* Able to allocate the per node structures */
3710 slab_state = PARTIAL;
3712 temp_kmem_cache = kmem_cache;
3713 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3714 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3715 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3716 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3719 * Allocate kmem_cache_node properly from the kmem_cache slab.
3720 * kmem_cache_node is separately allocated so no need to
3721 * update any list pointers.
3723 temp_kmem_cache_node = kmem_cache_node;
3725 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3726 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3728 kmem_cache_bootstrap_fixup(kmem_cache_node);
3731 kmem_cache_bootstrap_fixup(kmem_cache);
3733 /* Free temporary boot structure */
3734 free_pages((unsigned long)temp_kmem_cache, order);
3736 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3739 * Patch up the size_index table if we have strange large alignment
3740 * requirements for the kmalloc array. This is only the case for
3741 * MIPS it seems. The standard arches will not generate any code here.
3743 * Largest permitted alignment is 256 bytes due to the way we
3744 * handle the index determination for the smaller caches.
3746 * Make sure that nothing crazy happens if someone starts tinkering
3747 * around with ARCH_KMALLOC_MINALIGN
3749 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3750 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3752 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3753 int elem = size_index_elem(i);
3754 if (elem >= ARRAY_SIZE(size_index))
3756 size_index[elem] = KMALLOC_SHIFT_LOW;
3759 if (KMALLOC_MIN_SIZE == 64) {
3761 * The 96 byte size cache is not used if the alignment
3764 for (i = 64 + 8; i <= 96; i += 8)
3765 size_index[size_index_elem(i)] = 7;
3766 } else if (KMALLOC_MIN_SIZE == 128) {
3768 * The 192 byte sized cache is not used if the alignment
3769 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3772 for (i = 128 + 8; i <= 192; i += 8)
3773 size_index[size_index_elem(i)] = 8;
3776 /* Caches that are not of the two-to-the-power-of size */
3777 if (KMALLOC_MIN_SIZE <= 32) {
3778 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3782 if (KMALLOC_MIN_SIZE <= 64) {
3783 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3787 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3788 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3794 /* Provide the correct kmalloc names now that the caches are up */
3795 if (KMALLOC_MIN_SIZE <= 32) {
3796 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3797 BUG_ON(!kmalloc_caches[1]->name);
3800 if (KMALLOC_MIN_SIZE <= 64) {
3801 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3802 BUG_ON(!kmalloc_caches[2]->name);
3805 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3806 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3809 kmalloc_caches[i]->name = s;
3813 register_cpu_notifier(&slab_notifier);
3816 #ifdef CONFIG_ZONE_DMA
3817 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3818 struct kmem_cache *s = kmalloc_caches[i];
3821 char *name = kasprintf(GFP_NOWAIT,
3822 "dma-kmalloc-%d", s->objsize);
3825 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3826 s->objsize, SLAB_CACHE_DMA);
3831 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3832 " CPUs=%d, Nodes=%d\n",
3833 caches, cache_line_size(),
3834 slub_min_order, slub_max_order, slub_min_objects,
3835 nr_cpu_ids, nr_node_ids);
3838 void __init kmem_cache_init_late(void)
3843 * Find a mergeable slab cache
3845 static int slab_unmergeable(struct kmem_cache *s)
3847 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3854 * We may have set a slab to be unmergeable during bootstrap.
3856 if (s->refcount < 0)
3862 static struct kmem_cache *find_mergeable(size_t size,
3863 size_t align, unsigned long flags, const char *name,
3864 void (*ctor)(void *))
3866 struct kmem_cache *s;
3868 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3874 size = ALIGN(size, sizeof(void *));
3875 align = calculate_alignment(flags, align, size);
3876 size = ALIGN(size, align);
3877 flags = kmem_cache_flags(size, flags, name, NULL);
3879 list_for_each_entry(s, &slab_caches, list) {
3880 if (slab_unmergeable(s))
3886 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3889 * Check if alignment is compatible.
3890 * Courtesy of Adrian Drzewiecki
3892 if ((s->size & ~(align - 1)) != s->size)
3895 if (s->size - size >= sizeof(void *))
3903 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3904 size_t align, unsigned long flags, void (*ctor)(void *))
3906 struct kmem_cache *s;
3912 down_write(&slub_lock);
3913 s = find_mergeable(size, align, flags, name, ctor);
3917 * Adjust the object sizes so that we clear
3918 * the complete object on kzalloc.
3920 s->objsize = max(s->objsize, (int)size);
3921 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3923 if (sysfs_slab_alias(s, name)) {
3927 up_write(&slub_lock);
3931 n = kstrdup(name, GFP_KERNEL);
3935 s = kmalloc(kmem_size, GFP_KERNEL);
3937 if (kmem_cache_open(s, n,
3938 size, align, flags, ctor)) {
3939 list_add(&s->list, &slab_caches);
3940 up_write(&slub_lock);
3941 if (sysfs_slab_add(s)) {
3942 down_write(&slub_lock);
3954 up_write(&slub_lock);
3956 if (flags & SLAB_PANIC)
3957 panic("Cannot create slabcache %s\n", name);
3962 EXPORT_SYMBOL(kmem_cache_create);
3966 * Use the cpu notifier to insure that the cpu slabs are flushed when
3969 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3970 unsigned long action, void *hcpu)
3972 long cpu = (long)hcpu;
3973 struct kmem_cache *s;
3974 unsigned long flags;
3977 case CPU_UP_CANCELED:
3978 case CPU_UP_CANCELED_FROZEN:
3980 case CPU_DEAD_FROZEN:
3981 down_read(&slub_lock);
3982 list_for_each_entry(s, &slab_caches, list) {
3983 local_irq_save(flags);
3984 __flush_cpu_slab(s, cpu);
3985 local_irq_restore(flags);
3987 up_read(&slub_lock);
3995 static struct notifier_block __cpuinitdata slab_notifier = {
3996 .notifier_call = slab_cpuup_callback
4001 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4003 struct kmem_cache *s;
4006 if (unlikely(size > SLUB_MAX_SIZE))
4007 return kmalloc_large(size, gfpflags);
4009 s = get_slab(size, gfpflags);
4011 if (unlikely(ZERO_OR_NULL_PTR(s)))
4014 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4016 /* Honor the call site pointer we received. */
4017 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4023 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4024 int node, unsigned long caller)
4026 struct kmem_cache *s;
4029 if (unlikely(size > SLUB_MAX_SIZE)) {
4030 ret = kmalloc_large_node(size, gfpflags, node);
4032 trace_kmalloc_node(caller, ret,
4033 size, PAGE_SIZE << get_order(size),
4039 s = get_slab(size, gfpflags);
4041 if (unlikely(ZERO_OR_NULL_PTR(s)))
4044 ret = slab_alloc(s, gfpflags, node, caller);
4046 /* Honor the call site pointer we received. */
4047 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4054 static int count_inuse(struct page *page)
4059 static int count_total(struct page *page)
4061 return page->objects;
4065 #ifdef CONFIG_SLUB_DEBUG
4066 static int validate_slab(struct kmem_cache *s, struct page *page,
4070 void *addr = page_address(page);
4072 if (!check_slab(s, page) ||
4073 !on_freelist(s, page, NULL))
4076 /* Now we know that a valid freelist exists */
4077 bitmap_zero(map, page->objects);
4079 get_map(s, page, map);
4080 for_each_object(p, s, addr, page->objects) {
4081 if (test_bit(slab_index(p, s, addr), map))
4082 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4086 for_each_object(p, s, addr, page->objects)
4087 if (!test_bit(slab_index(p, s, addr), map))
4088 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4093 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4097 validate_slab(s, page, map);
4101 static int validate_slab_node(struct kmem_cache *s,
4102 struct kmem_cache_node *n, unsigned long *map)
4104 unsigned long count = 0;
4106 unsigned long flags;
4108 spin_lock_irqsave(&n->list_lock, flags);
4110 list_for_each_entry(page, &n->partial, lru) {
4111 validate_slab_slab(s, page, map);
4114 if (count != n->nr_partial)
4115 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4116 "counter=%ld\n", s->name, count, n->nr_partial);
4118 if (!(s->flags & SLAB_STORE_USER))
4121 list_for_each_entry(page, &n->full, lru) {
4122 validate_slab_slab(s, page, map);
4125 if (count != atomic_long_read(&n->nr_slabs))
4126 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4127 "counter=%ld\n", s->name, count,
4128 atomic_long_read(&n->nr_slabs));
4131 spin_unlock_irqrestore(&n->list_lock, flags);
4135 static long validate_slab_cache(struct kmem_cache *s)
4138 unsigned long count = 0;
4139 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4140 sizeof(unsigned long), GFP_KERNEL);
4146 for_each_node_state(node, N_NORMAL_MEMORY) {
4147 struct kmem_cache_node *n = get_node(s, node);
4149 count += validate_slab_node(s, n, map);
4155 * Generate lists of code addresses where slabcache objects are allocated
4160 unsigned long count;
4167 DECLARE_BITMAP(cpus, NR_CPUS);
4173 unsigned long count;
4174 struct location *loc;
4177 static void free_loc_track(struct loc_track *t)
4180 free_pages((unsigned long)t->loc,
4181 get_order(sizeof(struct location) * t->max));
4184 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4189 order = get_order(sizeof(struct location) * max);
4191 l = (void *)__get_free_pages(flags, order);
4196 memcpy(l, t->loc, sizeof(struct location) * t->count);
4204 static int add_location(struct loc_track *t, struct kmem_cache *s,
4205 const struct track *track)
4207 long start, end, pos;
4209 unsigned long caddr;
4210 unsigned long age = jiffies - track->when;
4216 pos = start + (end - start + 1) / 2;
4219 * There is nothing at "end". If we end up there
4220 * we need to add something to before end.
4225 caddr = t->loc[pos].addr;
4226 if (track->addr == caddr) {
4232 if (age < l->min_time)
4234 if (age > l->max_time)
4237 if (track->pid < l->min_pid)
4238 l->min_pid = track->pid;
4239 if (track->pid > l->max_pid)
4240 l->max_pid = track->pid;
4242 cpumask_set_cpu(track->cpu,
4243 to_cpumask(l->cpus));
4245 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4249 if (track->addr < caddr)
4256 * Not found. Insert new tracking element.
4258 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4264 (t->count - pos) * sizeof(struct location));
4267 l->addr = track->addr;
4271 l->min_pid = track->pid;
4272 l->max_pid = track->pid;
4273 cpumask_clear(to_cpumask(l->cpus));
4274 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4275 nodes_clear(l->nodes);
4276 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4280 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4281 struct page *page, enum track_item alloc,
4284 void *addr = page_address(page);
4287 bitmap_zero(map, page->objects);
4288 get_map(s, page, map);
4290 for_each_object(p, s, addr, page->objects)
4291 if (!test_bit(slab_index(p, s, addr), map))
4292 add_location(t, s, get_track(s, p, alloc));
4295 static int list_locations(struct kmem_cache *s, char *buf,
4296 enum track_item alloc)
4300 struct loc_track t = { 0, 0, NULL };
4302 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4303 sizeof(unsigned long), GFP_KERNEL);
4305 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4308 return sprintf(buf, "Out of memory\n");
4310 /* Push back cpu slabs */
4313 for_each_node_state(node, N_NORMAL_MEMORY) {
4314 struct kmem_cache_node *n = get_node(s, node);
4315 unsigned long flags;
4318 if (!atomic_long_read(&n->nr_slabs))
4321 spin_lock_irqsave(&n->list_lock, flags);
4322 list_for_each_entry(page, &n->partial, lru)
4323 process_slab(&t, s, page, alloc, map);
4324 list_for_each_entry(page, &n->full, lru)
4325 process_slab(&t, s, page, alloc, map);
4326 spin_unlock_irqrestore(&n->list_lock, flags);
4329 for (i = 0; i < t.count; i++) {
4330 struct location *l = &t.loc[i];
4332 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4334 len += sprintf(buf + len, "%7ld ", l->count);
4337 len += sprintf(buf + len, "%pS", (void *)l->addr);
4339 len += sprintf(buf + len, "<not-available>");
4341 if (l->sum_time != l->min_time) {
4342 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4344 (long)div_u64(l->sum_time, l->count),
4347 len += sprintf(buf + len, " age=%ld",
4350 if (l->min_pid != l->max_pid)
4351 len += sprintf(buf + len, " pid=%ld-%ld",
4352 l->min_pid, l->max_pid);
4354 len += sprintf(buf + len, " pid=%ld",
4357 if (num_online_cpus() > 1 &&
4358 !cpumask_empty(to_cpumask(l->cpus)) &&
4359 len < PAGE_SIZE - 60) {
4360 len += sprintf(buf + len, " cpus=");
4361 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4362 to_cpumask(l->cpus));
4365 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4366 len < PAGE_SIZE - 60) {
4367 len += sprintf(buf + len, " nodes=");
4368 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4372 len += sprintf(buf + len, "\n");
4378 len += sprintf(buf, "No data\n");
4383 #ifdef SLUB_RESILIENCY_TEST
4384 static void resiliency_test(void)
4388 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4390 printk(KERN_ERR "SLUB resiliency testing\n");
4391 printk(KERN_ERR "-----------------------\n");
4392 printk(KERN_ERR "A. Corruption after allocation\n");
4394 p = kzalloc(16, GFP_KERNEL);
4396 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4397 " 0x12->0x%p\n\n", p + 16);
4399 validate_slab_cache(kmalloc_caches[4]);
4401 /* Hmmm... The next two are dangerous */
4402 p = kzalloc(32, GFP_KERNEL);
4403 p[32 + sizeof(void *)] = 0x34;
4404 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4405 " 0x34 -> -0x%p\n", p);
4407 "If allocated object is overwritten then not detectable\n\n");
4409 validate_slab_cache(kmalloc_caches[5]);
4410 p = kzalloc(64, GFP_KERNEL);
4411 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4413 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4416 "If allocated object is overwritten then not detectable\n\n");
4417 validate_slab_cache(kmalloc_caches[6]);
4419 printk(KERN_ERR "\nB. Corruption after free\n");
4420 p = kzalloc(128, GFP_KERNEL);
4423 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4424 validate_slab_cache(kmalloc_caches[7]);
4426 p = kzalloc(256, GFP_KERNEL);
4429 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4431 validate_slab_cache(kmalloc_caches[8]);
4433 p = kzalloc(512, GFP_KERNEL);
4436 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4437 validate_slab_cache(kmalloc_caches[9]);
4441 static void resiliency_test(void) {};
4446 enum slab_stat_type {
4447 SL_ALL, /* All slabs */
4448 SL_PARTIAL, /* Only partially allocated slabs */
4449 SL_CPU, /* Only slabs used for cpu caches */
4450 SL_OBJECTS, /* Determine allocated objects not slabs */
4451 SL_TOTAL /* Determine object capacity not slabs */
4454 #define SO_ALL (1 << SL_ALL)
4455 #define SO_PARTIAL (1 << SL_PARTIAL)
4456 #define SO_CPU (1 << SL_CPU)
4457 #define SO_OBJECTS (1 << SL_OBJECTS)
4458 #define SO_TOTAL (1 << SL_TOTAL)
4460 static ssize_t show_slab_objects(struct kmem_cache *s,
4461 char *buf, unsigned long flags)
4463 unsigned long total = 0;
4466 unsigned long *nodes;
4467 unsigned long *per_cpu;
4469 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4472 per_cpu = nodes + nr_node_ids;
4474 if (flags & SO_CPU) {
4477 for_each_possible_cpu(cpu) {
4478 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4479 int node = ACCESS_ONCE(c->node);
4484 page = ACCESS_ONCE(c->page);
4486 if (flags & SO_TOTAL)
4488 else if (flags & SO_OBJECTS)
4507 lock_memory_hotplug();
4508 #ifdef CONFIG_SLUB_DEBUG
4509 if (flags & SO_ALL) {
4510 for_each_node_state(node, N_NORMAL_MEMORY) {
4511 struct kmem_cache_node *n = get_node(s, node);
4513 if (flags & SO_TOTAL)
4514 x = atomic_long_read(&n->total_objects);
4515 else if (flags & SO_OBJECTS)
4516 x = atomic_long_read(&n->total_objects) -
4517 count_partial(n, count_free);
4520 x = atomic_long_read(&n->nr_slabs);
4527 if (flags & SO_PARTIAL) {
4528 for_each_node_state(node, N_NORMAL_MEMORY) {
4529 struct kmem_cache_node *n = get_node(s, node);
4531 if (flags & SO_TOTAL)
4532 x = count_partial(n, count_total);
4533 else if (flags & SO_OBJECTS)
4534 x = count_partial(n, count_inuse);
4541 x = sprintf(buf, "%lu", total);
4543 for_each_node_state(node, N_NORMAL_MEMORY)
4545 x += sprintf(buf + x, " N%d=%lu",
4548 unlock_memory_hotplug();
4550 return x + sprintf(buf + x, "\n");
4553 #ifdef CONFIG_SLUB_DEBUG
4554 static int any_slab_objects(struct kmem_cache *s)
4558 for_each_online_node(node) {
4559 struct kmem_cache_node *n = get_node(s, node);
4564 if (atomic_long_read(&n->total_objects))
4571 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4572 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4574 struct slab_attribute {
4575 struct attribute attr;
4576 ssize_t (*show)(struct kmem_cache *s, char *buf);
4577 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4580 #define SLAB_ATTR_RO(_name) \
4581 static struct slab_attribute _name##_attr = \
4582 __ATTR(_name, 0400, _name##_show, NULL)
4584 #define SLAB_ATTR(_name) \
4585 static struct slab_attribute _name##_attr = \
4586 __ATTR(_name, 0600, _name##_show, _name##_store)
4588 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4590 return sprintf(buf, "%d\n", s->size);
4592 SLAB_ATTR_RO(slab_size);
4594 static ssize_t align_show(struct kmem_cache *s, char *buf)
4596 return sprintf(buf, "%d\n", s->align);
4598 SLAB_ATTR_RO(align);
4600 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4602 return sprintf(buf, "%d\n", s->objsize);
4604 SLAB_ATTR_RO(object_size);
4606 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4608 return sprintf(buf, "%d\n", oo_objects(s->oo));
4610 SLAB_ATTR_RO(objs_per_slab);
4612 static ssize_t order_store(struct kmem_cache *s,
4613 const char *buf, size_t length)
4615 unsigned long order;
4618 err = strict_strtoul(buf, 10, &order);
4622 if (order > slub_max_order || order < slub_min_order)
4625 calculate_sizes(s, order);
4629 static ssize_t order_show(struct kmem_cache *s, char *buf)
4631 return sprintf(buf, "%d\n", oo_order(s->oo));
4635 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4637 return sprintf(buf, "%lu\n", s->min_partial);
4640 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4646 err = strict_strtoul(buf, 10, &min);
4650 set_min_partial(s, min);
4653 SLAB_ATTR(min_partial);
4655 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4657 return sprintf(buf, "%u\n", s->cpu_partial);
4660 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4663 unsigned long objects;
4666 err = strict_strtoul(buf, 10, &objects);
4669 if (objects && kmem_cache_debug(s))
4672 s->cpu_partial = objects;
4676 SLAB_ATTR(cpu_partial);
4678 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4682 return sprintf(buf, "%pS\n", s->ctor);
4686 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4688 return sprintf(buf, "%d\n", s->refcount - 1);
4690 SLAB_ATTR_RO(aliases);
4692 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4694 return show_slab_objects(s, buf, SO_PARTIAL);
4696 SLAB_ATTR_RO(partial);
4698 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4700 return show_slab_objects(s, buf, SO_CPU);
4702 SLAB_ATTR_RO(cpu_slabs);
4704 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4706 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4708 SLAB_ATTR_RO(objects);
4710 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4712 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4714 SLAB_ATTR_RO(objects_partial);
4716 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4723 for_each_online_cpu(cpu) {
4724 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4727 pages += page->pages;
4728 objects += page->pobjects;
4732 len = sprintf(buf, "%d(%d)", objects, pages);
4735 for_each_online_cpu(cpu) {
4736 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4738 if (page && len < PAGE_SIZE - 20)
4739 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4740 page->pobjects, page->pages);
4743 return len + sprintf(buf + len, "\n");
4745 SLAB_ATTR_RO(slabs_cpu_partial);
4747 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4749 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4752 static ssize_t reclaim_account_store(struct kmem_cache *s,
4753 const char *buf, size_t length)
4755 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4757 s->flags |= SLAB_RECLAIM_ACCOUNT;
4760 SLAB_ATTR(reclaim_account);
4762 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4764 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4766 SLAB_ATTR_RO(hwcache_align);
4768 #ifdef CONFIG_ZONE_DMA
4769 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4771 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4773 SLAB_ATTR_RO(cache_dma);
4776 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4778 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4780 SLAB_ATTR_RO(destroy_by_rcu);
4782 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4784 return sprintf(buf, "%d\n", s->reserved);
4786 SLAB_ATTR_RO(reserved);
4788 #ifdef CONFIG_SLUB_DEBUG
4789 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4791 return show_slab_objects(s, buf, SO_ALL);
4793 SLAB_ATTR_RO(slabs);
4795 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4797 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4799 SLAB_ATTR_RO(total_objects);
4801 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4803 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4806 static ssize_t sanity_checks_store(struct kmem_cache *s,
4807 const char *buf, size_t length)
4809 s->flags &= ~SLAB_DEBUG_FREE;
4810 if (buf[0] == '1') {
4811 s->flags &= ~__CMPXCHG_DOUBLE;
4812 s->flags |= SLAB_DEBUG_FREE;
4816 SLAB_ATTR(sanity_checks);
4818 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4820 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4823 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4826 s->flags &= ~SLAB_TRACE;
4827 if (buf[0] == '1') {
4828 s->flags &= ~__CMPXCHG_DOUBLE;
4829 s->flags |= SLAB_TRACE;
4835 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4837 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4840 static ssize_t red_zone_store(struct kmem_cache *s,
4841 const char *buf, size_t length)
4843 if (any_slab_objects(s))
4846 s->flags &= ~SLAB_RED_ZONE;
4847 if (buf[0] == '1') {
4848 s->flags &= ~__CMPXCHG_DOUBLE;
4849 s->flags |= SLAB_RED_ZONE;
4851 calculate_sizes(s, -1);
4854 SLAB_ATTR(red_zone);
4856 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4858 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4861 static ssize_t poison_store(struct kmem_cache *s,
4862 const char *buf, size_t length)
4864 if (any_slab_objects(s))
4867 s->flags &= ~SLAB_POISON;
4868 if (buf[0] == '1') {
4869 s->flags &= ~__CMPXCHG_DOUBLE;
4870 s->flags |= SLAB_POISON;
4872 calculate_sizes(s, -1);
4877 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4879 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4882 static ssize_t store_user_store(struct kmem_cache *s,
4883 const char *buf, size_t length)
4885 if (any_slab_objects(s))
4888 s->flags &= ~SLAB_STORE_USER;
4889 if (buf[0] == '1') {
4890 s->flags &= ~__CMPXCHG_DOUBLE;
4891 s->flags |= SLAB_STORE_USER;
4893 calculate_sizes(s, -1);
4896 SLAB_ATTR(store_user);
4898 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4903 static ssize_t validate_store(struct kmem_cache *s,
4904 const char *buf, size_t length)
4908 if (buf[0] == '1') {
4909 ret = validate_slab_cache(s);
4915 SLAB_ATTR(validate);
4917 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4919 if (!(s->flags & SLAB_STORE_USER))
4921 return list_locations(s, buf, TRACK_ALLOC);
4923 SLAB_ATTR_RO(alloc_calls);
4925 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4927 if (!(s->flags & SLAB_STORE_USER))
4929 return list_locations(s, buf, TRACK_FREE);
4931 SLAB_ATTR_RO(free_calls);
4932 #endif /* CONFIG_SLUB_DEBUG */
4934 #ifdef CONFIG_FAILSLAB
4935 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4937 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4940 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4943 s->flags &= ~SLAB_FAILSLAB;
4945 s->flags |= SLAB_FAILSLAB;
4948 SLAB_ATTR(failslab);
4951 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4956 static ssize_t shrink_store(struct kmem_cache *s,
4957 const char *buf, size_t length)
4959 if (buf[0] == '1') {
4960 int rc = kmem_cache_shrink(s);
4971 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4973 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4976 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4977 const char *buf, size_t length)
4979 unsigned long ratio;
4982 err = strict_strtoul(buf, 10, &ratio);
4987 s->remote_node_defrag_ratio = ratio * 10;
4991 SLAB_ATTR(remote_node_defrag_ratio);
4994 #ifdef CONFIG_SLUB_STATS
4995 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4997 unsigned long sum = 0;
5000 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5005 for_each_online_cpu(cpu) {
5006 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5012 len = sprintf(buf, "%lu", sum);
5015 for_each_online_cpu(cpu) {
5016 if (data[cpu] && len < PAGE_SIZE - 20)
5017 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5021 return len + sprintf(buf + len, "\n");
5024 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5028 for_each_online_cpu(cpu)
5029 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5032 #define STAT_ATTR(si, text) \
5033 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5035 return show_stat(s, buf, si); \
5037 static ssize_t text##_store(struct kmem_cache *s, \
5038 const char *buf, size_t length) \
5040 if (buf[0] != '0') \
5042 clear_stat(s, si); \
5047 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5048 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5049 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5050 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5051 STAT_ATTR(FREE_FROZEN, free_frozen);
5052 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5053 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5054 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5055 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5056 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5057 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5058 STAT_ATTR(FREE_SLAB, free_slab);
5059 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5060 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5061 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5062 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5063 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5064 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5065 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5066 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5067 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5068 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5069 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5070 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5073 static struct attribute *slab_attrs[] = {
5074 &slab_size_attr.attr,
5075 &object_size_attr.attr,
5076 &objs_per_slab_attr.attr,
5078 &min_partial_attr.attr,
5079 &cpu_partial_attr.attr,
5081 &objects_partial_attr.attr,
5083 &cpu_slabs_attr.attr,
5087 &hwcache_align_attr.attr,
5088 &reclaim_account_attr.attr,
5089 &destroy_by_rcu_attr.attr,
5091 &reserved_attr.attr,
5092 &slabs_cpu_partial_attr.attr,
5093 #ifdef CONFIG_SLUB_DEBUG
5094 &total_objects_attr.attr,
5096 &sanity_checks_attr.attr,
5098 &red_zone_attr.attr,
5100 &store_user_attr.attr,
5101 &validate_attr.attr,
5102 &alloc_calls_attr.attr,
5103 &free_calls_attr.attr,
5105 #ifdef CONFIG_ZONE_DMA
5106 &cache_dma_attr.attr,
5109 &remote_node_defrag_ratio_attr.attr,
5111 #ifdef CONFIG_SLUB_STATS
5112 &alloc_fastpath_attr.attr,
5113 &alloc_slowpath_attr.attr,
5114 &free_fastpath_attr.attr,
5115 &free_slowpath_attr.attr,
5116 &free_frozen_attr.attr,
5117 &free_add_partial_attr.attr,
5118 &free_remove_partial_attr.attr,
5119 &alloc_from_partial_attr.attr,
5120 &alloc_slab_attr.attr,
5121 &alloc_refill_attr.attr,
5122 &alloc_node_mismatch_attr.attr,
5123 &free_slab_attr.attr,
5124 &cpuslab_flush_attr.attr,
5125 &deactivate_full_attr.attr,
5126 &deactivate_empty_attr.attr,
5127 &deactivate_to_head_attr.attr,
5128 &deactivate_to_tail_attr.attr,
5129 &deactivate_remote_frees_attr.attr,
5130 &deactivate_bypass_attr.attr,
5131 &order_fallback_attr.attr,
5132 &cmpxchg_double_fail_attr.attr,
5133 &cmpxchg_double_cpu_fail_attr.attr,
5134 &cpu_partial_alloc_attr.attr,
5135 &cpu_partial_free_attr.attr,
5137 #ifdef CONFIG_FAILSLAB
5138 &failslab_attr.attr,
5144 static struct attribute_group slab_attr_group = {
5145 .attrs = slab_attrs,
5148 static ssize_t slab_attr_show(struct kobject *kobj,
5149 struct attribute *attr,
5152 struct slab_attribute *attribute;
5153 struct kmem_cache *s;
5156 attribute = to_slab_attr(attr);
5159 if (!attribute->show)
5162 err = attribute->show(s, buf);
5167 static ssize_t slab_attr_store(struct kobject *kobj,
5168 struct attribute *attr,
5169 const char *buf, size_t len)
5171 struct slab_attribute *attribute;
5172 struct kmem_cache *s;
5175 attribute = to_slab_attr(attr);
5178 if (!attribute->store)
5181 err = attribute->store(s, buf, len);
5186 static void kmem_cache_release(struct kobject *kobj)
5188 struct kmem_cache *s = to_slab(kobj);
5194 static const struct sysfs_ops slab_sysfs_ops = {
5195 .show = slab_attr_show,
5196 .store = slab_attr_store,
5199 static struct kobj_type slab_ktype = {
5200 .sysfs_ops = &slab_sysfs_ops,
5201 .release = kmem_cache_release
5204 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5206 struct kobj_type *ktype = get_ktype(kobj);
5208 if (ktype == &slab_ktype)
5213 static const struct kset_uevent_ops slab_uevent_ops = {
5214 .filter = uevent_filter,
5217 static struct kset *slab_kset;
5219 #define ID_STR_LENGTH 64
5221 /* Create a unique string id for a slab cache:
5223 * Format :[flags-]size
5225 static char *create_unique_id(struct kmem_cache *s)
5227 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5234 * First flags affecting slabcache operations. We will only
5235 * get here for aliasable slabs so we do not need to support
5236 * too many flags. The flags here must cover all flags that
5237 * are matched during merging to guarantee that the id is
5240 if (s->flags & SLAB_CACHE_DMA)
5242 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5244 if (s->flags & SLAB_DEBUG_FREE)
5246 if (!(s->flags & SLAB_NOTRACK))
5250 p += sprintf(p, "%07d", s->size);
5251 BUG_ON(p > name + ID_STR_LENGTH - 1);
5255 static int sysfs_slab_add(struct kmem_cache *s)
5261 if (slab_state < SYSFS)
5262 /* Defer until later */
5265 unmergeable = slab_unmergeable(s);
5268 * Slabcache can never be merged so we can use the name proper.
5269 * This is typically the case for debug situations. In that
5270 * case we can catch duplicate names easily.
5272 sysfs_remove_link(&slab_kset->kobj, s->name);
5276 * Create a unique name for the slab as a target
5279 name = create_unique_id(s);
5282 s->kobj.kset = slab_kset;
5283 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5285 kobject_put(&s->kobj);
5289 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5291 kobject_del(&s->kobj);
5292 kobject_put(&s->kobj);
5295 kobject_uevent(&s->kobj, KOBJ_ADD);
5297 /* Setup first alias */
5298 sysfs_slab_alias(s, s->name);
5304 static void sysfs_slab_remove(struct kmem_cache *s)
5306 if (slab_state < SYSFS)
5308 * Sysfs has not been setup yet so no need to remove the
5313 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5314 kobject_del(&s->kobj);
5315 kobject_put(&s->kobj);
5319 * Need to buffer aliases during bootup until sysfs becomes
5320 * available lest we lose that information.
5322 struct saved_alias {
5323 struct kmem_cache *s;
5325 struct saved_alias *next;
5328 static struct saved_alias *alias_list;
5330 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5332 struct saved_alias *al;
5334 if (slab_state == SYSFS) {
5336 * If we have a leftover link then remove it.
5338 sysfs_remove_link(&slab_kset->kobj, name);
5339 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5342 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5348 al->next = alias_list;
5353 static int __init slab_sysfs_init(void)
5355 struct kmem_cache *s;
5358 down_write(&slub_lock);
5360 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5362 up_write(&slub_lock);
5363 printk(KERN_ERR "Cannot register slab subsystem.\n");
5369 list_for_each_entry(s, &slab_caches, list) {
5370 err = sysfs_slab_add(s);
5372 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5373 " to sysfs\n", s->name);
5376 while (alias_list) {
5377 struct saved_alias *al = alias_list;
5379 alias_list = alias_list->next;
5380 err = sysfs_slab_alias(al->s, al->name);
5382 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5383 " %s to sysfs\n", s->name);
5387 up_write(&slub_lock);
5392 __initcall(slab_sysfs_init);
5393 #endif /* CONFIG_SYSFS */
5396 * The /proc/slabinfo ABI
5398 #ifdef CONFIG_SLABINFO
5399 static void print_slabinfo_header(struct seq_file *m)
5401 seq_puts(m, "slabinfo - version: 2.1\n");
5402 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5403 "<objperslab> <pagesperslab>");
5404 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5405 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5409 static void *s_start(struct seq_file *m, loff_t *pos)
5413 down_read(&slub_lock);
5415 print_slabinfo_header(m);
5417 return seq_list_start(&slab_caches, *pos);
5420 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5422 return seq_list_next(p, &slab_caches, pos);
5425 static void s_stop(struct seq_file *m, void *p)
5427 up_read(&slub_lock);
5430 static int s_show(struct seq_file *m, void *p)
5432 unsigned long nr_partials = 0;
5433 unsigned long nr_slabs = 0;
5434 unsigned long nr_inuse = 0;
5435 unsigned long nr_objs = 0;
5436 unsigned long nr_free = 0;
5437 struct kmem_cache *s;
5440 s = list_entry(p, struct kmem_cache, list);
5442 for_each_online_node(node) {
5443 struct kmem_cache_node *n = get_node(s, node);
5448 nr_partials += n->nr_partial;
5449 nr_slabs += atomic_long_read(&n->nr_slabs);
5450 nr_objs += atomic_long_read(&n->total_objects);
5451 nr_free += count_partial(n, count_free);
5454 nr_inuse = nr_objs - nr_free;
5456 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5457 nr_objs, s->size, oo_objects(s->oo),
5458 (1 << oo_order(s->oo)));
5459 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5460 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5466 static const struct seq_operations slabinfo_op = {
5473 static int slabinfo_open(struct inode *inode, struct file *file)
5475 return seq_open(file, &slabinfo_op);
5478 static const struct file_operations proc_slabinfo_operations = {
5479 .open = slabinfo_open,
5481 .llseek = seq_lseek,
5482 .release = seq_release,
5485 static int __init slab_proc_init(void)
5487 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5490 module_init(slab_proc_init);
5491 #endif /* CONFIG_SLABINFO */