2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
136 * Issues still to be resolved:
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 * - Variable sizing of the per node arrays
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 #define MIN_PARTIAL 5
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
160 #define MAX_PARTIAL 10
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
173 * Set of flags that will prevent slab merging
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
183 #define OO_MASK ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON 0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
191 static struct notifier_block slab_notifier;
195 * Tracking user of a slab.
197 #define TRACK_ADDRS_COUNT 16
199 unsigned long addr; /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
208 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 static int sysfs_slab_add(struct kmem_cache *);
212 static int sysfs_slab_alias(struct kmem_cache *, const char *);
213 static void sysfs_slab_remove(struct kmem_cache *);
214 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
216 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
217 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
219 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
221 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void stat(const struct kmem_cache *s, enum stat_item si)
226 #ifdef CONFIG_SLUB_STATS
227 __this_cpu_inc(s->cpu_slab->stat[si]);
231 /********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
235 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237 return s->node[node];
240 /* Verify that a pointer has an address that is valid within a slab page */
241 static inline int check_valid_pointer(struct kmem_cache *s,
242 struct page *page, const void *object)
249 base = page_address(page);
250 if (object < base || object >= base + page->objects * s->size ||
251 (object - base) % s->size) {
258 static inline void *get_freepointer(struct kmem_cache *s, void *object)
260 return *(void **)(object + s->offset);
263 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
265 prefetch(object + s->offset);
268 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
272 #ifdef CONFIG_DEBUG_PAGEALLOC
273 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
275 p = get_freepointer(s, object);
280 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
282 *(void **)(object + s->offset) = fp;
285 /* Loop over all objects in a slab */
286 #define for_each_object(__p, __s, __addr, __objects) \
287 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 return (p - addr) / s->size;
296 static inline size_t slab_ksize(const struct kmem_cache *s)
298 #ifdef CONFIG_SLUB_DEBUG
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
304 return s->object_size;
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
315 * Else we can use all the padding etc for the allocation
320 static inline int order_objects(int order, unsigned long size, int reserved)
322 return ((PAGE_SIZE << order) - reserved) / size;
325 static inline struct kmem_cache_order_objects oo_make(int order,
326 unsigned long size, int reserved)
328 struct kmem_cache_order_objects x = {
329 (order << OO_SHIFT) + order_objects(order, size, reserved)
335 static inline int oo_order(struct kmem_cache_order_objects x)
337 return x.x >> OO_SHIFT;
340 static inline int oo_objects(struct kmem_cache_order_objects x)
342 return x.x & OO_MASK;
346 * Per slab locking using the pagelock
348 static __always_inline void slab_lock(struct page *page)
350 bit_spin_lock(PG_locked, &page->flags);
353 static __always_inline void slab_unlock(struct page *page)
355 __bit_spin_unlock(PG_locked, &page->flags);
358 /* Interrupts must be disabled (for the fallback code to work right) */
359 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
360 void *freelist_old, unsigned long counters_old,
361 void *freelist_new, unsigned long counters_new,
364 VM_BUG_ON(!irqs_disabled());
365 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
366 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
367 if (s->flags & __CMPXCHG_DOUBLE) {
368 if (cmpxchg_double(&page->freelist, &page->counters,
369 freelist_old, counters_old,
370 freelist_new, counters_new))
376 if (page->freelist == freelist_old && page->counters == counters_old) {
377 page->freelist = freelist_new;
378 page->counters = counters_new;
386 stat(s, CMPXCHG_DOUBLE_FAIL);
388 #ifdef SLUB_DEBUG_CMPXCHG
389 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
395 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
396 void *freelist_old, unsigned long counters_old,
397 void *freelist_new, unsigned long counters_new,
400 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
401 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
402 if (s->flags & __CMPXCHG_DOUBLE) {
403 if (cmpxchg_double(&page->freelist, &page->counters,
404 freelist_old, counters_old,
405 freelist_new, counters_new))
412 local_irq_save(flags);
414 if (page->freelist == freelist_old && page->counters == counters_old) {
415 page->freelist = freelist_new;
416 page->counters = counters_new;
418 local_irq_restore(flags);
422 local_irq_restore(flags);
426 stat(s, CMPXCHG_DOUBLE_FAIL);
428 #ifdef SLUB_DEBUG_CMPXCHG
429 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
435 #ifdef CONFIG_SLUB_DEBUG
437 * Determine a map of object in use on a page.
439 * Node listlock must be held to guarantee that the page does
440 * not vanish from under us.
442 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
445 void *addr = page_address(page);
447 for (p = page->freelist; p; p = get_freepointer(s, p))
448 set_bit(slab_index(p, s, addr), map);
454 #ifdef CONFIG_SLUB_DEBUG_ON
455 static int slub_debug = DEBUG_DEFAULT_FLAGS;
457 static int slub_debug;
460 static char *slub_debug_slabs;
461 static int disable_higher_order_debug;
466 static void print_section(char *text, u8 *addr, unsigned int length)
468 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
472 static struct track *get_track(struct kmem_cache *s, void *object,
473 enum track_item alloc)
478 p = object + s->offset + sizeof(void *);
480 p = object + s->inuse;
485 static void set_track(struct kmem_cache *s, void *object,
486 enum track_item alloc, unsigned long addr)
488 struct track *p = get_track(s, object, alloc);
491 #ifdef CONFIG_STACKTRACE
492 struct stack_trace trace;
495 trace.nr_entries = 0;
496 trace.max_entries = TRACK_ADDRS_COUNT;
497 trace.entries = p->addrs;
499 save_stack_trace(&trace);
501 /* See rant in lockdep.c */
502 if (trace.nr_entries != 0 &&
503 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
506 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
510 p->cpu = smp_processor_id();
511 p->pid = current->pid;
514 memset(p, 0, sizeof(struct track));
517 static void init_tracking(struct kmem_cache *s, void *object)
519 if (!(s->flags & SLAB_STORE_USER))
522 set_track(s, object, TRACK_FREE, 0UL);
523 set_track(s, object, TRACK_ALLOC, 0UL);
526 static void print_track(const char *s, struct track *t)
531 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
532 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
533 #ifdef CONFIG_STACKTRACE
536 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
538 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
545 static void print_tracking(struct kmem_cache *s, void *object)
547 if (!(s->flags & SLAB_STORE_USER))
550 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
551 print_track("Freed", get_track(s, object, TRACK_FREE));
554 static void print_page_info(struct page *page)
556 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
557 page, page->objects, page->inuse, page->freelist, page->flags);
561 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
567 vsnprintf(buf, sizeof(buf), fmt, args);
569 printk(KERN_ERR "========================================"
570 "=====================================\n");
571 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
572 printk(KERN_ERR "----------------------------------------"
573 "-------------------------------------\n\n");
575 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
578 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
584 vsnprintf(buf, sizeof(buf), fmt, args);
586 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
589 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
591 unsigned int off; /* Offset of last byte */
592 u8 *addr = page_address(page);
594 print_tracking(s, p);
596 print_page_info(page);
598 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
599 p, p - addr, get_freepointer(s, p));
602 print_section("Bytes b4 ", p - 16, 16);
604 print_section("Object ", p, min_t(unsigned long, s->object_size,
606 if (s->flags & SLAB_RED_ZONE)
607 print_section("Redzone ", p + s->object_size,
608 s->inuse - s->object_size);
611 off = s->offset + sizeof(void *);
615 if (s->flags & SLAB_STORE_USER)
616 off += 2 * sizeof(struct track);
619 /* Beginning of the filler is the free pointer */
620 print_section("Padding ", p + off, s->size - off);
625 static void object_err(struct kmem_cache *s, struct page *page,
626 u8 *object, char *reason)
628 slab_bug(s, "%s", reason);
629 print_trailer(s, page, object);
632 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
638 vsnprintf(buf, sizeof(buf), fmt, args);
640 slab_bug(s, "%s", buf);
641 print_page_info(page);
645 static void init_object(struct kmem_cache *s, void *object, u8 val)
649 if (s->flags & __OBJECT_POISON) {
650 memset(p, POISON_FREE, s->object_size - 1);
651 p[s->object_size - 1] = POISON_END;
654 if (s->flags & SLAB_RED_ZONE)
655 memset(p + s->object_size, val, s->inuse - s->object_size);
658 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
659 void *from, void *to)
661 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
662 memset(from, data, to - from);
665 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
666 u8 *object, char *what,
667 u8 *start, unsigned int value, unsigned int bytes)
672 fault = memchr_inv(start, value, bytes);
677 while (end > fault && end[-1] == value)
680 slab_bug(s, "%s overwritten", what);
681 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
682 fault, end - 1, fault[0], value);
683 print_trailer(s, page, object);
685 restore_bytes(s, what, value, fault, end);
693 * Bytes of the object to be managed.
694 * If the freepointer may overlay the object then the free
695 * pointer is the first word of the object.
697 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
700 * object + s->object_size
701 * Padding to reach word boundary. This is also used for Redzoning.
702 * Padding is extended by another word if Redzoning is enabled and
703 * object_size == inuse.
705 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
706 * 0xcc (RED_ACTIVE) for objects in use.
709 * Meta data starts here.
711 * A. Free pointer (if we cannot overwrite object on free)
712 * B. Tracking data for SLAB_STORE_USER
713 * C. Padding to reach required alignment boundary or at mininum
714 * one word if debugging is on to be able to detect writes
715 * before the word boundary.
717 * Padding is done using 0x5a (POISON_INUSE)
720 * Nothing is used beyond s->size.
722 * If slabcaches are merged then the object_size and inuse boundaries are mostly
723 * ignored. And therefore no slab options that rely on these boundaries
724 * may be used with merged slabcaches.
727 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
729 unsigned long off = s->inuse; /* The end of info */
732 /* Freepointer is placed after the object. */
733 off += sizeof(void *);
735 if (s->flags & SLAB_STORE_USER)
736 /* We also have user information there */
737 off += 2 * sizeof(struct track);
742 return check_bytes_and_report(s, page, p, "Object padding",
743 p + off, POISON_INUSE, s->size - off);
746 /* Check the pad bytes at the end of a slab page */
747 static int slab_pad_check(struct kmem_cache *s, struct page *page)
755 if (!(s->flags & SLAB_POISON))
758 start = page_address(page);
759 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
760 end = start + length;
761 remainder = length % s->size;
765 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
768 while (end > fault && end[-1] == POISON_INUSE)
771 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
772 print_section("Padding ", end - remainder, remainder);
774 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
778 static int check_object(struct kmem_cache *s, struct page *page,
779 void *object, u8 val)
782 u8 *endobject = object + s->object_size;
784 if (s->flags & SLAB_RED_ZONE) {
785 if (!check_bytes_and_report(s, page, object, "Redzone",
786 endobject, val, s->inuse - s->object_size))
789 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
790 check_bytes_and_report(s, page, p, "Alignment padding",
791 endobject, POISON_INUSE, s->inuse - s->object_size);
795 if (s->flags & SLAB_POISON) {
796 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
797 (!check_bytes_and_report(s, page, p, "Poison", p,
798 POISON_FREE, s->object_size - 1) ||
799 !check_bytes_and_report(s, page, p, "Poison",
800 p + s->object_size - 1, POISON_END, 1)))
803 * check_pad_bytes cleans up on its own.
805 check_pad_bytes(s, page, p);
808 if (!s->offset && val == SLUB_RED_ACTIVE)
810 * Object and freepointer overlap. Cannot check
811 * freepointer while object is allocated.
815 /* Check free pointer validity */
816 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
817 object_err(s, page, p, "Freepointer corrupt");
819 * No choice but to zap it and thus lose the remainder
820 * of the free objects in this slab. May cause
821 * another error because the object count is now wrong.
823 set_freepointer(s, p, NULL);
829 static int check_slab(struct kmem_cache *s, struct page *page)
833 VM_BUG_ON(!irqs_disabled());
835 if (!PageSlab(page)) {
836 slab_err(s, page, "Not a valid slab page");
840 maxobj = order_objects(compound_order(page), s->size, s->reserved);
841 if (page->objects > maxobj) {
842 slab_err(s, page, "objects %u > max %u",
843 s->name, page->objects, maxobj);
846 if (page->inuse > page->objects) {
847 slab_err(s, page, "inuse %u > max %u",
848 s->name, page->inuse, page->objects);
851 /* Slab_pad_check fixes things up after itself */
852 slab_pad_check(s, page);
857 * Determine if a certain object on a page is on the freelist. Must hold the
858 * slab lock to guarantee that the chains are in a consistent state.
860 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
865 unsigned long max_objects;
868 while (fp && nr <= page->objects) {
871 if (!check_valid_pointer(s, page, fp)) {
873 object_err(s, page, object,
874 "Freechain corrupt");
875 set_freepointer(s, object, NULL);
878 slab_err(s, page, "Freepointer corrupt");
879 page->freelist = NULL;
880 page->inuse = page->objects;
881 slab_fix(s, "Freelist cleared");
887 fp = get_freepointer(s, object);
891 max_objects = order_objects(compound_order(page), s->size, s->reserved);
892 if (max_objects > MAX_OBJS_PER_PAGE)
893 max_objects = MAX_OBJS_PER_PAGE;
895 if (page->objects != max_objects) {
896 slab_err(s, page, "Wrong number of objects. Found %d but "
897 "should be %d", page->objects, max_objects);
898 page->objects = max_objects;
899 slab_fix(s, "Number of objects adjusted.");
901 if (page->inuse != page->objects - nr) {
902 slab_err(s, page, "Wrong object count. Counter is %d but "
903 "counted were %d", page->inuse, page->objects - nr);
904 page->inuse = page->objects - nr;
905 slab_fix(s, "Object count adjusted.");
907 return search == NULL;
910 static void trace(struct kmem_cache *s, struct page *page, void *object,
913 if (s->flags & SLAB_TRACE) {
914 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
916 alloc ? "alloc" : "free",
921 print_section("Object ", (void *)object, s->object_size);
928 * Hooks for other subsystems that check memory allocations. In a typical
929 * production configuration these hooks all should produce no code at all.
931 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
933 kmemleak_alloc(ptr, size, 1, flags);
936 static inline void kfree_hook(const void *x)
941 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
943 flags &= gfp_allowed_mask;
944 lockdep_trace_alloc(flags);
945 might_sleep_if(flags & __GFP_WAIT);
947 return should_failslab(s->object_size, flags, s->flags);
950 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
952 flags &= gfp_allowed_mask;
953 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
954 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
957 static inline void slab_free_hook(struct kmem_cache *s, void *x)
959 kmemleak_free_recursive(x, s->flags);
962 * Trouble is that we may no longer disable interupts in the fast path
963 * So in order to make the debug calls that expect irqs to be
964 * disabled we need to disable interrupts temporarily.
966 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
970 local_irq_save(flags);
971 kmemcheck_slab_free(s, x, s->object_size);
972 debug_check_no_locks_freed(x, s->object_size);
973 local_irq_restore(flags);
976 if (!(s->flags & SLAB_DEBUG_OBJECTS))
977 debug_check_no_obj_freed(x, s->object_size);
981 * Tracking of fully allocated slabs for debugging purposes.
983 * list_lock must be held.
985 static void add_full(struct kmem_cache *s,
986 struct kmem_cache_node *n, struct page *page)
988 if (!(s->flags & SLAB_STORE_USER))
991 list_add(&page->lru, &n->full);
995 * list_lock must be held.
997 static void remove_full(struct kmem_cache *s, struct page *page)
999 if (!(s->flags & SLAB_STORE_USER))
1002 list_del(&page->lru);
1005 /* Tracking of the number of slabs for debugging purposes */
1006 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1008 struct kmem_cache_node *n = get_node(s, node);
1010 return atomic_long_read(&n->nr_slabs);
1013 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1015 return atomic_long_read(&n->nr_slabs);
1018 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1020 struct kmem_cache_node *n = get_node(s, node);
1023 * May be called early in order to allocate a slab for the
1024 * kmem_cache_node structure. Solve the chicken-egg
1025 * dilemma by deferring the increment of the count during
1026 * bootstrap (see early_kmem_cache_node_alloc).
1029 atomic_long_inc(&n->nr_slabs);
1030 atomic_long_add(objects, &n->total_objects);
1033 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1035 struct kmem_cache_node *n = get_node(s, node);
1037 atomic_long_dec(&n->nr_slabs);
1038 atomic_long_sub(objects, &n->total_objects);
1041 /* Object debug checks for alloc/free paths */
1042 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1045 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1048 init_object(s, object, SLUB_RED_INACTIVE);
1049 init_tracking(s, object);
1052 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1053 void *object, unsigned long addr)
1055 if (!check_slab(s, page))
1058 if (!check_valid_pointer(s, page, object)) {
1059 object_err(s, page, object, "Freelist Pointer check fails");
1063 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1066 /* Success perform special debug activities for allocs */
1067 if (s->flags & SLAB_STORE_USER)
1068 set_track(s, object, TRACK_ALLOC, addr);
1069 trace(s, page, object, 1);
1070 init_object(s, object, SLUB_RED_ACTIVE);
1074 if (PageSlab(page)) {
1076 * If this is a slab page then lets do the best we can
1077 * to avoid issues in the future. Marking all objects
1078 * as used avoids touching the remaining objects.
1080 slab_fix(s, "Marking all objects used");
1081 page->inuse = page->objects;
1082 page->freelist = NULL;
1087 static noinline struct kmem_cache_node *free_debug_processing(
1088 struct kmem_cache *s, struct page *page, void *object,
1089 unsigned long addr, unsigned long *flags)
1091 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1093 spin_lock_irqsave(&n->list_lock, *flags);
1096 if (!check_slab(s, page))
1099 if (!check_valid_pointer(s, page, object)) {
1100 slab_err(s, page, "Invalid object pointer 0x%p", object);
1104 if (on_freelist(s, page, object)) {
1105 object_err(s, page, object, "Object already free");
1109 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1112 if (unlikely(s != page->slab_cache)) {
1113 if (!PageSlab(page)) {
1114 slab_err(s, page, "Attempt to free object(0x%p) "
1115 "outside of slab", object);
1116 } else if (!page->slab_cache) {
1118 "SLUB <none>: no slab for object 0x%p.\n",
1122 object_err(s, page, object,
1123 "page slab pointer corrupt.");
1127 if (s->flags & SLAB_STORE_USER)
1128 set_track(s, object, TRACK_FREE, addr);
1129 trace(s, page, object, 0);
1130 init_object(s, object, SLUB_RED_INACTIVE);
1134 * Keep node_lock to preserve integrity
1135 * until the object is actually freed
1141 spin_unlock_irqrestore(&n->list_lock, *flags);
1142 slab_fix(s, "Object at 0x%p not freed", object);
1146 static int __init setup_slub_debug(char *str)
1148 slub_debug = DEBUG_DEFAULT_FLAGS;
1149 if (*str++ != '=' || !*str)
1151 * No options specified. Switch on full debugging.
1157 * No options but restriction on slabs. This means full
1158 * debugging for slabs matching a pattern.
1162 if (tolower(*str) == 'o') {
1164 * Avoid enabling debugging on caches if its minimum order
1165 * would increase as a result.
1167 disable_higher_order_debug = 1;
1174 * Switch off all debugging measures.
1179 * Determine which debug features should be switched on
1181 for (; *str && *str != ','; str++) {
1182 switch (tolower(*str)) {
1184 slub_debug |= SLAB_DEBUG_FREE;
1187 slub_debug |= SLAB_RED_ZONE;
1190 slub_debug |= SLAB_POISON;
1193 slub_debug |= SLAB_STORE_USER;
1196 slub_debug |= SLAB_TRACE;
1199 slub_debug |= SLAB_FAILSLAB;
1202 printk(KERN_ERR "slub_debug option '%c' "
1203 "unknown. skipped\n", *str);
1209 slub_debug_slabs = str + 1;
1214 __setup("slub_debug", setup_slub_debug);
1216 static unsigned long kmem_cache_flags(unsigned long object_size,
1217 unsigned long flags, const char *name,
1218 void (*ctor)(void *))
1221 * Enable debugging if selected on the kernel commandline.
1223 if (slub_debug && (!slub_debug_slabs || (name &&
1224 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1225 flags |= slub_debug;
1230 static inline void setup_object_debug(struct kmem_cache *s,
1231 struct page *page, void *object) {}
1233 static inline int alloc_debug_processing(struct kmem_cache *s,
1234 struct page *page, void *object, unsigned long addr) { return 0; }
1236 static inline struct kmem_cache_node *free_debug_processing(
1237 struct kmem_cache *s, struct page *page, void *object,
1238 unsigned long addr, unsigned long *flags) { return NULL; }
1240 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1242 static inline int check_object(struct kmem_cache *s, struct page *page,
1243 void *object, u8 val) { return 1; }
1244 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1245 struct page *page) {}
1246 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1247 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1248 unsigned long flags, const char *name,
1249 void (*ctor)(void *))
1253 #define slub_debug 0
1255 #define disable_higher_order_debug 0
1257 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1259 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1261 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1263 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1266 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1268 kmemleak_alloc(ptr, size, 1, flags);
1271 static inline void kfree_hook(const void *x)
1276 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1279 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1282 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1283 flags & gfp_allowed_mask);
1286 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1288 kmemleak_free_recursive(x, s->flags);
1291 #endif /* CONFIG_SLUB_DEBUG */
1294 * Slab allocation and freeing
1296 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1297 struct kmem_cache_order_objects oo)
1299 int order = oo_order(oo);
1301 flags |= __GFP_NOTRACK;
1303 if (node == NUMA_NO_NODE)
1304 return alloc_pages(flags, order);
1306 return alloc_pages_exact_node(node, flags, order);
1309 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1312 struct kmem_cache_order_objects oo = s->oo;
1315 flags &= gfp_allowed_mask;
1317 if (flags & __GFP_WAIT)
1320 flags |= s->allocflags;
1323 * Let the initial higher-order allocation fail under memory pressure
1324 * so we fall-back to the minimum order allocation.
1326 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1328 page = alloc_slab_page(alloc_gfp, node, oo);
1329 if (unlikely(!page)) {
1332 * Allocation may have failed due to fragmentation.
1333 * Try a lower order alloc if possible
1335 page = alloc_slab_page(flags, node, oo);
1338 stat(s, ORDER_FALLBACK);
1341 if (kmemcheck_enabled && page
1342 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1343 int pages = 1 << oo_order(oo);
1345 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1348 * Objects from caches that have a constructor don't get
1349 * cleared when they're allocated, so we need to do it here.
1352 kmemcheck_mark_uninitialized_pages(page, pages);
1354 kmemcheck_mark_unallocated_pages(page, pages);
1357 if (flags & __GFP_WAIT)
1358 local_irq_disable();
1362 page->objects = oo_objects(oo);
1363 mod_zone_page_state(page_zone(page),
1364 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1365 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1371 static void setup_object(struct kmem_cache *s, struct page *page,
1374 setup_object_debug(s, page, object);
1375 if (unlikely(s->ctor))
1379 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1387 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1389 page = allocate_slab(s,
1390 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1394 order = compound_order(page);
1395 inc_slabs_node(s, page_to_nid(page), page->objects);
1396 memcg_bind_pages(s, order);
1397 page->slab_cache = s;
1398 __SetPageSlab(page);
1399 if (page->pfmemalloc)
1400 SetPageSlabPfmemalloc(page);
1402 start = page_address(page);
1404 if (unlikely(s->flags & SLAB_POISON))
1405 memset(start, POISON_INUSE, PAGE_SIZE << order);
1408 for_each_object(p, s, start, page->objects) {
1409 setup_object(s, page, last);
1410 set_freepointer(s, last, p);
1413 setup_object(s, page, last);
1414 set_freepointer(s, last, NULL);
1416 page->freelist = start;
1417 page->inuse = page->objects;
1423 static void __free_slab(struct kmem_cache *s, struct page *page)
1425 int order = compound_order(page);
1426 int pages = 1 << order;
1428 if (kmem_cache_debug(s)) {
1431 slab_pad_check(s, page);
1432 for_each_object(p, s, page_address(page),
1434 check_object(s, page, p, SLUB_RED_INACTIVE);
1437 kmemcheck_free_shadow(page, compound_order(page));
1439 mod_zone_page_state(page_zone(page),
1440 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1441 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1444 __ClearPageSlabPfmemalloc(page);
1445 __ClearPageSlab(page);
1447 memcg_release_pages(s, order);
1448 page_mapcount_reset(page);
1449 if (current->reclaim_state)
1450 current->reclaim_state->reclaimed_slab += pages;
1451 __free_memcg_kmem_pages(page, order);
1454 #define need_reserve_slab_rcu \
1455 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1457 static void rcu_free_slab(struct rcu_head *h)
1461 if (need_reserve_slab_rcu)
1462 page = virt_to_head_page(h);
1464 page = container_of((struct list_head *)h, struct page, lru);
1466 __free_slab(page->slab_cache, page);
1469 static void free_slab(struct kmem_cache *s, struct page *page)
1471 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1472 struct rcu_head *head;
1474 if (need_reserve_slab_rcu) {
1475 int order = compound_order(page);
1476 int offset = (PAGE_SIZE << order) - s->reserved;
1478 VM_BUG_ON(s->reserved != sizeof(*head));
1479 head = page_address(page) + offset;
1482 * RCU free overloads the RCU head over the LRU
1484 head = (void *)&page->lru;
1487 call_rcu(head, rcu_free_slab);
1489 __free_slab(s, page);
1492 static void discard_slab(struct kmem_cache *s, struct page *page)
1494 dec_slabs_node(s, page_to_nid(page), page->objects);
1499 * Management of partially allocated slabs.
1501 * list_lock must be held.
1503 static inline void add_partial(struct kmem_cache_node *n,
1504 struct page *page, int tail)
1507 if (tail == DEACTIVATE_TO_TAIL)
1508 list_add_tail(&page->lru, &n->partial);
1510 list_add(&page->lru, &n->partial);
1514 * list_lock must be held.
1516 static inline void remove_partial(struct kmem_cache_node *n,
1519 list_del(&page->lru);
1524 * Remove slab from the partial list, freeze it and
1525 * return the pointer to the freelist.
1527 * Returns a list of objects or NULL if it fails.
1529 * Must hold list_lock since we modify the partial list.
1531 static inline void *acquire_slab(struct kmem_cache *s,
1532 struct kmem_cache_node *n, struct page *page,
1533 int mode, int *objects)
1536 unsigned long counters;
1540 * Zap the freelist and set the frozen bit.
1541 * The old freelist is the list of objects for the
1542 * per cpu allocation list.
1544 freelist = page->freelist;
1545 counters = page->counters;
1546 new.counters = counters;
1547 *objects = new.objects - new.inuse;
1549 new.inuse = page->objects;
1550 new.freelist = NULL;
1552 new.freelist = freelist;
1555 VM_BUG_ON(new.frozen);
1558 if (!__cmpxchg_double_slab(s, page,
1560 new.freelist, new.counters,
1564 remove_partial(n, page);
1569 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1570 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1573 * Try to allocate a partial slab from a specific node.
1575 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1576 struct kmem_cache_cpu *c, gfp_t flags)
1578 struct page *page, *page2;
1579 void *object = NULL;
1584 * Racy check. If we mistakenly see no partial slabs then we
1585 * just allocate an empty slab. If we mistakenly try to get a
1586 * partial slab and there is none available then get_partials()
1589 if (!n || !n->nr_partial)
1592 spin_lock(&n->list_lock);
1593 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1596 if (!pfmemalloc_match(page, flags))
1599 t = acquire_slab(s, n, page, object == NULL, &objects);
1603 available += objects;
1606 stat(s, ALLOC_FROM_PARTIAL);
1609 put_cpu_partial(s, page, 0);
1610 stat(s, CPU_PARTIAL_NODE);
1612 if (!kmem_cache_has_cpu_partial(s)
1613 || available > s->cpu_partial / 2)
1617 spin_unlock(&n->list_lock);
1622 * Get a page from somewhere. Search in increasing NUMA distances.
1624 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1625 struct kmem_cache_cpu *c)
1628 struct zonelist *zonelist;
1631 enum zone_type high_zoneidx = gfp_zone(flags);
1633 unsigned int cpuset_mems_cookie;
1636 * The defrag ratio allows a configuration of the tradeoffs between
1637 * inter node defragmentation and node local allocations. A lower
1638 * defrag_ratio increases the tendency to do local allocations
1639 * instead of attempting to obtain partial slabs from other nodes.
1641 * If the defrag_ratio is set to 0 then kmalloc() always
1642 * returns node local objects. If the ratio is higher then kmalloc()
1643 * may return off node objects because partial slabs are obtained
1644 * from other nodes and filled up.
1646 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1647 * defrag_ratio = 1000) then every (well almost) allocation will
1648 * first attempt to defrag slab caches on other nodes. This means
1649 * scanning over all nodes to look for partial slabs which may be
1650 * expensive if we do it every time we are trying to find a slab
1651 * with available objects.
1653 if (!s->remote_node_defrag_ratio ||
1654 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1658 cpuset_mems_cookie = get_mems_allowed();
1659 zonelist = node_zonelist(slab_node(), flags);
1660 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1661 struct kmem_cache_node *n;
1663 n = get_node(s, zone_to_nid(zone));
1665 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1666 n->nr_partial > s->min_partial) {
1667 object = get_partial_node(s, n, c, flags);
1670 * Return the object even if
1671 * put_mems_allowed indicated that
1672 * the cpuset mems_allowed was
1673 * updated in parallel. It's a
1674 * harmless race between the alloc
1675 * and the cpuset update.
1677 put_mems_allowed(cpuset_mems_cookie);
1682 } while (!put_mems_allowed(cpuset_mems_cookie));
1688 * Get a partial page, lock it and return it.
1690 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1691 struct kmem_cache_cpu *c)
1694 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1696 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1697 if (object || node != NUMA_NO_NODE)
1700 return get_any_partial(s, flags, c);
1703 #ifdef CONFIG_PREEMPT
1705 * Calculate the next globally unique transaction for disambiguiation
1706 * during cmpxchg. The transactions start with the cpu number and are then
1707 * incremented by CONFIG_NR_CPUS.
1709 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1712 * No preemption supported therefore also no need to check for
1718 static inline unsigned long next_tid(unsigned long tid)
1720 return tid + TID_STEP;
1723 static inline unsigned int tid_to_cpu(unsigned long tid)
1725 return tid % TID_STEP;
1728 static inline unsigned long tid_to_event(unsigned long tid)
1730 return tid / TID_STEP;
1733 static inline unsigned int init_tid(int cpu)
1738 static inline void note_cmpxchg_failure(const char *n,
1739 const struct kmem_cache *s, unsigned long tid)
1741 #ifdef SLUB_DEBUG_CMPXCHG
1742 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1744 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1746 #ifdef CONFIG_PREEMPT
1747 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1748 printk("due to cpu change %d -> %d\n",
1749 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1752 if (tid_to_event(tid) != tid_to_event(actual_tid))
1753 printk("due to cpu running other code. Event %ld->%ld\n",
1754 tid_to_event(tid), tid_to_event(actual_tid));
1756 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1757 actual_tid, tid, next_tid(tid));
1759 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1762 static void init_kmem_cache_cpus(struct kmem_cache *s)
1766 for_each_possible_cpu(cpu)
1767 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1771 * Remove the cpu slab
1773 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1775 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1776 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1778 enum slab_modes l = M_NONE, m = M_NONE;
1780 int tail = DEACTIVATE_TO_HEAD;
1784 if (page->freelist) {
1785 stat(s, DEACTIVATE_REMOTE_FREES);
1786 tail = DEACTIVATE_TO_TAIL;
1790 * Stage one: Free all available per cpu objects back
1791 * to the page freelist while it is still frozen. Leave the
1794 * There is no need to take the list->lock because the page
1797 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1799 unsigned long counters;
1802 prior = page->freelist;
1803 counters = page->counters;
1804 set_freepointer(s, freelist, prior);
1805 new.counters = counters;
1807 VM_BUG_ON(!new.frozen);
1809 } while (!__cmpxchg_double_slab(s, page,
1811 freelist, new.counters,
1812 "drain percpu freelist"));
1814 freelist = nextfree;
1818 * Stage two: Ensure that the page is unfrozen while the
1819 * list presence reflects the actual number of objects
1822 * We setup the list membership and then perform a cmpxchg
1823 * with the count. If there is a mismatch then the page
1824 * is not unfrozen but the page is on the wrong list.
1826 * Then we restart the process which may have to remove
1827 * the page from the list that we just put it on again
1828 * because the number of objects in the slab may have
1833 old.freelist = page->freelist;
1834 old.counters = page->counters;
1835 VM_BUG_ON(!old.frozen);
1837 /* Determine target state of the slab */
1838 new.counters = old.counters;
1841 set_freepointer(s, freelist, old.freelist);
1842 new.freelist = freelist;
1844 new.freelist = old.freelist;
1848 if (!new.inuse && n->nr_partial > s->min_partial)
1850 else if (new.freelist) {
1855 * Taking the spinlock removes the possiblity
1856 * that acquire_slab() will see a slab page that
1859 spin_lock(&n->list_lock);
1863 if (kmem_cache_debug(s) && !lock) {
1866 * This also ensures that the scanning of full
1867 * slabs from diagnostic functions will not see
1870 spin_lock(&n->list_lock);
1878 remove_partial(n, page);
1880 else if (l == M_FULL)
1882 remove_full(s, page);
1884 if (m == M_PARTIAL) {
1886 add_partial(n, page, tail);
1889 } else if (m == M_FULL) {
1891 stat(s, DEACTIVATE_FULL);
1892 add_full(s, n, page);
1898 if (!__cmpxchg_double_slab(s, page,
1899 old.freelist, old.counters,
1900 new.freelist, new.counters,
1905 spin_unlock(&n->list_lock);
1908 stat(s, DEACTIVATE_EMPTY);
1909 discard_slab(s, page);
1915 * Unfreeze all the cpu partial slabs.
1917 * This function must be called with interrupts disabled
1918 * for the cpu using c (or some other guarantee must be there
1919 * to guarantee no concurrent accesses).
1921 static void unfreeze_partials(struct kmem_cache *s,
1922 struct kmem_cache_cpu *c)
1924 #ifdef CONFIG_SLUB_CPU_PARTIAL
1925 struct kmem_cache_node *n = NULL, *n2 = NULL;
1926 struct page *page, *discard_page = NULL;
1928 while ((page = c->partial)) {
1932 c->partial = page->next;
1934 n2 = get_node(s, page_to_nid(page));
1937 spin_unlock(&n->list_lock);
1940 spin_lock(&n->list_lock);
1945 old.freelist = page->freelist;
1946 old.counters = page->counters;
1947 VM_BUG_ON(!old.frozen);
1949 new.counters = old.counters;
1950 new.freelist = old.freelist;
1954 } while (!__cmpxchg_double_slab(s, page,
1955 old.freelist, old.counters,
1956 new.freelist, new.counters,
1957 "unfreezing slab"));
1959 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1960 page->next = discard_page;
1961 discard_page = page;
1963 add_partial(n, page, DEACTIVATE_TO_TAIL);
1964 stat(s, FREE_ADD_PARTIAL);
1969 spin_unlock(&n->list_lock);
1971 while (discard_page) {
1972 page = discard_page;
1973 discard_page = discard_page->next;
1975 stat(s, DEACTIVATE_EMPTY);
1976 discard_slab(s, page);
1983 * Put a page that was just frozen (in __slab_free) into a partial page
1984 * slot if available. This is done without interrupts disabled and without
1985 * preemption disabled. The cmpxchg is racy and may put the partial page
1986 * onto a random cpus partial slot.
1988 * If we did not find a slot then simply move all the partials to the
1989 * per node partial list.
1991 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1993 #ifdef CONFIG_SLUB_CPU_PARTIAL
1994 struct page *oldpage;
2001 oldpage = this_cpu_read(s->cpu_slab->partial);
2004 pobjects = oldpage->pobjects;
2005 pages = oldpage->pages;
2006 if (drain && pobjects > s->cpu_partial) {
2007 unsigned long flags;
2009 * partial array is full. Move the existing
2010 * set to the per node partial list.
2012 local_irq_save(flags);
2013 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2014 local_irq_restore(flags);
2018 stat(s, CPU_PARTIAL_DRAIN);
2023 pobjects += page->objects - page->inuse;
2025 page->pages = pages;
2026 page->pobjects = pobjects;
2027 page->next = oldpage;
2029 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2033 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2035 stat(s, CPUSLAB_FLUSH);
2036 deactivate_slab(s, c->page, c->freelist);
2038 c->tid = next_tid(c->tid);
2046 * Called from IPI handler with interrupts disabled.
2048 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2050 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2056 unfreeze_partials(s, c);
2060 static void flush_cpu_slab(void *d)
2062 struct kmem_cache *s = d;
2064 __flush_cpu_slab(s, smp_processor_id());
2067 static bool has_cpu_slab(int cpu, void *info)
2069 struct kmem_cache *s = info;
2070 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2072 return c->page || c->partial;
2075 static void flush_all(struct kmem_cache *s)
2077 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2081 * Check if the objects in a per cpu structure fit numa
2082 * locality expectations.
2084 static inline int node_match(struct page *page, int node)
2087 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2093 static int count_free(struct page *page)
2095 return page->objects - page->inuse;
2098 static unsigned long count_partial(struct kmem_cache_node *n,
2099 int (*get_count)(struct page *))
2101 unsigned long flags;
2102 unsigned long x = 0;
2105 spin_lock_irqsave(&n->list_lock, flags);
2106 list_for_each_entry(page, &n->partial, lru)
2107 x += get_count(page);
2108 spin_unlock_irqrestore(&n->list_lock, flags);
2112 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2114 #ifdef CONFIG_SLUB_DEBUG
2115 return atomic_long_read(&n->total_objects);
2121 static noinline void
2122 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2127 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2129 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2130 "default order: %d, min order: %d\n", s->name, s->object_size,
2131 s->size, oo_order(s->oo), oo_order(s->min));
2133 if (oo_order(s->min) > get_order(s->object_size))
2134 printk(KERN_WARNING " %s debugging increased min order, use "
2135 "slub_debug=O to disable.\n", s->name);
2137 for_each_online_node(node) {
2138 struct kmem_cache_node *n = get_node(s, node);
2139 unsigned long nr_slabs;
2140 unsigned long nr_objs;
2141 unsigned long nr_free;
2146 nr_free = count_partial(n, count_free);
2147 nr_slabs = node_nr_slabs(n);
2148 nr_objs = node_nr_objs(n);
2151 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2152 node, nr_slabs, nr_objs, nr_free);
2156 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2157 int node, struct kmem_cache_cpu **pc)
2160 struct kmem_cache_cpu *c = *pc;
2163 freelist = get_partial(s, flags, node, c);
2168 page = new_slab(s, flags, node);
2170 c = __this_cpu_ptr(s->cpu_slab);
2175 * No other reference to the page yet so we can
2176 * muck around with it freely without cmpxchg
2178 freelist = page->freelist;
2179 page->freelist = NULL;
2181 stat(s, ALLOC_SLAB);
2190 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2192 if (unlikely(PageSlabPfmemalloc(page)))
2193 return gfp_pfmemalloc_allowed(gfpflags);
2199 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2200 * or deactivate the page.
2202 * The page is still frozen if the return value is not NULL.
2204 * If this function returns NULL then the page has been unfrozen.
2206 * This function must be called with interrupt disabled.
2208 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2211 unsigned long counters;
2215 freelist = page->freelist;
2216 counters = page->counters;
2218 new.counters = counters;
2219 VM_BUG_ON(!new.frozen);
2221 new.inuse = page->objects;
2222 new.frozen = freelist != NULL;
2224 } while (!__cmpxchg_double_slab(s, page,
2233 * Slow path. The lockless freelist is empty or we need to perform
2236 * Processing is still very fast if new objects have been freed to the
2237 * regular freelist. In that case we simply take over the regular freelist
2238 * as the lockless freelist and zap the regular freelist.
2240 * If that is not working then we fall back to the partial lists. We take the
2241 * first element of the freelist as the object to allocate now and move the
2242 * rest of the freelist to the lockless freelist.
2244 * And if we were unable to get a new slab from the partial slab lists then
2245 * we need to allocate a new slab. This is the slowest path since it involves
2246 * a call to the page allocator and the setup of a new slab.
2248 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2249 unsigned long addr, struct kmem_cache_cpu *c)
2253 unsigned long flags;
2255 local_irq_save(flags);
2256 #ifdef CONFIG_PREEMPT
2258 * We may have been preempted and rescheduled on a different
2259 * cpu before disabling interrupts. Need to reload cpu area
2262 c = this_cpu_ptr(s->cpu_slab);
2270 if (unlikely(!node_match(page, node))) {
2271 stat(s, ALLOC_NODE_MISMATCH);
2272 deactivate_slab(s, page, c->freelist);
2279 * By rights, we should be searching for a slab page that was
2280 * PFMEMALLOC but right now, we are losing the pfmemalloc
2281 * information when the page leaves the per-cpu allocator
2283 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2284 deactivate_slab(s, page, c->freelist);
2290 /* must check again c->freelist in case of cpu migration or IRQ */
2291 freelist = c->freelist;
2295 stat(s, ALLOC_SLOWPATH);
2297 freelist = get_freelist(s, page);
2301 stat(s, DEACTIVATE_BYPASS);
2305 stat(s, ALLOC_REFILL);
2309 * freelist is pointing to the list of objects to be used.
2310 * page is pointing to the page from which the objects are obtained.
2311 * That page must be frozen for per cpu allocations to work.
2313 VM_BUG_ON(!c->page->frozen);
2314 c->freelist = get_freepointer(s, freelist);
2315 c->tid = next_tid(c->tid);
2316 local_irq_restore(flags);
2322 page = c->page = c->partial;
2323 c->partial = page->next;
2324 stat(s, CPU_PARTIAL_ALLOC);
2329 freelist = new_slab_objects(s, gfpflags, node, &c);
2331 if (unlikely(!freelist)) {
2332 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2333 slab_out_of_memory(s, gfpflags, node);
2335 local_irq_restore(flags);
2340 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2343 /* Only entered in the debug case */
2344 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2345 goto new_slab; /* Slab failed checks. Next slab needed */
2347 deactivate_slab(s, page, get_freepointer(s, freelist));
2350 local_irq_restore(flags);
2355 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2356 * have the fastpath folded into their functions. So no function call
2357 * overhead for requests that can be satisfied on the fastpath.
2359 * The fastpath works by first checking if the lockless freelist can be used.
2360 * If not then __slab_alloc is called for slow processing.
2362 * Otherwise we can simply pick the next object from the lockless free list.
2364 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2365 gfp_t gfpflags, int node, unsigned long addr)
2368 struct kmem_cache_cpu *c;
2372 if (slab_pre_alloc_hook(s, gfpflags))
2375 s = memcg_kmem_get_cache(s, gfpflags);
2378 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2379 * enabled. We may switch back and forth between cpus while
2380 * reading from one cpu area. That does not matter as long
2381 * as we end up on the original cpu again when doing the cmpxchg.
2383 * Preemption is disabled for the retrieval of the tid because that
2384 * must occur from the current processor. We cannot allow rescheduling
2385 * on a different processor between the determination of the pointer
2386 * and the retrieval of the tid.
2389 c = __this_cpu_ptr(s->cpu_slab);
2392 * The transaction ids are globally unique per cpu and per operation on
2393 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2394 * occurs on the right processor and that there was no operation on the
2395 * linked list in between.
2400 object = c->freelist;
2402 if (unlikely(!object || !page || !node_match(page, node)))
2403 object = __slab_alloc(s, gfpflags, node, addr, c);
2406 void *next_object = get_freepointer_safe(s, object);
2409 * The cmpxchg will only match if there was no additional
2410 * operation and if we are on the right processor.
2412 * The cmpxchg does the following atomically (without lock semantics!)
2413 * 1. Relocate first pointer to the current per cpu area.
2414 * 2. Verify that tid and freelist have not been changed
2415 * 3. If they were not changed replace tid and freelist
2417 * Since this is without lock semantics the protection is only against
2418 * code executing on this cpu *not* from access by other cpus.
2420 if (unlikely(!this_cpu_cmpxchg_double(
2421 s->cpu_slab->freelist, s->cpu_slab->tid,
2423 next_object, next_tid(tid)))) {
2425 note_cmpxchg_failure("slab_alloc", s, tid);
2428 prefetch_freepointer(s, next_object);
2429 stat(s, ALLOC_FASTPATH);
2432 if (unlikely(gfpflags & __GFP_ZERO) && object)
2433 memset(object, 0, s->object_size);
2435 slab_post_alloc_hook(s, gfpflags, object);
2440 static __always_inline void *slab_alloc(struct kmem_cache *s,
2441 gfp_t gfpflags, unsigned long addr)
2443 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2446 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2448 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2450 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2454 EXPORT_SYMBOL(kmem_cache_alloc);
2456 #ifdef CONFIG_TRACING
2457 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2459 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2460 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2463 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2465 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2467 void *ret = kmalloc_order(size, flags, order);
2468 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2471 EXPORT_SYMBOL(kmalloc_order_trace);
2475 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2477 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2479 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2480 s->object_size, s->size, gfpflags, node);
2484 EXPORT_SYMBOL(kmem_cache_alloc_node);
2486 #ifdef CONFIG_TRACING
2487 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2489 int node, size_t size)
2491 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2493 trace_kmalloc_node(_RET_IP_, ret,
2494 size, s->size, gfpflags, node);
2497 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2502 * Slow patch handling. This may still be called frequently since objects
2503 * have a longer lifetime than the cpu slabs in most processing loads.
2505 * So we still attempt to reduce cache line usage. Just take the slab
2506 * lock and free the item. If there is no additional partial page
2507 * handling required then we can return immediately.
2509 static void __slab_free(struct kmem_cache *s, struct page *page,
2510 void *x, unsigned long addr)
2513 void **object = (void *)x;
2516 unsigned long counters;
2517 struct kmem_cache_node *n = NULL;
2518 unsigned long uninitialized_var(flags);
2520 stat(s, FREE_SLOWPATH);
2522 if (kmem_cache_debug(s) &&
2523 !(n = free_debug_processing(s, page, x, addr, &flags)))
2528 spin_unlock_irqrestore(&n->list_lock, flags);
2531 prior = page->freelist;
2532 counters = page->counters;
2533 set_freepointer(s, object, prior);
2534 new.counters = counters;
2535 was_frozen = new.frozen;
2537 if ((!new.inuse || !prior) && !was_frozen) {
2539 if (kmem_cache_has_cpu_partial(s) && !prior)
2542 * Slab was on no list before and will be partially empty
2543 * We can defer the list move and instead freeze it.
2547 else { /* Needs to be taken off a list */
2549 n = get_node(s, page_to_nid(page));
2551 * Speculatively acquire the list_lock.
2552 * If the cmpxchg does not succeed then we may
2553 * drop the list_lock without any processing.
2555 * Otherwise the list_lock will synchronize with
2556 * other processors updating the list of slabs.
2558 spin_lock_irqsave(&n->list_lock, flags);
2563 } while (!cmpxchg_double_slab(s, page,
2565 object, new.counters,
2571 * If we just froze the page then put it onto the
2572 * per cpu partial list.
2574 if (new.frozen && !was_frozen) {
2575 put_cpu_partial(s, page, 1);
2576 stat(s, CPU_PARTIAL_FREE);
2579 * The list lock was not taken therefore no list
2580 * activity can be necessary.
2583 stat(s, FREE_FROZEN);
2587 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2591 * Objects left in the slab. If it was not on the partial list before
2594 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2595 if (kmem_cache_debug(s))
2596 remove_full(s, page);
2597 add_partial(n, page, DEACTIVATE_TO_TAIL);
2598 stat(s, FREE_ADD_PARTIAL);
2600 spin_unlock_irqrestore(&n->list_lock, flags);
2606 * Slab on the partial list.
2608 remove_partial(n, page);
2609 stat(s, FREE_REMOVE_PARTIAL);
2611 /* Slab must be on the full list */
2612 remove_full(s, page);
2614 spin_unlock_irqrestore(&n->list_lock, flags);
2616 discard_slab(s, page);
2620 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2621 * can perform fastpath freeing without additional function calls.
2623 * The fastpath is only possible if we are freeing to the current cpu slab
2624 * of this processor. This typically the case if we have just allocated
2627 * If fastpath is not possible then fall back to __slab_free where we deal
2628 * with all sorts of special processing.
2630 static __always_inline void slab_free(struct kmem_cache *s,
2631 struct page *page, void *x, unsigned long addr)
2633 void **object = (void *)x;
2634 struct kmem_cache_cpu *c;
2637 slab_free_hook(s, x);
2641 * Determine the currently cpus per cpu slab.
2642 * The cpu may change afterward. However that does not matter since
2643 * data is retrieved via this pointer. If we are on the same cpu
2644 * during the cmpxchg then the free will succedd.
2647 c = __this_cpu_ptr(s->cpu_slab);
2652 if (likely(page == c->page)) {
2653 set_freepointer(s, object, c->freelist);
2655 if (unlikely(!this_cpu_cmpxchg_double(
2656 s->cpu_slab->freelist, s->cpu_slab->tid,
2658 object, next_tid(tid)))) {
2660 note_cmpxchg_failure("slab_free", s, tid);
2663 stat(s, FREE_FASTPATH);
2665 __slab_free(s, page, x, addr);
2669 void kmem_cache_free(struct kmem_cache *s, void *x)
2671 s = cache_from_obj(s, x);
2674 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2675 trace_kmem_cache_free(_RET_IP_, x);
2677 EXPORT_SYMBOL(kmem_cache_free);
2680 * Object placement in a slab is made very easy because we always start at
2681 * offset 0. If we tune the size of the object to the alignment then we can
2682 * get the required alignment by putting one properly sized object after
2685 * Notice that the allocation order determines the sizes of the per cpu
2686 * caches. Each processor has always one slab available for allocations.
2687 * Increasing the allocation order reduces the number of times that slabs
2688 * must be moved on and off the partial lists and is therefore a factor in
2693 * Mininum / Maximum order of slab pages. This influences locking overhead
2694 * and slab fragmentation. A higher order reduces the number of partial slabs
2695 * and increases the number of allocations possible without having to
2696 * take the list_lock.
2698 static int slub_min_order;
2699 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2700 static int slub_min_objects;
2703 * Merge control. If this is set then no merging of slab caches will occur.
2704 * (Could be removed. This was introduced to pacify the merge skeptics.)
2706 static int slub_nomerge;
2709 * Calculate the order of allocation given an slab object size.
2711 * The order of allocation has significant impact on performance and other
2712 * system components. Generally order 0 allocations should be preferred since
2713 * order 0 does not cause fragmentation in the page allocator. Larger objects
2714 * be problematic to put into order 0 slabs because there may be too much
2715 * unused space left. We go to a higher order if more than 1/16th of the slab
2718 * In order to reach satisfactory performance we must ensure that a minimum
2719 * number of objects is in one slab. Otherwise we may generate too much
2720 * activity on the partial lists which requires taking the list_lock. This is
2721 * less a concern for large slabs though which are rarely used.
2723 * slub_max_order specifies the order where we begin to stop considering the
2724 * number of objects in a slab as critical. If we reach slub_max_order then
2725 * we try to keep the page order as low as possible. So we accept more waste
2726 * of space in favor of a small page order.
2728 * Higher order allocations also allow the placement of more objects in a
2729 * slab and thereby reduce object handling overhead. If the user has
2730 * requested a higher mininum order then we start with that one instead of
2731 * the smallest order which will fit the object.
2733 static inline int slab_order(int size, int min_objects,
2734 int max_order, int fract_leftover, int reserved)
2738 int min_order = slub_min_order;
2740 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2741 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2743 for (order = max(min_order,
2744 fls(min_objects * size - 1) - PAGE_SHIFT);
2745 order <= max_order; order++) {
2747 unsigned long slab_size = PAGE_SIZE << order;
2749 if (slab_size < min_objects * size + reserved)
2752 rem = (slab_size - reserved) % size;
2754 if (rem <= slab_size / fract_leftover)
2762 static inline int calculate_order(int size, int reserved)
2770 * Attempt to find best configuration for a slab. This
2771 * works by first attempting to generate a layout with
2772 * the best configuration and backing off gradually.
2774 * First we reduce the acceptable waste in a slab. Then
2775 * we reduce the minimum objects required in a slab.
2777 min_objects = slub_min_objects;
2779 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2780 max_objects = order_objects(slub_max_order, size, reserved);
2781 min_objects = min(min_objects, max_objects);
2783 while (min_objects > 1) {
2785 while (fraction >= 4) {
2786 order = slab_order(size, min_objects,
2787 slub_max_order, fraction, reserved);
2788 if (order <= slub_max_order)
2796 * We were unable to place multiple objects in a slab. Now
2797 * lets see if we can place a single object there.
2799 order = slab_order(size, 1, slub_max_order, 1, reserved);
2800 if (order <= slub_max_order)
2804 * Doh this slab cannot be placed using slub_max_order.
2806 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2807 if (order < MAX_ORDER)
2813 init_kmem_cache_node(struct kmem_cache_node *n)
2816 spin_lock_init(&n->list_lock);
2817 INIT_LIST_HEAD(&n->partial);
2818 #ifdef CONFIG_SLUB_DEBUG
2819 atomic_long_set(&n->nr_slabs, 0);
2820 atomic_long_set(&n->total_objects, 0);
2821 INIT_LIST_HEAD(&n->full);
2825 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2827 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2828 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2831 * Must align to double word boundary for the double cmpxchg
2832 * instructions to work; see __pcpu_double_call_return_bool().
2834 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2835 2 * sizeof(void *));
2840 init_kmem_cache_cpus(s);
2845 static struct kmem_cache *kmem_cache_node;
2848 * No kmalloc_node yet so do it by hand. We know that this is the first
2849 * slab on the node for this slabcache. There are no concurrent accesses
2852 * Note that this function only works on the kmem_cache_node
2853 * when allocating for the kmem_cache_node. This is used for bootstrapping
2854 * memory on a fresh node that has no slab structures yet.
2856 static void early_kmem_cache_node_alloc(int node)
2859 struct kmem_cache_node *n;
2861 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2863 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2866 if (page_to_nid(page) != node) {
2867 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2869 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2870 "in order to be able to continue\n");
2875 page->freelist = get_freepointer(kmem_cache_node, n);
2878 kmem_cache_node->node[node] = n;
2879 #ifdef CONFIG_SLUB_DEBUG
2880 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2881 init_tracking(kmem_cache_node, n);
2883 init_kmem_cache_node(n);
2884 inc_slabs_node(kmem_cache_node, node, page->objects);
2886 add_partial(n, page, DEACTIVATE_TO_HEAD);
2889 static void free_kmem_cache_nodes(struct kmem_cache *s)
2893 for_each_node_state(node, N_NORMAL_MEMORY) {
2894 struct kmem_cache_node *n = s->node[node];
2897 kmem_cache_free(kmem_cache_node, n);
2899 s->node[node] = NULL;
2903 static int init_kmem_cache_nodes(struct kmem_cache *s)
2907 for_each_node_state(node, N_NORMAL_MEMORY) {
2908 struct kmem_cache_node *n;
2910 if (slab_state == DOWN) {
2911 early_kmem_cache_node_alloc(node);
2914 n = kmem_cache_alloc_node(kmem_cache_node,
2918 free_kmem_cache_nodes(s);
2923 init_kmem_cache_node(n);
2928 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2930 if (min < MIN_PARTIAL)
2932 else if (min > MAX_PARTIAL)
2934 s->min_partial = min;
2938 * calculate_sizes() determines the order and the distribution of data within
2941 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2943 unsigned long flags = s->flags;
2944 unsigned long size = s->object_size;
2948 * Round up object size to the next word boundary. We can only
2949 * place the free pointer at word boundaries and this determines
2950 * the possible location of the free pointer.
2952 size = ALIGN(size, sizeof(void *));
2954 #ifdef CONFIG_SLUB_DEBUG
2956 * Determine if we can poison the object itself. If the user of
2957 * the slab may touch the object after free or before allocation
2958 * then we should never poison the object itself.
2960 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2962 s->flags |= __OBJECT_POISON;
2964 s->flags &= ~__OBJECT_POISON;
2968 * If we are Redzoning then check if there is some space between the
2969 * end of the object and the free pointer. If not then add an
2970 * additional word to have some bytes to store Redzone information.
2972 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2973 size += sizeof(void *);
2977 * With that we have determined the number of bytes in actual use
2978 * by the object. This is the potential offset to the free pointer.
2982 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2985 * Relocate free pointer after the object if it is not
2986 * permitted to overwrite the first word of the object on
2989 * This is the case if we do RCU, have a constructor or
2990 * destructor or are poisoning the objects.
2993 size += sizeof(void *);
2996 #ifdef CONFIG_SLUB_DEBUG
2997 if (flags & SLAB_STORE_USER)
2999 * Need to store information about allocs and frees after
3002 size += 2 * sizeof(struct track);
3004 if (flags & SLAB_RED_ZONE)
3006 * Add some empty padding so that we can catch
3007 * overwrites from earlier objects rather than let
3008 * tracking information or the free pointer be
3009 * corrupted if a user writes before the start
3012 size += sizeof(void *);
3016 * SLUB stores one object immediately after another beginning from
3017 * offset 0. In order to align the objects we have to simply size
3018 * each object to conform to the alignment.
3020 size = ALIGN(size, s->align);
3022 if (forced_order >= 0)
3023 order = forced_order;
3025 order = calculate_order(size, s->reserved);
3032 s->allocflags |= __GFP_COMP;
3034 if (s->flags & SLAB_CACHE_DMA)
3035 s->allocflags |= GFP_DMA;
3037 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3038 s->allocflags |= __GFP_RECLAIMABLE;
3041 * Determine the number of objects per slab
3043 s->oo = oo_make(order, size, s->reserved);
3044 s->min = oo_make(get_order(size), size, s->reserved);
3045 if (oo_objects(s->oo) > oo_objects(s->max))
3048 return !!oo_objects(s->oo);
3051 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3053 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3056 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3057 s->reserved = sizeof(struct rcu_head);
3059 if (!calculate_sizes(s, -1))
3061 if (disable_higher_order_debug) {
3063 * Disable debugging flags that store metadata if the min slab
3066 if (get_order(s->size) > get_order(s->object_size)) {
3067 s->flags &= ~DEBUG_METADATA_FLAGS;
3069 if (!calculate_sizes(s, -1))
3074 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3075 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3076 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3077 /* Enable fast mode */
3078 s->flags |= __CMPXCHG_DOUBLE;
3082 * The larger the object size is, the more pages we want on the partial
3083 * list to avoid pounding the page allocator excessively.
3085 set_min_partial(s, ilog2(s->size) / 2);
3088 * cpu_partial determined the maximum number of objects kept in the
3089 * per cpu partial lists of a processor.
3091 * Per cpu partial lists mainly contain slabs that just have one
3092 * object freed. If they are used for allocation then they can be
3093 * filled up again with minimal effort. The slab will never hit the
3094 * per node partial lists and therefore no locking will be required.
3096 * This setting also determines
3098 * A) The number of objects from per cpu partial slabs dumped to the
3099 * per node list when we reach the limit.
3100 * B) The number of objects in cpu partial slabs to extract from the
3101 * per node list when we run out of per cpu objects. We only fetch 50%
3102 * to keep some capacity around for frees.
3104 if (!kmem_cache_has_cpu_partial(s))
3106 else if (s->size >= PAGE_SIZE)
3108 else if (s->size >= 1024)
3110 else if (s->size >= 256)
3111 s->cpu_partial = 13;
3113 s->cpu_partial = 30;
3116 s->remote_node_defrag_ratio = 1000;
3118 if (!init_kmem_cache_nodes(s))
3121 if (alloc_kmem_cache_cpus(s))
3124 free_kmem_cache_nodes(s);
3126 if (flags & SLAB_PANIC)
3127 panic("Cannot create slab %s size=%lu realsize=%u "
3128 "order=%u offset=%u flags=%lx\n",
3129 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3134 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3137 #ifdef CONFIG_SLUB_DEBUG
3138 void *addr = page_address(page);
3140 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3141 sizeof(long), GFP_ATOMIC);
3144 slab_err(s, page, text, s->name);
3147 get_map(s, page, map);
3148 for_each_object(p, s, addr, page->objects) {
3150 if (!test_bit(slab_index(p, s, addr), map)) {
3151 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3153 print_tracking(s, p);
3162 * Attempt to free all partial slabs on a node.
3163 * This is called from kmem_cache_close(). We must be the last thread
3164 * using the cache and therefore we do not need to lock anymore.
3166 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3168 struct page *page, *h;
3170 list_for_each_entry_safe(page, h, &n->partial, lru) {
3172 remove_partial(n, page);
3173 discard_slab(s, page);
3175 list_slab_objects(s, page,
3176 "Objects remaining in %s on kmem_cache_close()");
3182 * Release all resources used by a slab cache.
3184 static inline int kmem_cache_close(struct kmem_cache *s)
3189 /* Attempt to free all objects */
3190 for_each_node_state(node, N_NORMAL_MEMORY) {
3191 struct kmem_cache_node *n = get_node(s, node);
3194 if (n->nr_partial || slabs_node(s, node))
3197 free_percpu(s->cpu_slab);
3198 free_kmem_cache_nodes(s);
3202 int __kmem_cache_shutdown(struct kmem_cache *s)
3204 int rc = kmem_cache_close(s);
3208 * We do the same lock strategy around sysfs_slab_add, see
3209 * __kmem_cache_create. Because this is pretty much the last
3210 * operation we do and the lock will be released shortly after
3211 * that in slab_common.c, we could just move sysfs_slab_remove
3212 * to a later point in common code. We should do that when we
3213 * have a common sysfs framework for all allocators.
3215 mutex_unlock(&slab_mutex);
3216 sysfs_slab_remove(s);
3217 mutex_lock(&slab_mutex);
3223 /********************************************************************
3225 *******************************************************************/
3227 static int __init setup_slub_min_order(char *str)
3229 get_option(&str, &slub_min_order);
3234 __setup("slub_min_order=", setup_slub_min_order);
3236 static int __init setup_slub_max_order(char *str)
3238 get_option(&str, &slub_max_order);
3239 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3244 __setup("slub_max_order=", setup_slub_max_order);
3246 static int __init setup_slub_min_objects(char *str)
3248 get_option(&str, &slub_min_objects);
3253 __setup("slub_min_objects=", setup_slub_min_objects);
3255 static int __init setup_slub_nomerge(char *str)
3261 __setup("slub_nomerge", setup_slub_nomerge);
3263 void *__kmalloc(size_t size, gfp_t flags)
3265 struct kmem_cache *s;
3268 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3269 return kmalloc_large(size, flags);
3271 s = kmalloc_slab(size, flags);
3273 if (unlikely(ZERO_OR_NULL_PTR(s)))
3276 ret = slab_alloc(s, flags, _RET_IP_);
3278 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3282 EXPORT_SYMBOL(__kmalloc);
3285 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3290 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3291 page = alloc_pages_node(node, flags, get_order(size));
3293 ptr = page_address(page);
3295 kmalloc_large_node_hook(ptr, size, flags);
3299 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3301 struct kmem_cache *s;
3304 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3305 ret = kmalloc_large_node(size, flags, node);
3307 trace_kmalloc_node(_RET_IP_, ret,
3308 size, PAGE_SIZE << get_order(size),
3314 s = kmalloc_slab(size, flags);
3316 if (unlikely(ZERO_OR_NULL_PTR(s)))
3319 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3321 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3325 EXPORT_SYMBOL(__kmalloc_node);
3328 size_t ksize(const void *object)
3332 if (unlikely(object == ZERO_SIZE_PTR))
3335 page = virt_to_head_page(object);
3337 if (unlikely(!PageSlab(page))) {
3338 WARN_ON(!PageCompound(page));
3339 return PAGE_SIZE << compound_order(page);
3342 return slab_ksize(page->slab_cache);
3344 EXPORT_SYMBOL(ksize);
3346 #ifdef CONFIG_SLUB_DEBUG
3347 bool verify_mem_not_deleted(const void *x)
3350 void *object = (void *)x;
3351 unsigned long flags;
3354 if (unlikely(ZERO_OR_NULL_PTR(x)))
3357 local_irq_save(flags);
3359 page = virt_to_head_page(x);
3360 if (unlikely(!PageSlab(page))) {
3361 /* maybe it was from stack? */
3367 if (on_freelist(page->slab_cache, page, object)) {
3368 object_err(page->slab_cache, page, object, "Object is on free-list");
3376 local_irq_restore(flags);
3379 EXPORT_SYMBOL(verify_mem_not_deleted);
3382 void kfree(const void *x)
3385 void *object = (void *)x;
3387 trace_kfree(_RET_IP_, x);
3389 if (unlikely(ZERO_OR_NULL_PTR(x)))
3392 page = virt_to_head_page(x);
3393 if (unlikely(!PageSlab(page))) {
3394 BUG_ON(!PageCompound(page));
3396 __free_memcg_kmem_pages(page, compound_order(page));
3399 slab_free(page->slab_cache, page, object, _RET_IP_);
3401 EXPORT_SYMBOL(kfree);
3404 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3405 * the remaining slabs by the number of items in use. The slabs with the
3406 * most items in use come first. New allocations will then fill those up
3407 * and thus they can be removed from the partial lists.
3409 * The slabs with the least items are placed last. This results in them
3410 * being allocated from last increasing the chance that the last objects
3411 * are freed in them.
3413 int kmem_cache_shrink(struct kmem_cache *s)
3417 struct kmem_cache_node *n;
3420 int objects = oo_objects(s->max);
3421 struct list_head *slabs_by_inuse =
3422 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3423 unsigned long flags;
3425 if (!slabs_by_inuse)
3429 for_each_node_state(node, N_NORMAL_MEMORY) {
3430 n = get_node(s, node);
3435 for (i = 0; i < objects; i++)
3436 INIT_LIST_HEAD(slabs_by_inuse + i);
3438 spin_lock_irqsave(&n->list_lock, flags);
3441 * Build lists indexed by the items in use in each slab.
3443 * Note that concurrent frees may occur while we hold the
3444 * list_lock. page->inuse here is the upper limit.
3446 list_for_each_entry_safe(page, t, &n->partial, lru) {
3447 list_move(&page->lru, slabs_by_inuse + page->inuse);
3453 * Rebuild the partial list with the slabs filled up most
3454 * first and the least used slabs at the end.
3456 for (i = objects - 1; i > 0; i--)
3457 list_splice(slabs_by_inuse + i, n->partial.prev);
3459 spin_unlock_irqrestore(&n->list_lock, flags);
3461 /* Release empty slabs */
3462 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3463 discard_slab(s, page);
3466 kfree(slabs_by_inuse);
3469 EXPORT_SYMBOL(kmem_cache_shrink);
3471 static int slab_mem_going_offline_callback(void *arg)
3473 struct kmem_cache *s;
3475 mutex_lock(&slab_mutex);
3476 list_for_each_entry(s, &slab_caches, list)
3477 kmem_cache_shrink(s);
3478 mutex_unlock(&slab_mutex);
3483 static void slab_mem_offline_callback(void *arg)
3485 struct kmem_cache_node *n;
3486 struct kmem_cache *s;
3487 struct memory_notify *marg = arg;
3490 offline_node = marg->status_change_nid_normal;
3493 * If the node still has available memory. we need kmem_cache_node
3496 if (offline_node < 0)
3499 mutex_lock(&slab_mutex);
3500 list_for_each_entry(s, &slab_caches, list) {
3501 n = get_node(s, offline_node);
3504 * if n->nr_slabs > 0, slabs still exist on the node
3505 * that is going down. We were unable to free them,
3506 * and offline_pages() function shouldn't call this
3507 * callback. So, we must fail.
3509 BUG_ON(slabs_node(s, offline_node));
3511 s->node[offline_node] = NULL;
3512 kmem_cache_free(kmem_cache_node, n);
3515 mutex_unlock(&slab_mutex);
3518 static int slab_mem_going_online_callback(void *arg)
3520 struct kmem_cache_node *n;
3521 struct kmem_cache *s;
3522 struct memory_notify *marg = arg;
3523 int nid = marg->status_change_nid_normal;
3527 * If the node's memory is already available, then kmem_cache_node is
3528 * already created. Nothing to do.
3534 * We are bringing a node online. No memory is available yet. We must
3535 * allocate a kmem_cache_node structure in order to bring the node
3538 mutex_lock(&slab_mutex);
3539 list_for_each_entry(s, &slab_caches, list) {
3541 * XXX: kmem_cache_alloc_node will fallback to other nodes
3542 * since memory is not yet available from the node that
3545 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3550 init_kmem_cache_node(n);
3554 mutex_unlock(&slab_mutex);
3558 static int slab_memory_callback(struct notifier_block *self,
3559 unsigned long action, void *arg)
3564 case MEM_GOING_ONLINE:
3565 ret = slab_mem_going_online_callback(arg);
3567 case MEM_GOING_OFFLINE:
3568 ret = slab_mem_going_offline_callback(arg);
3571 case MEM_CANCEL_ONLINE:
3572 slab_mem_offline_callback(arg);
3575 case MEM_CANCEL_OFFLINE:
3579 ret = notifier_from_errno(ret);
3585 static struct notifier_block slab_memory_callback_nb = {
3586 .notifier_call = slab_memory_callback,
3587 .priority = SLAB_CALLBACK_PRI,
3590 /********************************************************************
3591 * Basic setup of slabs
3592 *******************************************************************/
3595 * Used for early kmem_cache structures that were allocated using
3596 * the page allocator. Allocate them properly then fix up the pointers
3597 * that may be pointing to the wrong kmem_cache structure.
3600 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3603 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3605 memcpy(s, static_cache, kmem_cache->object_size);
3608 * This runs very early, and only the boot processor is supposed to be
3609 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3612 __flush_cpu_slab(s, smp_processor_id());
3613 for_each_node_state(node, N_NORMAL_MEMORY) {
3614 struct kmem_cache_node *n = get_node(s, node);
3618 list_for_each_entry(p, &n->partial, lru)
3621 #ifdef CONFIG_SLUB_DEBUG
3622 list_for_each_entry(p, &n->full, lru)
3627 list_add(&s->list, &slab_caches);
3631 void __init kmem_cache_init(void)
3633 static __initdata struct kmem_cache boot_kmem_cache,
3634 boot_kmem_cache_node;
3636 if (debug_guardpage_minorder())
3639 kmem_cache_node = &boot_kmem_cache_node;
3640 kmem_cache = &boot_kmem_cache;
3642 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3643 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3645 register_hotmemory_notifier(&slab_memory_callback_nb);
3647 /* Able to allocate the per node structures */
3648 slab_state = PARTIAL;
3650 create_boot_cache(kmem_cache, "kmem_cache",
3651 offsetof(struct kmem_cache, node) +
3652 nr_node_ids * sizeof(struct kmem_cache_node *),
3653 SLAB_HWCACHE_ALIGN);
3655 kmem_cache = bootstrap(&boot_kmem_cache);
3658 * Allocate kmem_cache_node properly from the kmem_cache slab.
3659 * kmem_cache_node is separately allocated so no need to
3660 * update any list pointers.
3662 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3664 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3665 create_kmalloc_caches(0);
3668 register_cpu_notifier(&slab_notifier);
3672 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3673 " CPUs=%d, Nodes=%d\n",
3675 slub_min_order, slub_max_order, slub_min_objects,
3676 nr_cpu_ids, nr_node_ids);
3679 void __init kmem_cache_init_late(void)
3684 * Find a mergeable slab cache
3686 static int slab_unmergeable(struct kmem_cache *s)
3688 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3695 * We may have set a slab to be unmergeable during bootstrap.
3697 if (s->refcount < 0)
3703 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3704 size_t align, unsigned long flags, const char *name,
3705 void (*ctor)(void *))
3707 struct kmem_cache *s;
3709 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3715 size = ALIGN(size, sizeof(void *));
3716 align = calculate_alignment(flags, align, size);
3717 size = ALIGN(size, align);
3718 flags = kmem_cache_flags(size, flags, name, NULL);
3720 list_for_each_entry(s, &slab_caches, list) {
3721 if (slab_unmergeable(s))
3727 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3730 * Check if alignment is compatible.
3731 * Courtesy of Adrian Drzewiecki
3733 if ((s->size & ~(align - 1)) != s->size)
3736 if (s->size - size >= sizeof(void *))
3739 if (!cache_match_memcg(s, memcg))
3748 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3749 size_t align, unsigned long flags, void (*ctor)(void *))
3751 struct kmem_cache *s;
3753 s = find_mergeable(memcg, size, align, flags, name, ctor);
3757 * Adjust the object sizes so that we clear
3758 * the complete object on kzalloc.
3760 s->object_size = max(s->object_size, (int)size);
3761 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3763 if (sysfs_slab_alias(s, name)) {
3772 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3776 err = kmem_cache_open(s, flags);
3780 /* Mutex is not taken during early boot */
3781 if (slab_state <= UP)
3784 memcg_propagate_slab_attrs(s);
3785 mutex_unlock(&slab_mutex);
3786 err = sysfs_slab_add(s);
3787 mutex_lock(&slab_mutex);
3790 kmem_cache_close(s);
3797 * Use the cpu notifier to insure that the cpu slabs are flushed when
3800 static int slab_cpuup_callback(struct notifier_block *nfb,
3801 unsigned long action, void *hcpu)
3803 long cpu = (long)hcpu;
3804 struct kmem_cache *s;
3805 unsigned long flags;
3808 case CPU_UP_CANCELED:
3809 case CPU_UP_CANCELED_FROZEN:
3811 case CPU_DEAD_FROZEN:
3812 mutex_lock(&slab_mutex);
3813 list_for_each_entry(s, &slab_caches, list) {
3814 local_irq_save(flags);
3815 __flush_cpu_slab(s, cpu);
3816 local_irq_restore(flags);
3818 mutex_unlock(&slab_mutex);
3826 static struct notifier_block slab_notifier = {
3827 .notifier_call = slab_cpuup_callback
3832 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3834 struct kmem_cache *s;
3837 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3838 return kmalloc_large(size, gfpflags);
3840 s = kmalloc_slab(size, gfpflags);
3842 if (unlikely(ZERO_OR_NULL_PTR(s)))
3845 ret = slab_alloc(s, gfpflags, caller);
3847 /* Honor the call site pointer we received. */
3848 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3854 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3855 int node, unsigned long caller)
3857 struct kmem_cache *s;
3860 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3861 ret = kmalloc_large_node(size, gfpflags, node);
3863 trace_kmalloc_node(caller, ret,
3864 size, PAGE_SIZE << get_order(size),
3870 s = kmalloc_slab(size, gfpflags);
3872 if (unlikely(ZERO_OR_NULL_PTR(s)))
3875 ret = slab_alloc_node(s, gfpflags, node, caller);
3877 /* Honor the call site pointer we received. */
3878 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3885 static int count_inuse(struct page *page)
3890 static int count_total(struct page *page)
3892 return page->objects;
3896 #ifdef CONFIG_SLUB_DEBUG
3897 static int validate_slab(struct kmem_cache *s, struct page *page,
3901 void *addr = page_address(page);
3903 if (!check_slab(s, page) ||
3904 !on_freelist(s, page, NULL))
3907 /* Now we know that a valid freelist exists */
3908 bitmap_zero(map, page->objects);
3910 get_map(s, page, map);
3911 for_each_object(p, s, addr, page->objects) {
3912 if (test_bit(slab_index(p, s, addr), map))
3913 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3917 for_each_object(p, s, addr, page->objects)
3918 if (!test_bit(slab_index(p, s, addr), map))
3919 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3924 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3928 validate_slab(s, page, map);
3932 static int validate_slab_node(struct kmem_cache *s,
3933 struct kmem_cache_node *n, unsigned long *map)
3935 unsigned long count = 0;
3937 unsigned long flags;
3939 spin_lock_irqsave(&n->list_lock, flags);
3941 list_for_each_entry(page, &n->partial, lru) {
3942 validate_slab_slab(s, page, map);
3945 if (count != n->nr_partial)
3946 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3947 "counter=%ld\n", s->name, count, n->nr_partial);
3949 if (!(s->flags & SLAB_STORE_USER))
3952 list_for_each_entry(page, &n->full, lru) {
3953 validate_slab_slab(s, page, map);
3956 if (count != atomic_long_read(&n->nr_slabs))
3957 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3958 "counter=%ld\n", s->name, count,
3959 atomic_long_read(&n->nr_slabs));
3962 spin_unlock_irqrestore(&n->list_lock, flags);
3966 static long validate_slab_cache(struct kmem_cache *s)
3969 unsigned long count = 0;
3970 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3971 sizeof(unsigned long), GFP_KERNEL);
3977 for_each_node_state(node, N_NORMAL_MEMORY) {
3978 struct kmem_cache_node *n = get_node(s, node);
3980 count += validate_slab_node(s, n, map);
3986 * Generate lists of code addresses where slabcache objects are allocated
3991 unsigned long count;
3998 DECLARE_BITMAP(cpus, NR_CPUS);
4004 unsigned long count;
4005 struct location *loc;
4008 static void free_loc_track(struct loc_track *t)
4011 free_pages((unsigned long)t->loc,
4012 get_order(sizeof(struct location) * t->max));
4015 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4020 order = get_order(sizeof(struct location) * max);
4022 l = (void *)__get_free_pages(flags, order);
4027 memcpy(l, t->loc, sizeof(struct location) * t->count);
4035 static int add_location(struct loc_track *t, struct kmem_cache *s,
4036 const struct track *track)
4038 long start, end, pos;
4040 unsigned long caddr;
4041 unsigned long age = jiffies - track->when;
4047 pos = start + (end - start + 1) / 2;
4050 * There is nothing at "end". If we end up there
4051 * we need to add something to before end.
4056 caddr = t->loc[pos].addr;
4057 if (track->addr == caddr) {
4063 if (age < l->min_time)
4065 if (age > l->max_time)
4068 if (track->pid < l->min_pid)
4069 l->min_pid = track->pid;
4070 if (track->pid > l->max_pid)
4071 l->max_pid = track->pid;
4073 cpumask_set_cpu(track->cpu,
4074 to_cpumask(l->cpus));
4076 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4080 if (track->addr < caddr)
4087 * Not found. Insert new tracking element.
4089 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4095 (t->count - pos) * sizeof(struct location));
4098 l->addr = track->addr;
4102 l->min_pid = track->pid;
4103 l->max_pid = track->pid;
4104 cpumask_clear(to_cpumask(l->cpus));
4105 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4106 nodes_clear(l->nodes);
4107 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4111 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4112 struct page *page, enum track_item alloc,
4115 void *addr = page_address(page);
4118 bitmap_zero(map, page->objects);
4119 get_map(s, page, map);
4121 for_each_object(p, s, addr, page->objects)
4122 if (!test_bit(slab_index(p, s, addr), map))
4123 add_location(t, s, get_track(s, p, alloc));
4126 static int list_locations(struct kmem_cache *s, char *buf,
4127 enum track_item alloc)
4131 struct loc_track t = { 0, 0, NULL };
4133 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4134 sizeof(unsigned long), GFP_KERNEL);
4136 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4139 return sprintf(buf, "Out of memory\n");
4141 /* Push back cpu slabs */
4144 for_each_node_state(node, N_NORMAL_MEMORY) {
4145 struct kmem_cache_node *n = get_node(s, node);
4146 unsigned long flags;
4149 if (!atomic_long_read(&n->nr_slabs))
4152 spin_lock_irqsave(&n->list_lock, flags);
4153 list_for_each_entry(page, &n->partial, lru)
4154 process_slab(&t, s, page, alloc, map);
4155 list_for_each_entry(page, &n->full, lru)
4156 process_slab(&t, s, page, alloc, map);
4157 spin_unlock_irqrestore(&n->list_lock, flags);
4160 for (i = 0; i < t.count; i++) {
4161 struct location *l = &t.loc[i];
4163 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4165 len += sprintf(buf + len, "%7ld ", l->count);
4168 len += sprintf(buf + len, "%pS", (void *)l->addr);
4170 len += sprintf(buf + len, "<not-available>");
4172 if (l->sum_time != l->min_time) {
4173 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4175 (long)div_u64(l->sum_time, l->count),
4178 len += sprintf(buf + len, " age=%ld",
4181 if (l->min_pid != l->max_pid)
4182 len += sprintf(buf + len, " pid=%ld-%ld",
4183 l->min_pid, l->max_pid);
4185 len += sprintf(buf + len, " pid=%ld",
4188 if (num_online_cpus() > 1 &&
4189 !cpumask_empty(to_cpumask(l->cpus)) &&
4190 len < PAGE_SIZE - 60) {
4191 len += sprintf(buf + len, " cpus=");
4192 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4193 to_cpumask(l->cpus));
4196 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4197 len < PAGE_SIZE - 60) {
4198 len += sprintf(buf + len, " nodes=");
4199 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4203 len += sprintf(buf + len, "\n");
4209 len += sprintf(buf, "No data\n");
4214 #ifdef SLUB_RESILIENCY_TEST
4215 static void resiliency_test(void)
4219 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4221 printk(KERN_ERR "SLUB resiliency testing\n");
4222 printk(KERN_ERR "-----------------------\n");
4223 printk(KERN_ERR "A. Corruption after allocation\n");
4225 p = kzalloc(16, GFP_KERNEL);
4227 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4228 " 0x12->0x%p\n\n", p + 16);
4230 validate_slab_cache(kmalloc_caches[4]);
4232 /* Hmmm... The next two are dangerous */
4233 p = kzalloc(32, GFP_KERNEL);
4234 p[32 + sizeof(void *)] = 0x34;
4235 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4236 " 0x34 -> -0x%p\n", p);
4238 "If allocated object is overwritten then not detectable\n\n");
4240 validate_slab_cache(kmalloc_caches[5]);
4241 p = kzalloc(64, GFP_KERNEL);
4242 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4244 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4247 "If allocated object is overwritten then not detectable\n\n");
4248 validate_slab_cache(kmalloc_caches[6]);
4250 printk(KERN_ERR "\nB. Corruption after free\n");
4251 p = kzalloc(128, GFP_KERNEL);
4254 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4255 validate_slab_cache(kmalloc_caches[7]);
4257 p = kzalloc(256, GFP_KERNEL);
4260 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4262 validate_slab_cache(kmalloc_caches[8]);
4264 p = kzalloc(512, GFP_KERNEL);
4267 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4268 validate_slab_cache(kmalloc_caches[9]);
4272 static void resiliency_test(void) {};
4277 enum slab_stat_type {
4278 SL_ALL, /* All slabs */
4279 SL_PARTIAL, /* Only partially allocated slabs */
4280 SL_CPU, /* Only slabs used for cpu caches */
4281 SL_OBJECTS, /* Determine allocated objects not slabs */
4282 SL_TOTAL /* Determine object capacity not slabs */
4285 #define SO_ALL (1 << SL_ALL)
4286 #define SO_PARTIAL (1 << SL_PARTIAL)
4287 #define SO_CPU (1 << SL_CPU)
4288 #define SO_OBJECTS (1 << SL_OBJECTS)
4289 #define SO_TOTAL (1 << SL_TOTAL)
4291 static ssize_t show_slab_objects(struct kmem_cache *s,
4292 char *buf, unsigned long flags)
4294 unsigned long total = 0;
4297 unsigned long *nodes;
4298 unsigned long *per_cpu;
4300 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4303 per_cpu = nodes + nr_node_ids;
4305 if (flags & SO_CPU) {
4308 for_each_possible_cpu(cpu) {
4309 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4313 page = ACCESS_ONCE(c->page);
4317 node = page_to_nid(page);
4318 if (flags & SO_TOTAL)
4320 else if (flags & SO_OBJECTS)
4328 page = ACCESS_ONCE(c->partial);
4339 lock_memory_hotplug();
4340 #ifdef CONFIG_SLUB_DEBUG
4341 if (flags & SO_ALL) {
4342 for_each_node_state(node, N_NORMAL_MEMORY) {
4343 struct kmem_cache_node *n = get_node(s, node);
4345 if (flags & SO_TOTAL)
4346 x = atomic_long_read(&n->total_objects);
4347 else if (flags & SO_OBJECTS)
4348 x = atomic_long_read(&n->total_objects) -
4349 count_partial(n, count_free);
4352 x = atomic_long_read(&n->nr_slabs);
4359 if (flags & SO_PARTIAL) {
4360 for_each_node_state(node, N_NORMAL_MEMORY) {
4361 struct kmem_cache_node *n = get_node(s, node);
4363 if (flags & SO_TOTAL)
4364 x = count_partial(n, count_total);
4365 else if (flags & SO_OBJECTS)
4366 x = count_partial(n, count_inuse);
4373 x = sprintf(buf, "%lu", total);
4375 for_each_node_state(node, N_NORMAL_MEMORY)
4377 x += sprintf(buf + x, " N%d=%lu",
4380 unlock_memory_hotplug();
4382 return x + sprintf(buf + x, "\n");
4385 #ifdef CONFIG_SLUB_DEBUG
4386 static int any_slab_objects(struct kmem_cache *s)
4390 for_each_online_node(node) {
4391 struct kmem_cache_node *n = get_node(s, node);
4396 if (atomic_long_read(&n->total_objects))
4403 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4404 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4406 struct slab_attribute {
4407 struct attribute attr;
4408 ssize_t (*show)(struct kmem_cache *s, char *buf);
4409 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4412 #define SLAB_ATTR_RO(_name) \
4413 static struct slab_attribute _name##_attr = \
4414 __ATTR(_name, 0400, _name##_show, NULL)
4416 #define SLAB_ATTR(_name) \
4417 static struct slab_attribute _name##_attr = \
4418 __ATTR(_name, 0600, _name##_show, _name##_store)
4420 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4422 return sprintf(buf, "%d\n", s->size);
4424 SLAB_ATTR_RO(slab_size);
4426 static ssize_t align_show(struct kmem_cache *s, char *buf)
4428 return sprintf(buf, "%d\n", s->align);
4430 SLAB_ATTR_RO(align);
4432 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4434 return sprintf(buf, "%d\n", s->object_size);
4436 SLAB_ATTR_RO(object_size);
4438 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4440 return sprintf(buf, "%d\n", oo_objects(s->oo));
4442 SLAB_ATTR_RO(objs_per_slab);
4444 static ssize_t order_store(struct kmem_cache *s,
4445 const char *buf, size_t length)
4447 unsigned long order;
4450 err = strict_strtoul(buf, 10, &order);
4454 if (order > slub_max_order || order < slub_min_order)
4457 calculate_sizes(s, order);
4461 static ssize_t order_show(struct kmem_cache *s, char *buf)
4463 return sprintf(buf, "%d\n", oo_order(s->oo));
4467 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4469 return sprintf(buf, "%lu\n", s->min_partial);
4472 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4478 err = strict_strtoul(buf, 10, &min);
4482 set_min_partial(s, min);
4485 SLAB_ATTR(min_partial);
4487 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4489 return sprintf(buf, "%u\n", s->cpu_partial);
4492 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4495 unsigned long objects;
4498 err = strict_strtoul(buf, 10, &objects);
4501 if (objects && !kmem_cache_has_cpu_partial(s))
4504 s->cpu_partial = objects;
4508 SLAB_ATTR(cpu_partial);
4510 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4514 return sprintf(buf, "%pS\n", s->ctor);
4518 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4520 return sprintf(buf, "%d\n", s->refcount - 1);
4522 SLAB_ATTR_RO(aliases);
4524 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4526 return show_slab_objects(s, buf, SO_PARTIAL);
4528 SLAB_ATTR_RO(partial);
4530 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4532 return show_slab_objects(s, buf, SO_CPU);
4534 SLAB_ATTR_RO(cpu_slabs);
4536 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4538 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4540 SLAB_ATTR_RO(objects);
4542 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4544 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4546 SLAB_ATTR_RO(objects_partial);
4548 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4555 for_each_online_cpu(cpu) {
4556 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4559 pages += page->pages;
4560 objects += page->pobjects;
4564 len = sprintf(buf, "%d(%d)", objects, pages);
4567 for_each_online_cpu(cpu) {
4568 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4570 if (page && len < PAGE_SIZE - 20)
4571 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4572 page->pobjects, page->pages);
4575 return len + sprintf(buf + len, "\n");
4577 SLAB_ATTR_RO(slabs_cpu_partial);
4579 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4581 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4584 static ssize_t reclaim_account_store(struct kmem_cache *s,
4585 const char *buf, size_t length)
4587 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4589 s->flags |= SLAB_RECLAIM_ACCOUNT;
4592 SLAB_ATTR(reclaim_account);
4594 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4596 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4598 SLAB_ATTR_RO(hwcache_align);
4600 #ifdef CONFIG_ZONE_DMA
4601 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4603 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4605 SLAB_ATTR_RO(cache_dma);
4608 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4610 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4612 SLAB_ATTR_RO(destroy_by_rcu);
4614 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4616 return sprintf(buf, "%d\n", s->reserved);
4618 SLAB_ATTR_RO(reserved);
4620 #ifdef CONFIG_SLUB_DEBUG
4621 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4623 return show_slab_objects(s, buf, SO_ALL);
4625 SLAB_ATTR_RO(slabs);
4627 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4629 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4631 SLAB_ATTR_RO(total_objects);
4633 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4635 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4638 static ssize_t sanity_checks_store(struct kmem_cache *s,
4639 const char *buf, size_t length)
4641 s->flags &= ~SLAB_DEBUG_FREE;
4642 if (buf[0] == '1') {
4643 s->flags &= ~__CMPXCHG_DOUBLE;
4644 s->flags |= SLAB_DEBUG_FREE;
4648 SLAB_ATTR(sanity_checks);
4650 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4652 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4655 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4658 s->flags &= ~SLAB_TRACE;
4659 if (buf[0] == '1') {
4660 s->flags &= ~__CMPXCHG_DOUBLE;
4661 s->flags |= SLAB_TRACE;
4667 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4669 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4672 static ssize_t red_zone_store(struct kmem_cache *s,
4673 const char *buf, size_t length)
4675 if (any_slab_objects(s))
4678 s->flags &= ~SLAB_RED_ZONE;
4679 if (buf[0] == '1') {
4680 s->flags &= ~__CMPXCHG_DOUBLE;
4681 s->flags |= SLAB_RED_ZONE;
4683 calculate_sizes(s, -1);
4686 SLAB_ATTR(red_zone);
4688 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4690 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4693 static ssize_t poison_store(struct kmem_cache *s,
4694 const char *buf, size_t length)
4696 if (any_slab_objects(s))
4699 s->flags &= ~SLAB_POISON;
4700 if (buf[0] == '1') {
4701 s->flags &= ~__CMPXCHG_DOUBLE;
4702 s->flags |= SLAB_POISON;
4704 calculate_sizes(s, -1);
4709 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4711 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4714 static ssize_t store_user_store(struct kmem_cache *s,
4715 const char *buf, size_t length)
4717 if (any_slab_objects(s))
4720 s->flags &= ~SLAB_STORE_USER;
4721 if (buf[0] == '1') {
4722 s->flags &= ~__CMPXCHG_DOUBLE;
4723 s->flags |= SLAB_STORE_USER;
4725 calculate_sizes(s, -1);
4728 SLAB_ATTR(store_user);
4730 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4735 static ssize_t validate_store(struct kmem_cache *s,
4736 const char *buf, size_t length)
4740 if (buf[0] == '1') {
4741 ret = validate_slab_cache(s);
4747 SLAB_ATTR(validate);
4749 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4751 if (!(s->flags & SLAB_STORE_USER))
4753 return list_locations(s, buf, TRACK_ALLOC);
4755 SLAB_ATTR_RO(alloc_calls);
4757 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4759 if (!(s->flags & SLAB_STORE_USER))
4761 return list_locations(s, buf, TRACK_FREE);
4763 SLAB_ATTR_RO(free_calls);
4764 #endif /* CONFIG_SLUB_DEBUG */
4766 #ifdef CONFIG_FAILSLAB
4767 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4769 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4772 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4775 s->flags &= ~SLAB_FAILSLAB;
4777 s->flags |= SLAB_FAILSLAB;
4780 SLAB_ATTR(failslab);
4783 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4788 static ssize_t shrink_store(struct kmem_cache *s,
4789 const char *buf, size_t length)
4791 if (buf[0] == '1') {
4792 int rc = kmem_cache_shrink(s);
4803 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4805 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4808 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4809 const char *buf, size_t length)
4811 unsigned long ratio;
4814 err = strict_strtoul(buf, 10, &ratio);
4819 s->remote_node_defrag_ratio = ratio * 10;
4823 SLAB_ATTR(remote_node_defrag_ratio);
4826 #ifdef CONFIG_SLUB_STATS
4827 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4829 unsigned long sum = 0;
4832 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4837 for_each_online_cpu(cpu) {
4838 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4844 len = sprintf(buf, "%lu", sum);
4847 for_each_online_cpu(cpu) {
4848 if (data[cpu] && len < PAGE_SIZE - 20)
4849 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4853 return len + sprintf(buf + len, "\n");
4856 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4860 for_each_online_cpu(cpu)
4861 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4864 #define STAT_ATTR(si, text) \
4865 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4867 return show_stat(s, buf, si); \
4869 static ssize_t text##_store(struct kmem_cache *s, \
4870 const char *buf, size_t length) \
4872 if (buf[0] != '0') \
4874 clear_stat(s, si); \
4879 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4880 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4881 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4882 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4883 STAT_ATTR(FREE_FROZEN, free_frozen);
4884 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4885 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4886 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4887 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4888 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4889 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4890 STAT_ATTR(FREE_SLAB, free_slab);
4891 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4892 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4893 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4894 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4895 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4896 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4897 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4898 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4899 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4900 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4901 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4902 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4903 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4904 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4907 static struct attribute *slab_attrs[] = {
4908 &slab_size_attr.attr,
4909 &object_size_attr.attr,
4910 &objs_per_slab_attr.attr,
4912 &min_partial_attr.attr,
4913 &cpu_partial_attr.attr,
4915 &objects_partial_attr.attr,
4917 &cpu_slabs_attr.attr,
4921 &hwcache_align_attr.attr,
4922 &reclaim_account_attr.attr,
4923 &destroy_by_rcu_attr.attr,
4925 &reserved_attr.attr,
4926 &slabs_cpu_partial_attr.attr,
4927 #ifdef CONFIG_SLUB_DEBUG
4928 &total_objects_attr.attr,
4930 &sanity_checks_attr.attr,
4932 &red_zone_attr.attr,
4934 &store_user_attr.attr,
4935 &validate_attr.attr,
4936 &alloc_calls_attr.attr,
4937 &free_calls_attr.attr,
4939 #ifdef CONFIG_ZONE_DMA
4940 &cache_dma_attr.attr,
4943 &remote_node_defrag_ratio_attr.attr,
4945 #ifdef CONFIG_SLUB_STATS
4946 &alloc_fastpath_attr.attr,
4947 &alloc_slowpath_attr.attr,
4948 &free_fastpath_attr.attr,
4949 &free_slowpath_attr.attr,
4950 &free_frozen_attr.attr,
4951 &free_add_partial_attr.attr,
4952 &free_remove_partial_attr.attr,
4953 &alloc_from_partial_attr.attr,
4954 &alloc_slab_attr.attr,
4955 &alloc_refill_attr.attr,
4956 &alloc_node_mismatch_attr.attr,
4957 &free_slab_attr.attr,
4958 &cpuslab_flush_attr.attr,
4959 &deactivate_full_attr.attr,
4960 &deactivate_empty_attr.attr,
4961 &deactivate_to_head_attr.attr,
4962 &deactivate_to_tail_attr.attr,
4963 &deactivate_remote_frees_attr.attr,
4964 &deactivate_bypass_attr.attr,
4965 &order_fallback_attr.attr,
4966 &cmpxchg_double_fail_attr.attr,
4967 &cmpxchg_double_cpu_fail_attr.attr,
4968 &cpu_partial_alloc_attr.attr,
4969 &cpu_partial_free_attr.attr,
4970 &cpu_partial_node_attr.attr,
4971 &cpu_partial_drain_attr.attr,
4973 #ifdef CONFIG_FAILSLAB
4974 &failslab_attr.attr,
4980 static struct attribute_group slab_attr_group = {
4981 .attrs = slab_attrs,
4984 static ssize_t slab_attr_show(struct kobject *kobj,
4985 struct attribute *attr,
4988 struct slab_attribute *attribute;
4989 struct kmem_cache *s;
4992 attribute = to_slab_attr(attr);
4995 if (!attribute->show)
4998 err = attribute->show(s, buf);
5003 static ssize_t slab_attr_store(struct kobject *kobj,
5004 struct attribute *attr,
5005 const char *buf, size_t len)
5007 struct slab_attribute *attribute;
5008 struct kmem_cache *s;
5011 attribute = to_slab_attr(attr);
5014 if (!attribute->store)
5017 err = attribute->store(s, buf, len);
5018 #ifdef CONFIG_MEMCG_KMEM
5019 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5022 mutex_lock(&slab_mutex);
5023 if (s->max_attr_size < len)
5024 s->max_attr_size = len;
5027 * This is a best effort propagation, so this function's return
5028 * value will be determined by the parent cache only. This is
5029 * basically because not all attributes will have a well
5030 * defined semantics for rollbacks - most of the actions will
5031 * have permanent effects.
5033 * Returning the error value of any of the children that fail
5034 * is not 100 % defined, in the sense that users seeing the
5035 * error code won't be able to know anything about the state of
5038 * Only returning the error code for the parent cache at least
5039 * has well defined semantics. The cache being written to
5040 * directly either failed or succeeded, in which case we loop
5041 * through the descendants with best-effort propagation.
5043 for_each_memcg_cache_index(i) {
5044 struct kmem_cache *c = cache_from_memcg(s, i);
5046 attribute->store(c, buf, len);
5048 mutex_unlock(&slab_mutex);
5054 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5056 #ifdef CONFIG_MEMCG_KMEM
5058 char *buffer = NULL;
5060 if (!is_root_cache(s))
5064 * This mean this cache had no attribute written. Therefore, no point
5065 * in copying default values around
5067 if (!s->max_attr_size)
5070 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5073 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5075 if (!attr || !attr->store || !attr->show)
5079 * It is really bad that we have to allocate here, so we will
5080 * do it only as a fallback. If we actually allocate, though,
5081 * we can just use the allocated buffer until the end.
5083 * Most of the slub attributes will tend to be very small in
5084 * size, but sysfs allows buffers up to a page, so they can
5085 * theoretically happen.
5089 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5092 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5093 if (WARN_ON(!buffer))
5098 attr->show(s->memcg_params->root_cache, buf);
5099 attr->store(s, buf, strlen(buf));
5103 free_page((unsigned long)buffer);
5107 static const struct sysfs_ops slab_sysfs_ops = {
5108 .show = slab_attr_show,
5109 .store = slab_attr_store,
5112 static struct kobj_type slab_ktype = {
5113 .sysfs_ops = &slab_sysfs_ops,
5116 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5118 struct kobj_type *ktype = get_ktype(kobj);
5120 if (ktype == &slab_ktype)
5125 static const struct kset_uevent_ops slab_uevent_ops = {
5126 .filter = uevent_filter,
5129 static struct kset *slab_kset;
5131 #define ID_STR_LENGTH 64
5133 /* Create a unique string id for a slab cache:
5135 * Format :[flags-]size
5137 static char *create_unique_id(struct kmem_cache *s)
5139 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5146 * First flags affecting slabcache operations. We will only
5147 * get here for aliasable slabs so we do not need to support
5148 * too many flags. The flags here must cover all flags that
5149 * are matched during merging to guarantee that the id is
5152 if (s->flags & SLAB_CACHE_DMA)
5154 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5156 if (s->flags & SLAB_DEBUG_FREE)
5158 if (!(s->flags & SLAB_NOTRACK))
5162 p += sprintf(p, "%07d", s->size);
5164 #ifdef CONFIG_MEMCG_KMEM
5165 if (!is_root_cache(s))
5166 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5169 BUG_ON(p > name + ID_STR_LENGTH - 1);
5173 static int sysfs_slab_add(struct kmem_cache *s)
5177 int unmergeable = slab_unmergeable(s);
5181 * Slabcache can never be merged so we can use the name proper.
5182 * This is typically the case for debug situations. In that
5183 * case we can catch duplicate names easily.
5185 sysfs_remove_link(&slab_kset->kobj, s->name);
5189 * Create a unique name for the slab as a target
5192 name = create_unique_id(s);
5195 s->kobj.kset = slab_kset;
5196 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5198 kobject_put(&s->kobj);
5202 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5204 kobject_del(&s->kobj);
5205 kobject_put(&s->kobj);
5208 kobject_uevent(&s->kobj, KOBJ_ADD);
5210 /* Setup first alias */
5211 sysfs_slab_alias(s, s->name);
5217 static void sysfs_slab_remove(struct kmem_cache *s)
5219 if (slab_state < FULL)
5221 * Sysfs has not been setup yet so no need to remove the
5226 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5227 kobject_del(&s->kobj);
5228 kobject_put(&s->kobj);
5232 * Need to buffer aliases during bootup until sysfs becomes
5233 * available lest we lose that information.
5235 struct saved_alias {
5236 struct kmem_cache *s;
5238 struct saved_alias *next;
5241 static struct saved_alias *alias_list;
5243 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5245 struct saved_alias *al;
5247 if (slab_state == FULL) {
5249 * If we have a leftover link then remove it.
5251 sysfs_remove_link(&slab_kset->kobj, name);
5252 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5255 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5261 al->next = alias_list;
5266 static int __init slab_sysfs_init(void)
5268 struct kmem_cache *s;
5271 mutex_lock(&slab_mutex);
5273 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5275 mutex_unlock(&slab_mutex);
5276 printk(KERN_ERR "Cannot register slab subsystem.\n");
5282 list_for_each_entry(s, &slab_caches, list) {
5283 err = sysfs_slab_add(s);
5285 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5286 " to sysfs\n", s->name);
5289 while (alias_list) {
5290 struct saved_alias *al = alias_list;
5292 alias_list = alias_list->next;
5293 err = sysfs_slab_alias(al->s, al->name);
5295 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5296 " %s to sysfs\n", al->name);
5300 mutex_unlock(&slab_mutex);
5305 __initcall(slab_sysfs_init);
5306 #endif /* CONFIG_SYSFS */
5309 * The /proc/slabinfo ABI
5311 #ifdef CONFIG_SLABINFO
5312 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5314 unsigned long nr_slabs = 0;
5315 unsigned long nr_objs = 0;
5316 unsigned long nr_free = 0;
5319 for_each_online_node(node) {
5320 struct kmem_cache_node *n = get_node(s, node);
5325 nr_slabs += node_nr_slabs(n);
5326 nr_objs += node_nr_objs(n);
5327 nr_free += count_partial(n, count_free);
5330 sinfo->active_objs = nr_objs - nr_free;
5331 sinfo->num_objs = nr_objs;
5332 sinfo->active_slabs = nr_slabs;
5333 sinfo->num_slabs = nr_slabs;
5334 sinfo->objects_per_slab = oo_objects(s->oo);
5335 sinfo->cache_order = oo_order(s->oo);
5338 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5342 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5343 size_t count, loff_t *ppos)
5347 #endif /* CONFIG_SLABINFO */