2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/seq_file.h>
22 #include <linux/kmemcheck.h>
23 #include <linux/cpu.h>
24 #include <linux/cpuset.h>
25 #include <linux/mempolicy.h>
26 #include <linux/ctype.h>
27 #include <linux/debugobjects.h>
28 #include <linux/kallsyms.h>
29 #include <linux/memory.h>
30 #include <linux/math64.h>
31 #include <linux/fault-inject.h>
32 #include <linux/stacktrace.h>
33 #include <linux/prefetch.h>
34 #include <linux/memcontrol.h>
36 #include <trace/events/kmem.h>
42 * 1. slab_mutex (Global Mutex)
44 * 3. slab_lock(page) (Only on some arches and for debugging)
48 * The role of the slab_mutex is to protect the list of all the slabs
49 * and to synchronize major metadata changes to slab cache structures.
51 * The slab_lock is only used for debugging and on arches that do not
52 * have the ability to do a cmpxchg_double. It only protects the second
53 * double word in the page struct. Meaning
54 * A. page->freelist -> List of object free in a page
55 * B. page->counters -> Counters of objects
56 * C. page->frozen -> frozen state
58 * If a slab is frozen then it is exempt from list management. It is not
59 * on any list. The processor that froze the slab is the one who can
60 * perform list operations on the page. Other processors may put objects
61 * onto the freelist but the processor that froze the slab is the only
62 * one that can retrieve the objects from the page's freelist.
64 * The list_lock protects the partial and full list on each node and
65 * the partial slab counter. If taken then no new slabs may be added or
66 * removed from the lists nor make the number of partial slabs be modified.
67 * (Note that the total number of slabs is an atomic value that may be
68 * modified without taking the list lock).
70 * The list_lock is a centralized lock and thus we avoid taking it as
71 * much as possible. As long as SLUB does not have to handle partial
72 * slabs, operations can continue without any centralized lock. F.e.
73 * allocating a long series of objects that fill up slabs does not require
75 * Interrupts are disabled during allocation and deallocation in order to
76 * make the slab allocator safe to use in the context of an irq. In addition
77 * interrupts are disabled to ensure that the processor does not change
78 * while handling per_cpu slabs, due to kernel preemption.
80 * SLUB assigns one slab for allocation to each processor.
81 * Allocations only occur from these slabs called cpu slabs.
83 * Slabs with free elements are kept on a partial list and during regular
84 * operations no list for full slabs is used. If an object in a full slab is
85 * freed then the slab will show up again on the partial lists.
86 * We track full slabs for debugging purposes though because otherwise we
87 * cannot scan all objects.
89 * Slabs are freed when they become empty. Teardown and setup is
90 * minimal so we rely on the page allocators per cpu caches for
91 * fast frees and allocs.
93 * Overloading of page flags that are otherwise used for LRU management.
95 * PageActive The slab is frozen and exempt from list processing.
96 * This means that the slab is dedicated to a purpose
97 * such as satisfying allocations for a specific
98 * processor. Objects may be freed in the slab while
99 * it is frozen but slab_free will then skip the usual
100 * list operations. It is up to the processor holding
101 * the slab to integrate the slab into the slab lists
102 * when the slab is no longer needed.
104 * One use of this flag is to mark slabs that are
105 * used for allocations. Then such a slab becomes a cpu
106 * slab. The cpu slab may be equipped with an additional
107 * freelist that allows lockless access to
108 * free objects in addition to the regular freelist
109 * that requires the slab lock.
111 * PageError Slab requires special handling due to debug
112 * options set. This moves slab handling out of
113 * the fast path and disables lockless freelists.
116 static inline int kmem_cache_debug(struct kmem_cache *s)
118 #ifdef CONFIG_SLUB_DEBUG
119 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 * Issues still to be resolved:
128 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
130 * - Variable sizing of the per node arrays
133 /* Enable to test recovery from slab corruption on boot */
134 #undef SLUB_RESILIENCY_TEST
136 /* Enable to log cmpxchg failures */
137 #undef SLUB_DEBUG_CMPXCHG
140 * Mininum number of partial slabs. These will be left on the partial
141 * lists even if they are empty. kmem_cache_shrink may reclaim them.
143 #define MIN_PARTIAL 5
146 * Maximum number of desirable partial slabs.
147 * The existence of more partial slabs makes kmem_cache_shrink
148 * sort the partial list by the number of objects in the.
150 #define MAX_PARTIAL 10
152 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
153 SLAB_POISON | SLAB_STORE_USER)
156 * Debugging flags that require metadata to be stored in the slab. These get
157 * disabled when slub_debug=O is used and a cache's min order increases with
160 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
163 * Set of flags that will prevent slab merging
165 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
166 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
169 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
170 SLAB_CACHE_DMA | SLAB_NOTRACK)
173 #define OO_MASK ((1 << OO_SHIFT) - 1)
174 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
176 /* Internal SLUB flags */
177 #define __OBJECT_POISON 0x80000000UL /* Poison object */
178 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
181 static struct notifier_block slab_notifier;
185 * Tracking user of a slab.
187 #define TRACK_ADDRS_COUNT 16
189 unsigned long addr; /* Called from address */
190 #ifdef CONFIG_STACKTRACE
191 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
193 int cpu; /* Was running on cpu */
194 int pid; /* Pid context */
195 unsigned long when; /* When did the operation occur */
198 enum track_item { TRACK_ALLOC, TRACK_FREE };
201 static int sysfs_slab_add(struct kmem_cache *);
202 static int sysfs_slab_alias(struct kmem_cache *, const char *);
203 static void sysfs_slab_remove(struct kmem_cache *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
211 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
214 static inline void stat(const struct kmem_cache *s, enum stat_item si)
216 #ifdef CONFIG_SLUB_STATS
217 __this_cpu_inc(s->cpu_slab->stat[si]);
221 /********************************************************************
222 * Core slab cache functions
223 *******************************************************************/
225 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
227 return s->node[node];
230 /* Verify that a pointer has an address that is valid within a slab page */
231 static inline int check_valid_pointer(struct kmem_cache *s,
232 struct page *page, const void *object)
239 base = page_address(page);
240 if (object < base || object >= base + page->objects * s->size ||
241 (object - base) % s->size) {
248 static inline void *get_freepointer(struct kmem_cache *s, void *object)
250 return *(void **)(object + s->offset);
253 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
255 prefetch(object + s->offset);
258 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
262 #ifdef CONFIG_DEBUG_PAGEALLOC
263 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
265 p = get_freepointer(s, object);
270 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
272 *(void **)(object + s->offset) = fp;
275 /* Loop over all objects in a slab */
276 #define for_each_object(__p, __s, __addr, __objects) \
277 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
280 /* Determine object index from a given position */
281 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
283 return (p - addr) / s->size;
286 static inline size_t slab_ksize(const struct kmem_cache *s)
288 #ifdef CONFIG_SLUB_DEBUG
290 * Debugging requires use of the padding between object
291 * and whatever may come after it.
293 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
294 return s->object_size;
298 * If we have the need to store the freelist pointer
299 * back there or track user information then we can
300 * only use the space before that information.
302 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
305 * Else we can use all the padding etc for the allocation
310 static inline int order_objects(int order, unsigned long size, int reserved)
312 return ((PAGE_SIZE << order) - reserved) / size;
315 static inline struct kmem_cache_order_objects oo_make(int order,
316 unsigned long size, int reserved)
318 struct kmem_cache_order_objects x = {
319 (order << OO_SHIFT) + order_objects(order, size, reserved)
325 static inline int oo_order(struct kmem_cache_order_objects x)
327 return x.x >> OO_SHIFT;
330 static inline int oo_objects(struct kmem_cache_order_objects x)
332 return x.x & OO_MASK;
336 * Per slab locking using the pagelock
338 static __always_inline void slab_lock(struct page *page)
340 bit_spin_lock(PG_locked, &page->flags);
343 static __always_inline void slab_unlock(struct page *page)
345 __bit_spin_unlock(PG_locked, &page->flags);
348 /* Interrupts must be disabled (for the fallback code to work right) */
349 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
350 void *freelist_old, unsigned long counters_old,
351 void *freelist_new, unsigned long counters_new,
354 VM_BUG_ON(!irqs_disabled());
355 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
356 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
357 if (s->flags & __CMPXCHG_DOUBLE) {
358 if (cmpxchg_double(&page->freelist, &page->counters,
359 freelist_old, counters_old,
360 freelist_new, counters_new))
366 if (page->freelist == freelist_old && page->counters == counters_old) {
367 page->freelist = freelist_new;
368 page->counters = counters_new;
376 stat(s, CMPXCHG_DOUBLE_FAIL);
378 #ifdef SLUB_DEBUG_CMPXCHG
379 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
385 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
386 void *freelist_old, unsigned long counters_old,
387 void *freelist_new, unsigned long counters_new,
390 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
391 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
392 if (s->flags & __CMPXCHG_DOUBLE) {
393 if (cmpxchg_double(&page->freelist, &page->counters,
394 freelist_old, counters_old,
395 freelist_new, counters_new))
402 local_irq_save(flags);
404 if (page->freelist == freelist_old && page->counters == counters_old) {
405 page->freelist = freelist_new;
406 page->counters = counters_new;
408 local_irq_restore(flags);
412 local_irq_restore(flags);
416 stat(s, CMPXCHG_DOUBLE_FAIL);
418 #ifdef SLUB_DEBUG_CMPXCHG
419 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
425 #ifdef CONFIG_SLUB_DEBUG
427 * Determine a map of object in use on a page.
429 * Node listlock must be held to guarantee that the page does
430 * not vanish from under us.
432 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
435 void *addr = page_address(page);
437 for (p = page->freelist; p; p = get_freepointer(s, p))
438 set_bit(slab_index(p, s, addr), map);
444 #ifdef CONFIG_SLUB_DEBUG_ON
445 static int slub_debug = DEBUG_DEFAULT_FLAGS;
447 static int slub_debug;
450 static char *slub_debug_slabs;
451 static int disable_higher_order_debug;
456 static void print_section(char *text, u8 *addr, unsigned int length)
458 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
462 static struct track *get_track(struct kmem_cache *s, void *object,
463 enum track_item alloc)
468 p = object + s->offset + sizeof(void *);
470 p = object + s->inuse;
475 static void set_track(struct kmem_cache *s, void *object,
476 enum track_item alloc, unsigned long addr)
478 struct track *p = get_track(s, object, alloc);
481 #ifdef CONFIG_STACKTRACE
482 struct stack_trace trace;
485 trace.nr_entries = 0;
486 trace.max_entries = TRACK_ADDRS_COUNT;
487 trace.entries = p->addrs;
489 save_stack_trace(&trace);
491 /* See rant in lockdep.c */
492 if (trace.nr_entries != 0 &&
493 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
496 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
500 p->cpu = smp_processor_id();
501 p->pid = current->pid;
504 memset(p, 0, sizeof(struct track));
507 static void init_tracking(struct kmem_cache *s, void *object)
509 if (!(s->flags & SLAB_STORE_USER))
512 set_track(s, object, TRACK_FREE, 0UL);
513 set_track(s, object, TRACK_ALLOC, 0UL);
516 static void print_track(const char *s, struct track *t)
521 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
522 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
523 #ifdef CONFIG_STACKTRACE
526 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
528 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
535 static void print_tracking(struct kmem_cache *s, void *object)
537 if (!(s->flags & SLAB_STORE_USER))
540 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
541 print_track("Freed", get_track(s, object, TRACK_FREE));
544 static void print_page_info(struct page *page)
546 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
547 page, page->objects, page->inuse, page->freelist, page->flags);
551 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
557 vsnprintf(buf, sizeof(buf), fmt, args);
559 printk(KERN_ERR "========================================"
560 "=====================================\n");
561 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
562 printk(KERN_ERR "----------------------------------------"
563 "-------------------------------------\n\n");
565 add_taint(TAINT_BAD_PAGE);
568 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
574 vsnprintf(buf, sizeof(buf), fmt, args);
576 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
579 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
581 unsigned int off; /* Offset of last byte */
582 u8 *addr = page_address(page);
584 print_tracking(s, p);
586 print_page_info(page);
588 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
589 p, p - addr, get_freepointer(s, p));
592 print_section("Bytes b4 ", p - 16, 16);
594 print_section("Object ", p, min_t(unsigned long, s->object_size,
596 if (s->flags & SLAB_RED_ZONE)
597 print_section("Redzone ", p + s->object_size,
598 s->inuse - s->object_size);
601 off = s->offset + sizeof(void *);
605 if (s->flags & SLAB_STORE_USER)
606 off += 2 * sizeof(struct track);
609 /* Beginning of the filler is the free pointer */
610 print_section("Padding ", p + off, s->size - off);
615 static void object_err(struct kmem_cache *s, struct page *page,
616 u8 *object, char *reason)
618 slab_bug(s, "%s", reason);
619 print_trailer(s, page, object);
622 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
628 vsnprintf(buf, sizeof(buf), fmt, args);
630 slab_bug(s, "%s", buf);
631 print_page_info(page);
635 static void init_object(struct kmem_cache *s, void *object, u8 val)
639 if (s->flags & __OBJECT_POISON) {
640 memset(p, POISON_FREE, s->object_size - 1);
641 p[s->object_size - 1] = POISON_END;
644 if (s->flags & SLAB_RED_ZONE)
645 memset(p + s->object_size, val, s->inuse - s->object_size);
648 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
649 void *from, void *to)
651 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
652 memset(from, data, to - from);
655 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
656 u8 *object, char *what,
657 u8 *start, unsigned int value, unsigned int bytes)
662 fault = memchr_inv(start, value, bytes);
667 while (end > fault && end[-1] == value)
670 slab_bug(s, "%s overwritten", what);
671 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
672 fault, end - 1, fault[0], value);
673 print_trailer(s, page, object);
675 restore_bytes(s, what, value, fault, end);
683 * Bytes of the object to be managed.
684 * If the freepointer may overlay the object then the free
685 * pointer is the first word of the object.
687 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
690 * object + s->object_size
691 * Padding to reach word boundary. This is also used for Redzoning.
692 * Padding is extended by another word if Redzoning is enabled and
693 * object_size == inuse.
695 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
696 * 0xcc (RED_ACTIVE) for objects in use.
699 * Meta data starts here.
701 * A. Free pointer (if we cannot overwrite object on free)
702 * B. Tracking data for SLAB_STORE_USER
703 * C. Padding to reach required alignment boundary or at mininum
704 * one word if debugging is on to be able to detect writes
705 * before the word boundary.
707 * Padding is done using 0x5a (POISON_INUSE)
710 * Nothing is used beyond s->size.
712 * If slabcaches are merged then the object_size and inuse boundaries are mostly
713 * ignored. And therefore no slab options that rely on these boundaries
714 * may be used with merged slabcaches.
717 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
719 unsigned long off = s->inuse; /* The end of info */
722 /* Freepointer is placed after the object. */
723 off += sizeof(void *);
725 if (s->flags & SLAB_STORE_USER)
726 /* We also have user information there */
727 off += 2 * sizeof(struct track);
732 return check_bytes_and_report(s, page, p, "Object padding",
733 p + off, POISON_INUSE, s->size - off);
736 /* Check the pad bytes at the end of a slab page */
737 static int slab_pad_check(struct kmem_cache *s, struct page *page)
745 if (!(s->flags & SLAB_POISON))
748 start = page_address(page);
749 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
750 end = start + length;
751 remainder = length % s->size;
755 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
758 while (end > fault && end[-1] == POISON_INUSE)
761 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
762 print_section("Padding ", end - remainder, remainder);
764 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
768 static int check_object(struct kmem_cache *s, struct page *page,
769 void *object, u8 val)
772 u8 *endobject = object + s->object_size;
774 if (s->flags & SLAB_RED_ZONE) {
775 if (!check_bytes_and_report(s, page, object, "Redzone",
776 endobject, val, s->inuse - s->object_size))
779 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
780 check_bytes_and_report(s, page, p, "Alignment padding",
781 endobject, POISON_INUSE, s->inuse - s->object_size);
785 if (s->flags & SLAB_POISON) {
786 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
787 (!check_bytes_and_report(s, page, p, "Poison", p,
788 POISON_FREE, s->object_size - 1) ||
789 !check_bytes_and_report(s, page, p, "Poison",
790 p + s->object_size - 1, POISON_END, 1)))
793 * check_pad_bytes cleans up on its own.
795 check_pad_bytes(s, page, p);
798 if (!s->offset && val == SLUB_RED_ACTIVE)
800 * Object and freepointer overlap. Cannot check
801 * freepointer while object is allocated.
805 /* Check free pointer validity */
806 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
807 object_err(s, page, p, "Freepointer corrupt");
809 * No choice but to zap it and thus lose the remainder
810 * of the free objects in this slab. May cause
811 * another error because the object count is now wrong.
813 set_freepointer(s, p, NULL);
819 static int check_slab(struct kmem_cache *s, struct page *page)
823 VM_BUG_ON(!irqs_disabled());
825 if (!PageSlab(page)) {
826 slab_err(s, page, "Not a valid slab page");
830 maxobj = order_objects(compound_order(page), s->size, s->reserved);
831 if (page->objects > maxobj) {
832 slab_err(s, page, "objects %u > max %u",
833 s->name, page->objects, maxobj);
836 if (page->inuse > page->objects) {
837 slab_err(s, page, "inuse %u > max %u",
838 s->name, page->inuse, page->objects);
841 /* Slab_pad_check fixes things up after itself */
842 slab_pad_check(s, page);
847 * Determine if a certain object on a page is on the freelist. Must hold the
848 * slab lock to guarantee that the chains are in a consistent state.
850 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
855 unsigned long max_objects;
858 while (fp && nr <= page->objects) {
861 if (!check_valid_pointer(s, page, fp)) {
863 object_err(s, page, object,
864 "Freechain corrupt");
865 set_freepointer(s, object, NULL);
868 slab_err(s, page, "Freepointer corrupt");
869 page->freelist = NULL;
870 page->inuse = page->objects;
871 slab_fix(s, "Freelist cleared");
877 fp = get_freepointer(s, object);
881 max_objects = order_objects(compound_order(page), s->size, s->reserved);
882 if (max_objects > MAX_OBJS_PER_PAGE)
883 max_objects = MAX_OBJS_PER_PAGE;
885 if (page->objects != max_objects) {
886 slab_err(s, page, "Wrong number of objects. Found %d but "
887 "should be %d", page->objects, max_objects);
888 page->objects = max_objects;
889 slab_fix(s, "Number of objects adjusted.");
891 if (page->inuse != page->objects - nr) {
892 slab_err(s, page, "Wrong object count. Counter is %d but "
893 "counted were %d", page->inuse, page->objects - nr);
894 page->inuse = page->objects - nr;
895 slab_fix(s, "Object count adjusted.");
897 return search == NULL;
900 static void trace(struct kmem_cache *s, struct page *page, void *object,
903 if (s->flags & SLAB_TRACE) {
904 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
906 alloc ? "alloc" : "free",
911 print_section("Object ", (void *)object, s->object_size);
918 * Hooks for other subsystems that check memory allocations. In a typical
919 * production configuration these hooks all should produce no code at all.
921 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
923 flags &= gfp_allowed_mask;
924 lockdep_trace_alloc(flags);
925 might_sleep_if(flags & __GFP_WAIT);
927 return should_failslab(s->object_size, flags, s->flags);
930 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
932 flags &= gfp_allowed_mask;
933 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
934 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
937 static inline void slab_free_hook(struct kmem_cache *s, void *x)
939 kmemleak_free_recursive(x, s->flags);
942 * Trouble is that we may no longer disable interupts in the fast path
943 * So in order to make the debug calls that expect irqs to be
944 * disabled we need to disable interrupts temporarily.
946 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
950 local_irq_save(flags);
951 kmemcheck_slab_free(s, x, s->object_size);
952 debug_check_no_locks_freed(x, s->object_size);
953 local_irq_restore(flags);
956 if (!(s->flags & SLAB_DEBUG_OBJECTS))
957 debug_check_no_obj_freed(x, s->object_size);
961 * Tracking of fully allocated slabs for debugging purposes.
963 * list_lock must be held.
965 static void add_full(struct kmem_cache *s,
966 struct kmem_cache_node *n, struct page *page)
968 if (!(s->flags & SLAB_STORE_USER))
971 list_add(&page->lru, &n->full);
975 * list_lock must be held.
977 static void remove_full(struct kmem_cache *s, struct page *page)
979 if (!(s->flags & SLAB_STORE_USER))
982 list_del(&page->lru);
985 /* Tracking of the number of slabs for debugging purposes */
986 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
988 struct kmem_cache_node *n = get_node(s, node);
990 return atomic_long_read(&n->nr_slabs);
993 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
995 return atomic_long_read(&n->nr_slabs);
998 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1000 struct kmem_cache_node *n = get_node(s, node);
1003 * May be called early in order to allocate a slab for the
1004 * kmem_cache_node structure. Solve the chicken-egg
1005 * dilemma by deferring the increment of the count during
1006 * bootstrap (see early_kmem_cache_node_alloc).
1009 atomic_long_inc(&n->nr_slabs);
1010 atomic_long_add(objects, &n->total_objects);
1013 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1015 struct kmem_cache_node *n = get_node(s, node);
1017 atomic_long_dec(&n->nr_slabs);
1018 atomic_long_sub(objects, &n->total_objects);
1021 /* Object debug checks for alloc/free paths */
1022 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1025 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1028 init_object(s, object, SLUB_RED_INACTIVE);
1029 init_tracking(s, object);
1032 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1033 void *object, unsigned long addr)
1035 if (!check_slab(s, page))
1038 if (!check_valid_pointer(s, page, object)) {
1039 object_err(s, page, object, "Freelist Pointer check fails");
1043 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1046 /* Success perform special debug activities for allocs */
1047 if (s->flags & SLAB_STORE_USER)
1048 set_track(s, object, TRACK_ALLOC, addr);
1049 trace(s, page, object, 1);
1050 init_object(s, object, SLUB_RED_ACTIVE);
1054 if (PageSlab(page)) {
1056 * If this is a slab page then lets do the best we can
1057 * to avoid issues in the future. Marking all objects
1058 * as used avoids touching the remaining objects.
1060 slab_fix(s, "Marking all objects used");
1061 page->inuse = page->objects;
1062 page->freelist = NULL;
1067 static noinline struct kmem_cache_node *free_debug_processing(
1068 struct kmem_cache *s, struct page *page, void *object,
1069 unsigned long addr, unsigned long *flags)
1071 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1073 spin_lock_irqsave(&n->list_lock, *flags);
1076 if (!check_slab(s, page))
1079 if (!check_valid_pointer(s, page, object)) {
1080 slab_err(s, page, "Invalid object pointer 0x%p", object);
1084 if (on_freelist(s, page, object)) {
1085 object_err(s, page, object, "Object already free");
1089 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1092 if (unlikely(s != page->slab_cache)) {
1093 if (!PageSlab(page)) {
1094 slab_err(s, page, "Attempt to free object(0x%p) "
1095 "outside of slab", object);
1096 } else if (!page->slab_cache) {
1098 "SLUB <none>: no slab for object 0x%p.\n",
1102 object_err(s, page, object,
1103 "page slab pointer corrupt.");
1107 if (s->flags & SLAB_STORE_USER)
1108 set_track(s, object, TRACK_FREE, addr);
1109 trace(s, page, object, 0);
1110 init_object(s, object, SLUB_RED_INACTIVE);
1114 * Keep node_lock to preserve integrity
1115 * until the object is actually freed
1121 spin_unlock_irqrestore(&n->list_lock, *flags);
1122 slab_fix(s, "Object at 0x%p not freed", object);
1126 static int __init setup_slub_debug(char *str)
1128 slub_debug = DEBUG_DEFAULT_FLAGS;
1129 if (*str++ != '=' || !*str)
1131 * No options specified. Switch on full debugging.
1137 * No options but restriction on slabs. This means full
1138 * debugging for slabs matching a pattern.
1142 if (tolower(*str) == 'o') {
1144 * Avoid enabling debugging on caches if its minimum order
1145 * would increase as a result.
1147 disable_higher_order_debug = 1;
1154 * Switch off all debugging measures.
1159 * Determine which debug features should be switched on
1161 for (; *str && *str != ','; str++) {
1162 switch (tolower(*str)) {
1164 slub_debug |= SLAB_DEBUG_FREE;
1167 slub_debug |= SLAB_RED_ZONE;
1170 slub_debug |= SLAB_POISON;
1173 slub_debug |= SLAB_STORE_USER;
1176 slub_debug |= SLAB_TRACE;
1179 slub_debug |= SLAB_FAILSLAB;
1182 printk(KERN_ERR "slub_debug option '%c' "
1183 "unknown. skipped\n", *str);
1189 slub_debug_slabs = str + 1;
1194 __setup("slub_debug", setup_slub_debug);
1196 static unsigned long kmem_cache_flags(unsigned long object_size,
1197 unsigned long flags, const char *name,
1198 void (*ctor)(void *))
1201 * Enable debugging if selected on the kernel commandline.
1203 if (slub_debug && (!slub_debug_slabs ||
1204 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1205 flags |= slub_debug;
1210 static inline void setup_object_debug(struct kmem_cache *s,
1211 struct page *page, void *object) {}
1213 static inline int alloc_debug_processing(struct kmem_cache *s,
1214 struct page *page, void *object, unsigned long addr) { return 0; }
1216 static inline struct kmem_cache_node *free_debug_processing(
1217 struct kmem_cache *s, struct page *page, void *object,
1218 unsigned long addr, unsigned long *flags) { return NULL; }
1220 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1222 static inline int check_object(struct kmem_cache *s, struct page *page,
1223 void *object, u8 val) { return 1; }
1224 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1225 struct page *page) {}
1226 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1227 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1228 unsigned long flags, const char *name,
1229 void (*ctor)(void *))
1233 #define slub_debug 0
1235 #define disable_higher_order_debug 0
1237 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1239 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1241 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1243 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1246 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1249 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1252 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1254 #endif /* CONFIG_SLUB_DEBUG */
1257 * Slab allocation and freeing
1259 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1260 struct kmem_cache_order_objects oo)
1262 int order = oo_order(oo);
1264 flags |= __GFP_NOTRACK;
1266 if (node == NUMA_NO_NODE)
1267 return alloc_pages(flags, order);
1269 return alloc_pages_exact_node(node, flags, order);
1272 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1275 struct kmem_cache_order_objects oo = s->oo;
1278 flags &= gfp_allowed_mask;
1280 if (flags & __GFP_WAIT)
1283 flags |= s->allocflags;
1286 * Let the initial higher-order allocation fail under memory pressure
1287 * so we fall-back to the minimum order allocation.
1289 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1291 page = alloc_slab_page(alloc_gfp, node, oo);
1292 if (unlikely(!page)) {
1295 * Allocation may have failed due to fragmentation.
1296 * Try a lower order alloc if possible
1298 page = alloc_slab_page(flags, node, oo);
1301 stat(s, ORDER_FALLBACK);
1304 if (kmemcheck_enabled && page
1305 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1306 int pages = 1 << oo_order(oo);
1308 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1311 * Objects from caches that have a constructor don't get
1312 * cleared when they're allocated, so we need to do it here.
1315 kmemcheck_mark_uninitialized_pages(page, pages);
1317 kmemcheck_mark_unallocated_pages(page, pages);
1320 if (flags & __GFP_WAIT)
1321 local_irq_disable();
1325 page->objects = oo_objects(oo);
1326 mod_zone_page_state(page_zone(page),
1327 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1328 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1334 static void setup_object(struct kmem_cache *s, struct page *page,
1337 setup_object_debug(s, page, object);
1338 if (unlikely(s->ctor))
1342 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1350 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1352 page = allocate_slab(s,
1353 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1357 order = compound_order(page);
1358 inc_slabs_node(s, page_to_nid(page), page->objects);
1359 memcg_bind_pages(s, order);
1360 page->slab_cache = s;
1361 __SetPageSlab(page);
1362 if (page->pfmemalloc)
1363 SetPageSlabPfmemalloc(page);
1365 start = page_address(page);
1367 if (unlikely(s->flags & SLAB_POISON))
1368 memset(start, POISON_INUSE, PAGE_SIZE << order);
1371 for_each_object(p, s, start, page->objects) {
1372 setup_object(s, page, last);
1373 set_freepointer(s, last, p);
1376 setup_object(s, page, last);
1377 set_freepointer(s, last, NULL);
1379 page->freelist = start;
1380 page->inuse = page->objects;
1386 static void __free_slab(struct kmem_cache *s, struct page *page)
1388 int order = compound_order(page);
1389 int pages = 1 << order;
1391 if (kmem_cache_debug(s)) {
1394 slab_pad_check(s, page);
1395 for_each_object(p, s, page_address(page),
1397 check_object(s, page, p, SLUB_RED_INACTIVE);
1400 kmemcheck_free_shadow(page, compound_order(page));
1402 mod_zone_page_state(page_zone(page),
1403 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1404 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1407 __ClearPageSlabPfmemalloc(page);
1408 __ClearPageSlab(page);
1410 memcg_release_pages(s, order);
1411 reset_page_mapcount(page);
1412 if (current->reclaim_state)
1413 current->reclaim_state->reclaimed_slab += pages;
1414 __free_memcg_kmem_pages(page, order);
1417 #define need_reserve_slab_rcu \
1418 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1420 static void rcu_free_slab(struct rcu_head *h)
1424 if (need_reserve_slab_rcu)
1425 page = virt_to_head_page(h);
1427 page = container_of((struct list_head *)h, struct page, lru);
1429 __free_slab(page->slab_cache, page);
1432 static void free_slab(struct kmem_cache *s, struct page *page)
1434 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1435 struct rcu_head *head;
1437 if (need_reserve_slab_rcu) {
1438 int order = compound_order(page);
1439 int offset = (PAGE_SIZE << order) - s->reserved;
1441 VM_BUG_ON(s->reserved != sizeof(*head));
1442 head = page_address(page) + offset;
1445 * RCU free overloads the RCU head over the LRU
1447 head = (void *)&page->lru;
1450 call_rcu(head, rcu_free_slab);
1452 __free_slab(s, page);
1455 static void discard_slab(struct kmem_cache *s, struct page *page)
1457 dec_slabs_node(s, page_to_nid(page), page->objects);
1462 * Management of partially allocated slabs.
1464 * list_lock must be held.
1466 static inline void add_partial(struct kmem_cache_node *n,
1467 struct page *page, int tail)
1470 if (tail == DEACTIVATE_TO_TAIL)
1471 list_add_tail(&page->lru, &n->partial);
1473 list_add(&page->lru, &n->partial);
1477 * list_lock must be held.
1479 static inline void remove_partial(struct kmem_cache_node *n,
1482 list_del(&page->lru);
1487 * Remove slab from the partial list, freeze it and
1488 * return the pointer to the freelist.
1490 * Returns a list of objects or NULL if it fails.
1492 * Must hold list_lock since we modify the partial list.
1494 static inline void *acquire_slab(struct kmem_cache *s,
1495 struct kmem_cache_node *n, struct page *page,
1496 int mode, int *objects)
1499 unsigned long counters;
1503 * Zap the freelist and set the frozen bit.
1504 * The old freelist is the list of objects for the
1505 * per cpu allocation list.
1507 freelist = page->freelist;
1508 counters = page->counters;
1509 new.counters = counters;
1510 *objects = new.objects - new.inuse;
1512 new.inuse = page->objects;
1513 new.freelist = NULL;
1515 new.freelist = freelist;
1518 VM_BUG_ON(new.frozen);
1521 if (!__cmpxchg_double_slab(s, page,
1523 new.freelist, new.counters,
1527 remove_partial(n, page);
1532 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1533 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1536 * Try to allocate a partial slab from a specific node.
1538 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1539 struct kmem_cache_cpu *c, gfp_t flags)
1541 struct page *page, *page2;
1542 void *object = NULL;
1547 * Racy check. If we mistakenly see no partial slabs then we
1548 * just allocate an empty slab. If we mistakenly try to get a
1549 * partial slab and there is none available then get_partials()
1552 if (!n || !n->nr_partial)
1555 spin_lock(&n->list_lock);
1556 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1559 if (!pfmemalloc_match(page, flags))
1562 t = acquire_slab(s, n, page, object == NULL, &objects);
1566 available += objects;
1569 stat(s, ALLOC_FROM_PARTIAL);
1572 put_cpu_partial(s, page, 0);
1573 stat(s, CPU_PARTIAL_NODE);
1575 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1579 spin_unlock(&n->list_lock);
1584 * Get a page from somewhere. Search in increasing NUMA distances.
1586 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1587 struct kmem_cache_cpu *c)
1590 struct zonelist *zonelist;
1593 enum zone_type high_zoneidx = gfp_zone(flags);
1595 unsigned int cpuset_mems_cookie;
1598 * The defrag ratio allows a configuration of the tradeoffs between
1599 * inter node defragmentation and node local allocations. A lower
1600 * defrag_ratio increases the tendency to do local allocations
1601 * instead of attempting to obtain partial slabs from other nodes.
1603 * If the defrag_ratio is set to 0 then kmalloc() always
1604 * returns node local objects. If the ratio is higher then kmalloc()
1605 * may return off node objects because partial slabs are obtained
1606 * from other nodes and filled up.
1608 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1609 * defrag_ratio = 1000) then every (well almost) allocation will
1610 * first attempt to defrag slab caches on other nodes. This means
1611 * scanning over all nodes to look for partial slabs which may be
1612 * expensive if we do it every time we are trying to find a slab
1613 * with available objects.
1615 if (!s->remote_node_defrag_ratio ||
1616 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1620 cpuset_mems_cookie = get_mems_allowed();
1621 zonelist = node_zonelist(slab_node(), flags);
1622 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1623 struct kmem_cache_node *n;
1625 n = get_node(s, zone_to_nid(zone));
1627 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1628 n->nr_partial > s->min_partial) {
1629 object = get_partial_node(s, n, c, flags);
1632 * Return the object even if
1633 * put_mems_allowed indicated that
1634 * the cpuset mems_allowed was
1635 * updated in parallel. It's a
1636 * harmless race between the alloc
1637 * and the cpuset update.
1639 put_mems_allowed(cpuset_mems_cookie);
1644 } while (!put_mems_allowed(cpuset_mems_cookie));
1650 * Get a partial page, lock it and return it.
1652 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1653 struct kmem_cache_cpu *c)
1656 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1658 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1659 if (object || node != NUMA_NO_NODE)
1662 return get_any_partial(s, flags, c);
1665 #ifdef CONFIG_PREEMPT
1667 * Calculate the next globally unique transaction for disambiguiation
1668 * during cmpxchg. The transactions start with the cpu number and are then
1669 * incremented by CONFIG_NR_CPUS.
1671 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1674 * No preemption supported therefore also no need to check for
1680 static inline unsigned long next_tid(unsigned long tid)
1682 return tid + TID_STEP;
1685 static inline unsigned int tid_to_cpu(unsigned long tid)
1687 return tid % TID_STEP;
1690 static inline unsigned long tid_to_event(unsigned long tid)
1692 return tid / TID_STEP;
1695 static inline unsigned int init_tid(int cpu)
1700 static inline void note_cmpxchg_failure(const char *n,
1701 const struct kmem_cache *s, unsigned long tid)
1703 #ifdef SLUB_DEBUG_CMPXCHG
1704 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1706 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1708 #ifdef CONFIG_PREEMPT
1709 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1710 printk("due to cpu change %d -> %d\n",
1711 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1714 if (tid_to_event(tid) != tid_to_event(actual_tid))
1715 printk("due to cpu running other code. Event %ld->%ld\n",
1716 tid_to_event(tid), tid_to_event(actual_tid));
1718 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1719 actual_tid, tid, next_tid(tid));
1721 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1724 static void init_kmem_cache_cpus(struct kmem_cache *s)
1728 for_each_possible_cpu(cpu)
1729 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1733 * Remove the cpu slab
1735 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1737 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1738 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1740 enum slab_modes l = M_NONE, m = M_NONE;
1742 int tail = DEACTIVATE_TO_HEAD;
1746 if (page->freelist) {
1747 stat(s, DEACTIVATE_REMOTE_FREES);
1748 tail = DEACTIVATE_TO_TAIL;
1752 * Stage one: Free all available per cpu objects back
1753 * to the page freelist while it is still frozen. Leave the
1756 * There is no need to take the list->lock because the page
1759 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1761 unsigned long counters;
1764 prior = page->freelist;
1765 counters = page->counters;
1766 set_freepointer(s, freelist, prior);
1767 new.counters = counters;
1769 VM_BUG_ON(!new.frozen);
1771 } while (!__cmpxchg_double_slab(s, page,
1773 freelist, new.counters,
1774 "drain percpu freelist"));
1776 freelist = nextfree;
1780 * Stage two: Ensure that the page is unfrozen while the
1781 * list presence reflects the actual number of objects
1784 * We setup the list membership and then perform a cmpxchg
1785 * with the count. If there is a mismatch then the page
1786 * is not unfrozen but the page is on the wrong list.
1788 * Then we restart the process which may have to remove
1789 * the page from the list that we just put it on again
1790 * because the number of objects in the slab may have
1795 old.freelist = page->freelist;
1796 old.counters = page->counters;
1797 VM_BUG_ON(!old.frozen);
1799 /* Determine target state of the slab */
1800 new.counters = old.counters;
1803 set_freepointer(s, freelist, old.freelist);
1804 new.freelist = freelist;
1806 new.freelist = old.freelist;
1810 if (!new.inuse && n->nr_partial > s->min_partial)
1812 else if (new.freelist) {
1817 * Taking the spinlock removes the possiblity
1818 * that acquire_slab() will see a slab page that
1821 spin_lock(&n->list_lock);
1825 if (kmem_cache_debug(s) && !lock) {
1828 * This also ensures that the scanning of full
1829 * slabs from diagnostic functions will not see
1832 spin_lock(&n->list_lock);
1840 remove_partial(n, page);
1842 else if (l == M_FULL)
1844 remove_full(s, page);
1846 if (m == M_PARTIAL) {
1848 add_partial(n, page, tail);
1851 } else if (m == M_FULL) {
1853 stat(s, DEACTIVATE_FULL);
1854 add_full(s, n, page);
1860 if (!__cmpxchg_double_slab(s, page,
1861 old.freelist, old.counters,
1862 new.freelist, new.counters,
1867 spin_unlock(&n->list_lock);
1870 stat(s, DEACTIVATE_EMPTY);
1871 discard_slab(s, page);
1877 * Unfreeze all the cpu partial slabs.
1879 * This function must be called with interrupts disabled
1880 * for the cpu using c (or some other guarantee must be there
1881 * to guarantee no concurrent accesses).
1883 static void unfreeze_partials(struct kmem_cache *s,
1884 struct kmem_cache_cpu *c)
1886 struct kmem_cache_node *n = NULL, *n2 = NULL;
1887 struct page *page, *discard_page = NULL;
1889 while ((page = c->partial)) {
1893 c->partial = page->next;
1895 n2 = get_node(s, page_to_nid(page));
1898 spin_unlock(&n->list_lock);
1901 spin_lock(&n->list_lock);
1906 old.freelist = page->freelist;
1907 old.counters = page->counters;
1908 VM_BUG_ON(!old.frozen);
1910 new.counters = old.counters;
1911 new.freelist = old.freelist;
1915 } while (!__cmpxchg_double_slab(s, page,
1916 old.freelist, old.counters,
1917 new.freelist, new.counters,
1918 "unfreezing slab"));
1920 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1921 page->next = discard_page;
1922 discard_page = page;
1924 add_partial(n, page, DEACTIVATE_TO_TAIL);
1925 stat(s, FREE_ADD_PARTIAL);
1930 spin_unlock(&n->list_lock);
1932 while (discard_page) {
1933 page = discard_page;
1934 discard_page = discard_page->next;
1936 stat(s, DEACTIVATE_EMPTY);
1937 discard_slab(s, page);
1943 * Put a page that was just frozen (in __slab_free) into a partial page
1944 * slot if available. This is done without interrupts disabled and without
1945 * preemption disabled. The cmpxchg is racy and may put the partial page
1946 * onto a random cpus partial slot.
1948 * If we did not find a slot then simply move all the partials to the
1949 * per node partial list.
1951 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1953 struct page *oldpage;
1960 oldpage = this_cpu_read(s->cpu_slab->partial);
1963 pobjects = oldpage->pobjects;
1964 pages = oldpage->pages;
1965 if (drain && pobjects > s->cpu_partial) {
1966 unsigned long flags;
1968 * partial array is full. Move the existing
1969 * set to the per node partial list.
1971 local_irq_save(flags);
1972 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1973 local_irq_restore(flags);
1977 stat(s, CPU_PARTIAL_DRAIN);
1982 pobjects += page->objects - page->inuse;
1984 page->pages = pages;
1985 page->pobjects = pobjects;
1986 page->next = oldpage;
1988 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1991 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1993 stat(s, CPUSLAB_FLUSH);
1994 deactivate_slab(s, c->page, c->freelist);
1996 c->tid = next_tid(c->tid);
2004 * Called from IPI handler with interrupts disabled.
2006 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2008 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2014 unfreeze_partials(s, c);
2018 static void flush_cpu_slab(void *d)
2020 struct kmem_cache *s = d;
2022 __flush_cpu_slab(s, smp_processor_id());
2025 static bool has_cpu_slab(int cpu, void *info)
2027 struct kmem_cache *s = info;
2028 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2030 return c->page || c->partial;
2033 static void flush_all(struct kmem_cache *s)
2035 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2039 * Check if the objects in a per cpu structure fit numa
2040 * locality expectations.
2042 static inline int node_match(struct page *page, int node)
2045 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2051 static int count_free(struct page *page)
2053 return page->objects - page->inuse;
2056 static unsigned long count_partial(struct kmem_cache_node *n,
2057 int (*get_count)(struct page *))
2059 unsigned long flags;
2060 unsigned long x = 0;
2063 spin_lock_irqsave(&n->list_lock, flags);
2064 list_for_each_entry(page, &n->partial, lru)
2065 x += get_count(page);
2066 spin_unlock_irqrestore(&n->list_lock, flags);
2070 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2072 #ifdef CONFIG_SLUB_DEBUG
2073 return atomic_long_read(&n->total_objects);
2079 static noinline void
2080 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2085 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2087 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2088 "default order: %d, min order: %d\n", s->name, s->object_size,
2089 s->size, oo_order(s->oo), oo_order(s->min));
2091 if (oo_order(s->min) > get_order(s->object_size))
2092 printk(KERN_WARNING " %s debugging increased min order, use "
2093 "slub_debug=O to disable.\n", s->name);
2095 for_each_online_node(node) {
2096 struct kmem_cache_node *n = get_node(s, node);
2097 unsigned long nr_slabs;
2098 unsigned long nr_objs;
2099 unsigned long nr_free;
2104 nr_free = count_partial(n, count_free);
2105 nr_slabs = node_nr_slabs(n);
2106 nr_objs = node_nr_objs(n);
2109 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2110 node, nr_slabs, nr_objs, nr_free);
2114 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2115 int node, struct kmem_cache_cpu **pc)
2118 struct kmem_cache_cpu *c = *pc;
2121 freelist = get_partial(s, flags, node, c);
2126 page = new_slab(s, flags, node);
2128 c = __this_cpu_ptr(s->cpu_slab);
2133 * No other reference to the page yet so we can
2134 * muck around with it freely without cmpxchg
2136 freelist = page->freelist;
2137 page->freelist = NULL;
2139 stat(s, ALLOC_SLAB);
2148 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2150 if (unlikely(PageSlabPfmemalloc(page)))
2151 return gfp_pfmemalloc_allowed(gfpflags);
2157 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2158 * or deactivate the page.
2160 * The page is still frozen if the return value is not NULL.
2162 * If this function returns NULL then the page has been unfrozen.
2164 * This function must be called with interrupt disabled.
2166 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2169 unsigned long counters;
2173 freelist = page->freelist;
2174 counters = page->counters;
2176 new.counters = counters;
2177 VM_BUG_ON(!new.frozen);
2179 new.inuse = page->objects;
2180 new.frozen = freelist != NULL;
2182 } while (!__cmpxchg_double_slab(s, page,
2191 * Slow path. The lockless freelist is empty or we need to perform
2194 * Processing is still very fast if new objects have been freed to the
2195 * regular freelist. In that case we simply take over the regular freelist
2196 * as the lockless freelist and zap the regular freelist.
2198 * If that is not working then we fall back to the partial lists. We take the
2199 * first element of the freelist as the object to allocate now and move the
2200 * rest of the freelist to the lockless freelist.
2202 * And if we were unable to get a new slab from the partial slab lists then
2203 * we need to allocate a new slab. This is the slowest path since it involves
2204 * a call to the page allocator and the setup of a new slab.
2206 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2207 unsigned long addr, struct kmem_cache_cpu *c)
2211 unsigned long flags;
2213 local_irq_save(flags);
2214 #ifdef CONFIG_PREEMPT
2216 * We may have been preempted and rescheduled on a different
2217 * cpu before disabling interrupts. Need to reload cpu area
2220 c = this_cpu_ptr(s->cpu_slab);
2228 if (unlikely(!node_match(page, node))) {
2229 stat(s, ALLOC_NODE_MISMATCH);
2230 deactivate_slab(s, page, c->freelist);
2237 * By rights, we should be searching for a slab page that was
2238 * PFMEMALLOC but right now, we are losing the pfmemalloc
2239 * information when the page leaves the per-cpu allocator
2241 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2242 deactivate_slab(s, page, c->freelist);
2248 /* must check again c->freelist in case of cpu migration or IRQ */
2249 freelist = c->freelist;
2253 stat(s, ALLOC_SLOWPATH);
2255 freelist = get_freelist(s, page);
2259 stat(s, DEACTIVATE_BYPASS);
2263 stat(s, ALLOC_REFILL);
2267 * freelist is pointing to the list of objects to be used.
2268 * page is pointing to the page from which the objects are obtained.
2269 * That page must be frozen for per cpu allocations to work.
2271 VM_BUG_ON(!c->page->frozen);
2272 c->freelist = get_freepointer(s, freelist);
2273 c->tid = next_tid(c->tid);
2274 local_irq_restore(flags);
2280 page = c->page = c->partial;
2281 c->partial = page->next;
2282 stat(s, CPU_PARTIAL_ALLOC);
2287 freelist = new_slab_objects(s, gfpflags, node, &c);
2289 if (unlikely(!freelist)) {
2290 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2291 slab_out_of_memory(s, gfpflags, node);
2293 local_irq_restore(flags);
2298 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2301 /* Only entered in the debug case */
2302 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2303 goto new_slab; /* Slab failed checks. Next slab needed */
2305 deactivate_slab(s, page, get_freepointer(s, freelist));
2308 local_irq_restore(flags);
2313 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2314 * have the fastpath folded into their functions. So no function call
2315 * overhead for requests that can be satisfied on the fastpath.
2317 * The fastpath works by first checking if the lockless freelist can be used.
2318 * If not then __slab_alloc is called for slow processing.
2320 * Otherwise we can simply pick the next object from the lockless free list.
2322 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2323 gfp_t gfpflags, int node, unsigned long addr)
2326 struct kmem_cache_cpu *c;
2330 if (slab_pre_alloc_hook(s, gfpflags))
2333 s = memcg_kmem_get_cache(s, gfpflags);
2337 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2338 * enabled. We may switch back and forth between cpus while
2339 * reading from one cpu area. That does not matter as long
2340 * as we end up on the original cpu again when doing the cmpxchg.
2342 c = __this_cpu_ptr(s->cpu_slab);
2345 * The transaction ids are globally unique per cpu and per operation on
2346 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2347 * occurs on the right processor and that there was no operation on the
2348 * linked list in between.
2353 object = c->freelist;
2355 if (unlikely(!object || !node_match(page, node)))
2356 object = __slab_alloc(s, gfpflags, node, addr, c);
2359 void *next_object = get_freepointer_safe(s, object);
2362 * The cmpxchg will only match if there was no additional
2363 * operation and if we are on the right processor.
2365 * The cmpxchg does the following atomically (without lock semantics!)
2366 * 1. Relocate first pointer to the current per cpu area.
2367 * 2. Verify that tid and freelist have not been changed
2368 * 3. If they were not changed replace tid and freelist
2370 * Since this is without lock semantics the protection is only against
2371 * code executing on this cpu *not* from access by other cpus.
2373 if (unlikely(!this_cpu_cmpxchg_double(
2374 s->cpu_slab->freelist, s->cpu_slab->tid,
2376 next_object, next_tid(tid)))) {
2378 note_cmpxchg_failure("slab_alloc", s, tid);
2381 prefetch_freepointer(s, next_object);
2382 stat(s, ALLOC_FASTPATH);
2385 if (unlikely(gfpflags & __GFP_ZERO) && object)
2386 memset(object, 0, s->object_size);
2388 slab_post_alloc_hook(s, gfpflags, object);
2393 static __always_inline void *slab_alloc(struct kmem_cache *s,
2394 gfp_t gfpflags, unsigned long addr)
2396 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2399 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2401 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2403 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2407 EXPORT_SYMBOL(kmem_cache_alloc);
2409 #ifdef CONFIG_TRACING
2410 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2412 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2413 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2416 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2418 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2420 void *ret = kmalloc_order(size, flags, order);
2421 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2424 EXPORT_SYMBOL(kmalloc_order_trace);
2428 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2430 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2432 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2433 s->object_size, s->size, gfpflags, node);
2437 EXPORT_SYMBOL(kmem_cache_alloc_node);
2439 #ifdef CONFIG_TRACING
2440 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2442 int node, size_t size)
2444 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2446 trace_kmalloc_node(_RET_IP_, ret,
2447 size, s->size, gfpflags, node);
2450 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2455 * Slow patch handling. This may still be called frequently since objects
2456 * have a longer lifetime than the cpu slabs in most processing loads.
2458 * So we still attempt to reduce cache line usage. Just take the slab
2459 * lock and free the item. If there is no additional partial page
2460 * handling required then we can return immediately.
2462 static void __slab_free(struct kmem_cache *s, struct page *page,
2463 void *x, unsigned long addr)
2466 void **object = (void *)x;
2469 unsigned long counters;
2470 struct kmem_cache_node *n = NULL;
2471 unsigned long uninitialized_var(flags);
2473 stat(s, FREE_SLOWPATH);
2475 if (kmem_cache_debug(s) &&
2476 !(n = free_debug_processing(s, page, x, addr, &flags)))
2481 spin_unlock_irqrestore(&n->list_lock, flags);
2484 prior = page->freelist;
2485 counters = page->counters;
2486 set_freepointer(s, object, prior);
2487 new.counters = counters;
2488 was_frozen = new.frozen;
2490 if ((!new.inuse || !prior) && !was_frozen) {
2492 if (!kmem_cache_debug(s) && !prior)
2495 * Slab was on no list before and will be partially empty
2496 * We can defer the list move and instead freeze it.
2500 else { /* Needs to be taken off a list */
2502 n = get_node(s, page_to_nid(page));
2504 * Speculatively acquire the list_lock.
2505 * If the cmpxchg does not succeed then we may
2506 * drop the list_lock without any processing.
2508 * Otherwise the list_lock will synchronize with
2509 * other processors updating the list of slabs.
2511 spin_lock_irqsave(&n->list_lock, flags);
2516 } while (!cmpxchg_double_slab(s, page,
2518 object, new.counters,
2524 * If we just froze the page then put it onto the
2525 * per cpu partial list.
2527 if (new.frozen && !was_frozen) {
2528 put_cpu_partial(s, page, 1);
2529 stat(s, CPU_PARTIAL_FREE);
2532 * The list lock was not taken therefore no list
2533 * activity can be necessary.
2536 stat(s, FREE_FROZEN);
2540 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2544 * Objects left in the slab. If it was not on the partial list before
2547 if (kmem_cache_debug(s) && unlikely(!prior)) {
2548 remove_full(s, page);
2549 add_partial(n, page, DEACTIVATE_TO_TAIL);
2550 stat(s, FREE_ADD_PARTIAL);
2552 spin_unlock_irqrestore(&n->list_lock, flags);
2558 * Slab on the partial list.
2560 remove_partial(n, page);
2561 stat(s, FREE_REMOVE_PARTIAL);
2563 /* Slab must be on the full list */
2564 remove_full(s, page);
2566 spin_unlock_irqrestore(&n->list_lock, flags);
2568 discard_slab(s, page);
2572 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2573 * can perform fastpath freeing without additional function calls.
2575 * The fastpath is only possible if we are freeing to the current cpu slab
2576 * of this processor. This typically the case if we have just allocated
2579 * If fastpath is not possible then fall back to __slab_free where we deal
2580 * with all sorts of special processing.
2582 static __always_inline void slab_free(struct kmem_cache *s,
2583 struct page *page, void *x, unsigned long addr)
2585 void **object = (void *)x;
2586 struct kmem_cache_cpu *c;
2589 slab_free_hook(s, x);
2593 * Determine the currently cpus per cpu slab.
2594 * The cpu may change afterward. However that does not matter since
2595 * data is retrieved via this pointer. If we are on the same cpu
2596 * during the cmpxchg then the free will succedd.
2598 c = __this_cpu_ptr(s->cpu_slab);
2603 if (likely(page == c->page)) {
2604 set_freepointer(s, object, c->freelist);
2606 if (unlikely(!this_cpu_cmpxchg_double(
2607 s->cpu_slab->freelist, s->cpu_slab->tid,
2609 object, next_tid(tid)))) {
2611 note_cmpxchg_failure("slab_free", s, tid);
2614 stat(s, FREE_FASTPATH);
2616 __slab_free(s, page, x, addr);
2620 void kmem_cache_free(struct kmem_cache *s, void *x)
2622 s = cache_from_obj(s, x);
2625 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2626 trace_kmem_cache_free(_RET_IP_, x);
2628 EXPORT_SYMBOL(kmem_cache_free);
2631 * Object placement in a slab is made very easy because we always start at
2632 * offset 0. If we tune the size of the object to the alignment then we can
2633 * get the required alignment by putting one properly sized object after
2636 * Notice that the allocation order determines the sizes of the per cpu
2637 * caches. Each processor has always one slab available for allocations.
2638 * Increasing the allocation order reduces the number of times that slabs
2639 * must be moved on and off the partial lists and is therefore a factor in
2644 * Mininum / Maximum order of slab pages. This influences locking overhead
2645 * and slab fragmentation. A higher order reduces the number of partial slabs
2646 * and increases the number of allocations possible without having to
2647 * take the list_lock.
2649 static int slub_min_order;
2650 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2651 static int slub_min_objects;
2654 * Merge control. If this is set then no merging of slab caches will occur.
2655 * (Could be removed. This was introduced to pacify the merge skeptics.)
2657 static int slub_nomerge;
2660 * Calculate the order of allocation given an slab object size.
2662 * The order of allocation has significant impact on performance and other
2663 * system components. Generally order 0 allocations should be preferred since
2664 * order 0 does not cause fragmentation in the page allocator. Larger objects
2665 * be problematic to put into order 0 slabs because there may be too much
2666 * unused space left. We go to a higher order if more than 1/16th of the slab
2669 * In order to reach satisfactory performance we must ensure that a minimum
2670 * number of objects is in one slab. Otherwise we may generate too much
2671 * activity on the partial lists which requires taking the list_lock. This is
2672 * less a concern for large slabs though which are rarely used.
2674 * slub_max_order specifies the order where we begin to stop considering the
2675 * number of objects in a slab as critical. If we reach slub_max_order then
2676 * we try to keep the page order as low as possible. So we accept more waste
2677 * of space in favor of a small page order.
2679 * Higher order allocations also allow the placement of more objects in a
2680 * slab and thereby reduce object handling overhead. If the user has
2681 * requested a higher mininum order then we start with that one instead of
2682 * the smallest order which will fit the object.
2684 static inline int slab_order(int size, int min_objects,
2685 int max_order, int fract_leftover, int reserved)
2689 int min_order = slub_min_order;
2691 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2692 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2694 for (order = max(min_order,
2695 fls(min_objects * size - 1) - PAGE_SHIFT);
2696 order <= max_order; order++) {
2698 unsigned long slab_size = PAGE_SIZE << order;
2700 if (slab_size < min_objects * size + reserved)
2703 rem = (slab_size - reserved) % size;
2705 if (rem <= slab_size / fract_leftover)
2713 static inline int calculate_order(int size, int reserved)
2721 * Attempt to find best configuration for a slab. This
2722 * works by first attempting to generate a layout with
2723 * the best configuration and backing off gradually.
2725 * First we reduce the acceptable waste in a slab. Then
2726 * we reduce the minimum objects required in a slab.
2728 min_objects = slub_min_objects;
2730 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2731 max_objects = order_objects(slub_max_order, size, reserved);
2732 min_objects = min(min_objects, max_objects);
2734 while (min_objects > 1) {
2736 while (fraction >= 4) {
2737 order = slab_order(size, min_objects,
2738 slub_max_order, fraction, reserved);
2739 if (order <= slub_max_order)
2747 * We were unable to place multiple objects in a slab. Now
2748 * lets see if we can place a single object there.
2750 order = slab_order(size, 1, slub_max_order, 1, reserved);
2751 if (order <= slub_max_order)
2755 * Doh this slab cannot be placed using slub_max_order.
2757 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2758 if (order < MAX_ORDER)
2764 init_kmem_cache_node(struct kmem_cache_node *n)
2767 spin_lock_init(&n->list_lock);
2768 INIT_LIST_HEAD(&n->partial);
2769 #ifdef CONFIG_SLUB_DEBUG
2770 atomic_long_set(&n->nr_slabs, 0);
2771 atomic_long_set(&n->total_objects, 0);
2772 INIT_LIST_HEAD(&n->full);
2776 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2778 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2779 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2782 * Must align to double word boundary for the double cmpxchg
2783 * instructions to work; see __pcpu_double_call_return_bool().
2785 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2786 2 * sizeof(void *));
2791 init_kmem_cache_cpus(s);
2796 static struct kmem_cache *kmem_cache_node;
2799 * No kmalloc_node yet so do it by hand. We know that this is the first
2800 * slab on the node for this slabcache. There are no concurrent accesses
2803 * Note that this function only works on the kmalloc_node_cache
2804 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2805 * memory on a fresh node that has no slab structures yet.
2807 static void early_kmem_cache_node_alloc(int node)
2810 struct kmem_cache_node *n;
2812 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2814 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2817 if (page_to_nid(page) != node) {
2818 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2820 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2821 "in order to be able to continue\n");
2826 page->freelist = get_freepointer(kmem_cache_node, n);
2829 kmem_cache_node->node[node] = n;
2830 #ifdef CONFIG_SLUB_DEBUG
2831 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2832 init_tracking(kmem_cache_node, n);
2834 init_kmem_cache_node(n);
2835 inc_slabs_node(kmem_cache_node, node, page->objects);
2837 add_partial(n, page, DEACTIVATE_TO_HEAD);
2840 static void free_kmem_cache_nodes(struct kmem_cache *s)
2844 for_each_node_state(node, N_NORMAL_MEMORY) {
2845 struct kmem_cache_node *n = s->node[node];
2848 kmem_cache_free(kmem_cache_node, n);
2850 s->node[node] = NULL;
2854 static int init_kmem_cache_nodes(struct kmem_cache *s)
2858 for_each_node_state(node, N_NORMAL_MEMORY) {
2859 struct kmem_cache_node *n;
2861 if (slab_state == DOWN) {
2862 early_kmem_cache_node_alloc(node);
2865 n = kmem_cache_alloc_node(kmem_cache_node,
2869 free_kmem_cache_nodes(s);
2874 init_kmem_cache_node(n);
2879 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2881 if (min < MIN_PARTIAL)
2883 else if (min > MAX_PARTIAL)
2885 s->min_partial = min;
2889 * calculate_sizes() determines the order and the distribution of data within
2892 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2894 unsigned long flags = s->flags;
2895 unsigned long size = s->object_size;
2899 * Round up object size to the next word boundary. We can only
2900 * place the free pointer at word boundaries and this determines
2901 * the possible location of the free pointer.
2903 size = ALIGN(size, sizeof(void *));
2905 #ifdef CONFIG_SLUB_DEBUG
2907 * Determine if we can poison the object itself. If the user of
2908 * the slab may touch the object after free or before allocation
2909 * then we should never poison the object itself.
2911 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2913 s->flags |= __OBJECT_POISON;
2915 s->flags &= ~__OBJECT_POISON;
2919 * If we are Redzoning then check if there is some space between the
2920 * end of the object and the free pointer. If not then add an
2921 * additional word to have some bytes to store Redzone information.
2923 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2924 size += sizeof(void *);
2928 * With that we have determined the number of bytes in actual use
2929 * by the object. This is the potential offset to the free pointer.
2933 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2936 * Relocate free pointer after the object if it is not
2937 * permitted to overwrite the first word of the object on
2940 * This is the case if we do RCU, have a constructor or
2941 * destructor or are poisoning the objects.
2944 size += sizeof(void *);
2947 #ifdef CONFIG_SLUB_DEBUG
2948 if (flags & SLAB_STORE_USER)
2950 * Need to store information about allocs and frees after
2953 size += 2 * sizeof(struct track);
2955 if (flags & SLAB_RED_ZONE)
2957 * Add some empty padding so that we can catch
2958 * overwrites from earlier objects rather than let
2959 * tracking information or the free pointer be
2960 * corrupted if a user writes before the start
2963 size += sizeof(void *);
2967 * SLUB stores one object immediately after another beginning from
2968 * offset 0. In order to align the objects we have to simply size
2969 * each object to conform to the alignment.
2971 size = ALIGN(size, s->align);
2973 if (forced_order >= 0)
2974 order = forced_order;
2976 order = calculate_order(size, s->reserved);
2983 s->allocflags |= __GFP_COMP;
2985 if (s->flags & SLAB_CACHE_DMA)
2986 s->allocflags |= GFP_DMA;
2988 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2989 s->allocflags |= __GFP_RECLAIMABLE;
2992 * Determine the number of objects per slab
2994 s->oo = oo_make(order, size, s->reserved);
2995 s->min = oo_make(get_order(size), size, s->reserved);
2996 if (oo_objects(s->oo) > oo_objects(s->max))
2999 return !!oo_objects(s->oo);
3002 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3004 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3007 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3008 s->reserved = sizeof(struct rcu_head);
3010 if (!calculate_sizes(s, -1))
3012 if (disable_higher_order_debug) {
3014 * Disable debugging flags that store metadata if the min slab
3017 if (get_order(s->size) > get_order(s->object_size)) {
3018 s->flags &= ~DEBUG_METADATA_FLAGS;
3020 if (!calculate_sizes(s, -1))
3025 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3026 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3027 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3028 /* Enable fast mode */
3029 s->flags |= __CMPXCHG_DOUBLE;
3033 * The larger the object size is, the more pages we want on the partial
3034 * list to avoid pounding the page allocator excessively.
3036 set_min_partial(s, ilog2(s->size) / 2);
3039 * cpu_partial determined the maximum number of objects kept in the
3040 * per cpu partial lists of a processor.
3042 * Per cpu partial lists mainly contain slabs that just have one
3043 * object freed. If they are used for allocation then they can be
3044 * filled up again with minimal effort. The slab will never hit the
3045 * per node partial lists and therefore no locking will be required.
3047 * This setting also determines
3049 * A) The number of objects from per cpu partial slabs dumped to the
3050 * per node list when we reach the limit.
3051 * B) The number of objects in cpu partial slabs to extract from the
3052 * per node list when we run out of per cpu objects. We only fetch 50%
3053 * to keep some capacity around for frees.
3055 if (kmem_cache_debug(s))
3057 else if (s->size >= PAGE_SIZE)
3059 else if (s->size >= 1024)
3061 else if (s->size >= 256)
3062 s->cpu_partial = 13;
3064 s->cpu_partial = 30;
3067 s->remote_node_defrag_ratio = 1000;
3069 if (!init_kmem_cache_nodes(s))
3072 if (alloc_kmem_cache_cpus(s))
3075 free_kmem_cache_nodes(s);
3077 if (flags & SLAB_PANIC)
3078 panic("Cannot create slab %s size=%lu realsize=%u "
3079 "order=%u offset=%u flags=%lx\n",
3080 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3085 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3088 #ifdef CONFIG_SLUB_DEBUG
3089 void *addr = page_address(page);
3091 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3092 sizeof(long), GFP_ATOMIC);
3095 slab_err(s, page, text, s->name);
3098 get_map(s, page, map);
3099 for_each_object(p, s, addr, page->objects) {
3101 if (!test_bit(slab_index(p, s, addr), map)) {
3102 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3104 print_tracking(s, p);
3113 * Attempt to free all partial slabs on a node.
3114 * This is called from kmem_cache_close(). We must be the last thread
3115 * using the cache and therefore we do not need to lock anymore.
3117 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3119 struct page *page, *h;
3121 list_for_each_entry_safe(page, h, &n->partial, lru) {
3123 remove_partial(n, page);
3124 discard_slab(s, page);
3126 list_slab_objects(s, page,
3127 "Objects remaining in %s on kmem_cache_close()");
3133 * Release all resources used by a slab cache.
3135 static inline int kmem_cache_close(struct kmem_cache *s)
3140 /* Attempt to free all objects */
3141 for_each_node_state(node, N_NORMAL_MEMORY) {
3142 struct kmem_cache_node *n = get_node(s, node);
3145 if (n->nr_partial || slabs_node(s, node))
3148 free_percpu(s->cpu_slab);
3149 free_kmem_cache_nodes(s);
3153 int __kmem_cache_shutdown(struct kmem_cache *s)
3155 int rc = kmem_cache_close(s);
3159 * We do the same lock strategy around sysfs_slab_add, see
3160 * __kmem_cache_create. Because this is pretty much the last
3161 * operation we do and the lock will be released shortly after
3162 * that in slab_common.c, we could just move sysfs_slab_remove
3163 * to a later point in common code. We should do that when we
3164 * have a common sysfs framework for all allocators.
3166 mutex_unlock(&slab_mutex);
3167 sysfs_slab_remove(s);
3168 mutex_lock(&slab_mutex);
3174 /********************************************************************
3176 *******************************************************************/
3178 static int __init setup_slub_min_order(char *str)
3180 get_option(&str, &slub_min_order);
3185 __setup("slub_min_order=", setup_slub_min_order);
3187 static int __init setup_slub_max_order(char *str)
3189 get_option(&str, &slub_max_order);
3190 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3195 __setup("slub_max_order=", setup_slub_max_order);
3197 static int __init setup_slub_min_objects(char *str)
3199 get_option(&str, &slub_min_objects);
3204 __setup("slub_min_objects=", setup_slub_min_objects);
3206 static int __init setup_slub_nomerge(char *str)
3212 __setup("slub_nomerge", setup_slub_nomerge);
3214 void *__kmalloc(size_t size, gfp_t flags)
3216 struct kmem_cache *s;
3219 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3220 return kmalloc_large(size, flags);
3222 s = kmalloc_slab(size, flags);
3224 if (unlikely(ZERO_OR_NULL_PTR(s)))
3227 ret = slab_alloc(s, flags, _RET_IP_);
3229 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3233 EXPORT_SYMBOL(__kmalloc);
3236 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3241 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3242 page = alloc_pages_node(node, flags, get_order(size));
3244 ptr = page_address(page);
3246 kmemleak_alloc(ptr, size, 1, flags);
3250 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3252 struct kmem_cache *s;
3255 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3256 ret = kmalloc_large_node(size, flags, node);
3258 trace_kmalloc_node(_RET_IP_, ret,
3259 size, PAGE_SIZE << get_order(size),
3265 s = kmalloc_slab(size, flags);
3267 if (unlikely(ZERO_OR_NULL_PTR(s)))
3270 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3272 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3276 EXPORT_SYMBOL(__kmalloc_node);
3279 size_t ksize(const void *object)
3283 if (unlikely(object == ZERO_SIZE_PTR))
3286 page = virt_to_head_page(object);
3288 if (unlikely(!PageSlab(page))) {
3289 WARN_ON(!PageCompound(page));
3290 return PAGE_SIZE << compound_order(page);
3293 return slab_ksize(page->slab_cache);
3295 EXPORT_SYMBOL(ksize);
3297 #ifdef CONFIG_SLUB_DEBUG
3298 bool verify_mem_not_deleted(const void *x)
3301 void *object = (void *)x;
3302 unsigned long flags;
3305 if (unlikely(ZERO_OR_NULL_PTR(x)))
3308 local_irq_save(flags);
3310 page = virt_to_head_page(x);
3311 if (unlikely(!PageSlab(page))) {
3312 /* maybe it was from stack? */
3318 if (on_freelist(page->slab_cache, page, object)) {
3319 object_err(page->slab_cache, page, object, "Object is on free-list");
3327 local_irq_restore(flags);
3330 EXPORT_SYMBOL(verify_mem_not_deleted);
3333 void kfree(const void *x)
3336 void *object = (void *)x;
3338 trace_kfree(_RET_IP_, x);
3340 if (unlikely(ZERO_OR_NULL_PTR(x)))
3343 page = virt_to_head_page(x);
3344 if (unlikely(!PageSlab(page))) {
3345 BUG_ON(!PageCompound(page));
3347 __free_memcg_kmem_pages(page, compound_order(page));
3350 slab_free(page->slab_cache, page, object, _RET_IP_);
3352 EXPORT_SYMBOL(kfree);
3355 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3356 * the remaining slabs by the number of items in use. The slabs with the
3357 * most items in use come first. New allocations will then fill those up
3358 * and thus they can be removed from the partial lists.
3360 * The slabs with the least items are placed last. This results in them
3361 * being allocated from last increasing the chance that the last objects
3362 * are freed in them.
3364 int kmem_cache_shrink(struct kmem_cache *s)
3368 struct kmem_cache_node *n;
3371 int objects = oo_objects(s->max);
3372 struct list_head *slabs_by_inuse =
3373 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3374 unsigned long flags;
3376 if (!slabs_by_inuse)
3380 for_each_node_state(node, N_NORMAL_MEMORY) {
3381 n = get_node(s, node);
3386 for (i = 0; i < objects; i++)
3387 INIT_LIST_HEAD(slabs_by_inuse + i);
3389 spin_lock_irqsave(&n->list_lock, flags);
3392 * Build lists indexed by the items in use in each slab.
3394 * Note that concurrent frees may occur while we hold the
3395 * list_lock. page->inuse here is the upper limit.
3397 list_for_each_entry_safe(page, t, &n->partial, lru) {
3398 list_move(&page->lru, slabs_by_inuse + page->inuse);
3404 * Rebuild the partial list with the slabs filled up most
3405 * first and the least used slabs at the end.
3407 for (i = objects - 1; i > 0; i--)
3408 list_splice(slabs_by_inuse + i, n->partial.prev);
3410 spin_unlock_irqrestore(&n->list_lock, flags);
3412 /* Release empty slabs */
3413 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3414 discard_slab(s, page);
3417 kfree(slabs_by_inuse);
3420 EXPORT_SYMBOL(kmem_cache_shrink);
3422 #if defined(CONFIG_MEMORY_HOTPLUG)
3423 static int slab_mem_going_offline_callback(void *arg)
3425 struct kmem_cache *s;
3427 mutex_lock(&slab_mutex);
3428 list_for_each_entry(s, &slab_caches, list)
3429 kmem_cache_shrink(s);
3430 mutex_unlock(&slab_mutex);
3435 static void slab_mem_offline_callback(void *arg)
3437 struct kmem_cache_node *n;
3438 struct kmem_cache *s;
3439 struct memory_notify *marg = arg;
3442 offline_node = marg->status_change_nid_normal;
3445 * If the node still has available memory. we need kmem_cache_node
3448 if (offline_node < 0)
3451 mutex_lock(&slab_mutex);
3452 list_for_each_entry(s, &slab_caches, list) {
3453 n = get_node(s, offline_node);
3456 * if n->nr_slabs > 0, slabs still exist on the node
3457 * that is going down. We were unable to free them,
3458 * and offline_pages() function shouldn't call this
3459 * callback. So, we must fail.
3461 BUG_ON(slabs_node(s, offline_node));
3463 s->node[offline_node] = NULL;
3464 kmem_cache_free(kmem_cache_node, n);
3467 mutex_unlock(&slab_mutex);
3470 static int slab_mem_going_online_callback(void *arg)
3472 struct kmem_cache_node *n;
3473 struct kmem_cache *s;
3474 struct memory_notify *marg = arg;
3475 int nid = marg->status_change_nid_normal;
3479 * If the node's memory is already available, then kmem_cache_node is
3480 * already created. Nothing to do.
3486 * We are bringing a node online. No memory is available yet. We must
3487 * allocate a kmem_cache_node structure in order to bring the node
3490 mutex_lock(&slab_mutex);
3491 list_for_each_entry(s, &slab_caches, list) {
3493 * XXX: kmem_cache_alloc_node will fallback to other nodes
3494 * since memory is not yet available from the node that
3497 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3502 init_kmem_cache_node(n);
3506 mutex_unlock(&slab_mutex);
3510 static int slab_memory_callback(struct notifier_block *self,
3511 unsigned long action, void *arg)
3516 case MEM_GOING_ONLINE:
3517 ret = slab_mem_going_online_callback(arg);
3519 case MEM_GOING_OFFLINE:
3520 ret = slab_mem_going_offline_callback(arg);
3523 case MEM_CANCEL_ONLINE:
3524 slab_mem_offline_callback(arg);
3527 case MEM_CANCEL_OFFLINE:
3531 ret = notifier_from_errno(ret);
3537 #endif /* CONFIG_MEMORY_HOTPLUG */
3539 /********************************************************************
3540 * Basic setup of slabs
3541 *******************************************************************/
3544 * Used for early kmem_cache structures that were allocated using
3545 * the page allocator. Allocate them properly then fix up the pointers
3546 * that may be pointing to the wrong kmem_cache structure.
3549 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3552 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3554 memcpy(s, static_cache, kmem_cache->object_size);
3557 * This runs very early, and only the boot processor is supposed to be
3558 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3561 __flush_cpu_slab(s, smp_processor_id());
3562 for_each_node_state(node, N_NORMAL_MEMORY) {
3563 struct kmem_cache_node *n = get_node(s, node);
3567 list_for_each_entry(p, &n->partial, lru)
3570 #ifdef CONFIG_SLUB_DEBUG
3571 list_for_each_entry(p, &n->full, lru)
3576 list_add(&s->list, &slab_caches);
3580 void __init kmem_cache_init(void)
3582 static __initdata struct kmem_cache boot_kmem_cache,
3583 boot_kmem_cache_node;
3585 if (debug_guardpage_minorder())
3588 kmem_cache_node = &boot_kmem_cache_node;
3589 kmem_cache = &boot_kmem_cache;
3591 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3592 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3594 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3596 /* Able to allocate the per node structures */
3597 slab_state = PARTIAL;
3599 create_boot_cache(kmem_cache, "kmem_cache",
3600 offsetof(struct kmem_cache, node) +
3601 nr_node_ids * sizeof(struct kmem_cache_node *),
3602 SLAB_HWCACHE_ALIGN);
3604 kmem_cache = bootstrap(&boot_kmem_cache);
3607 * Allocate kmem_cache_node properly from the kmem_cache slab.
3608 * kmem_cache_node is separately allocated so no need to
3609 * update any list pointers.
3611 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3613 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3614 create_kmalloc_caches(0);
3617 register_cpu_notifier(&slab_notifier);
3621 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3622 " CPUs=%d, Nodes=%d\n",
3624 slub_min_order, slub_max_order, slub_min_objects,
3625 nr_cpu_ids, nr_node_ids);
3628 void __init kmem_cache_init_late(void)
3633 * Find a mergeable slab cache
3635 static int slab_unmergeable(struct kmem_cache *s)
3637 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3644 * We may have set a slab to be unmergeable during bootstrap.
3646 if (s->refcount < 0)
3652 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3653 size_t align, unsigned long flags, const char *name,
3654 void (*ctor)(void *))
3656 struct kmem_cache *s;
3658 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3664 size = ALIGN(size, sizeof(void *));
3665 align = calculate_alignment(flags, align, size);
3666 size = ALIGN(size, align);
3667 flags = kmem_cache_flags(size, flags, name, NULL);
3669 list_for_each_entry(s, &slab_caches, list) {
3670 if (slab_unmergeable(s))
3676 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3679 * Check if alignment is compatible.
3680 * Courtesy of Adrian Drzewiecki
3682 if ((s->size & ~(align - 1)) != s->size)
3685 if (s->size - size >= sizeof(void *))
3688 if (!cache_match_memcg(s, memcg))
3697 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3698 size_t align, unsigned long flags, void (*ctor)(void *))
3700 struct kmem_cache *s;
3702 s = find_mergeable(memcg, size, align, flags, name, ctor);
3706 * Adjust the object sizes so that we clear
3707 * the complete object on kzalloc.
3709 s->object_size = max(s->object_size, (int)size);
3710 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3712 if (sysfs_slab_alias(s, name)) {
3721 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3725 err = kmem_cache_open(s, flags);
3729 /* Mutex is not taken during early boot */
3730 if (slab_state <= UP)
3733 memcg_propagate_slab_attrs(s);
3734 mutex_unlock(&slab_mutex);
3735 err = sysfs_slab_add(s);
3736 mutex_lock(&slab_mutex);
3739 kmem_cache_close(s);
3746 * Use the cpu notifier to insure that the cpu slabs are flushed when
3749 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3750 unsigned long action, void *hcpu)
3752 long cpu = (long)hcpu;
3753 struct kmem_cache *s;
3754 unsigned long flags;
3757 case CPU_UP_CANCELED:
3758 case CPU_UP_CANCELED_FROZEN:
3760 case CPU_DEAD_FROZEN:
3761 mutex_lock(&slab_mutex);
3762 list_for_each_entry(s, &slab_caches, list) {
3763 local_irq_save(flags);
3764 __flush_cpu_slab(s, cpu);
3765 local_irq_restore(flags);
3767 mutex_unlock(&slab_mutex);
3775 static struct notifier_block __cpuinitdata slab_notifier = {
3776 .notifier_call = slab_cpuup_callback
3781 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3783 struct kmem_cache *s;
3786 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3787 return kmalloc_large(size, gfpflags);
3789 s = kmalloc_slab(size, gfpflags);
3791 if (unlikely(ZERO_OR_NULL_PTR(s)))
3794 ret = slab_alloc(s, gfpflags, caller);
3796 /* Honor the call site pointer we received. */
3797 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3803 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3804 int node, unsigned long caller)
3806 struct kmem_cache *s;
3809 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3810 ret = kmalloc_large_node(size, gfpflags, node);
3812 trace_kmalloc_node(caller, ret,
3813 size, PAGE_SIZE << get_order(size),
3819 s = kmalloc_slab(size, gfpflags);
3821 if (unlikely(ZERO_OR_NULL_PTR(s)))
3824 ret = slab_alloc_node(s, gfpflags, node, caller);
3826 /* Honor the call site pointer we received. */
3827 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3834 static int count_inuse(struct page *page)
3839 static int count_total(struct page *page)
3841 return page->objects;
3845 #ifdef CONFIG_SLUB_DEBUG
3846 static int validate_slab(struct kmem_cache *s, struct page *page,
3850 void *addr = page_address(page);
3852 if (!check_slab(s, page) ||
3853 !on_freelist(s, page, NULL))
3856 /* Now we know that a valid freelist exists */
3857 bitmap_zero(map, page->objects);
3859 get_map(s, page, map);
3860 for_each_object(p, s, addr, page->objects) {
3861 if (test_bit(slab_index(p, s, addr), map))
3862 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3866 for_each_object(p, s, addr, page->objects)
3867 if (!test_bit(slab_index(p, s, addr), map))
3868 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3873 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3877 validate_slab(s, page, map);
3881 static int validate_slab_node(struct kmem_cache *s,
3882 struct kmem_cache_node *n, unsigned long *map)
3884 unsigned long count = 0;
3886 unsigned long flags;
3888 spin_lock_irqsave(&n->list_lock, flags);
3890 list_for_each_entry(page, &n->partial, lru) {
3891 validate_slab_slab(s, page, map);
3894 if (count != n->nr_partial)
3895 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3896 "counter=%ld\n", s->name, count, n->nr_partial);
3898 if (!(s->flags & SLAB_STORE_USER))
3901 list_for_each_entry(page, &n->full, lru) {
3902 validate_slab_slab(s, page, map);
3905 if (count != atomic_long_read(&n->nr_slabs))
3906 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3907 "counter=%ld\n", s->name, count,
3908 atomic_long_read(&n->nr_slabs));
3911 spin_unlock_irqrestore(&n->list_lock, flags);
3915 static long validate_slab_cache(struct kmem_cache *s)
3918 unsigned long count = 0;
3919 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3920 sizeof(unsigned long), GFP_KERNEL);
3926 for_each_node_state(node, N_NORMAL_MEMORY) {
3927 struct kmem_cache_node *n = get_node(s, node);
3929 count += validate_slab_node(s, n, map);
3935 * Generate lists of code addresses where slabcache objects are allocated
3940 unsigned long count;
3947 DECLARE_BITMAP(cpus, NR_CPUS);
3953 unsigned long count;
3954 struct location *loc;
3957 static void free_loc_track(struct loc_track *t)
3960 free_pages((unsigned long)t->loc,
3961 get_order(sizeof(struct location) * t->max));
3964 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3969 order = get_order(sizeof(struct location) * max);
3971 l = (void *)__get_free_pages(flags, order);
3976 memcpy(l, t->loc, sizeof(struct location) * t->count);
3984 static int add_location(struct loc_track *t, struct kmem_cache *s,
3985 const struct track *track)
3987 long start, end, pos;
3989 unsigned long caddr;
3990 unsigned long age = jiffies - track->when;
3996 pos = start + (end - start + 1) / 2;
3999 * There is nothing at "end". If we end up there
4000 * we need to add something to before end.
4005 caddr = t->loc[pos].addr;
4006 if (track->addr == caddr) {
4012 if (age < l->min_time)
4014 if (age > l->max_time)
4017 if (track->pid < l->min_pid)
4018 l->min_pid = track->pid;
4019 if (track->pid > l->max_pid)
4020 l->max_pid = track->pid;
4022 cpumask_set_cpu(track->cpu,
4023 to_cpumask(l->cpus));
4025 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4029 if (track->addr < caddr)
4036 * Not found. Insert new tracking element.
4038 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4044 (t->count - pos) * sizeof(struct location));
4047 l->addr = track->addr;
4051 l->min_pid = track->pid;
4052 l->max_pid = track->pid;
4053 cpumask_clear(to_cpumask(l->cpus));
4054 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4055 nodes_clear(l->nodes);
4056 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4060 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4061 struct page *page, enum track_item alloc,
4064 void *addr = page_address(page);
4067 bitmap_zero(map, page->objects);
4068 get_map(s, page, map);
4070 for_each_object(p, s, addr, page->objects)
4071 if (!test_bit(slab_index(p, s, addr), map))
4072 add_location(t, s, get_track(s, p, alloc));
4075 static int list_locations(struct kmem_cache *s, char *buf,
4076 enum track_item alloc)
4080 struct loc_track t = { 0, 0, NULL };
4082 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4083 sizeof(unsigned long), GFP_KERNEL);
4085 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4088 return sprintf(buf, "Out of memory\n");
4090 /* Push back cpu slabs */
4093 for_each_node_state(node, N_NORMAL_MEMORY) {
4094 struct kmem_cache_node *n = get_node(s, node);
4095 unsigned long flags;
4098 if (!atomic_long_read(&n->nr_slabs))
4101 spin_lock_irqsave(&n->list_lock, flags);
4102 list_for_each_entry(page, &n->partial, lru)
4103 process_slab(&t, s, page, alloc, map);
4104 list_for_each_entry(page, &n->full, lru)
4105 process_slab(&t, s, page, alloc, map);
4106 spin_unlock_irqrestore(&n->list_lock, flags);
4109 for (i = 0; i < t.count; i++) {
4110 struct location *l = &t.loc[i];
4112 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4114 len += sprintf(buf + len, "%7ld ", l->count);
4117 len += sprintf(buf + len, "%pS", (void *)l->addr);
4119 len += sprintf(buf + len, "<not-available>");
4121 if (l->sum_time != l->min_time) {
4122 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4124 (long)div_u64(l->sum_time, l->count),
4127 len += sprintf(buf + len, " age=%ld",
4130 if (l->min_pid != l->max_pid)
4131 len += sprintf(buf + len, " pid=%ld-%ld",
4132 l->min_pid, l->max_pid);
4134 len += sprintf(buf + len, " pid=%ld",
4137 if (num_online_cpus() > 1 &&
4138 !cpumask_empty(to_cpumask(l->cpus)) &&
4139 len < PAGE_SIZE - 60) {
4140 len += sprintf(buf + len, " cpus=");
4141 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4142 to_cpumask(l->cpus));
4145 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4146 len < PAGE_SIZE - 60) {
4147 len += sprintf(buf + len, " nodes=");
4148 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4152 len += sprintf(buf + len, "\n");
4158 len += sprintf(buf, "No data\n");
4163 #ifdef SLUB_RESILIENCY_TEST
4164 static void resiliency_test(void)
4168 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4170 printk(KERN_ERR "SLUB resiliency testing\n");
4171 printk(KERN_ERR "-----------------------\n");
4172 printk(KERN_ERR "A. Corruption after allocation\n");
4174 p = kzalloc(16, GFP_KERNEL);
4176 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4177 " 0x12->0x%p\n\n", p + 16);
4179 validate_slab_cache(kmalloc_caches[4]);
4181 /* Hmmm... The next two are dangerous */
4182 p = kzalloc(32, GFP_KERNEL);
4183 p[32 + sizeof(void *)] = 0x34;
4184 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4185 " 0x34 -> -0x%p\n", p);
4187 "If allocated object is overwritten then not detectable\n\n");
4189 validate_slab_cache(kmalloc_caches[5]);
4190 p = kzalloc(64, GFP_KERNEL);
4191 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4193 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4196 "If allocated object is overwritten then not detectable\n\n");
4197 validate_slab_cache(kmalloc_caches[6]);
4199 printk(KERN_ERR "\nB. Corruption after free\n");
4200 p = kzalloc(128, GFP_KERNEL);
4203 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4204 validate_slab_cache(kmalloc_caches[7]);
4206 p = kzalloc(256, GFP_KERNEL);
4209 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4211 validate_slab_cache(kmalloc_caches[8]);
4213 p = kzalloc(512, GFP_KERNEL);
4216 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4217 validate_slab_cache(kmalloc_caches[9]);
4221 static void resiliency_test(void) {};
4226 enum slab_stat_type {
4227 SL_ALL, /* All slabs */
4228 SL_PARTIAL, /* Only partially allocated slabs */
4229 SL_CPU, /* Only slabs used for cpu caches */
4230 SL_OBJECTS, /* Determine allocated objects not slabs */
4231 SL_TOTAL /* Determine object capacity not slabs */
4234 #define SO_ALL (1 << SL_ALL)
4235 #define SO_PARTIAL (1 << SL_PARTIAL)
4236 #define SO_CPU (1 << SL_CPU)
4237 #define SO_OBJECTS (1 << SL_OBJECTS)
4238 #define SO_TOTAL (1 << SL_TOTAL)
4240 static ssize_t show_slab_objects(struct kmem_cache *s,
4241 char *buf, unsigned long flags)
4243 unsigned long total = 0;
4246 unsigned long *nodes;
4247 unsigned long *per_cpu;
4249 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4252 per_cpu = nodes + nr_node_ids;
4254 if (flags & SO_CPU) {
4257 for_each_possible_cpu(cpu) {
4258 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4262 page = ACCESS_ONCE(c->page);
4266 node = page_to_nid(page);
4267 if (flags & SO_TOTAL)
4269 else if (flags & SO_OBJECTS)
4277 page = ACCESS_ONCE(c->partial);
4288 lock_memory_hotplug();
4289 #ifdef CONFIG_SLUB_DEBUG
4290 if (flags & SO_ALL) {
4291 for_each_node_state(node, N_NORMAL_MEMORY) {
4292 struct kmem_cache_node *n = get_node(s, node);
4294 if (flags & SO_TOTAL)
4295 x = atomic_long_read(&n->total_objects);
4296 else if (flags & SO_OBJECTS)
4297 x = atomic_long_read(&n->total_objects) -
4298 count_partial(n, count_free);
4301 x = atomic_long_read(&n->nr_slabs);
4308 if (flags & SO_PARTIAL) {
4309 for_each_node_state(node, N_NORMAL_MEMORY) {
4310 struct kmem_cache_node *n = get_node(s, node);
4312 if (flags & SO_TOTAL)
4313 x = count_partial(n, count_total);
4314 else if (flags & SO_OBJECTS)
4315 x = count_partial(n, count_inuse);
4322 x = sprintf(buf, "%lu", total);
4324 for_each_node_state(node, N_NORMAL_MEMORY)
4326 x += sprintf(buf + x, " N%d=%lu",
4329 unlock_memory_hotplug();
4331 return x + sprintf(buf + x, "\n");
4334 #ifdef CONFIG_SLUB_DEBUG
4335 static int any_slab_objects(struct kmem_cache *s)
4339 for_each_online_node(node) {
4340 struct kmem_cache_node *n = get_node(s, node);
4345 if (atomic_long_read(&n->total_objects))
4352 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4353 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4355 struct slab_attribute {
4356 struct attribute attr;
4357 ssize_t (*show)(struct kmem_cache *s, char *buf);
4358 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4361 #define SLAB_ATTR_RO(_name) \
4362 static struct slab_attribute _name##_attr = \
4363 __ATTR(_name, 0400, _name##_show, NULL)
4365 #define SLAB_ATTR(_name) \
4366 static struct slab_attribute _name##_attr = \
4367 __ATTR(_name, 0600, _name##_show, _name##_store)
4369 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4371 return sprintf(buf, "%d\n", s->size);
4373 SLAB_ATTR_RO(slab_size);
4375 static ssize_t align_show(struct kmem_cache *s, char *buf)
4377 return sprintf(buf, "%d\n", s->align);
4379 SLAB_ATTR_RO(align);
4381 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4383 return sprintf(buf, "%d\n", s->object_size);
4385 SLAB_ATTR_RO(object_size);
4387 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4389 return sprintf(buf, "%d\n", oo_objects(s->oo));
4391 SLAB_ATTR_RO(objs_per_slab);
4393 static ssize_t order_store(struct kmem_cache *s,
4394 const char *buf, size_t length)
4396 unsigned long order;
4399 err = strict_strtoul(buf, 10, &order);
4403 if (order > slub_max_order || order < slub_min_order)
4406 calculate_sizes(s, order);
4410 static ssize_t order_show(struct kmem_cache *s, char *buf)
4412 return sprintf(buf, "%d\n", oo_order(s->oo));
4416 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4418 return sprintf(buf, "%lu\n", s->min_partial);
4421 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4427 err = strict_strtoul(buf, 10, &min);
4431 set_min_partial(s, min);
4434 SLAB_ATTR(min_partial);
4436 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4438 return sprintf(buf, "%u\n", s->cpu_partial);
4441 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4444 unsigned long objects;
4447 err = strict_strtoul(buf, 10, &objects);
4450 if (objects && kmem_cache_debug(s))
4453 s->cpu_partial = objects;
4457 SLAB_ATTR(cpu_partial);
4459 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4463 return sprintf(buf, "%pS\n", s->ctor);
4467 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4469 return sprintf(buf, "%d\n", s->refcount - 1);
4471 SLAB_ATTR_RO(aliases);
4473 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4475 return show_slab_objects(s, buf, SO_PARTIAL);
4477 SLAB_ATTR_RO(partial);
4479 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4481 return show_slab_objects(s, buf, SO_CPU);
4483 SLAB_ATTR_RO(cpu_slabs);
4485 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4487 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4489 SLAB_ATTR_RO(objects);
4491 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4493 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4495 SLAB_ATTR_RO(objects_partial);
4497 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4504 for_each_online_cpu(cpu) {
4505 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4508 pages += page->pages;
4509 objects += page->pobjects;
4513 len = sprintf(buf, "%d(%d)", objects, pages);
4516 for_each_online_cpu(cpu) {
4517 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4519 if (page && len < PAGE_SIZE - 20)
4520 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4521 page->pobjects, page->pages);
4524 return len + sprintf(buf + len, "\n");
4526 SLAB_ATTR_RO(slabs_cpu_partial);
4528 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4530 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4533 static ssize_t reclaim_account_store(struct kmem_cache *s,
4534 const char *buf, size_t length)
4536 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4538 s->flags |= SLAB_RECLAIM_ACCOUNT;
4541 SLAB_ATTR(reclaim_account);
4543 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4545 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4547 SLAB_ATTR_RO(hwcache_align);
4549 #ifdef CONFIG_ZONE_DMA
4550 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4552 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4554 SLAB_ATTR_RO(cache_dma);
4557 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4559 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4561 SLAB_ATTR_RO(destroy_by_rcu);
4563 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4565 return sprintf(buf, "%d\n", s->reserved);
4567 SLAB_ATTR_RO(reserved);
4569 #ifdef CONFIG_SLUB_DEBUG
4570 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4572 return show_slab_objects(s, buf, SO_ALL);
4574 SLAB_ATTR_RO(slabs);
4576 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4578 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4580 SLAB_ATTR_RO(total_objects);
4582 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4584 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4587 static ssize_t sanity_checks_store(struct kmem_cache *s,
4588 const char *buf, size_t length)
4590 s->flags &= ~SLAB_DEBUG_FREE;
4591 if (buf[0] == '1') {
4592 s->flags &= ~__CMPXCHG_DOUBLE;
4593 s->flags |= SLAB_DEBUG_FREE;
4597 SLAB_ATTR(sanity_checks);
4599 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4601 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4604 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4607 s->flags &= ~SLAB_TRACE;
4608 if (buf[0] == '1') {
4609 s->flags &= ~__CMPXCHG_DOUBLE;
4610 s->flags |= SLAB_TRACE;
4616 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4618 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4621 static ssize_t red_zone_store(struct kmem_cache *s,
4622 const char *buf, size_t length)
4624 if (any_slab_objects(s))
4627 s->flags &= ~SLAB_RED_ZONE;
4628 if (buf[0] == '1') {
4629 s->flags &= ~__CMPXCHG_DOUBLE;
4630 s->flags |= SLAB_RED_ZONE;
4632 calculate_sizes(s, -1);
4635 SLAB_ATTR(red_zone);
4637 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4639 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4642 static ssize_t poison_store(struct kmem_cache *s,
4643 const char *buf, size_t length)
4645 if (any_slab_objects(s))
4648 s->flags &= ~SLAB_POISON;
4649 if (buf[0] == '1') {
4650 s->flags &= ~__CMPXCHG_DOUBLE;
4651 s->flags |= SLAB_POISON;
4653 calculate_sizes(s, -1);
4658 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4660 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4663 static ssize_t store_user_store(struct kmem_cache *s,
4664 const char *buf, size_t length)
4666 if (any_slab_objects(s))
4669 s->flags &= ~SLAB_STORE_USER;
4670 if (buf[0] == '1') {
4671 s->flags &= ~__CMPXCHG_DOUBLE;
4672 s->flags |= SLAB_STORE_USER;
4674 calculate_sizes(s, -1);
4677 SLAB_ATTR(store_user);
4679 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4684 static ssize_t validate_store(struct kmem_cache *s,
4685 const char *buf, size_t length)
4689 if (buf[0] == '1') {
4690 ret = validate_slab_cache(s);
4696 SLAB_ATTR(validate);
4698 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4700 if (!(s->flags & SLAB_STORE_USER))
4702 return list_locations(s, buf, TRACK_ALLOC);
4704 SLAB_ATTR_RO(alloc_calls);
4706 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4708 if (!(s->flags & SLAB_STORE_USER))
4710 return list_locations(s, buf, TRACK_FREE);
4712 SLAB_ATTR_RO(free_calls);
4713 #endif /* CONFIG_SLUB_DEBUG */
4715 #ifdef CONFIG_FAILSLAB
4716 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4718 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4721 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4724 s->flags &= ~SLAB_FAILSLAB;
4726 s->flags |= SLAB_FAILSLAB;
4729 SLAB_ATTR(failslab);
4732 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4737 static ssize_t shrink_store(struct kmem_cache *s,
4738 const char *buf, size_t length)
4740 if (buf[0] == '1') {
4741 int rc = kmem_cache_shrink(s);
4752 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4754 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4757 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4758 const char *buf, size_t length)
4760 unsigned long ratio;
4763 err = strict_strtoul(buf, 10, &ratio);
4768 s->remote_node_defrag_ratio = ratio * 10;
4772 SLAB_ATTR(remote_node_defrag_ratio);
4775 #ifdef CONFIG_SLUB_STATS
4776 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4778 unsigned long sum = 0;
4781 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4786 for_each_online_cpu(cpu) {
4787 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4793 len = sprintf(buf, "%lu", sum);
4796 for_each_online_cpu(cpu) {
4797 if (data[cpu] && len < PAGE_SIZE - 20)
4798 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4802 return len + sprintf(buf + len, "\n");
4805 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4809 for_each_online_cpu(cpu)
4810 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4813 #define STAT_ATTR(si, text) \
4814 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4816 return show_stat(s, buf, si); \
4818 static ssize_t text##_store(struct kmem_cache *s, \
4819 const char *buf, size_t length) \
4821 if (buf[0] != '0') \
4823 clear_stat(s, si); \
4828 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4829 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4830 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4831 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4832 STAT_ATTR(FREE_FROZEN, free_frozen);
4833 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4834 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4835 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4836 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4837 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4838 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4839 STAT_ATTR(FREE_SLAB, free_slab);
4840 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4841 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4842 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4843 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4844 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4845 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4846 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4847 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4848 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4849 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4850 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4851 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4852 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4853 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4856 static struct attribute *slab_attrs[] = {
4857 &slab_size_attr.attr,
4858 &object_size_attr.attr,
4859 &objs_per_slab_attr.attr,
4861 &min_partial_attr.attr,
4862 &cpu_partial_attr.attr,
4864 &objects_partial_attr.attr,
4866 &cpu_slabs_attr.attr,
4870 &hwcache_align_attr.attr,
4871 &reclaim_account_attr.attr,
4872 &destroy_by_rcu_attr.attr,
4874 &reserved_attr.attr,
4875 &slabs_cpu_partial_attr.attr,
4876 #ifdef CONFIG_SLUB_DEBUG
4877 &total_objects_attr.attr,
4879 &sanity_checks_attr.attr,
4881 &red_zone_attr.attr,
4883 &store_user_attr.attr,
4884 &validate_attr.attr,
4885 &alloc_calls_attr.attr,
4886 &free_calls_attr.attr,
4888 #ifdef CONFIG_ZONE_DMA
4889 &cache_dma_attr.attr,
4892 &remote_node_defrag_ratio_attr.attr,
4894 #ifdef CONFIG_SLUB_STATS
4895 &alloc_fastpath_attr.attr,
4896 &alloc_slowpath_attr.attr,
4897 &free_fastpath_attr.attr,
4898 &free_slowpath_attr.attr,
4899 &free_frozen_attr.attr,
4900 &free_add_partial_attr.attr,
4901 &free_remove_partial_attr.attr,
4902 &alloc_from_partial_attr.attr,
4903 &alloc_slab_attr.attr,
4904 &alloc_refill_attr.attr,
4905 &alloc_node_mismatch_attr.attr,
4906 &free_slab_attr.attr,
4907 &cpuslab_flush_attr.attr,
4908 &deactivate_full_attr.attr,
4909 &deactivate_empty_attr.attr,
4910 &deactivate_to_head_attr.attr,
4911 &deactivate_to_tail_attr.attr,
4912 &deactivate_remote_frees_attr.attr,
4913 &deactivate_bypass_attr.attr,
4914 &order_fallback_attr.attr,
4915 &cmpxchg_double_fail_attr.attr,
4916 &cmpxchg_double_cpu_fail_attr.attr,
4917 &cpu_partial_alloc_attr.attr,
4918 &cpu_partial_free_attr.attr,
4919 &cpu_partial_node_attr.attr,
4920 &cpu_partial_drain_attr.attr,
4922 #ifdef CONFIG_FAILSLAB
4923 &failslab_attr.attr,
4929 static struct attribute_group slab_attr_group = {
4930 .attrs = slab_attrs,
4933 static ssize_t slab_attr_show(struct kobject *kobj,
4934 struct attribute *attr,
4937 struct slab_attribute *attribute;
4938 struct kmem_cache *s;
4941 attribute = to_slab_attr(attr);
4944 if (!attribute->show)
4947 err = attribute->show(s, buf);
4952 static ssize_t slab_attr_store(struct kobject *kobj,
4953 struct attribute *attr,
4954 const char *buf, size_t len)
4956 struct slab_attribute *attribute;
4957 struct kmem_cache *s;
4960 attribute = to_slab_attr(attr);
4963 if (!attribute->store)
4966 err = attribute->store(s, buf, len);
4967 #ifdef CONFIG_MEMCG_KMEM
4968 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4971 mutex_lock(&slab_mutex);
4972 if (s->max_attr_size < len)
4973 s->max_attr_size = len;
4976 * This is a best effort propagation, so this function's return
4977 * value will be determined by the parent cache only. This is
4978 * basically because not all attributes will have a well
4979 * defined semantics for rollbacks - most of the actions will
4980 * have permanent effects.
4982 * Returning the error value of any of the children that fail
4983 * is not 100 % defined, in the sense that users seeing the
4984 * error code won't be able to know anything about the state of
4987 * Only returning the error code for the parent cache at least
4988 * has well defined semantics. The cache being written to
4989 * directly either failed or succeeded, in which case we loop
4990 * through the descendants with best-effort propagation.
4992 for_each_memcg_cache_index(i) {
4993 struct kmem_cache *c = cache_from_memcg(s, i);
4995 attribute->store(c, buf, len);
4997 mutex_unlock(&slab_mutex);
5003 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5005 #ifdef CONFIG_MEMCG_KMEM
5007 char *buffer = NULL;
5009 if (!is_root_cache(s))
5013 * This mean this cache had no attribute written. Therefore, no point
5014 * in copying default values around
5016 if (!s->max_attr_size)
5019 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5022 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5024 if (!attr || !attr->store || !attr->show)
5028 * It is really bad that we have to allocate here, so we will
5029 * do it only as a fallback. If we actually allocate, though,
5030 * we can just use the allocated buffer until the end.
5032 * Most of the slub attributes will tend to be very small in
5033 * size, but sysfs allows buffers up to a page, so they can
5034 * theoretically happen.
5038 else if (s->max_attr_size < ARRAY_SIZE(mbuf))
5041 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5042 if (WARN_ON(!buffer))
5047 attr->show(s->memcg_params->root_cache, buf);
5048 attr->store(s, buf, strlen(buf));
5052 free_page((unsigned long)buffer);
5056 static const struct sysfs_ops slab_sysfs_ops = {
5057 .show = slab_attr_show,
5058 .store = slab_attr_store,
5061 static struct kobj_type slab_ktype = {
5062 .sysfs_ops = &slab_sysfs_ops,
5065 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5067 struct kobj_type *ktype = get_ktype(kobj);
5069 if (ktype == &slab_ktype)
5074 static const struct kset_uevent_ops slab_uevent_ops = {
5075 .filter = uevent_filter,
5078 static struct kset *slab_kset;
5080 #define ID_STR_LENGTH 64
5082 /* Create a unique string id for a slab cache:
5084 * Format :[flags-]size
5086 static char *create_unique_id(struct kmem_cache *s)
5088 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5095 * First flags affecting slabcache operations. We will only
5096 * get here for aliasable slabs so we do not need to support
5097 * too many flags. The flags here must cover all flags that
5098 * are matched during merging to guarantee that the id is
5101 if (s->flags & SLAB_CACHE_DMA)
5103 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5105 if (s->flags & SLAB_DEBUG_FREE)
5107 if (!(s->flags & SLAB_NOTRACK))
5111 p += sprintf(p, "%07d", s->size);
5113 #ifdef CONFIG_MEMCG_KMEM
5114 if (!is_root_cache(s))
5115 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5118 BUG_ON(p > name + ID_STR_LENGTH - 1);
5122 static int sysfs_slab_add(struct kmem_cache *s)
5126 int unmergeable = slab_unmergeable(s);
5130 * Slabcache can never be merged so we can use the name proper.
5131 * This is typically the case for debug situations. In that
5132 * case we can catch duplicate names easily.
5134 sysfs_remove_link(&slab_kset->kobj, s->name);
5138 * Create a unique name for the slab as a target
5141 name = create_unique_id(s);
5144 s->kobj.kset = slab_kset;
5145 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5147 kobject_put(&s->kobj);
5151 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5153 kobject_del(&s->kobj);
5154 kobject_put(&s->kobj);
5157 kobject_uevent(&s->kobj, KOBJ_ADD);
5159 /* Setup first alias */
5160 sysfs_slab_alias(s, s->name);
5166 static void sysfs_slab_remove(struct kmem_cache *s)
5168 if (slab_state < FULL)
5170 * Sysfs has not been setup yet so no need to remove the
5175 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5176 kobject_del(&s->kobj);
5177 kobject_put(&s->kobj);
5181 * Need to buffer aliases during bootup until sysfs becomes
5182 * available lest we lose that information.
5184 struct saved_alias {
5185 struct kmem_cache *s;
5187 struct saved_alias *next;
5190 static struct saved_alias *alias_list;
5192 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5194 struct saved_alias *al;
5196 if (slab_state == FULL) {
5198 * If we have a leftover link then remove it.
5200 sysfs_remove_link(&slab_kset->kobj, name);
5201 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5204 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5210 al->next = alias_list;
5215 static int __init slab_sysfs_init(void)
5217 struct kmem_cache *s;
5220 mutex_lock(&slab_mutex);
5222 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5224 mutex_unlock(&slab_mutex);
5225 printk(KERN_ERR "Cannot register slab subsystem.\n");
5231 list_for_each_entry(s, &slab_caches, list) {
5232 err = sysfs_slab_add(s);
5234 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5235 " to sysfs\n", s->name);
5238 while (alias_list) {
5239 struct saved_alias *al = alias_list;
5241 alias_list = alias_list->next;
5242 err = sysfs_slab_alias(al->s, al->name);
5244 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5245 " %s to sysfs\n", al->name);
5249 mutex_unlock(&slab_mutex);
5254 __initcall(slab_sysfs_init);
5255 #endif /* CONFIG_SYSFS */
5258 * The /proc/slabinfo ABI
5260 #ifdef CONFIG_SLABINFO
5261 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5263 unsigned long nr_partials = 0;
5264 unsigned long nr_slabs = 0;
5265 unsigned long nr_objs = 0;
5266 unsigned long nr_free = 0;
5269 for_each_online_node(node) {
5270 struct kmem_cache_node *n = get_node(s, node);
5275 nr_partials += n->nr_partial;
5276 nr_slabs += atomic_long_read(&n->nr_slabs);
5277 nr_objs += atomic_long_read(&n->total_objects);
5278 nr_free += count_partial(n, count_free);
5281 sinfo->active_objs = nr_objs - nr_free;
5282 sinfo->num_objs = nr_objs;
5283 sinfo->active_slabs = nr_slabs;
5284 sinfo->num_slabs = nr_slabs;
5285 sinfo->objects_per_slab = oo_objects(s->oo);
5286 sinfo->cache_order = oo_order(s->oo);
5289 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5293 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5294 size_t count, loff_t *ppos)
5298 #endif /* CONFIG_SLABINFO */