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 *);
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) { }
213 static inline void stat(const struct kmem_cache *s, enum stat_item si)
215 #ifdef CONFIG_SLUB_STATS
216 __this_cpu_inc(s->cpu_slab->stat[si]);
220 /********************************************************************
221 * Core slab cache functions
222 *******************************************************************/
224 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
226 return s->node[node];
229 /* Verify that a pointer has an address that is valid within a slab page */
230 static inline int check_valid_pointer(struct kmem_cache *s,
231 struct page *page, const void *object)
238 base = page_address(page);
239 if (object < base || object >= base + page->objects * s->size ||
240 (object - base) % s->size) {
247 static inline void *get_freepointer(struct kmem_cache *s, void *object)
249 return *(void **)(object + s->offset);
252 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
254 prefetch(object + s->offset);
257 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
261 #ifdef CONFIG_DEBUG_PAGEALLOC
262 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
264 p = get_freepointer(s, object);
269 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
271 *(void **)(object + s->offset) = fp;
274 /* Loop over all objects in a slab */
275 #define for_each_object(__p, __s, __addr, __objects) \
276 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
279 /* Determine object index from a given position */
280 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282 return (p - addr) / s->size;
285 static inline size_t slab_ksize(const struct kmem_cache *s)
287 #ifdef CONFIG_SLUB_DEBUG
289 * Debugging requires use of the padding between object
290 * and whatever may come after it.
292 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
293 return s->object_size;
297 * If we have the need to store the freelist pointer
298 * back there or track user information then we can
299 * only use the space before that information.
301 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
304 * Else we can use all the padding etc for the allocation
309 static inline int order_objects(int order, unsigned long size, int reserved)
311 return ((PAGE_SIZE << order) - reserved) / size;
314 static inline struct kmem_cache_order_objects oo_make(int order,
315 unsigned long size, int reserved)
317 struct kmem_cache_order_objects x = {
318 (order << OO_SHIFT) + order_objects(order, size, reserved)
324 static inline int oo_order(struct kmem_cache_order_objects x)
326 return x.x >> OO_SHIFT;
329 static inline int oo_objects(struct kmem_cache_order_objects x)
331 return x.x & OO_MASK;
335 * Per slab locking using the pagelock
337 static __always_inline void slab_lock(struct page *page)
339 bit_spin_lock(PG_locked, &page->flags);
342 static __always_inline void slab_unlock(struct page *page)
344 __bit_spin_unlock(PG_locked, &page->flags);
347 /* Interrupts must be disabled (for the fallback code to work right) */
348 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
349 void *freelist_old, unsigned long counters_old,
350 void *freelist_new, unsigned long counters_new,
353 VM_BUG_ON(!irqs_disabled());
354 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
355 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
356 if (s->flags & __CMPXCHG_DOUBLE) {
357 if (cmpxchg_double(&page->freelist, &page->counters,
358 freelist_old, counters_old,
359 freelist_new, counters_new))
365 if (page->freelist == freelist_old && page->counters == counters_old) {
366 page->freelist = freelist_new;
367 page->counters = counters_new;
375 stat(s, CMPXCHG_DOUBLE_FAIL);
377 #ifdef SLUB_DEBUG_CMPXCHG
378 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
384 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
385 void *freelist_old, unsigned long counters_old,
386 void *freelist_new, unsigned long counters_new,
389 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
390 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
391 if (s->flags & __CMPXCHG_DOUBLE) {
392 if (cmpxchg_double(&page->freelist, &page->counters,
393 freelist_old, counters_old,
394 freelist_new, counters_new))
401 local_irq_save(flags);
403 if (page->freelist == freelist_old && page->counters == counters_old) {
404 page->freelist = freelist_new;
405 page->counters = counters_new;
407 local_irq_restore(flags);
411 local_irq_restore(flags);
415 stat(s, CMPXCHG_DOUBLE_FAIL);
417 #ifdef SLUB_DEBUG_CMPXCHG
418 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
424 #ifdef CONFIG_SLUB_DEBUG
426 * Determine a map of object in use on a page.
428 * Node listlock must be held to guarantee that the page does
429 * not vanish from under us.
431 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
434 void *addr = page_address(page);
436 for (p = page->freelist; p; p = get_freepointer(s, p))
437 set_bit(slab_index(p, s, addr), map);
443 #ifdef CONFIG_SLUB_DEBUG_ON
444 static int slub_debug = DEBUG_DEFAULT_FLAGS;
446 static int slub_debug;
449 static char *slub_debug_slabs;
450 static int disable_higher_order_debug;
455 static void print_section(char *text, u8 *addr, unsigned int length)
457 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
461 static struct track *get_track(struct kmem_cache *s, void *object,
462 enum track_item alloc)
467 p = object + s->offset + sizeof(void *);
469 p = object + s->inuse;
474 static void set_track(struct kmem_cache *s, void *object,
475 enum track_item alloc, unsigned long addr)
477 struct track *p = get_track(s, object, alloc);
480 #ifdef CONFIG_STACKTRACE
481 struct stack_trace trace;
484 trace.nr_entries = 0;
485 trace.max_entries = TRACK_ADDRS_COUNT;
486 trace.entries = p->addrs;
488 save_stack_trace(&trace);
490 /* See rant in lockdep.c */
491 if (trace.nr_entries != 0 &&
492 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
495 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
499 p->cpu = smp_processor_id();
500 p->pid = current->pid;
503 memset(p, 0, sizeof(struct track));
506 static void init_tracking(struct kmem_cache *s, void *object)
508 if (!(s->flags & SLAB_STORE_USER))
511 set_track(s, object, TRACK_FREE, 0UL);
512 set_track(s, object, TRACK_ALLOC, 0UL);
515 static void print_track(const char *s, struct track *t)
520 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
521 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
522 #ifdef CONFIG_STACKTRACE
525 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
527 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
534 static void print_tracking(struct kmem_cache *s, void *object)
536 if (!(s->flags & SLAB_STORE_USER))
539 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
540 print_track("Freed", get_track(s, object, TRACK_FREE));
543 static void print_page_info(struct page *page)
545 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
546 page, page->objects, page->inuse, page->freelist, page->flags);
550 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
556 vsnprintf(buf, sizeof(buf), fmt, args);
558 printk(KERN_ERR "========================================"
559 "=====================================\n");
560 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
561 printk(KERN_ERR "----------------------------------------"
562 "-------------------------------------\n\n");
564 add_taint(TAINT_BAD_PAGE);
567 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
573 vsnprintf(buf, sizeof(buf), fmt, args);
575 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
578 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
580 unsigned int off; /* Offset of last byte */
581 u8 *addr = page_address(page);
583 print_tracking(s, p);
585 print_page_info(page);
587 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
588 p, p - addr, get_freepointer(s, p));
591 print_section("Bytes b4 ", p - 16, 16);
593 print_section("Object ", p, min_t(unsigned long, s->object_size,
595 if (s->flags & SLAB_RED_ZONE)
596 print_section("Redzone ", p + s->object_size,
597 s->inuse - s->object_size);
600 off = s->offset + sizeof(void *);
604 if (s->flags & SLAB_STORE_USER)
605 off += 2 * sizeof(struct track);
608 /* Beginning of the filler is the free pointer */
609 print_section("Padding ", p + off, s->size - off);
614 static void object_err(struct kmem_cache *s, struct page *page,
615 u8 *object, char *reason)
617 slab_bug(s, "%s", reason);
618 print_trailer(s, page, object);
621 static void slab_err(struct kmem_cache *s, struct page *page, const char *fmt, ...)
627 vsnprintf(buf, sizeof(buf), fmt, args);
629 slab_bug(s, "%s", buf);
630 print_page_info(page);
634 static void init_object(struct kmem_cache *s, void *object, u8 val)
638 if (s->flags & __OBJECT_POISON) {
639 memset(p, POISON_FREE, s->object_size - 1);
640 p[s->object_size - 1] = POISON_END;
643 if (s->flags & SLAB_RED_ZONE)
644 memset(p + s->object_size, val, s->inuse - s->object_size);
647 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
648 void *from, void *to)
650 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
651 memset(from, data, to - from);
654 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
655 u8 *object, char *what,
656 u8 *start, unsigned int value, unsigned int bytes)
661 fault = memchr_inv(start, value, bytes);
666 while (end > fault && end[-1] == value)
669 slab_bug(s, "%s overwritten", what);
670 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
671 fault, end - 1, fault[0], value);
672 print_trailer(s, page, object);
674 restore_bytes(s, what, value, fault, end);
682 * Bytes of the object to be managed.
683 * If the freepointer may overlay the object then the free
684 * pointer is the first word of the object.
686 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
689 * object + s->object_size
690 * Padding to reach word boundary. This is also used for Redzoning.
691 * Padding is extended by another word if Redzoning is enabled and
692 * object_size == inuse.
694 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
695 * 0xcc (RED_ACTIVE) for objects in use.
698 * Meta data starts here.
700 * A. Free pointer (if we cannot overwrite object on free)
701 * B. Tracking data for SLAB_STORE_USER
702 * C. Padding to reach required alignment boundary or at mininum
703 * one word if debugging is on to be able to detect writes
704 * before the word boundary.
706 * Padding is done using 0x5a (POISON_INUSE)
709 * Nothing is used beyond s->size.
711 * If slabcaches are merged then the object_size and inuse boundaries are mostly
712 * ignored. And therefore no slab options that rely on these boundaries
713 * may be used with merged slabcaches.
716 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
718 unsigned long off = s->inuse; /* The end of info */
721 /* Freepointer is placed after the object. */
722 off += sizeof(void *);
724 if (s->flags & SLAB_STORE_USER)
725 /* We also have user information there */
726 off += 2 * sizeof(struct track);
731 return check_bytes_and_report(s, page, p, "Object padding",
732 p + off, POISON_INUSE, s->size - off);
735 /* Check the pad bytes at the end of a slab page */
736 static int slab_pad_check(struct kmem_cache *s, struct page *page)
744 if (!(s->flags & SLAB_POISON))
747 start = page_address(page);
748 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
749 end = start + length;
750 remainder = length % s->size;
754 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
757 while (end > fault && end[-1] == POISON_INUSE)
760 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
761 print_section("Padding ", end - remainder, remainder);
763 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
767 static int check_object(struct kmem_cache *s, struct page *page,
768 void *object, u8 val)
771 u8 *endobject = object + s->object_size;
773 if (s->flags & SLAB_RED_ZONE) {
774 if (!check_bytes_and_report(s, page, object, "Redzone",
775 endobject, val, s->inuse - s->object_size))
778 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
779 check_bytes_and_report(s, page, p, "Alignment padding",
780 endobject, POISON_INUSE, s->inuse - s->object_size);
784 if (s->flags & SLAB_POISON) {
785 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
786 (!check_bytes_and_report(s, page, p, "Poison", p,
787 POISON_FREE, s->object_size - 1) ||
788 !check_bytes_and_report(s, page, p, "Poison",
789 p + s->object_size - 1, POISON_END, 1)))
792 * check_pad_bytes cleans up on its own.
794 check_pad_bytes(s, page, p);
797 if (!s->offset && val == SLUB_RED_ACTIVE)
799 * Object and freepointer overlap. Cannot check
800 * freepointer while object is allocated.
804 /* Check free pointer validity */
805 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
806 object_err(s, page, p, "Freepointer corrupt");
808 * No choice but to zap it and thus lose the remainder
809 * of the free objects in this slab. May cause
810 * another error because the object count is now wrong.
812 set_freepointer(s, p, NULL);
818 static int check_slab(struct kmem_cache *s, struct page *page)
822 VM_BUG_ON(!irqs_disabled());
824 if (!PageSlab(page)) {
825 slab_err(s, page, "Not a valid slab page");
829 maxobj = order_objects(compound_order(page), s->size, s->reserved);
830 if (page->objects > maxobj) {
831 slab_err(s, page, "objects %u > max %u",
832 s->name, page->objects, maxobj);
835 if (page->inuse > page->objects) {
836 slab_err(s, page, "inuse %u > max %u",
837 s->name, page->inuse, page->objects);
840 /* Slab_pad_check fixes things up after itself */
841 slab_pad_check(s, page);
846 * Determine if a certain object on a page is on the freelist. Must hold the
847 * slab lock to guarantee that the chains are in a consistent state.
849 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
854 unsigned long max_objects;
857 while (fp && nr <= page->objects) {
860 if (!check_valid_pointer(s, page, fp)) {
862 object_err(s, page, object,
863 "Freechain corrupt");
864 set_freepointer(s, object, NULL);
867 slab_err(s, page, "Freepointer corrupt");
868 page->freelist = NULL;
869 page->inuse = page->objects;
870 slab_fix(s, "Freelist cleared");
876 fp = get_freepointer(s, object);
880 max_objects = order_objects(compound_order(page), s->size, s->reserved);
881 if (max_objects > MAX_OBJS_PER_PAGE)
882 max_objects = MAX_OBJS_PER_PAGE;
884 if (page->objects != max_objects) {
885 slab_err(s, page, "Wrong number of objects. Found %d but "
886 "should be %d", page->objects, max_objects);
887 page->objects = max_objects;
888 slab_fix(s, "Number of objects adjusted.");
890 if (page->inuse != page->objects - nr) {
891 slab_err(s, page, "Wrong object count. Counter is %d but "
892 "counted were %d", page->inuse, page->objects - nr);
893 page->inuse = page->objects - nr;
894 slab_fix(s, "Object count adjusted.");
896 return search == NULL;
899 static void trace(struct kmem_cache *s, struct page *page, void *object,
902 if (s->flags & SLAB_TRACE) {
903 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
905 alloc ? "alloc" : "free",
910 print_section("Object ", (void *)object, s->object_size);
917 * Hooks for other subsystems that check memory allocations. In a typical
918 * production configuration these hooks all should produce no code at all.
920 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
922 flags &= gfp_allowed_mask;
923 lockdep_trace_alloc(flags);
924 might_sleep_if(flags & __GFP_WAIT);
926 return should_failslab(s->object_size, flags, s->flags);
929 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
931 flags &= gfp_allowed_mask;
932 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
933 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
936 static inline void slab_free_hook(struct kmem_cache *s, void *x)
938 kmemleak_free_recursive(x, s->flags);
941 * Trouble is that we may no longer disable interupts in the fast path
942 * So in order to make the debug calls that expect irqs to be
943 * disabled we need to disable interrupts temporarily.
945 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
949 local_irq_save(flags);
950 kmemcheck_slab_free(s, x, s->object_size);
951 debug_check_no_locks_freed(x, s->object_size);
952 local_irq_restore(flags);
955 if (!(s->flags & SLAB_DEBUG_OBJECTS))
956 debug_check_no_obj_freed(x, s->object_size);
960 * Tracking of fully allocated slabs for debugging purposes.
962 * list_lock must be held.
964 static void add_full(struct kmem_cache *s,
965 struct kmem_cache_node *n, struct page *page)
967 if (!(s->flags & SLAB_STORE_USER))
970 list_add(&page->lru, &n->full);
974 * list_lock must be held.
976 static void remove_full(struct kmem_cache *s, struct page *page)
978 if (!(s->flags & SLAB_STORE_USER))
981 list_del(&page->lru);
984 /* Tracking of the number of slabs for debugging purposes */
985 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
987 struct kmem_cache_node *n = get_node(s, node);
989 return atomic_long_read(&n->nr_slabs);
992 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
994 return atomic_long_read(&n->nr_slabs);
997 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
999 struct kmem_cache_node *n = get_node(s, node);
1002 * May be called early in order to allocate a slab for the
1003 * kmem_cache_node structure. Solve the chicken-egg
1004 * dilemma by deferring the increment of the count during
1005 * bootstrap (see early_kmem_cache_node_alloc).
1008 atomic_long_inc(&n->nr_slabs);
1009 atomic_long_add(objects, &n->total_objects);
1012 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1014 struct kmem_cache_node *n = get_node(s, node);
1016 atomic_long_dec(&n->nr_slabs);
1017 atomic_long_sub(objects, &n->total_objects);
1020 /* Object debug checks for alloc/free paths */
1021 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1024 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1027 init_object(s, object, SLUB_RED_INACTIVE);
1028 init_tracking(s, object);
1031 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1032 void *object, unsigned long addr)
1034 if (!check_slab(s, page))
1037 if (!check_valid_pointer(s, page, object)) {
1038 object_err(s, page, object, "Freelist Pointer check fails");
1042 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1045 /* Success perform special debug activities for allocs */
1046 if (s->flags & SLAB_STORE_USER)
1047 set_track(s, object, TRACK_ALLOC, addr);
1048 trace(s, page, object, 1);
1049 init_object(s, object, SLUB_RED_ACTIVE);
1053 if (PageSlab(page)) {
1055 * If this is a slab page then lets do the best we can
1056 * to avoid issues in the future. Marking all objects
1057 * as used avoids touching the remaining objects.
1059 slab_fix(s, "Marking all objects used");
1060 page->inuse = page->objects;
1061 page->freelist = NULL;
1066 static noinline struct kmem_cache_node *free_debug_processing(
1067 struct kmem_cache *s, struct page *page, void *object,
1068 unsigned long addr, unsigned long *flags)
1070 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1072 spin_lock_irqsave(&n->list_lock, *flags);
1075 if (!check_slab(s, page))
1078 if (!check_valid_pointer(s, page, object)) {
1079 slab_err(s, page, "Invalid object pointer 0x%p", object);
1083 if (on_freelist(s, page, object)) {
1084 object_err(s, page, object, "Object already free");
1088 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1091 if (unlikely(s != page->slab_cache)) {
1092 if (!PageSlab(page)) {
1093 slab_err(s, page, "Attempt to free object(0x%p) "
1094 "outside of slab", object);
1095 } else if (!page->slab_cache) {
1097 "SLUB <none>: no slab for object 0x%p.\n",
1101 object_err(s, page, object,
1102 "page slab pointer corrupt.");
1106 if (s->flags & SLAB_STORE_USER)
1107 set_track(s, object, TRACK_FREE, addr);
1108 trace(s, page, object, 0);
1109 init_object(s, object, SLUB_RED_INACTIVE);
1113 * Keep node_lock to preserve integrity
1114 * until the object is actually freed
1120 spin_unlock_irqrestore(&n->list_lock, *flags);
1121 slab_fix(s, "Object at 0x%p not freed", object);
1125 static int __init setup_slub_debug(char *str)
1127 slub_debug = DEBUG_DEFAULT_FLAGS;
1128 if (*str++ != '=' || !*str)
1130 * No options specified. Switch on full debugging.
1136 * No options but restriction on slabs. This means full
1137 * debugging for slabs matching a pattern.
1141 if (tolower(*str) == 'o') {
1143 * Avoid enabling debugging on caches if its minimum order
1144 * would increase as a result.
1146 disable_higher_order_debug = 1;
1153 * Switch off all debugging measures.
1158 * Determine which debug features should be switched on
1160 for (; *str && *str != ','; str++) {
1161 switch (tolower(*str)) {
1163 slub_debug |= SLAB_DEBUG_FREE;
1166 slub_debug |= SLAB_RED_ZONE;
1169 slub_debug |= SLAB_POISON;
1172 slub_debug |= SLAB_STORE_USER;
1175 slub_debug |= SLAB_TRACE;
1178 slub_debug |= SLAB_FAILSLAB;
1181 printk(KERN_ERR "slub_debug option '%c' "
1182 "unknown. skipped\n", *str);
1188 slub_debug_slabs = str + 1;
1193 __setup("slub_debug", setup_slub_debug);
1195 static unsigned long kmem_cache_flags(unsigned long object_size,
1196 unsigned long flags, const char *name,
1197 void (*ctor)(void *))
1200 * Enable debugging if selected on the kernel commandline.
1202 if (slub_debug && (!slub_debug_slabs ||
1203 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1204 flags |= slub_debug;
1209 static inline void setup_object_debug(struct kmem_cache *s,
1210 struct page *page, void *object) {}
1212 static inline int alloc_debug_processing(struct kmem_cache *s,
1213 struct page *page, void *object, unsigned long addr) { return 0; }
1215 static inline struct kmem_cache_node *free_debug_processing(
1216 struct kmem_cache *s, struct page *page, void *object,
1217 unsigned long addr, unsigned long *flags) { return NULL; }
1219 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1221 static inline int check_object(struct kmem_cache *s, struct page *page,
1222 void *object, u8 val) { return 1; }
1223 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1224 struct page *page) {}
1225 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1226 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1227 unsigned long flags, const char *name,
1228 void (*ctor)(void *))
1232 #define slub_debug 0
1234 #define disable_higher_order_debug 0
1236 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1238 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1240 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1242 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1245 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1248 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1251 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1253 #endif /* CONFIG_SLUB_DEBUG */
1256 * Slab allocation and freeing
1258 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1259 struct kmem_cache_order_objects oo)
1261 int order = oo_order(oo);
1263 flags |= __GFP_NOTRACK;
1265 if (node == NUMA_NO_NODE)
1266 return alloc_pages(flags, order);
1268 return alloc_pages_exact_node(node, flags, order);
1271 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1274 struct kmem_cache_order_objects oo = s->oo;
1277 flags &= gfp_allowed_mask;
1279 if (flags & __GFP_WAIT)
1282 flags |= s->allocflags;
1285 * Let the initial higher-order allocation fail under memory pressure
1286 * so we fall-back to the minimum order allocation.
1288 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1290 page = alloc_slab_page(alloc_gfp, node, oo);
1291 if (unlikely(!page)) {
1294 * Allocation may have failed due to fragmentation.
1295 * Try a lower order alloc if possible
1297 page = alloc_slab_page(flags, node, oo);
1300 stat(s, ORDER_FALLBACK);
1303 if (kmemcheck_enabled && page
1304 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1305 int pages = 1 << oo_order(oo);
1307 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1310 * Objects from caches that have a constructor don't get
1311 * cleared when they're allocated, so we need to do it here.
1314 kmemcheck_mark_uninitialized_pages(page, pages);
1316 kmemcheck_mark_unallocated_pages(page, pages);
1319 if (flags & __GFP_WAIT)
1320 local_irq_disable();
1324 page->objects = oo_objects(oo);
1325 mod_zone_page_state(page_zone(page),
1326 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1327 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1333 static void setup_object(struct kmem_cache *s, struct page *page,
1336 setup_object_debug(s, page, object);
1337 if (unlikely(s->ctor))
1341 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1349 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1351 page = allocate_slab(s,
1352 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1356 order = compound_order(page);
1357 inc_slabs_node(s, page_to_nid(page), page->objects);
1358 memcg_bind_pages(s, order);
1359 page->slab_cache = s;
1360 __SetPageSlab(page);
1361 if (page->pfmemalloc)
1362 SetPageSlabPfmemalloc(page);
1364 start = page_address(page);
1366 if (unlikely(s->flags & SLAB_POISON))
1367 memset(start, POISON_INUSE, PAGE_SIZE << order);
1370 for_each_object(p, s, start, page->objects) {
1371 setup_object(s, page, last);
1372 set_freepointer(s, last, p);
1375 setup_object(s, page, last);
1376 set_freepointer(s, last, NULL);
1378 page->freelist = start;
1379 page->inuse = page->objects;
1385 static void __free_slab(struct kmem_cache *s, struct page *page)
1387 int order = compound_order(page);
1388 int pages = 1 << order;
1390 if (kmem_cache_debug(s)) {
1393 slab_pad_check(s, page);
1394 for_each_object(p, s, page_address(page),
1396 check_object(s, page, p, SLUB_RED_INACTIVE);
1399 kmemcheck_free_shadow(page, compound_order(page));
1401 mod_zone_page_state(page_zone(page),
1402 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1403 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1406 __ClearPageSlabPfmemalloc(page);
1407 __ClearPageSlab(page);
1409 memcg_release_pages(s, order);
1410 reset_page_mapcount(page);
1411 if (current->reclaim_state)
1412 current->reclaim_state->reclaimed_slab += pages;
1413 __free_memcg_kmem_pages(page, order);
1416 #define need_reserve_slab_rcu \
1417 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1419 static void rcu_free_slab(struct rcu_head *h)
1423 if (need_reserve_slab_rcu)
1424 page = virt_to_head_page(h);
1426 page = container_of((struct list_head *)h, struct page, lru);
1428 __free_slab(page->slab_cache, page);
1431 static void free_slab(struct kmem_cache *s, struct page *page)
1433 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1434 struct rcu_head *head;
1436 if (need_reserve_slab_rcu) {
1437 int order = compound_order(page);
1438 int offset = (PAGE_SIZE << order) - s->reserved;
1440 VM_BUG_ON(s->reserved != sizeof(*head));
1441 head = page_address(page) + offset;
1444 * RCU free overloads the RCU head over the LRU
1446 head = (void *)&page->lru;
1449 call_rcu(head, rcu_free_slab);
1451 __free_slab(s, page);
1454 static void discard_slab(struct kmem_cache *s, struct page *page)
1456 dec_slabs_node(s, page_to_nid(page), page->objects);
1461 * Management of partially allocated slabs.
1463 * list_lock must be held.
1465 static inline void add_partial(struct kmem_cache_node *n,
1466 struct page *page, int tail)
1469 if (tail == DEACTIVATE_TO_TAIL)
1470 list_add_tail(&page->lru, &n->partial);
1472 list_add(&page->lru, &n->partial);
1476 * list_lock must be held.
1478 static inline void remove_partial(struct kmem_cache_node *n,
1481 list_del(&page->lru);
1486 * Remove slab from the partial list, freeze it and
1487 * return the pointer to the freelist.
1489 * Returns a list of objects or NULL if it fails.
1491 * Must hold list_lock since we modify the partial list.
1493 static inline void *acquire_slab(struct kmem_cache *s,
1494 struct kmem_cache_node *n, struct page *page,
1498 unsigned long counters;
1502 * Zap the freelist and set the frozen bit.
1503 * The old freelist is the list of objects for the
1504 * per cpu allocation list.
1506 freelist = page->freelist;
1507 counters = page->counters;
1508 new.counters = counters;
1510 new.inuse = page->objects;
1511 new.freelist = NULL;
1513 new.freelist = freelist;
1516 VM_BUG_ON(new.frozen);
1519 if (!__cmpxchg_double_slab(s, page,
1521 new.freelist, new.counters,
1525 remove_partial(n, page);
1530 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1531 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1534 * Try to allocate a partial slab from a specific node.
1536 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1537 struct kmem_cache_cpu *c, gfp_t flags)
1539 struct page *page, *page2;
1540 void *object = NULL;
1543 * Racy check. If we mistakenly see no partial slabs then we
1544 * just allocate an empty slab. If we mistakenly try to get a
1545 * partial slab and there is none available then get_partials()
1548 if (!n || !n->nr_partial)
1551 spin_lock(&n->list_lock);
1552 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1556 if (!pfmemalloc_match(page, flags))
1559 t = acquire_slab(s, n, page, object == NULL);
1565 stat(s, ALLOC_FROM_PARTIAL);
1567 available = page->objects - page->inuse;
1569 available = put_cpu_partial(s, page, 0);
1570 stat(s, CPU_PARTIAL_NODE);
1572 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1576 spin_unlock(&n->list_lock);
1581 * Get a page from somewhere. Search in increasing NUMA distances.
1583 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1584 struct kmem_cache_cpu *c)
1587 struct zonelist *zonelist;
1590 enum zone_type high_zoneidx = gfp_zone(flags);
1592 unsigned int cpuset_mems_cookie;
1595 * The defrag ratio allows a configuration of the tradeoffs between
1596 * inter node defragmentation and node local allocations. A lower
1597 * defrag_ratio increases the tendency to do local allocations
1598 * instead of attempting to obtain partial slabs from other nodes.
1600 * If the defrag_ratio is set to 0 then kmalloc() always
1601 * returns node local objects. If the ratio is higher then kmalloc()
1602 * may return off node objects because partial slabs are obtained
1603 * from other nodes and filled up.
1605 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1606 * defrag_ratio = 1000) then every (well almost) allocation will
1607 * first attempt to defrag slab caches on other nodes. This means
1608 * scanning over all nodes to look for partial slabs which may be
1609 * expensive if we do it every time we are trying to find a slab
1610 * with available objects.
1612 if (!s->remote_node_defrag_ratio ||
1613 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1617 cpuset_mems_cookie = get_mems_allowed();
1618 zonelist = node_zonelist(slab_node(), flags);
1619 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1620 struct kmem_cache_node *n;
1622 n = get_node(s, zone_to_nid(zone));
1624 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1625 n->nr_partial > s->min_partial) {
1626 object = get_partial_node(s, n, c, flags);
1629 * Return the object even if
1630 * put_mems_allowed indicated that
1631 * the cpuset mems_allowed was
1632 * updated in parallel. It's a
1633 * harmless race between the alloc
1634 * and the cpuset update.
1636 put_mems_allowed(cpuset_mems_cookie);
1641 } while (!put_mems_allowed(cpuset_mems_cookie));
1647 * Get a partial page, lock it and return it.
1649 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1650 struct kmem_cache_cpu *c)
1653 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1655 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1656 if (object || node != NUMA_NO_NODE)
1659 return get_any_partial(s, flags, c);
1662 #ifdef CONFIG_PREEMPT
1664 * Calculate the next globally unique transaction for disambiguiation
1665 * during cmpxchg. The transactions start with the cpu number and are then
1666 * incremented by CONFIG_NR_CPUS.
1668 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1671 * No preemption supported therefore also no need to check for
1677 static inline unsigned long next_tid(unsigned long tid)
1679 return tid + TID_STEP;
1682 static inline unsigned int tid_to_cpu(unsigned long tid)
1684 return tid % TID_STEP;
1687 static inline unsigned long tid_to_event(unsigned long tid)
1689 return tid / TID_STEP;
1692 static inline unsigned int init_tid(int cpu)
1697 static inline void note_cmpxchg_failure(const char *n,
1698 const struct kmem_cache *s, unsigned long tid)
1700 #ifdef SLUB_DEBUG_CMPXCHG
1701 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1703 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1705 #ifdef CONFIG_PREEMPT
1706 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1707 printk("due to cpu change %d -> %d\n",
1708 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1711 if (tid_to_event(tid) != tid_to_event(actual_tid))
1712 printk("due to cpu running other code. Event %ld->%ld\n",
1713 tid_to_event(tid), tid_to_event(actual_tid));
1715 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1716 actual_tid, tid, next_tid(tid));
1718 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1721 static void init_kmem_cache_cpus(struct kmem_cache *s)
1725 for_each_possible_cpu(cpu)
1726 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1730 * Remove the cpu slab
1732 static void deactivate_slab(struct kmem_cache *s, struct page *page, void *freelist)
1734 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1735 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1737 enum slab_modes l = M_NONE, m = M_NONE;
1739 int tail = DEACTIVATE_TO_HEAD;
1743 if (page->freelist) {
1744 stat(s, DEACTIVATE_REMOTE_FREES);
1745 tail = DEACTIVATE_TO_TAIL;
1749 * Stage one: Free all available per cpu objects back
1750 * to the page freelist while it is still frozen. Leave the
1753 * There is no need to take the list->lock because the page
1756 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1758 unsigned long counters;
1761 prior = page->freelist;
1762 counters = page->counters;
1763 set_freepointer(s, freelist, prior);
1764 new.counters = counters;
1766 VM_BUG_ON(!new.frozen);
1768 } while (!__cmpxchg_double_slab(s, page,
1770 freelist, new.counters,
1771 "drain percpu freelist"));
1773 freelist = nextfree;
1777 * Stage two: Ensure that the page is unfrozen while the
1778 * list presence reflects the actual number of objects
1781 * We setup the list membership and then perform a cmpxchg
1782 * with the count. If there is a mismatch then the page
1783 * is not unfrozen but the page is on the wrong list.
1785 * Then we restart the process which may have to remove
1786 * the page from the list that we just put it on again
1787 * because the number of objects in the slab may have
1792 old.freelist = page->freelist;
1793 old.counters = page->counters;
1794 VM_BUG_ON(!old.frozen);
1796 /* Determine target state of the slab */
1797 new.counters = old.counters;
1800 set_freepointer(s, freelist, old.freelist);
1801 new.freelist = freelist;
1803 new.freelist = old.freelist;
1807 if (!new.inuse && n->nr_partial > s->min_partial)
1809 else if (new.freelist) {
1814 * Taking the spinlock removes the possiblity
1815 * that acquire_slab() will see a slab page that
1818 spin_lock(&n->list_lock);
1822 if (kmem_cache_debug(s) && !lock) {
1825 * This also ensures that the scanning of full
1826 * slabs from diagnostic functions will not see
1829 spin_lock(&n->list_lock);
1837 remove_partial(n, page);
1839 else if (l == M_FULL)
1841 remove_full(s, page);
1843 if (m == M_PARTIAL) {
1845 add_partial(n, page, tail);
1848 } else if (m == M_FULL) {
1850 stat(s, DEACTIVATE_FULL);
1851 add_full(s, n, page);
1857 if (!__cmpxchg_double_slab(s, page,
1858 old.freelist, old.counters,
1859 new.freelist, new.counters,
1864 spin_unlock(&n->list_lock);
1867 stat(s, DEACTIVATE_EMPTY);
1868 discard_slab(s, page);
1874 * Unfreeze all the cpu partial slabs.
1876 * This function must be called with interrupts disabled
1877 * for the cpu using c (or some other guarantee must be there
1878 * to guarantee no concurrent accesses).
1880 static void unfreeze_partials(struct kmem_cache *s,
1881 struct kmem_cache_cpu *c)
1883 struct kmem_cache_node *n = NULL, *n2 = NULL;
1884 struct page *page, *discard_page = NULL;
1886 while ((page = c->partial)) {
1890 c->partial = page->next;
1892 n2 = get_node(s, page_to_nid(page));
1895 spin_unlock(&n->list_lock);
1898 spin_lock(&n->list_lock);
1903 old.freelist = page->freelist;
1904 old.counters = page->counters;
1905 VM_BUG_ON(!old.frozen);
1907 new.counters = old.counters;
1908 new.freelist = old.freelist;
1912 } while (!__cmpxchg_double_slab(s, page,
1913 old.freelist, old.counters,
1914 new.freelist, new.counters,
1915 "unfreezing slab"));
1917 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1918 page->next = discard_page;
1919 discard_page = page;
1921 add_partial(n, page, DEACTIVATE_TO_TAIL);
1922 stat(s, FREE_ADD_PARTIAL);
1927 spin_unlock(&n->list_lock);
1929 while (discard_page) {
1930 page = discard_page;
1931 discard_page = discard_page->next;
1933 stat(s, DEACTIVATE_EMPTY);
1934 discard_slab(s, page);
1940 * Put a page that was just frozen (in __slab_free) into a partial page
1941 * slot if available. This is done without interrupts disabled and without
1942 * preemption disabled. The cmpxchg is racy and may put the partial page
1943 * onto a random cpus partial slot.
1945 * If we did not find a slot then simply move all the partials to the
1946 * per node partial list.
1948 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1950 struct page *oldpage;
1957 oldpage = this_cpu_read(s->cpu_slab->partial);
1960 pobjects = oldpage->pobjects;
1961 pages = oldpage->pages;
1962 if (drain && pobjects > s->cpu_partial) {
1963 unsigned long flags;
1965 * partial array is full. Move the existing
1966 * set to the per node partial list.
1968 local_irq_save(flags);
1969 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
1970 local_irq_restore(flags);
1974 stat(s, CPU_PARTIAL_DRAIN);
1979 pobjects += page->objects - page->inuse;
1981 page->pages = pages;
1982 page->pobjects = pobjects;
1983 page->next = oldpage;
1985 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1989 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1991 stat(s, CPUSLAB_FLUSH);
1992 deactivate_slab(s, c->page, c->freelist);
1994 c->tid = next_tid(c->tid);
2002 * Called from IPI handler with interrupts disabled.
2004 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2006 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2012 unfreeze_partials(s, c);
2016 static void flush_cpu_slab(void *d)
2018 struct kmem_cache *s = d;
2020 __flush_cpu_slab(s, smp_processor_id());
2023 static bool has_cpu_slab(int cpu, void *info)
2025 struct kmem_cache *s = info;
2026 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2028 return c->page || c->partial;
2031 static void flush_all(struct kmem_cache *s)
2033 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2037 * Check if the objects in a per cpu structure fit numa
2038 * locality expectations.
2040 static inline int node_match(struct page *page, int node)
2043 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2049 static int count_free(struct page *page)
2051 return page->objects - page->inuse;
2054 static unsigned long count_partial(struct kmem_cache_node *n,
2055 int (*get_count)(struct page *))
2057 unsigned long flags;
2058 unsigned long x = 0;
2061 spin_lock_irqsave(&n->list_lock, flags);
2062 list_for_each_entry(page, &n->partial, lru)
2063 x += get_count(page);
2064 spin_unlock_irqrestore(&n->list_lock, flags);
2068 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2070 #ifdef CONFIG_SLUB_DEBUG
2071 return atomic_long_read(&n->total_objects);
2077 static noinline void
2078 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2083 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2085 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2086 "default order: %d, min order: %d\n", s->name, s->object_size,
2087 s->size, oo_order(s->oo), oo_order(s->min));
2089 if (oo_order(s->min) > get_order(s->object_size))
2090 printk(KERN_WARNING " %s debugging increased min order, use "
2091 "slub_debug=O to disable.\n", s->name);
2093 for_each_online_node(node) {
2094 struct kmem_cache_node *n = get_node(s, node);
2095 unsigned long nr_slabs;
2096 unsigned long nr_objs;
2097 unsigned long nr_free;
2102 nr_free = count_partial(n, count_free);
2103 nr_slabs = node_nr_slabs(n);
2104 nr_objs = node_nr_objs(n);
2107 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2108 node, nr_slabs, nr_objs, nr_free);
2112 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2113 int node, struct kmem_cache_cpu **pc)
2116 struct kmem_cache_cpu *c = *pc;
2119 freelist = get_partial(s, flags, node, c);
2124 page = new_slab(s, flags, node);
2126 c = __this_cpu_ptr(s->cpu_slab);
2131 * No other reference to the page yet so we can
2132 * muck around with it freely without cmpxchg
2134 freelist = page->freelist;
2135 page->freelist = NULL;
2137 stat(s, ALLOC_SLAB);
2146 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2148 if (unlikely(PageSlabPfmemalloc(page)))
2149 return gfp_pfmemalloc_allowed(gfpflags);
2155 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2156 * or deactivate the page.
2158 * The page is still frozen if the return value is not NULL.
2160 * If this function returns NULL then the page has been unfrozen.
2162 * This function must be called with interrupt disabled.
2164 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2167 unsigned long counters;
2171 freelist = page->freelist;
2172 counters = page->counters;
2174 new.counters = counters;
2175 VM_BUG_ON(!new.frozen);
2177 new.inuse = page->objects;
2178 new.frozen = freelist != NULL;
2180 } while (!__cmpxchg_double_slab(s, page,
2189 * Slow path. The lockless freelist is empty or we need to perform
2192 * Processing is still very fast if new objects have been freed to the
2193 * regular freelist. In that case we simply take over the regular freelist
2194 * as the lockless freelist and zap the regular freelist.
2196 * If that is not working then we fall back to the partial lists. We take the
2197 * first element of the freelist as the object to allocate now and move the
2198 * rest of the freelist to the lockless freelist.
2200 * And if we were unable to get a new slab from the partial slab lists then
2201 * we need to allocate a new slab. This is the slowest path since it involves
2202 * a call to the page allocator and the setup of a new slab.
2204 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2205 unsigned long addr, struct kmem_cache_cpu *c)
2209 unsigned long flags;
2211 local_irq_save(flags);
2212 #ifdef CONFIG_PREEMPT
2214 * We may have been preempted and rescheduled on a different
2215 * cpu before disabling interrupts. Need to reload cpu area
2218 c = this_cpu_ptr(s->cpu_slab);
2226 if (unlikely(!node_match(page, node))) {
2227 stat(s, ALLOC_NODE_MISMATCH);
2228 deactivate_slab(s, page, c->freelist);
2235 * By rights, we should be searching for a slab page that was
2236 * PFMEMALLOC but right now, we are losing the pfmemalloc
2237 * information when the page leaves the per-cpu allocator
2239 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2240 deactivate_slab(s, page, c->freelist);
2246 /* must check again c->freelist in case of cpu migration or IRQ */
2247 freelist = c->freelist;
2251 stat(s, ALLOC_SLOWPATH);
2253 freelist = get_freelist(s, page);
2257 stat(s, DEACTIVATE_BYPASS);
2261 stat(s, ALLOC_REFILL);
2265 * freelist is pointing to the list of objects to be used.
2266 * page is pointing to the page from which the objects are obtained.
2267 * That page must be frozen for per cpu allocations to work.
2269 VM_BUG_ON(!c->page->frozen);
2270 c->freelist = get_freepointer(s, freelist);
2271 c->tid = next_tid(c->tid);
2272 local_irq_restore(flags);
2278 page = c->page = c->partial;
2279 c->partial = page->next;
2280 stat(s, CPU_PARTIAL_ALLOC);
2285 freelist = new_slab_objects(s, gfpflags, node, &c);
2287 if (unlikely(!freelist)) {
2288 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2289 slab_out_of_memory(s, gfpflags, node);
2291 local_irq_restore(flags);
2296 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2299 /* Only entered in the debug case */
2300 if (kmem_cache_debug(s) && !alloc_debug_processing(s, page, freelist, addr))
2301 goto new_slab; /* Slab failed checks. Next slab needed */
2303 deactivate_slab(s, page, get_freepointer(s, freelist));
2306 local_irq_restore(flags);
2311 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2312 * have the fastpath folded into their functions. So no function call
2313 * overhead for requests that can be satisfied on the fastpath.
2315 * The fastpath works by first checking if the lockless freelist can be used.
2316 * If not then __slab_alloc is called for slow processing.
2318 * Otherwise we can simply pick the next object from the lockless free list.
2320 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2321 gfp_t gfpflags, int node, unsigned long addr)
2324 struct kmem_cache_cpu *c;
2328 if (slab_pre_alloc_hook(s, gfpflags))
2331 s = memcg_kmem_get_cache(s, gfpflags);
2335 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2336 * enabled. We may switch back and forth between cpus while
2337 * reading from one cpu area. That does not matter as long
2338 * as we end up on the original cpu again when doing the cmpxchg.
2340 c = __this_cpu_ptr(s->cpu_slab);
2343 * The transaction ids are globally unique per cpu and per operation on
2344 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2345 * occurs on the right processor and that there was no operation on the
2346 * linked list in between.
2351 object = c->freelist;
2353 if (unlikely(!object || !node_match(page, node)))
2354 object = __slab_alloc(s, gfpflags, node, addr, c);
2357 void *next_object = get_freepointer_safe(s, object);
2360 * The cmpxchg will only match if there was no additional
2361 * operation and if we are on the right processor.
2363 * The cmpxchg does the following atomically (without lock semantics!)
2364 * 1. Relocate first pointer to the current per cpu area.
2365 * 2. Verify that tid and freelist have not been changed
2366 * 3. If they were not changed replace tid and freelist
2368 * Since this is without lock semantics the protection is only against
2369 * code executing on this cpu *not* from access by other cpus.
2371 if (unlikely(!this_cpu_cmpxchg_double(
2372 s->cpu_slab->freelist, s->cpu_slab->tid,
2374 next_object, next_tid(tid)))) {
2376 note_cmpxchg_failure("slab_alloc", s, tid);
2379 prefetch_freepointer(s, next_object);
2380 stat(s, ALLOC_FASTPATH);
2383 if (unlikely(gfpflags & __GFP_ZERO) && object)
2384 memset(object, 0, s->object_size);
2386 slab_post_alloc_hook(s, gfpflags, object);
2391 static __always_inline void *slab_alloc(struct kmem_cache *s,
2392 gfp_t gfpflags, unsigned long addr)
2394 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2397 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2399 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2401 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size, s->size, gfpflags);
2405 EXPORT_SYMBOL(kmem_cache_alloc);
2407 #ifdef CONFIG_TRACING
2408 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2410 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2411 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2414 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2416 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2418 void *ret = kmalloc_order(size, flags, order);
2419 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2422 EXPORT_SYMBOL(kmalloc_order_trace);
2426 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2428 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2430 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2431 s->object_size, s->size, gfpflags, node);
2435 EXPORT_SYMBOL(kmem_cache_alloc_node);
2437 #ifdef CONFIG_TRACING
2438 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2440 int node, size_t size)
2442 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2444 trace_kmalloc_node(_RET_IP_, ret,
2445 size, s->size, gfpflags, node);
2448 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2453 * Slow patch handling. This may still be called frequently since objects
2454 * have a longer lifetime than the cpu slabs in most processing loads.
2456 * So we still attempt to reduce cache line usage. Just take the slab
2457 * lock and free the item. If there is no additional partial page
2458 * handling required then we can return immediately.
2460 static void __slab_free(struct kmem_cache *s, struct page *page,
2461 void *x, unsigned long addr)
2464 void **object = (void *)x;
2467 unsigned long counters;
2468 struct kmem_cache_node *n = NULL;
2469 unsigned long uninitialized_var(flags);
2471 stat(s, FREE_SLOWPATH);
2473 if (kmem_cache_debug(s) &&
2474 !(n = free_debug_processing(s, page, x, addr, &flags)))
2479 spin_unlock_irqrestore(&n->list_lock, flags);
2482 prior = page->freelist;
2483 counters = page->counters;
2484 set_freepointer(s, object, prior);
2485 new.counters = counters;
2486 was_frozen = new.frozen;
2488 if ((!new.inuse || !prior) && !was_frozen) {
2490 if (!kmem_cache_debug(s) && !prior)
2493 * Slab was on no list before and will be partially empty
2494 * We can defer the list move and instead freeze it.
2498 else { /* Needs to be taken off a list */
2500 n = get_node(s, page_to_nid(page));
2502 * Speculatively acquire the list_lock.
2503 * If the cmpxchg does not succeed then we may
2504 * drop the list_lock without any processing.
2506 * Otherwise the list_lock will synchronize with
2507 * other processors updating the list of slabs.
2509 spin_lock_irqsave(&n->list_lock, flags);
2514 } while (!cmpxchg_double_slab(s, page,
2516 object, new.counters,
2522 * If we just froze the page then put it onto the
2523 * per cpu partial list.
2525 if (new.frozen && !was_frozen) {
2526 put_cpu_partial(s, page, 1);
2527 stat(s, CPU_PARTIAL_FREE);
2530 * The list lock was not taken therefore no list
2531 * activity can be necessary.
2534 stat(s, FREE_FROZEN);
2538 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2542 * Objects left in the slab. If it was not on the partial list before
2545 if (kmem_cache_debug(s) && unlikely(!prior)) {
2546 remove_full(s, page);
2547 add_partial(n, page, DEACTIVATE_TO_TAIL);
2548 stat(s, FREE_ADD_PARTIAL);
2550 spin_unlock_irqrestore(&n->list_lock, flags);
2556 * Slab on the partial list.
2558 remove_partial(n, page);
2559 stat(s, FREE_REMOVE_PARTIAL);
2561 /* Slab must be on the full list */
2562 remove_full(s, page);
2564 spin_unlock_irqrestore(&n->list_lock, flags);
2566 discard_slab(s, page);
2570 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2571 * can perform fastpath freeing without additional function calls.
2573 * The fastpath is only possible if we are freeing to the current cpu slab
2574 * of this processor. This typically the case if we have just allocated
2577 * If fastpath is not possible then fall back to __slab_free where we deal
2578 * with all sorts of special processing.
2580 static __always_inline void slab_free(struct kmem_cache *s,
2581 struct page *page, void *x, unsigned long addr)
2583 void **object = (void *)x;
2584 struct kmem_cache_cpu *c;
2587 slab_free_hook(s, x);
2591 * Determine the currently cpus per cpu slab.
2592 * The cpu may change afterward. However that does not matter since
2593 * data is retrieved via this pointer. If we are on the same cpu
2594 * during the cmpxchg then the free will succedd.
2596 c = __this_cpu_ptr(s->cpu_slab);
2601 if (likely(page == c->page)) {
2602 set_freepointer(s, object, c->freelist);
2604 if (unlikely(!this_cpu_cmpxchg_double(
2605 s->cpu_slab->freelist, s->cpu_slab->tid,
2607 object, next_tid(tid)))) {
2609 note_cmpxchg_failure("slab_free", s, tid);
2612 stat(s, FREE_FASTPATH);
2614 __slab_free(s, page, x, addr);
2618 void kmem_cache_free(struct kmem_cache *s, void *x)
2620 s = cache_from_obj(s, x);
2623 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2624 trace_kmem_cache_free(_RET_IP_, x);
2626 EXPORT_SYMBOL(kmem_cache_free);
2629 * Object placement in a slab is made very easy because we always start at
2630 * offset 0. If we tune the size of the object to the alignment then we can
2631 * get the required alignment by putting one properly sized object after
2634 * Notice that the allocation order determines the sizes of the per cpu
2635 * caches. Each processor has always one slab available for allocations.
2636 * Increasing the allocation order reduces the number of times that slabs
2637 * must be moved on and off the partial lists and is therefore a factor in
2642 * Mininum / Maximum order of slab pages. This influences locking overhead
2643 * and slab fragmentation. A higher order reduces the number of partial slabs
2644 * and increases the number of allocations possible without having to
2645 * take the list_lock.
2647 static int slub_min_order;
2648 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2649 static int slub_min_objects;
2652 * Merge control. If this is set then no merging of slab caches will occur.
2653 * (Could be removed. This was introduced to pacify the merge skeptics.)
2655 static int slub_nomerge;
2658 * Calculate the order of allocation given an slab object size.
2660 * The order of allocation has significant impact on performance and other
2661 * system components. Generally order 0 allocations should be preferred since
2662 * order 0 does not cause fragmentation in the page allocator. Larger objects
2663 * be problematic to put into order 0 slabs because there may be too much
2664 * unused space left. We go to a higher order if more than 1/16th of the slab
2667 * In order to reach satisfactory performance we must ensure that a minimum
2668 * number of objects is in one slab. Otherwise we may generate too much
2669 * activity on the partial lists which requires taking the list_lock. This is
2670 * less a concern for large slabs though which are rarely used.
2672 * slub_max_order specifies the order where we begin to stop considering the
2673 * number of objects in a slab as critical. If we reach slub_max_order then
2674 * we try to keep the page order as low as possible. So we accept more waste
2675 * of space in favor of a small page order.
2677 * Higher order allocations also allow the placement of more objects in a
2678 * slab and thereby reduce object handling overhead. If the user has
2679 * requested a higher mininum order then we start with that one instead of
2680 * the smallest order which will fit the object.
2682 static inline int slab_order(int size, int min_objects,
2683 int max_order, int fract_leftover, int reserved)
2687 int min_order = slub_min_order;
2689 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2690 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2692 for (order = max(min_order,
2693 fls(min_objects * size - 1) - PAGE_SHIFT);
2694 order <= max_order; order++) {
2696 unsigned long slab_size = PAGE_SIZE << order;
2698 if (slab_size < min_objects * size + reserved)
2701 rem = (slab_size - reserved) % size;
2703 if (rem <= slab_size / fract_leftover)
2711 static inline int calculate_order(int size, int reserved)
2719 * Attempt to find best configuration for a slab. This
2720 * works by first attempting to generate a layout with
2721 * the best configuration and backing off gradually.
2723 * First we reduce the acceptable waste in a slab. Then
2724 * we reduce the minimum objects required in a slab.
2726 min_objects = slub_min_objects;
2728 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2729 max_objects = order_objects(slub_max_order, size, reserved);
2730 min_objects = min(min_objects, max_objects);
2732 while (min_objects > 1) {
2734 while (fraction >= 4) {
2735 order = slab_order(size, min_objects,
2736 slub_max_order, fraction, reserved);
2737 if (order <= slub_max_order)
2745 * We were unable to place multiple objects in a slab. Now
2746 * lets see if we can place a single object there.
2748 order = slab_order(size, 1, slub_max_order, 1, reserved);
2749 if (order <= slub_max_order)
2753 * Doh this slab cannot be placed using slub_max_order.
2755 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2756 if (order < MAX_ORDER)
2762 init_kmem_cache_node(struct kmem_cache_node *n)
2765 spin_lock_init(&n->list_lock);
2766 INIT_LIST_HEAD(&n->partial);
2767 #ifdef CONFIG_SLUB_DEBUG
2768 atomic_long_set(&n->nr_slabs, 0);
2769 atomic_long_set(&n->total_objects, 0);
2770 INIT_LIST_HEAD(&n->full);
2774 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2776 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2777 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2780 * Must align to double word boundary for the double cmpxchg
2781 * instructions to work; see __pcpu_double_call_return_bool().
2783 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2784 2 * sizeof(void *));
2789 init_kmem_cache_cpus(s);
2794 static struct kmem_cache *kmem_cache_node;
2797 * No kmalloc_node yet so do it by hand. We know that this is the first
2798 * slab on the node for this slabcache. There are no concurrent accesses
2801 * Note that this function only works on the kmalloc_node_cache
2802 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2803 * memory on a fresh node that has no slab structures yet.
2805 static void early_kmem_cache_node_alloc(int node)
2808 struct kmem_cache_node *n;
2810 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2812 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2815 if (page_to_nid(page) != node) {
2816 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2818 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2819 "in order to be able to continue\n");
2824 page->freelist = get_freepointer(kmem_cache_node, n);
2827 kmem_cache_node->node[node] = n;
2828 #ifdef CONFIG_SLUB_DEBUG
2829 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2830 init_tracking(kmem_cache_node, n);
2832 init_kmem_cache_node(n);
2833 inc_slabs_node(kmem_cache_node, node, page->objects);
2835 add_partial(n, page, DEACTIVATE_TO_HEAD);
2838 static void free_kmem_cache_nodes(struct kmem_cache *s)
2842 for_each_node_state(node, N_NORMAL_MEMORY) {
2843 struct kmem_cache_node *n = s->node[node];
2846 kmem_cache_free(kmem_cache_node, n);
2848 s->node[node] = NULL;
2852 static int init_kmem_cache_nodes(struct kmem_cache *s)
2856 for_each_node_state(node, N_NORMAL_MEMORY) {
2857 struct kmem_cache_node *n;
2859 if (slab_state == DOWN) {
2860 early_kmem_cache_node_alloc(node);
2863 n = kmem_cache_alloc_node(kmem_cache_node,
2867 free_kmem_cache_nodes(s);
2872 init_kmem_cache_node(n);
2877 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2879 if (min < MIN_PARTIAL)
2881 else if (min > MAX_PARTIAL)
2883 s->min_partial = min;
2887 * calculate_sizes() determines the order and the distribution of data within
2890 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2892 unsigned long flags = s->flags;
2893 unsigned long size = s->object_size;
2897 * Round up object size to the next word boundary. We can only
2898 * place the free pointer at word boundaries and this determines
2899 * the possible location of the free pointer.
2901 size = ALIGN(size, sizeof(void *));
2903 #ifdef CONFIG_SLUB_DEBUG
2905 * Determine if we can poison the object itself. If the user of
2906 * the slab may touch the object after free or before allocation
2907 * then we should never poison the object itself.
2909 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2911 s->flags |= __OBJECT_POISON;
2913 s->flags &= ~__OBJECT_POISON;
2917 * If we are Redzoning then check if there is some space between the
2918 * end of the object and the free pointer. If not then add an
2919 * additional word to have some bytes to store Redzone information.
2921 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2922 size += sizeof(void *);
2926 * With that we have determined the number of bytes in actual use
2927 * by the object. This is the potential offset to the free pointer.
2931 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2934 * Relocate free pointer after the object if it is not
2935 * permitted to overwrite the first word of the object on
2938 * This is the case if we do RCU, have a constructor or
2939 * destructor or are poisoning the objects.
2942 size += sizeof(void *);
2945 #ifdef CONFIG_SLUB_DEBUG
2946 if (flags & SLAB_STORE_USER)
2948 * Need to store information about allocs and frees after
2951 size += 2 * sizeof(struct track);
2953 if (flags & SLAB_RED_ZONE)
2955 * Add some empty padding so that we can catch
2956 * overwrites from earlier objects rather than let
2957 * tracking information or the free pointer be
2958 * corrupted if a user writes before the start
2961 size += sizeof(void *);
2965 * SLUB stores one object immediately after another beginning from
2966 * offset 0. In order to align the objects we have to simply size
2967 * each object to conform to the alignment.
2969 size = ALIGN(size, s->align);
2971 if (forced_order >= 0)
2972 order = forced_order;
2974 order = calculate_order(size, s->reserved);
2981 s->allocflags |= __GFP_COMP;
2983 if (s->flags & SLAB_CACHE_DMA)
2984 s->allocflags |= SLUB_DMA;
2986 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2987 s->allocflags |= __GFP_RECLAIMABLE;
2990 * Determine the number of objects per slab
2992 s->oo = oo_make(order, size, s->reserved);
2993 s->min = oo_make(get_order(size), size, s->reserved);
2994 if (oo_objects(s->oo) > oo_objects(s->max))
2997 return !!oo_objects(s->oo);
3000 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3002 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3005 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3006 s->reserved = sizeof(struct rcu_head);
3008 if (!calculate_sizes(s, -1))
3010 if (disable_higher_order_debug) {
3012 * Disable debugging flags that store metadata if the min slab
3015 if (get_order(s->size) > get_order(s->object_size)) {
3016 s->flags &= ~DEBUG_METADATA_FLAGS;
3018 if (!calculate_sizes(s, -1))
3023 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3024 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3025 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3026 /* Enable fast mode */
3027 s->flags |= __CMPXCHG_DOUBLE;
3031 * The larger the object size is, the more pages we want on the partial
3032 * list to avoid pounding the page allocator excessively.
3034 set_min_partial(s, ilog2(s->size) / 2);
3037 * cpu_partial determined the maximum number of objects kept in the
3038 * per cpu partial lists of a processor.
3040 * Per cpu partial lists mainly contain slabs that just have one
3041 * object freed. If they are used for allocation then they can be
3042 * filled up again with minimal effort. The slab will never hit the
3043 * per node partial lists and therefore no locking will be required.
3045 * This setting also determines
3047 * A) The number of objects from per cpu partial slabs dumped to the
3048 * per node list when we reach the limit.
3049 * B) The number of objects in cpu partial slabs to extract from the
3050 * per node list when we run out of per cpu objects. We only fetch 50%
3051 * to keep some capacity around for frees.
3053 if (kmem_cache_debug(s))
3055 else if (s->size >= PAGE_SIZE)
3057 else if (s->size >= 1024)
3059 else if (s->size >= 256)
3060 s->cpu_partial = 13;
3062 s->cpu_partial = 30;
3065 s->remote_node_defrag_ratio = 1000;
3067 if (!init_kmem_cache_nodes(s))
3070 if (alloc_kmem_cache_cpus(s))
3073 free_kmem_cache_nodes(s);
3075 if (flags & SLAB_PANIC)
3076 panic("Cannot create slab %s size=%lu realsize=%u "
3077 "order=%u offset=%u flags=%lx\n",
3078 s->name, (unsigned long)s->size, s->size, oo_order(s->oo),
3083 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3086 #ifdef CONFIG_SLUB_DEBUG
3087 void *addr = page_address(page);
3089 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3090 sizeof(long), GFP_ATOMIC);
3093 slab_err(s, page, text, s->name);
3096 get_map(s, page, map);
3097 for_each_object(p, s, addr, page->objects) {
3099 if (!test_bit(slab_index(p, s, addr), map)) {
3100 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3102 print_tracking(s, p);
3111 * Attempt to free all partial slabs on a node.
3112 * This is called from kmem_cache_close(). We must be the last thread
3113 * using the cache and therefore we do not need to lock anymore.
3115 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3117 struct page *page, *h;
3119 list_for_each_entry_safe(page, h, &n->partial, lru) {
3121 remove_partial(n, page);
3122 discard_slab(s, page);
3124 list_slab_objects(s, page,
3125 "Objects remaining in %s on kmem_cache_close()");
3131 * Release all resources used by a slab cache.
3133 static inline int kmem_cache_close(struct kmem_cache *s)
3138 /* Attempt to free all objects */
3139 for_each_node_state(node, N_NORMAL_MEMORY) {
3140 struct kmem_cache_node *n = get_node(s, node);
3143 if (n->nr_partial || slabs_node(s, node))
3146 free_percpu(s->cpu_slab);
3147 free_kmem_cache_nodes(s);
3151 int __kmem_cache_shutdown(struct kmem_cache *s)
3153 int rc = kmem_cache_close(s);
3156 sysfs_slab_remove(s);
3161 /********************************************************************
3163 *******************************************************************/
3165 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3166 EXPORT_SYMBOL(kmalloc_caches);
3168 #ifdef CONFIG_ZONE_DMA
3169 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3172 static int __init setup_slub_min_order(char *str)
3174 get_option(&str, &slub_min_order);
3179 __setup("slub_min_order=", setup_slub_min_order);
3181 static int __init setup_slub_max_order(char *str)
3183 get_option(&str, &slub_max_order);
3184 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3189 __setup("slub_max_order=", setup_slub_max_order);
3191 static int __init setup_slub_min_objects(char *str)
3193 get_option(&str, &slub_min_objects);
3198 __setup("slub_min_objects=", setup_slub_min_objects);
3200 static int __init setup_slub_nomerge(char *str)
3206 __setup("slub_nomerge", setup_slub_nomerge);
3209 * Conversion table for small slabs sizes / 8 to the index in the
3210 * kmalloc array. This is necessary for slabs < 192 since we have non power
3211 * of two cache sizes there. The size of larger slabs can be determined using
3214 static s8 size_index[24] = {
3241 static inline int size_index_elem(size_t bytes)
3243 return (bytes - 1) / 8;
3246 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3252 return ZERO_SIZE_PTR;
3254 index = size_index[size_index_elem(size)];
3256 index = fls(size - 1);
3258 #ifdef CONFIG_ZONE_DMA
3259 if (unlikely((flags & SLUB_DMA)))
3260 return kmalloc_dma_caches[index];
3263 return kmalloc_caches[index];
3266 void *__kmalloc(size_t size, gfp_t flags)
3268 struct kmem_cache *s;
3271 if (unlikely(size > SLUB_MAX_SIZE))
3272 return kmalloc_large(size, flags);
3274 s = get_slab(size, flags);
3276 if (unlikely(ZERO_OR_NULL_PTR(s)))
3279 ret = slab_alloc(s, flags, _RET_IP_);
3281 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3285 EXPORT_SYMBOL(__kmalloc);
3288 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3293 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3294 page = alloc_pages_node(node, flags, get_order(size));
3296 ptr = page_address(page);
3298 kmemleak_alloc(ptr, size, 1, flags);
3302 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3304 struct kmem_cache *s;
3307 if (unlikely(size > SLUB_MAX_SIZE)) {
3308 ret = kmalloc_large_node(size, flags, node);
3310 trace_kmalloc_node(_RET_IP_, ret,
3311 size, PAGE_SIZE << get_order(size),
3317 s = get_slab(size, flags);
3319 if (unlikely(ZERO_OR_NULL_PTR(s)))
3322 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3324 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3328 EXPORT_SYMBOL(__kmalloc_node);
3331 size_t ksize(const void *object)
3335 if (unlikely(object == ZERO_SIZE_PTR))
3338 page = virt_to_head_page(object);
3340 if (unlikely(!PageSlab(page))) {
3341 WARN_ON(!PageCompound(page));
3342 return PAGE_SIZE << compound_order(page);
3345 return slab_ksize(page->slab_cache);
3347 EXPORT_SYMBOL(ksize);
3349 #ifdef CONFIG_SLUB_DEBUG
3350 bool verify_mem_not_deleted(const void *x)
3353 void *object = (void *)x;
3354 unsigned long flags;
3357 if (unlikely(ZERO_OR_NULL_PTR(x)))
3360 local_irq_save(flags);
3362 page = virt_to_head_page(x);
3363 if (unlikely(!PageSlab(page))) {
3364 /* maybe it was from stack? */
3370 if (on_freelist(page->slab_cache, page, object)) {
3371 object_err(page->slab_cache, page, object, "Object is on free-list");
3379 local_irq_restore(flags);
3382 EXPORT_SYMBOL(verify_mem_not_deleted);
3385 void kfree(const void *x)
3388 void *object = (void *)x;
3390 trace_kfree(_RET_IP_, x);
3392 if (unlikely(ZERO_OR_NULL_PTR(x)))
3395 page = virt_to_head_page(x);
3396 if (unlikely(!PageSlab(page))) {
3397 BUG_ON(!PageCompound(page));
3399 __free_memcg_kmem_pages(page, compound_order(page));
3402 slab_free(page->slab_cache, page, object, _RET_IP_);
3404 EXPORT_SYMBOL(kfree);
3407 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3408 * the remaining slabs by the number of items in use. The slabs with the
3409 * most items in use come first. New allocations will then fill those up
3410 * and thus they can be removed from the partial lists.
3412 * The slabs with the least items are placed last. This results in them
3413 * being allocated from last increasing the chance that the last objects
3414 * are freed in them.
3416 int kmem_cache_shrink(struct kmem_cache *s)
3420 struct kmem_cache_node *n;
3423 int objects = oo_objects(s->max);
3424 struct list_head *slabs_by_inuse =
3425 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3426 unsigned long flags;
3428 if (!slabs_by_inuse)
3432 for_each_node_state(node, N_NORMAL_MEMORY) {
3433 n = get_node(s, node);
3438 for (i = 0; i < objects; i++)
3439 INIT_LIST_HEAD(slabs_by_inuse + i);
3441 spin_lock_irqsave(&n->list_lock, flags);
3444 * Build lists indexed by the items in use in each slab.
3446 * Note that concurrent frees may occur while we hold the
3447 * list_lock. page->inuse here is the upper limit.
3449 list_for_each_entry_safe(page, t, &n->partial, lru) {
3450 list_move(&page->lru, slabs_by_inuse + page->inuse);
3456 * Rebuild the partial list with the slabs filled up most
3457 * first and the least used slabs at the end.
3459 for (i = objects - 1; i > 0; i--)
3460 list_splice(slabs_by_inuse + i, n->partial.prev);
3462 spin_unlock_irqrestore(&n->list_lock, flags);
3464 /* Release empty slabs */
3465 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3466 discard_slab(s, page);
3469 kfree(slabs_by_inuse);
3472 EXPORT_SYMBOL(kmem_cache_shrink);
3474 #if defined(CONFIG_MEMORY_HOTPLUG)
3475 static int slab_mem_going_offline_callback(void *arg)
3477 struct kmem_cache *s;
3479 mutex_lock(&slab_mutex);
3480 list_for_each_entry(s, &slab_caches, list)
3481 kmem_cache_shrink(s);
3482 mutex_unlock(&slab_mutex);
3487 static void slab_mem_offline_callback(void *arg)
3489 struct kmem_cache_node *n;
3490 struct kmem_cache *s;
3491 struct memory_notify *marg = arg;
3494 offline_node = marg->status_change_nid_normal;
3497 * If the node still has available memory. we need kmem_cache_node
3500 if (offline_node < 0)
3503 mutex_lock(&slab_mutex);
3504 list_for_each_entry(s, &slab_caches, list) {
3505 n = get_node(s, offline_node);
3508 * if n->nr_slabs > 0, slabs still exist on the node
3509 * that is going down. We were unable to free them,
3510 * and offline_pages() function shouldn't call this
3511 * callback. So, we must fail.
3513 BUG_ON(slabs_node(s, offline_node));
3515 s->node[offline_node] = NULL;
3516 kmem_cache_free(kmem_cache_node, n);
3519 mutex_unlock(&slab_mutex);
3522 static int slab_mem_going_online_callback(void *arg)
3524 struct kmem_cache_node *n;
3525 struct kmem_cache *s;
3526 struct memory_notify *marg = arg;
3527 int nid = marg->status_change_nid_normal;
3531 * If the node's memory is already available, then kmem_cache_node is
3532 * already created. Nothing to do.
3538 * We are bringing a node online. No memory is available yet. We must
3539 * allocate a kmem_cache_node structure in order to bring the node
3542 mutex_lock(&slab_mutex);
3543 list_for_each_entry(s, &slab_caches, list) {
3545 * XXX: kmem_cache_alloc_node will fallback to other nodes
3546 * since memory is not yet available from the node that
3549 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3554 init_kmem_cache_node(n);
3558 mutex_unlock(&slab_mutex);
3562 static int slab_memory_callback(struct notifier_block *self,
3563 unsigned long action, void *arg)
3568 case MEM_GOING_ONLINE:
3569 ret = slab_mem_going_online_callback(arg);
3571 case MEM_GOING_OFFLINE:
3572 ret = slab_mem_going_offline_callback(arg);
3575 case MEM_CANCEL_ONLINE:
3576 slab_mem_offline_callback(arg);
3579 case MEM_CANCEL_OFFLINE:
3583 ret = notifier_from_errno(ret);
3589 #endif /* CONFIG_MEMORY_HOTPLUG */
3591 /********************************************************************
3592 * Basic setup of slabs
3593 *******************************************************************/
3596 * Used for early kmem_cache structures that were allocated using
3597 * the page allocator. Allocate them properly then fix up the pointers
3598 * that may be pointing to the wrong kmem_cache structure.
3601 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3604 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3606 memcpy(s, static_cache, kmem_cache->object_size);
3608 for_each_node_state(node, N_NORMAL_MEMORY) {
3609 struct kmem_cache_node *n = get_node(s, node);
3613 list_for_each_entry(p, &n->partial, lru)
3616 #ifdef CONFIG_SLUB_DEBUG
3617 list_for_each_entry(p, &n->full, lru)
3622 list_add(&s->list, &slab_caches);
3626 void __init kmem_cache_init(void)
3628 static __initdata struct kmem_cache boot_kmem_cache,
3629 boot_kmem_cache_node;
3633 if (debug_guardpage_minorder())
3636 kmem_cache_node = &boot_kmem_cache_node;
3637 kmem_cache = &boot_kmem_cache;
3639 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3640 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3642 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3644 /* Able to allocate the per node structures */
3645 slab_state = PARTIAL;
3647 create_boot_cache(kmem_cache, "kmem_cache",
3648 offsetof(struct kmem_cache, node) +
3649 nr_node_ids * sizeof(struct kmem_cache_node *),
3650 SLAB_HWCACHE_ALIGN);
3652 kmem_cache = bootstrap(&boot_kmem_cache);
3655 * Allocate kmem_cache_node properly from the kmem_cache slab.
3656 * kmem_cache_node is separately allocated so no need to
3657 * update any list pointers.
3659 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3661 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3664 * Patch up the size_index table if we have strange large alignment
3665 * requirements for the kmalloc array. This is only the case for
3666 * MIPS it seems. The standard arches will not generate any code here.
3668 * Largest permitted alignment is 256 bytes due to the way we
3669 * handle the index determination for the smaller caches.
3671 * Make sure that nothing crazy happens if someone starts tinkering
3672 * around with ARCH_KMALLOC_MINALIGN
3674 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3675 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3677 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3678 int elem = size_index_elem(i);
3679 if (elem >= ARRAY_SIZE(size_index))
3681 size_index[elem] = KMALLOC_SHIFT_LOW;
3684 if (KMALLOC_MIN_SIZE == 64) {
3686 * The 96 byte size cache is not used if the alignment
3689 for (i = 64 + 8; i <= 96; i += 8)
3690 size_index[size_index_elem(i)] = 7;
3691 } else if (KMALLOC_MIN_SIZE == 128) {
3693 * The 192 byte sized cache is not used if the alignment
3694 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3697 for (i = 128 + 8; i <= 192; i += 8)
3698 size_index[size_index_elem(i)] = 8;
3701 /* Caches that are not of the two-to-the-power-of size */
3702 if (KMALLOC_MIN_SIZE <= 32) {
3703 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3707 if (KMALLOC_MIN_SIZE <= 64) {
3708 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3712 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3713 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3719 /* Provide the correct kmalloc names now that the caches are up */
3720 if (KMALLOC_MIN_SIZE <= 32) {
3721 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3722 BUG_ON(!kmalloc_caches[1]->name);
3725 if (KMALLOC_MIN_SIZE <= 64) {
3726 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3727 BUG_ON(!kmalloc_caches[2]->name);
3730 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3731 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3734 kmalloc_caches[i]->name = s;
3738 register_cpu_notifier(&slab_notifier);
3741 #ifdef CONFIG_ZONE_DMA
3742 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3743 struct kmem_cache *s = kmalloc_caches[i];
3746 char *name = kasprintf(GFP_NOWAIT,
3747 "dma-kmalloc-%d", s->object_size);
3750 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3751 s->object_size, SLAB_CACHE_DMA);
3756 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3757 " CPUs=%d, Nodes=%d\n",
3758 caches, cache_line_size(),
3759 slub_min_order, slub_max_order, slub_min_objects,
3760 nr_cpu_ids, nr_node_ids);
3763 void __init kmem_cache_init_late(void)
3768 * Find a mergeable slab cache
3770 static int slab_unmergeable(struct kmem_cache *s)
3772 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3779 * We may have set a slab to be unmergeable during bootstrap.
3781 if (s->refcount < 0)
3787 static struct kmem_cache *find_mergeable(struct mem_cgroup *memcg, size_t size,
3788 size_t align, unsigned long flags, const char *name,
3789 void (*ctor)(void *))
3791 struct kmem_cache *s;
3793 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3799 size = ALIGN(size, sizeof(void *));
3800 align = calculate_alignment(flags, align, size);
3801 size = ALIGN(size, align);
3802 flags = kmem_cache_flags(size, flags, name, NULL);
3804 list_for_each_entry(s, &slab_caches, list) {
3805 if (slab_unmergeable(s))
3811 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3814 * Check if alignment is compatible.
3815 * Courtesy of Adrian Drzewiecki
3817 if ((s->size & ~(align - 1)) != s->size)
3820 if (s->size - size >= sizeof(void *))
3823 if (!cache_match_memcg(s, memcg))
3832 __kmem_cache_alias(struct mem_cgroup *memcg, const char *name, size_t size,
3833 size_t align, unsigned long flags, void (*ctor)(void *))
3835 struct kmem_cache *s;
3837 s = find_mergeable(memcg, size, align, flags, name, ctor);
3841 * Adjust the object sizes so that we clear
3842 * the complete object on kzalloc.
3844 s->object_size = max(s->object_size, (int)size);
3845 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3847 if (sysfs_slab_alias(s, name)) {
3856 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3860 err = kmem_cache_open(s, flags);
3864 /* Mutex is not taken during early boot */
3865 if (slab_state <= UP)
3868 mutex_unlock(&slab_mutex);
3869 err = sysfs_slab_add(s);
3870 mutex_lock(&slab_mutex);
3873 kmem_cache_close(s);
3880 * Use the cpu notifier to insure that the cpu slabs are flushed when
3883 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3884 unsigned long action, void *hcpu)
3886 long cpu = (long)hcpu;
3887 struct kmem_cache *s;
3888 unsigned long flags;
3891 case CPU_UP_CANCELED:
3892 case CPU_UP_CANCELED_FROZEN:
3894 case CPU_DEAD_FROZEN:
3895 mutex_lock(&slab_mutex);
3896 list_for_each_entry(s, &slab_caches, list) {
3897 local_irq_save(flags);
3898 __flush_cpu_slab(s, cpu);
3899 local_irq_restore(flags);
3901 mutex_unlock(&slab_mutex);
3909 static struct notifier_block __cpuinitdata slab_notifier = {
3910 .notifier_call = slab_cpuup_callback
3915 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3917 struct kmem_cache *s;
3920 if (unlikely(size > SLUB_MAX_SIZE))
3921 return kmalloc_large(size, gfpflags);
3923 s = get_slab(size, gfpflags);
3925 if (unlikely(ZERO_OR_NULL_PTR(s)))
3928 ret = slab_alloc(s, gfpflags, caller);
3930 /* Honor the call site pointer we received. */
3931 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3937 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3938 int node, unsigned long caller)
3940 struct kmem_cache *s;
3943 if (unlikely(size > SLUB_MAX_SIZE)) {
3944 ret = kmalloc_large_node(size, gfpflags, node);
3946 trace_kmalloc_node(caller, ret,
3947 size, PAGE_SIZE << get_order(size),
3953 s = get_slab(size, gfpflags);
3955 if (unlikely(ZERO_OR_NULL_PTR(s)))
3958 ret = slab_alloc_node(s, gfpflags, node, caller);
3960 /* Honor the call site pointer we received. */
3961 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3968 static int count_inuse(struct page *page)
3973 static int count_total(struct page *page)
3975 return page->objects;
3979 #ifdef CONFIG_SLUB_DEBUG
3980 static int validate_slab(struct kmem_cache *s, struct page *page,
3984 void *addr = page_address(page);
3986 if (!check_slab(s, page) ||
3987 !on_freelist(s, page, NULL))
3990 /* Now we know that a valid freelist exists */
3991 bitmap_zero(map, page->objects);
3993 get_map(s, page, map);
3994 for_each_object(p, s, addr, page->objects) {
3995 if (test_bit(slab_index(p, s, addr), map))
3996 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4000 for_each_object(p, s, addr, page->objects)
4001 if (!test_bit(slab_index(p, s, addr), map))
4002 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4007 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4011 validate_slab(s, page, map);
4015 static int validate_slab_node(struct kmem_cache *s,
4016 struct kmem_cache_node *n, unsigned long *map)
4018 unsigned long count = 0;
4020 unsigned long flags;
4022 spin_lock_irqsave(&n->list_lock, flags);
4024 list_for_each_entry(page, &n->partial, lru) {
4025 validate_slab_slab(s, page, map);
4028 if (count != n->nr_partial)
4029 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4030 "counter=%ld\n", s->name, count, n->nr_partial);
4032 if (!(s->flags & SLAB_STORE_USER))
4035 list_for_each_entry(page, &n->full, lru) {
4036 validate_slab_slab(s, page, map);
4039 if (count != atomic_long_read(&n->nr_slabs))
4040 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4041 "counter=%ld\n", s->name, count,
4042 atomic_long_read(&n->nr_slabs));
4045 spin_unlock_irqrestore(&n->list_lock, flags);
4049 static long validate_slab_cache(struct kmem_cache *s)
4052 unsigned long count = 0;
4053 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4054 sizeof(unsigned long), GFP_KERNEL);
4060 for_each_node_state(node, N_NORMAL_MEMORY) {
4061 struct kmem_cache_node *n = get_node(s, node);
4063 count += validate_slab_node(s, n, map);
4069 * Generate lists of code addresses where slabcache objects are allocated
4074 unsigned long count;
4081 DECLARE_BITMAP(cpus, NR_CPUS);
4087 unsigned long count;
4088 struct location *loc;
4091 static void free_loc_track(struct loc_track *t)
4094 free_pages((unsigned long)t->loc,
4095 get_order(sizeof(struct location) * t->max));
4098 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4103 order = get_order(sizeof(struct location) * max);
4105 l = (void *)__get_free_pages(flags, order);
4110 memcpy(l, t->loc, sizeof(struct location) * t->count);
4118 static int add_location(struct loc_track *t, struct kmem_cache *s,
4119 const struct track *track)
4121 long start, end, pos;
4123 unsigned long caddr;
4124 unsigned long age = jiffies - track->when;
4130 pos = start + (end - start + 1) / 2;
4133 * There is nothing at "end". If we end up there
4134 * we need to add something to before end.
4139 caddr = t->loc[pos].addr;
4140 if (track->addr == caddr) {
4146 if (age < l->min_time)
4148 if (age > l->max_time)
4151 if (track->pid < l->min_pid)
4152 l->min_pid = track->pid;
4153 if (track->pid > l->max_pid)
4154 l->max_pid = track->pid;
4156 cpumask_set_cpu(track->cpu,
4157 to_cpumask(l->cpus));
4159 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4163 if (track->addr < caddr)
4170 * Not found. Insert new tracking element.
4172 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4178 (t->count - pos) * sizeof(struct location));
4181 l->addr = track->addr;
4185 l->min_pid = track->pid;
4186 l->max_pid = track->pid;
4187 cpumask_clear(to_cpumask(l->cpus));
4188 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4189 nodes_clear(l->nodes);
4190 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4194 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4195 struct page *page, enum track_item alloc,
4198 void *addr = page_address(page);
4201 bitmap_zero(map, page->objects);
4202 get_map(s, page, map);
4204 for_each_object(p, s, addr, page->objects)
4205 if (!test_bit(slab_index(p, s, addr), map))
4206 add_location(t, s, get_track(s, p, alloc));
4209 static int list_locations(struct kmem_cache *s, char *buf,
4210 enum track_item alloc)
4214 struct loc_track t = { 0, 0, NULL };
4216 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4217 sizeof(unsigned long), GFP_KERNEL);
4219 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4222 return sprintf(buf, "Out of memory\n");
4224 /* Push back cpu slabs */
4227 for_each_node_state(node, N_NORMAL_MEMORY) {
4228 struct kmem_cache_node *n = get_node(s, node);
4229 unsigned long flags;
4232 if (!atomic_long_read(&n->nr_slabs))
4235 spin_lock_irqsave(&n->list_lock, flags);
4236 list_for_each_entry(page, &n->partial, lru)
4237 process_slab(&t, s, page, alloc, map);
4238 list_for_each_entry(page, &n->full, lru)
4239 process_slab(&t, s, page, alloc, map);
4240 spin_unlock_irqrestore(&n->list_lock, flags);
4243 for (i = 0; i < t.count; i++) {
4244 struct location *l = &t.loc[i];
4246 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4248 len += sprintf(buf + len, "%7ld ", l->count);
4251 len += sprintf(buf + len, "%pS", (void *)l->addr);
4253 len += sprintf(buf + len, "<not-available>");
4255 if (l->sum_time != l->min_time) {
4256 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4258 (long)div_u64(l->sum_time, l->count),
4261 len += sprintf(buf + len, " age=%ld",
4264 if (l->min_pid != l->max_pid)
4265 len += sprintf(buf + len, " pid=%ld-%ld",
4266 l->min_pid, l->max_pid);
4268 len += sprintf(buf + len, " pid=%ld",
4271 if (num_online_cpus() > 1 &&
4272 !cpumask_empty(to_cpumask(l->cpus)) &&
4273 len < PAGE_SIZE - 60) {
4274 len += sprintf(buf + len, " cpus=");
4275 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4276 to_cpumask(l->cpus));
4279 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4280 len < PAGE_SIZE - 60) {
4281 len += sprintf(buf + len, " nodes=");
4282 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4286 len += sprintf(buf + len, "\n");
4292 len += sprintf(buf, "No data\n");
4297 #ifdef SLUB_RESILIENCY_TEST
4298 static void resiliency_test(void)
4302 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4304 printk(KERN_ERR "SLUB resiliency testing\n");
4305 printk(KERN_ERR "-----------------------\n");
4306 printk(KERN_ERR "A. Corruption after allocation\n");
4308 p = kzalloc(16, GFP_KERNEL);
4310 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4311 " 0x12->0x%p\n\n", p + 16);
4313 validate_slab_cache(kmalloc_caches[4]);
4315 /* Hmmm... The next two are dangerous */
4316 p = kzalloc(32, GFP_KERNEL);
4317 p[32 + sizeof(void *)] = 0x34;
4318 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4319 " 0x34 -> -0x%p\n", p);
4321 "If allocated object is overwritten then not detectable\n\n");
4323 validate_slab_cache(kmalloc_caches[5]);
4324 p = kzalloc(64, GFP_KERNEL);
4325 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4327 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4330 "If allocated object is overwritten then not detectable\n\n");
4331 validate_slab_cache(kmalloc_caches[6]);
4333 printk(KERN_ERR "\nB. Corruption after free\n");
4334 p = kzalloc(128, GFP_KERNEL);
4337 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4338 validate_slab_cache(kmalloc_caches[7]);
4340 p = kzalloc(256, GFP_KERNEL);
4343 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4345 validate_slab_cache(kmalloc_caches[8]);
4347 p = kzalloc(512, GFP_KERNEL);
4350 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4351 validate_slab_cache(kmalloc_caches[9]);
4355 static void resiliency_test(void) {};
4360 enum slab_stat_type {
4361 SL_ALL, /* All slabs */
4362 SL_PARTIAL, /* Only partially allocated slabs */
4363 SL_CPU, /* Only slabs used for cpu caches */
4364 SL_OBJECTS, /* Determine allocated objects not slabs */
4365 SL_TOTAL /* Determine object capacity not slabs */
4368 #define SO_ALL (1 << SL_ALL)
4369 #define SO_PARTIAL (1 << SL_PARTIAL)
4370 #define SO_CPU (1 << SL_CPU)
4371 #define SO_OBJECTS (1 << SL_OBJECTS)
4372 #define SO_TOTAL (1 << SL_TOTAL)
4374 static ssize_t show_slab_objects(struct kmem_cache *s,
4375 char *buf, unsigned long flags)
4377 unsigned long total = 0;
4380 unsigned long *nodes;
4381 unsigned long *per_cpu;
4383 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4386 per_cpu = nodes + nr_node_ids;
4388 if (flags & SO_CPU) {
4391 for_each_possible_cpu(cpu) {
4392 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4396 page = ACCESS_ONCE(c->page);
4400 node = page_to_nid(page);
4401 if (flags & SO_TOTAL)
4403 else if (flags & SO_OBJECTS)
4411 page = ACCESS_ONCE(c->partial);
4422 lock_memory_hotplug();
4423 #ifdef CONFIG_SLUB_DEBUG
4424 if (flags & SO_ALL) {
4425 for_each_node_state(node, N_NORMAL_MEMORY) {
4426 struct kmem_cache_node *n = get_node(s, node);
4428 if (flags & SO_TOTAL)
4429 x = atomic_long_read(&n->total_objects);
4430 else if (flags & SO_OBJECTS)
4431 x = atomic_long_read(&n->total_objects) -
4432 count_partial(n, count_free);
4435 x = atomic_long_read(&n->nr_slabs);
4442 if (flags & SO_PARTIAL) {
4443 for_each_node_state(node, N_NORMAL_MEMORY) {
4444 struct kmem_cache_node *n = get_node(s, node);
4446 if (flags & SO_TOTAL)
4447 x = count_partial(n, count_total);
4448 else if (flags & SO_OBJECTS)
4449 x = count_partial(n, count_inuse);
4456 x = sprintf(buf, "%lu", total);
4458 for_each_node_state(node, N_NORMAL_MEMORY)
4460 x += sprintf(buf + x, " N%d=%lu",
4463 unlock_memory_hotplug();
4465 return x + sprintf(buf + x, "\n");
4468 #ifdef CONFIG_SLUB_DEBUG
4469 static int any_slab_objects(struct kmem_cache *s)
4473 for_each_online_node(node) {
4474 struct kmem_cache_node *n = get_node(s, node);
4479 if (atomic_long_read(&n->total_objects))
4486 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4487 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4489 struct slab_attribute {
4490 struct attribute attr;
4491 ssize_t (*show)(struct kmem_cache *s, char *buf);
4492 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4495 #define SLAB_ATTR_RO(_name) \
4496 static struct slab_attribute _name##_attr = \
4497 __ATTR(_name, 0400, _name##_show, NULL)
4499 #define SLAB_ATTR(_name) \
4500 static struct slab_attribute _name##_attr = \
4501 __ATTR(_name, 0600, _name##_show, _name##_store)
4503 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4505 return sprintf(buf, "%d\n", s->size);
4507 SLAB_ATTR_RO(slab_size);
4509 static ssize_t align_show(struct kmem_cache *s, char *buf)
4511 return sprintf(buf, "%d\n", s->align);
4513 SLAB_ATTR_RO(align);
4515 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4517 return sprintf(buf, "%d\n", s->object_size);
4519 SLAB_ATTR_RO(object_size);
4521 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4523 return sprintf(buf, "%d\n", oo_objects(s->oo));
4525 SLAB_ATTR_RO(objs_per_slab);
4527 static ssize_t order_store(struct kmem_cache *s,
4528 const char *buf, size_t length)
4530 unsigned long order;
4533 err = strict_strtoul(buf, 10, &order);
4537 if (order > slub_max_order || order < slub_min_order)
4540 calculate_sizes(s, order);
4544 static ssize_t order_show(struct kmem_cache *s, char *buf)
4546 return sprintf(buf, "%d\n", oo_order(s->oo));
4550 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4552 return sprintf(buf, "%lu\n", s->min_partial);
4555 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4561 err = strict_strtoul(buf, 10, &min);
4565 set_min_partial(s, min);
4568 SLAB_ATTR(min_partial);
4570 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4572 return sprintf(buf, "%u\n", s->cpu_partial);
4575 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4578 unsigned long objects;
4581 err = strict_strtoul(buf, 10, &objects);
4584 if (objects && kmem_cache_debug(s))
4587 s->cpu_partial = objects;
4591 SLAB_ATTR(cpu_partial);
4593 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4597 return sprintf(buf, "%pS\n", s->ctor);
4601 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4603 return sprintf(buf, "%d\n", s->refcount - 1);
4605 SLAB_ATTR_RO(aliases);
4607 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4609 return show_slab_objects(s, buf, SO_PARTIAL);
4611 SLAB_ATTR_RO(partial);
4613 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4615 return show_slab_objects(s, buf, SO_CPU);
4617 SLAB_ATTR_RO(cpu_slabs);
4619 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4621 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4623 SLAB_ATTR_RO(objects);
4625 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4627 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4629 SLAB_ATTR_RO(objects_partial);
4631 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4638 for_each_online_cpu(cpu) {
4639 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4642 pages += page->pages;
4643 objects += page->pobjects;
4647 len = sprintf(buf, "%d(%d)", objects, pages);
4650 for_each_online_cpu(cpu) {
4651 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4653 if (page && len < PAGE_SIZE - 20)
4654 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4655 page->pobjects, page->pages);
4658 return len + sprintf(buf + len, "\n");
4660 SLAB_ATTR_RO(slabs_cpu_partial);
4662 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4664 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4667 static ssize_t reclaim_account_store(struct kmem_cache *s,
4668 const char *buf, size_t length)
4670 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4672 s->flags |= SLAB_RECLAIM_ACCOUNT;
4675 SLAB_ATTR(reclaim_account);
4677 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4679 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4681 SLAB_ATTR_RO(hwcache_align);
4683 #ifdef CONFIG_ZONE_DMA
4684 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4686 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4688 SLAB_ATTR_RO(cache_dma);
4691 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4693 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4695 SLAB_ATTR_RO(destroy_by_rcu);
4697 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4699 return sprintf(buf, "%d\n", s->reserved);
4701 SLAB_ATTR_RO(reserved);
4703 #ifdef CONFIG_SLUB_DEBUG
4704 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4706 return show_slab_objects(s, buf, SO_ALL);
4708 SLAB_ATTR_RO(slabs);
4710 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4712 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4714 SLAB_ATTR_RO(total_objects);
4716 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4718 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4721 static ssize_t sanity_checks_store(struct kmem_cache *s,
4722 const char *buf, size_t length)
4724 s->flags &= ~SLAB_DEBUG_FREE;
4725 if (buf[0] == '1') {
4726 s->flags &= ~__CMPXCHG_DOUBLE;
4727 s->flags |= SLAB_DEBUG_FREE;
4731 SLAB_ATTR(sanity_checks);
4733 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4735 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4738 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4741 s->flags &= ~SLAB_TRACE;
4742 if (buf[0] == '1') {
4743 s->flags &= ~__CMPXCHG_DOUBLE;
4744 s->flags |= SLAB_TRACE;
4750 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4752 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4755 static ssize_t red_zone_store(struct kmem_cache *s,
4756 const char *buf, size_t length)
4758 if (any_slab_objects(s))
4761 s->flags &= ~SLAB_RED_ZONE;
4762 if (buf[0] == '1') {
4763 s->flags &= ~__CMPXCHG_DOUBLE;
4764 s->flags |= SLAB_RED_ZONE;
4766 calculate_sizes(s, -1);
4769 SLAB_ATTR(red_zone);
4771 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4773 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4776 static ssize_t poison_store(struct kmem_cache *s,
4777 const char *buf, size_t length)
4779 if (any_slab_objects(s))
4782 s->flags &= ~SLAB_POISON;
4783 if (buf[0] == '1') {
4784 s->flags &= ~__CMPXCHG_DOUBLE;
4785 s->flags |= SLAB_POISON;
4787 calculate_sizes(s, -1);
4792 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4794 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4797 static ssize_t store_user_store(struct kmem_cache *s,
4798 const char *buf, size_t length)
4800 if (any_slab_objects(s))
4803 s->flags &= ~SLAB_STORE_USER;
4804 if (buf[0] == '1') {
4805 s->flags &= ~__CMPXCHG_DOUBLE;
4806 s->flags |= SLAB_STORE_USER;
4808 calculate_sizes(s, -1);
4811 SLAB_ATTR(store_user);
4813 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4818 static ssize_t validate_store(struct kmem_cache *s,
4819 const char *buf, size_t length)
4823 if (buf[0] == '1') {
4824 ret = validate_slab_cache(s);
4830 SLAB_ATTR(validate);
4832 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4834 if (!(s->flags & SLAB_STORE_USER))
4836 return list_locations(s, buf, TRACK_ALLOC);
4838 SLAB_ATTR_RO(alloc_calls);
4840 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4842 if (!(s->flags & SLAB_STORE_USER))
4844 return list_locations(s, buf, TRACK_FREE);
4846 SLAB_ATTR_RO(free_calls);
4847 #endif /* CONFIG_SLUB_DEBUG */
4849 #ifdef CONFIG_FAILSLAB
4850 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4852 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4855 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4858 s->flags &= ~SLAB_FAILSLAB;
4860 s->flags |= SLAB_FAILSLAB;
4863 SLAB_ATTR(failslab);
4866 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4871 static ssize_t shrink_store(struct kmem_cache *s,
4872 const char *buf, size_t length)
4874 if (buf[0] == '1') {
4875 int rc = kmem_cache_shrink(s);
4886 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4888 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4891 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4892 const char *buf, size_t length)
4894 unsigned long ratio;
4897 err = strict_strtoul(buf, 10, &ratio);
4902 s->remote_node_defrag_ratio = ratio * 10;
4906 SLAB_ATTR(remote_node_defrag_ratio);
4909 #ifdef CONFIG_SLUB_STATS
4910 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4912 unsigned long sum = 0;
4915 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4920 for_each_online_cpu(cpu) {
4921 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4927 len = sprintf(buf, "%lu", sum);
4930 for_each_online_cpu(cpu) {
4931 if (data[cpu] && len < PAGE_SIZE - 20)
4932 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4936 return len + sprintf(buf + len, "\n");
4939 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4943 for_each_online_cpu(cpu)
4944 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4947 #define STAT_ATTR(si, text) \
4948 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4950 return show_stat(s, buf, si); \
4952 static ssize_t text##_store(struct kmem_cache *s, \
4953 const char *buf, size_t length) \
4955 if (buf[0] != '0') \
4957 clear_stat(s, si); \
4962 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4963 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4964 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4965 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4966 STAT_ATTR(FREE_FROZEN, free_frozen);
4967 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4968 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4969 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4970 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4971 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4972 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4973 STAT_ATTR(FREE_SLAB, free_slab);
4974 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4975 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4976 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4977 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4978 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4979 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4980 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4981 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4982 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4983 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4984 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4985 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4986 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4987 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4990 static struct attribute *slab_attrs[] = {
4991 &slab_size_attr.attr,
4992 &object_size_attr.attr,
4993 &objs_per_slab_attr.attr,
4995 &min_partial_attr.attr,
4996 &cpu_partial_attr.attr,
4998 &objects_partial_attr.attr,
5000 &cpu_slabs_attr.attr,
5004 &hwcache_align_attr.attr,
5005 &reclaim_account_attr.attr,
5006 &destroy_by_rcu_attr.attr,
5008 &reserved_attr.attr,
5009 &slabs_cpu_partial_attr.attr,
5010 #ifdef CONFIG_SLUB_DEBUG
5011 &total_objects_attr.attr,
5013 &sanity_checks_attr.attr,
5015 &red_zone_attr.attr,
5017 &store_user_attr.attr,
5018 &validate_attr.attr,
5019 &alloc_calls_attr.attr,
5020 &free_calls_attr.attr,
5022 #ifdef CONFIG_ZONE_DMA
5023 &cache_dma_attr.attr,
5026 &remote_node_defrag_ratio_attr.attr,
5028 #ifdef CONFIG_SLUB_STATS
5029 &alloc_fastpath_attr.attr,
5030 &alloc_slowpath_attr.attr,
5031 &free_fastpath_attr.attr,
5032 &free_slowpath_attr.attr,
5033 &free_frozen_attr.attr,
5034 &free_add_partial_attr.attr,
5035 &free_remove_partial_attr.attr,
5036 &alloc_from_partial_attr.attr,
5037 &alloc_slab_attr.attr,
5038 &alloc_refill_attr.attr,
5039 &alloc_node_mismatch_attr.attr,
5040 &free_slab_attr.attr,
5041 &cpuslab_flush_attr.attr,
5042 &deactivate_full_attr.attr,
5043 &deactivate_empty_attr.attr,
5044 &deactivate_to_head_attr.attr,
5045 &deactivate_to_tail_attr.attr,
5046 &deactivate_remote_frees_attr.attr,
5047 &deactivate_bypass_attr.attr,
5048 &order_fallback_attr.attr,
5049 &cmpxchg_double_fail_attr.attr,
5050 &cmpxchg_double_cpu_fail_attr.attr,
5051 &cpu_partial_alloc_attr.attr,
5052 &cpu_partial_free_attr.attr,
5053 &cpu_partial_node_attr.attr,
5054 &cpu_partial_drain_attr.attr,
5056 #ifdef CONFIG_FAILSLAB
5057 &failslab_attr.attr,
5063 static struct attribute_group slab_attr_group = {
5064 .attrs = slab_attrs,
5067 static ssize_t slab_attr_show(struct kobject *kobj,
5068 struct attribute *attr,
5071 struct slab_attribute *attribute;
5072 struct kmem_cache *s;
5075 attribute = to_slab_attr(attr);
5078 if (!attribute->show)
5081 err = attribute->show(s, buf);
5086 static ssize_t slab_attr_store(struct kobject *kobj,
5087 struct attribute *attr,
5088 const char *buf, size_t len)
5090 struct slab_attribute *attribute;
5091 struct kmem_cache *s;
5094 attribute = to_slab_attr(attr);
5097 if (!attribute->store)
5100 err = attribute->store(s, buf, len);
5105 static const struct sysfs_ops slab_sysfs_ops = {
5106 .show = slab_attr_show,
5107 .store = slab_attr_store,
5110 static struct kobj_type slab_ktype = {
5111 .sysfs_ops = &slab_sysfs_ops,
5114 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5116 struct kobj_type *ktype = get_ktype(kobj);
5118 if (ktype == &slab_ktype)
5123 static const struct kset_uevent_ops slab_uevent_ops = {
5124 .filter = uevent_filter,
5127 static struct kset *slab_kset;
5129 #define ID_STR_LENGTH 64
5131 /* Create a unique string id for a slab cache:
5133 * Format :[flags-]size
5135 static char *create_unique_id(struct kmem_cache *s)
5137 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5144 * First flags affecting slabcache operations. We will only
5145 * get here for aliasable slabs so we do not need to support
5146 * too many flags. The flags here must cover all flags that
5147 * are matched during merging to guarantee that the id is
5150 if (s->flags & SLAB_CACHE_DMA)
5152 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5154 if (s->flags & SLAB_DEBUG_FREE)
5156 if (!(s->flags & SLAB_NOTRACK))
5160 p += sprintf(p, "%07d", s->size);
5162 #ifdef CONFIG_MEMCG_KMEM
5163 if (!is_root_cache(s))
5164 p += sprintf(p, "-%08d", memcg_cache_id(s->memcg_params->memcg));
5167 BUG_ON(p > name + ID_STR_LENGTH - 1);
5171 static int sysfs_slab_add(struct kmem_cache *s)
5175 int unmergeable = slab_unmergeable(s);
5179 * Slabcache can never be merged so we can use the name proper.
5180 * This is typically the case for debug situations. In that
5181 * case we can catch duplicate names easily.
5183 sysfs_remove_link(&slab_kset->kobj, s->name);
5187 * Create a unique name for the slab as a target
5190 name = create_unique_id(s);
5193 s->kobj.kset = slab_kset;
5194 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5196 kobject_put(&s->kobj);
5200 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5202 kobject_del(&s->kobj);
5203 kobject_put(&s->kobj);
5206 kobject_uevent(&s->kobj, KOBJ_ADD);
5208 /* Setup first alias */
5209 sysfs_slab_alias(s, s->name);
5215 static void sysfs_slab_remove(struct kmem_cache *s)
5217 if (slab_state < FULL)
5219 * Sysfs has not been setup yet so no need to remove the
5224 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5225 kobject_del(&s->kobj);
5226 kobject_put(&s->kobj);
5230 * Need to buffer aliases during bootup until sysfs becomes
5231 * available lest we lose that information.
5233 struct saved_alias {
5234 struct kmem_cache *s;
5236 struct saved_alias *next;
5239 static struct saved_alias *alias_list;
5241 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5243 struct saved_alias *al;
5245 if (slab_state == FULL) {
5247 * If we have a leftover link then remove it.
5249 sysfs_remove_link(&slab_kset->kobj, name);
5250 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5253 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5259 al->next = alias_list;
5264 static int __init slab_sysfs_init(void)
5266 struct kmem_cache *s;
5269 mutex_lock(&slab_mutex);
5271 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5273 mutex_unlock(&slab_mutex);
5274 printk(KERN_ERR "Cannot register slab subsystem.\n");
5280 list_for_each_entry(s, &slab_caches, list) {
5281 err = sysfs_slab_add(s);
5283 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5284 " to sysfs\n", s->name);
5287 while (alias_list) {
5288 struct saved_alias *al = alias_list;
5290 alias_list = alias_list->next;
5291 err = sysfs_slab_alias(al->s, al->name);
5293 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5294 " %s to sysfs\n", al->name);
5298 mutex_unlock(&slab_mutex);
5303 __initcall(slab_sysfs_init);
5304 #endif /* CONFIG_SYSFS */
5307 * The /proc/slabinfo ABI
5309 #ifdef CONFIG_SLABINFO
5310 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5312 unsigned long nr_partials = 0;
5313 unsigned long nr_slabs = 0;
5314 unsigned long nr_objs = 0;
5315 unsigned long nr_free = 0;
5318 for_each_online_node(node) {
5319 struct kmem_cache_node *n = get_node(s, node);
5324 nr_partials += n->nr_partial;
5325 nr_slabs += atomic_long_read(&n->nr_slabs);
5326 nr_objs += atomic_long_read(&n->total_objects);
5327 nr_free += count_partial(n, count_free);
5330 sinfo->active_objs = nr_objs - nr_free;
5331 sinfo->num_objs = nr_objs;
5332 sinfo->active_slabs = nr_slabs;
5333 sinfo->num_slabs = nr_slabs;
5334 sinfo->objects_per_slab = oo_objects(s->oo);
5335 sinfo->cache_order = oo_order(s->oo);
5338 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5342 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5343 size_t count, loff_t *ppos)
5347 #endif /* CONFIG_SLABINFO */