2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
219 raw_cpu_inc(s->cpu_slab->stat[si]);
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
236 base = page_address(page);
237 if (object < base || object >= base + page->objects * s->size ||
238 (object - base) % s->size) {
245 static inline void *get_freepointer(struct kmem_cache *s, void *object)
247 return *(void **)(object + s->offset);
250 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
252 prefetch(object + s->offset);
255 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
262 p = get_freepointer(s, object);
267 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
269 *(void **)(object + s->offset) = fp;
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 bit_spin_lock(PG_locked, &page->flags);
344 static __always_inline void slab_unlock(struct page *page)
346 __bit_spin_unlock(PG_locked, &page->flags);
349 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
352 tmp.counters = counters_new;
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
359 page->frozen = tmp.frozen;
360 page->inuse = tmp.inuse;
361 page->objects = tmp.objects;
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 set_page_slub_counters(page, counters_new);
393 stat(s, CMPXCHG_DOUBLE_FAIL);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
419 local_irq_save(flags);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 set_page_slub_counters(page, counters_new);
426 local_irq_restore(flags);
430 local_irq_restore(flags);
434 stat(s, CMPXCHG_DOUBLE_FAIL);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 void *addr = page_address(page);
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
462 #if defined(CONFIG_SLUB_DEBUG_ON)
463 static int slub_debug = DEBUG_DEFAULT_FLAGS;
464 #elif defined(CONFIG_KASAN)
465 static int slub_debug = SLAB_STORE_USER;
467 static int slub_debug;
470 static char *slub_debug_slabs;
471 static int disable_higher_order_debug;
474 * slub is about to manipulate internal object metadata. This memory lies
475 * outside the range of the allocated object, so accessing it would normally
476 * be reported by kasan as a bounds error. metadata_access_enable() is used
477 * to tell kasan that these accesses are OK.
479 static inline void metadata_access_enable(void)
481 kasan_disable_current();
484 static inline void metadata_access_disable(void)
486 kasan_enable_current();
492 static void print_section(char *text, u8 *addr, unsigned int length)
494 metadata_access_enable();
495 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
497 metadata_access_disable();
500 static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
506 p = object + s->offset + sizeof(void *);
508 p = object + s->inuse;
513 static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
516 struct track *p = get_track(s, object, alloc);
519 #ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
527 metadata_access_enable();
528 save_stack_trace(&trace);
529 metadata_access_disable();
531 /* See rant in lockdep.c */
532 if (trace.nr_entries != 0 &&
533 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
536 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
540 p->cpu = smp_processor_id();
541 p->pid = current->pid;
544 memset(p, 0, sizeof(struct track));
547 static void init_tracking(struct kmem_cache *s, void *object)
549 if (!(s->flags & SLAB_STORE_USER))
552 set_track(s, object, TRACK_FREE, 0UL);
553 set_track(s, object, TRACK_ALLOC, 0UL);
556 static void print_track(const char *s, struct track *t)
561 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
562 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
563 #ifdef CONFIG_STACKTRACE
566 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
568 pr_err("\t%pS\n", (void *)t->addrs[i]);
575 static void print_tracking(struct kmem_cache *s, void *object)
577 if (!(s->flags & SLAB_STORE_USER))
580 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
581 print_track("Freed", get_track(s, object, TRACK_FREE));
584 static void print_page_info(struct page *page)
586 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
587 page, page->objects, page->inuse, page->freelist, page->flags);
591 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
593 struct va_format vaf;
599 pr_err("=============================================================================\n");
600 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
601 pr_err("-----------------------------------------------------------------------------\n\n");
603 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
607 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
609 struct va_format vaf;
615 pr_err("FIX %s: %pV\n", s->name, &vaf);
619 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
621 unsigned int off; /* Offset of last byte */
622 u8 *addr = page_address(page);
624 print_tracking(s, p);
626 print_page_info(page);
628 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
629 p, p - addr, get_freepointer(s, p));
632 print_section("Bytes b4 ", p - 16, 16);
634 print_section("Object ", p, min_t(unsigned long, s->object_size,
636 if (s->flags & SLAB_RED_ZONE)
637 print_section("Redzone ", p + s->object_size,
638 s->inuse - s->object_size);
641 off = s->offset + sizeof(void *);
645 if (s->flags & SLAB_STORE_USER)
646 off += 2 * sizeof(struct track);
649 /* Beginning of the filler is the free pointer */
650 print_section("Padding ", p + off, s->size - off);
655 void object_err(struct kmem_cache *s, struct page *page,
656 u8 *object, char *reason)
658 slab_bug(s, "%s", reason);
659 print_trailer(s, page, object);
662 static void slab_err(struct kmem_cache *s, struct page *page,
663 const char *fmt, ...)
669 vsnprintf(buf, sizeof(buf), fmt, args);
671 slab_bug(s, "%s", buf);
672 print_page_info(page);
676 static void init_object(struct kmem_cache *s, void *object, u8 val)
680 if (s->flags & __OBJECT_POISON) {
681 memset(p, POISON_FREE, s->object_size - 1);
682 p[s->object_size - 1] = POISON_END;
685 if (s->flags & SLAB_RED_ZONE)
686 memset(p + s->object_size, val, s->inuse - s->object_size);
689 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
690 void *from, void *to)
692 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
693 memset(from, data, to - from);
696 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
697 u8 *object, char *what,
698 u8 *start, unsigned int value, unsigned int bytes)
703 metadata_access_enable();
704 fault = memchr_inv(start, value, bytes);
705 metadata_access_disable();
710 while (end > fault && end[-1] == value)
713 slab_bug(s, "%s overwritten", what);
714 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
715 fault, end - 1, fault[0], value);
716 print_trailer(s, page, object);
718 restore_bytes(s, what, value, fault, end);
726 * Bytes of the object to be managed.
727 * If the freepointer may overlay the object then the free
728 * pointer is the first word of the object.
730 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
733 * object + s->object_size
734 * Padding to reach word boundary. This is also used for Redzoning.
735 * Padding is extended by another word if Redzoning is enabled and
736 * object_size == inuse.
738 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
739 * 0xcc (RED_ACTIVE) for objects in use.
742 * Meta data starts here.
744 * A. Free pointer (if we cannot overwrite object on free)
745 * B. Tracking data for SLAB_STORE_USER
746 * C. Padding to reach required alignment boundary or at mininum
747 * one word if debugging is on to be able to detect writes
748 * before the word boundary.
750 * Padding is done using 0x5a (POISON_INUSE)
753 * Nothing is used beyond s->size.
755 * If slabcaches are merged then the object_size and inuse boundaries are mostly
756 * ignored. And therefore no slab options that rely on these boundaries
757 * may be used with merged slabcaches.
760 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
762 unsigned long off = s->inuse; /* The end of info */
765 /* Freepointer is placed after the object. */
766 off += sizeof(void *);
768 if (s->flags & SLAB_STORE_USER)
769 /* We also have user information there */
770 off += 2 * sizeof(struct track);
775 return check_bytes_and_report(s, page, p, "Object padding",
776 p + off, POISON_INUSE, s->size - off);
779 /* Check the pad bytes at the end of a slab page */
780 static int slab_pad_check(struct kmem_cache *s, struct page *page)
788 if (!(s->flags & SLAB_POISON))
791 start = page_address(page);
792 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
793 end = start + length;
794 remainder = length % s->size;
798 metadata_access_enable();
799 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
800 metadata_access_disable();
803 while (end > fault && end[-1] == POISON_INUSE)
806 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
807 print_section("Padding ", end - remainder, remainder);
809 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
813 static int check_object(struct kmem_cache *s, struct page *page,
814 void *object, u8 val)
817 u8 *endobject = object + s->object_size;
819 if (s->flags & SLAB_RED_ZONE) {
820 if (!check_bytes_and_report(s, page, object, "Redzone",
821 endobject, val, s->inuse - s->object_size))
824 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
825 check_bytes_and_report(s, page, p, "Alignment padding",
826 endobject, POISON_INUSE,
827 s->inuse - s->object_size);
831 if (s->flags & SLAB_POISON) {
832 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
833 (!check_bytes_and_report(s, page, p, "Poison", p,
834 POISON_FREE, s->object_size - 1) ||
835 !check_bytes_and_report(s, page, p, "Poison",
836 p + s->object_size - 1, POISON_END, 1)))
839 * check_pad_bytes cleans up on its own.
841 check_pad_bytes(s, page, p);
844 if (!s->offset && val == SLUB_RED_ACTIVE)
846 * Object and freepointer overlap. Cannot check
847 * freepointer while object is allocated.
851 /* Check free pointer validity */
852 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
853 object_err(s, page, p, "Freepointer corrupt");
855 * No choice but to zap it and thus lose the remainder
856 * of the free objects in this slab. May cause
857 * another error because the object count is now wrong.
859 set_freepointer(s, p, NULL);
865 static int check_slab(struct kmem_cache *s, struct page *page)
869 VM_BUG_ON(!irqs_disabled());
871 if (!PageSlab(page)) {
872 slab_err(s, page, "Not a valid slab page");
876 maxobj = order_objects(compound_order(page), s->size, s->reserved);
877 if (page->objects > maxobj) {
878 slab_err(s, page, "objects %u > max %u",
879 page->objects, maxobj);
882 if (page->inuse > page->objects) {
883 slab_err(s, page, "inuse %u > max %u",
884 page->inuse, page->objects);
887 /* Slab_pad_check fixes things up after itself */
888 slab_pad_check(s, page);
893 * Determine if a certain object on a page is on the freelist. Must hold the
894 * slab lock to guarantee that the chains are in a consistent state.
896 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
904 while (fp && nr <= page->objects) {
907 if (!check_valid_pointer(s, page, fp)) {
909 object_err(s, page, object,
910 "Freechain corrupt");
911 set_freepointer(s, object, NULL);
913 slab_err(s, page, "Freepointer corrupt");
914 page->freelist = NULL;
915 page->inuse = page->objects;
916 slab_fix(s, "Freelist cleared");
922 fp = get_freepointer(s, object);
926 max_objects = order_objects(compound_order(page), s->size, s->reserved);
927 if (max_objects > MAX_OBJS_PER_PAGE)
928 max_objects = MAX_OBJS_PER_PAGE;
930 if (page->objects != max_objects) {
931 slab_err(s, page, "Wrong number of objects. Found %d but "
932 "should be %d", page->objects, max_objects);
933 page->objects = max_objects;
934 slab_fix(s, "Number of objects adjusted.");
936 if (page->inuse != page->objects - nr) {
937 slab_err(s, page, "Wrong object count. Counter is %d but "
938 "counted were %d", page->inuse, page->objects - nr);
939 page->inuse = page->objects - nr;
940 slab_fix(s, "Object count adjusted.");
942 return search == NULL;
945 static void trace(struct kmem_cache *s, struct page *page, void *object,
948 if (s->flags & SLAB_TRACE) {
949 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
951 alloc ? "alloc" : "free",
956 print_section("Object ", (void *)object,
964 * Tracking of fully allocated slabs for debugging purposes.
966 static void add_full(struct kmem_cache *s,
967 struct kmem_cache_node *n, struct page *page)
969 if (!(s->flags & SLAB_STORE_USER))
972 lockdep_assert_held(&n->list_lock);
973 list_add(&page->lru, &n->full);
976 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
978 if (!(s->flags & SLAB_STORE_USER))
981 lockdep_assert_held(&n->list_lock);
982 list_del(&page->lru);
985 /* Tracking of the number of slabs for debugging purposes */
986 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
988 struct kmem_cache_node *n = get_node(s, node);
990 return atomic_long_read(&n->nr_slabs);
993 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
995 return atomic_long_read(&n->nr_slabs);
998 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1000 struct kmem_cache_node *n = get_node(s, node);
1003 * May be called early in order to allocate a slab for the
1004 * kmem_cache_node structure. Solve the chicken-egg
1005 * dilemma by deferring the increment of the count during
1006 * bootstrap (see early_kmem_cache_node_alloc).
1009 atomic_long_inc(&n->nr_slabs);
1010 atomic_long_add(objects, &n->total_objects);
1013 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1015 struct kmem_cache_node *n = get_node(s, node);
1017 atomic_long_dec(&n->nr_slabs);
1018 atomic_long_sub(objects, &n->total_objects);
1021 /* Object debug checks for alloc/free paths */
1022 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1025 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1028 init_object(s, object, SLUB_RED_INACTIVE);
1029 init_tracking(s, object);
1032 static noinline int alloc_debug_processing(struct kmem_cache *s,
1034 void *object, unsigned long addr)
1036 if (!check_slab(s, page))
1039 if (!check_valid_pointer(s, page, object)) {
1040 object_err(s, page, object, "Freelist Pointer check fails");
1044 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1047 /* Success perform special debug activities for allocs */
1048 if (s->flags & SLAB_STORE_USER)
1049 set_track(s, object, TRACK_ALLOC, addr);
1050 trace(s, page, object, 1);
1051 init_object(s, object, SLUB_RED_ACTIVE);
1055 if (PageSlab(page)) {
1057 * If this is a slab page then lets do the best we can
1058 * to avoid issues in the future. Marking all objects
1059 * as used avoids touching the remaining objects.
1061 slab_fix(s, "Marking all objects used");
1062 page->inuse = page->objects;
1063 page->freelist = NULL;
1068 /* Supports checking bulk free of a constructed freelist */
1069 static noinline struct kmem_cache_node *free_debug_processing(
1070 struct kmem_cache *s, struct page *page,
1071 void *head, void *tail, int bulk_cnt,
1072 unsigned long addr, unsigned long *flags)
1074 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1075 void *object = head;
1078 spin_lock_irqsave(&n->list_lock, *flags);
1081 if (!check_slab(s, page))
1087 if (!check_valid_pointer(s, page, object)) {
1088 slab_err(s, page, "Invalid object pointer 0x%p", object);
1092 if (on_freelist(s, page, object)) {
1093 object_err(s, page, object, "Object already free");
1097 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1100 if (unlikely(s != page->slab_cache)) {
1101 if (!PageSlab(page)) {
1102 slab_err(s, page, "Attempt to free object(0x%p) "
1103 "outside of slab", object);
1104 } else if (!page->slab_cache) {
1105 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1109 object_err(s, page, object,
1110 "page slab pointer corrupt.");
1114 if (s->flags & SLAB_STORE_USER)
1115 set_track(s, object, TRACK_FREE, addr);
1116 trace(s, page, object, 0);
1117 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1118 init_object(s, object, SLUB_RED_INACTIVE);
1120 /* Reached end of constructed freelist yet? */
1121 if (object != tail) {
1122 object = get_freepointer(s, object);
1126 if (cnt != bulk_cnt)
1127 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1132 * Keep node_lock to preserve integrity
1133 * until the object is actually freed
1139 spin_unlock_irqrestore(&n->list_lock, *flags);
1140 slab_fix(s, "Object at 0x%p not freed", object);
1144 static int __init setup_slub_debug(char *str)
1146 slub_debug = DEBUG_DEFAULT_FLAGS;
1147 if (*str++ != '=' || !*str)
1149 * No options specified. Switch on full debugging.
1155 * No options but restriction on slabs. This means full
1156 * debugging for slabs matching a pattern.
1163 * Switch off all debugging measures.
1168 * Determine which debug features should be switched on
1170 for (; *str && *str != ','; str++) {
1171 switch (tolower(*str)) {
1173 slub_debug |= SLAB_DEBUG_FREE;
1176 slub_debug |= SLAB_RED_ZONE;
1179 slub_debug |= SLAB_POISON;
1182 slub_debug |= SLAB_STORE_USER;
1185 slub_debug |= SLAB_TRACE;
1188 slub_debug |= SLAB_FAILSLAB;
1192 * Avoid enabling debugging on caches if its minimum
1193 * order would increase as a result.
1195 disable_higher_order_debug = 1;
1198 pr_err("slub_debug option '%c' unknown. skipped\n",
1205 slub_debug_slabs = str + 1;
1210 __setup("slub_debug", setup_slub_debug);
1212 unsigned long kmem_cache_flags(unsigned long object_size,
1213 unsigned long flags, const char *name,
1214 void (*ctor)(void *))
1217 * Enable debugging if selected on the kernel commandline.
1219 if (slub_debug && (!slub_debug_slabs || (name &&
1220 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1221 flags |= slub_debug;
1225 #else /* !CONFIG_SLUB_DEBUG */
1226 static inline void setup_object_debug(struct kmem_cache *s,
1227 struct page *page, void *object) {}
1229 static inline int alloc_debug_processing(struct kmem_cache *s,
1230 struct page *page, void *object, unsigned long addr) { return 0; }
1232 static inline struct kmem_cache_node *free_debug_processing(
1233 struct kmem_cache *s, struct page *page,
1234 void *head, void *tail, int bulk_cnt,
1235 unsigned long addr, unsigned long *flags) { return NULL; }
1237 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1239 static inline int check_object(struct kmem_cache *s, struct page *page,
1240 void *object, u8 val) { return 1; }
1241 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1242 struct page *page) {}
1243 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1244 struct page *page) {}
1245 unsigned long kmem_cache_flags(unsigned long object_size,
1246 unsigned long flags, const char *name,
1247 void (*ctor)(void *))
1251 #define slub_debug 0
1253 #define disable_higher_order_debug 0
1255 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1257 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1259 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1261 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1264 #endif /* CONFIG_SLUB_DEBUG */
1267 * Hooks for other subsystems that check memory allocations. In a typical
1268 * production configuration these hooks all should produce no code at all.
1270 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1272 kmemleak_alloc(ptr, size, 1, flags);
1273 kasan_kmalloc_large(ptr, size);
1276 static inline void kfree_hook(const void *x)
1279 kasan_kfree_large(x);
1282 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1285 flags &= gfp_allowed_mask;
1286 lockdep_trace_alloc(flags);
1287 might_sleep_if(gfpflags_allow_blocking(flags));
1289 if (should_failslab(s->object_size, flags, s->flags))
1292 return memcg_kmem_get_cache(s, flags);
1295 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1296 gfp_t flags, void *object)
1298 flags &= gfp_allowed_mask;
1299 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1300 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1301 memcg_kmem_put_cache(s);
1302 kasan_slab_alloc(s, object);
1305 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1307 kmemleak_free_recursive(x, s->flags);
1310 * Trouble is that we may no longer disable interrupts in the fast path
1311 * So in order to make the debug calls that expect irqs to be
1312 * disabled we need to disable interrupts temporarily.
1314 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1316 unsigned long flags;
1318 local_irq_save(flags);
1319 kmemcheck_slab_free(s, x, s->object_size);
1320 debug_check_no_locks_freed(x, s->object_size);
1321 local_irq_restore(flags);
1324 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1325 debug_check_no_obj_freed(x, s->object_size);
1327 kasan_slab_free(s, x);
1330 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1331 void *head, void *tail)
1334 * Compiler cannot detect this function can be removed if slab_free_hook()
1335 * evaluates to nothing. Thus, catch all relevant config debug options here.
1337 #if defined(CONFIG_KMEMCHECK) || \
1338 defined(CONFIG_LOCKDEP) || \
1339 defined(CONFIG_DEBUG_KMEMLEAK) || \
1340 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1341 defined(CONFIG_KASAN)
1343 void *object = head;
1344 void *tail_obj = tail ? : head;
1347 slab_free_hook(s, object);
1348 } while ((object != tail_obj) &&
1349 (object = get_freepointer(s, object)));
1353 static void setup_object(struct kmem_cache *s, struct page *page,
1356 setup_object_debug(s, page, object);
1357 if (unlikely(s->ctor)) {
1358 kasan_unpoison_object_data(s, object);
1360 kasan_poison_object_data(s, object);
1365 * Slab allocation and freeing
1367 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1368 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1371 int order = oo_order(oo);
1373 flags |= __GFP_NOTRACK;
1375 if (node == NUMA_NO_NODE)
1376 page = alloc_pages(flags, order);
1378 page = __alloc_pages_node(node, flags, order);
1380 if (page && memcg_charge_slab(page, flags, order, s)) {
1381 __free_pages(page, order);
1388 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1391 struct kmem_cache_order_objects oo = s->oo;
1396 flags &= gfp_allowed_mask;
1398 if (gfpflags_allow_blocking(flags))
1401 flags |= s->allocflags;
1404 * Let the initial higher-order allocation fail under memory pressure
1405 * so we fall-back to the minimum order allocation.
1407 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1408 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1409 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1411 page = alloc_slab_page(s, alloc_gfp, node, oo);
1412 if (unlikely(!page)) {
1416 * Allocation may have failed due to fragmentation.
1417 * Try a lower order alloc if possible
1419 page = alloc_slab_page(s, alloc_gfp, node, oo);
1420 if (unlikely(!page))
1422 stat(s, ORDER_FALLBACK);
1425 if (kmemcheck_enabled &&
1426 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1427 int pages = 1 << oo_order(oo);
1429 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1432 * Objects from caches that have a constructor don't get
1433 * cleared when they're allocated, so we need to do it here.
1436 kmemcheck_mark_uninitialized_pages(page, pages);
1438 kmemcheck_mark_unallocated_pages(page, pages);
1441 page->objects = oo_objects(oo);
1443 order = compound_order(page);
1444 page->slab_cache = s;
1445 __SetPageSlab(page);
1446 if (page_is_pfmemalloc(page))
1447 SetPageSlabPfmemalloc(page);
1449 start = page_address(page);
1451 if (unlikely(s->flags & SLAB_POISON))
1452 memset(start, POISON_INUSE, PAGE_SIZE << order);
1454 kasan_poison_slab(page);
1456 for_each_object_idx(p, idx, s, start, page->objects) {
1457 setup_object(s, page, p);
1458 if (likely(idx < page->objects))
1459 set_freepointer(s, p, p + s->size);
1461 set_freepointer(s, p, NULL);
1464 page->freelist = start;
1465 page->inuse = page->objects;
1469 if (gfpflags_allow_blocking(flags))
1470 local_irq_disable();
1474 mod_zone_page_state(page_zone(page),
1475 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1476 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1479 inc_slabs_node(s, page_to_nid(page), page->objects);
1484 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1486 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1487 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1491 return allocate_slab(s,
1492 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1495 static void __free_slab(struct kmem_cache *s, struct page *page)
1497 int order = compound_order(page);
1498 int pages = 1 << order;
1500 if (kmem_cache_debug(s)) {
1503 slab_pad_check(s, page);
1504 for_each_object(p, s, page_address(page),
1506 check_object(s, page, p, SLUB_RED_INACTIVE);
1509 kmemcheck_free_shadow(page, compound_order(page));
1511 mod_zone_page_state(page_zone(page),
1512 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1513 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1516 __ClearPageSlabPfmemalloc(page);
1517 __ClearPageSlab(page);
1519 page_mapcount_reset(page);
1520 if (current->reclaim_state)
1521 current->reclaim_state->reclaimed_slab += pages;
1522 __free_kmem_pages(page, order);
1525 #define need_reserve_slab_rcu \
1526 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1528 static void rcu_free_slab(struct rcu_head *h)
1532 if (need_reserve_slab_rcu)
1533 page = virt_to_head_page(h);
1535 page = container_of((struct list_head *)h, struct page, lru);
1537 __free_slab(page->slab_cache, page);
1540 static void free_slab(struct kmem_cache *s, struct page *page)
1542 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1543 struct rcu_head *head;
1545 if (need_reserve_slab_rcu) {
1546 int order = compound_order(page);
1547 int offset = (PAGE_SIZE << order) - s->reserved;
1549 VM_BUG_ON(s->reserved != sizeof(*head));
1550 head = page_address(page) + offset;
1552 head = &page->rcu_head;
1555 call_rcu(head, rcu_free_slab);
1557 __free_slab(s, page);
1560 static void discard_slab(struct kmem_cache *s, struct page *page)
1562 dec_slabs_node(s, page_to_nid(page), page->objects);
1567 * Management of partially allocated slabs.
1570 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1573 if (tail == DEACTIVATE_TO_TAIL)
1574 list_add_tail(&page->lru, &n->partial);
1576 list_add(&page->lru, &n->partial);
1579 static inline void add_partial(struct kmem_cache_node *n,
1580 struct page *page, int tail)
1582 lockdep_assert_held(&n->list_lock);
1583 __add_partial(n, page, tail);
1587 __remove_partial(struct kmem_cache_node *n, struct page *page)
1589 list_del(&page->lru);
1593 static inline void remove_partial(struct kmem_cache_node *n,
1596 lockdep_assert_held(&n->list_lock);
1597 __remove_partial(n, page);
1601 * Remove slab from the partial list, freeze it and
1602 * return the pointer to the freelist.
1604 * Returns a list of objects or NULL if it fails.
1606 static inline void *acquire_slab(struct kmem_cache *s,
1607 struct kmem_cache_node *n, struct page *page,
1608 int mode, int *objects)
1611 unsigned long counters;
1614 lockdep_assert_held(&n->list_lock);
1617 * Zap the freelist and set the frozen bit.
1618 * The old freelist is the list of objects for the
1619 * per cpu allocation list.
1621 freelist = page->freelist;
1622 counters = page->counters;
1623 new.counters = counters;
1624 *objects = new.objects - new.inuse;
1626 new.inuse = page->objects;
1627 new.freelist = NULL;
1629 new.freelist = freelist;
1632 VM_BUG_ON(new.frozen);
1635 if (!__cmpxchg_double_slab(s, page,
1637 new.freelist, new.counters,
1641 remove_partial(n, page);
1646 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1647 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1650 * Try to allocate a partial slab from a specific node.
1652 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1653 struct kmem_cache_cpu *c, gfp_t flags)
1655 struct page *page, *page2;
1656 void *object = NULL;
1661 * Racy check. If we mistakenly see no partial slabs then we
1662 * just allocate an empty slab. If we mistakenly try to get a
1663 * partial slab and there is none available then get_partials()
1666 if (!n || !n->nr_partial)
1669 spin_lock(&n->list_lock);
1670 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1673 if (!pfmemalloc_match(page, flags))
1676 t = acquire_slab(s, n, page, object == NULL, &objects);
1680 available += objects;
1683 stat(s, ALLOC_FROM_PARTIAL);
1686 put_cpu_partial(s, page, 0);
1687 stat(s, CPU_PARTIAL_NODE);
1689 if (!kmem_cache_has_cpu_partial(s)
1690 || available > s->cpu_partial / 2)
1694 spin_unlock(&n->list_lock);
1699 * Get a page from somewhere. Search in increasing NUMA distances.
1701 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1702 struct kmem_cache_cpu *c)
1705 struct zonelist *zonelist;
1708 enum zone_type high_zoneidx = gfp_zone(flags);
1710 unsigned int cpuset_mems_cookie;
1713 * The defrag ratio allows a configuration of the tradeoffs between
1714 * inter node defragmentation and node local allocations. A lower
1715 * defrag_ratio increases the tendency to do local allocations
1716 * instead of attempting to obtain partial slabs from other nodes.
1718 * If the defrag_ratio is set to 0 then kmalloc() always
1719 * returns node local objects. If the ratio is higher then kmalloc()
1720 * may return off node objects because partial slabs are obtained
1721 * from other nodes and filled up.
1723 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1724 * defrag_ratio = 1000) then every (well almost) allocation will
1725 * first attempt to defrag slab caches on other nodes. This means
1726 * scanning over all nodes to look for partial slabs which may be
1727 * expensive if we do it every time we are trying to find a slab
1728 * with available objects.
1730 if (!s->remote_node_defrag_ratio ||
1731 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1735 cpuset_mems_cookie = read_mems_allowed_begin();
1736 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1737 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1738 struct kmem_cache_node *n;
1740 n = get_node(s, zone_to_nid(zone));
1742 if (n && cpuset_zone_allowed(zone, flags) &&
1743 n->nr_partial > s->min_partial) {
1744 object = get_partial_node(s, n, c, flags);
1747 * Don't check read_mems_allowed_retry()
1748 * here - if mems_allowed was updated in
1749 * parallel, that was a harmless race
1750 * between allocation and the cpuset
1757 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1763 * Get a partial page, lock it and return it.
1765 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1766 struct kmem_cache_cpu *c)
1769 int searchnode = node;
1771 if (node == NUMA_NO_NODE)
1772 searchnode = numa_mem_id();
1773 else if (!node_present_pages(node))
1774 searchnode = node_to_mem_node(node);
1776 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1777 if (object || node != NUMA_NO_NODE)
1780 return get_any_partial(s, flags, c);
1783 #ifdef CONFIG_PREEMPT
1785 * Calculate the next globally unique transaction for disambiguiation
1786 * during cmpxchg. The transactions start with the cpu number and are then
1787 * incremented by CONFIG_NR_CPUS.
1789 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1792 * No preemption supported therefore also no need to check for
1798 static inline unsigned long next_tid(unsigned long tid)
1800 return tid + TID_STEP;
1803 static inline unsigned int tid_to_cpu(unsigned long tid)
1805 return tid % TID_STEP;
1808 static inline unsigned long tid_to_event(unsigned long tid)
1810 return tid / TID_STEP;
1813 static inline unsigned int init_tid(int cpu)
1818 static inline void note_cmpxchg_failure(const char *n,
1819 const struct kmem_cache *s, unsigned long tid)
1821 #ifdef SLUB_DEBUG_CMPXCHG
1822 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1824 pr_info("%s %s: cmpxchg redo ", n, s->name);
1826 #ifdef CONFIG_PREEMPT
1827 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1828 pr_warn("due to cpu change %d -> %d\n",
1829 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1832 if (tid_to_event(tid) != tid_to_event(actual_tid))
1833 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1834 tid_to_event(tid), tid_to_event(actual_tid));
1836 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1837 actual_tid, tid, next_tid(tid));
1839 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1842 static void init_kmem_cache_cpus(struct kmem_cache *s)
1846 for_each_possible_cpu(cpu)
1847 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1851 * Remove the cpu slab
1853 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1856 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1857 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1859 enum slab_modes l = M_NONE, m = M_NONE;
1861 int tail = DEACTIVATE_TO_HEAD;
1865 if (page->freelist) {
1866 stat(s, DEACTIVATE_REMOTE_FREES);
1867 tail = DEACTIVATE_TO_TAIL;
1871 * Stage one: Free all available per cpu objects back
1872 * to the page freelist while it is still frozen. Leave the
1875 * There is no need to take the list->lock because the page
1878 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1880 unsigned long counters;
1883 prior = page->freelist;
1884 counters = page->counters;
1885 set_freepointer(s, freelist, prior);
1886 new.counters = counters;
1888 VM_BUG_ON(!new.frozen);
1890 } while (!__cmpxchg_double_slab(s, page,
1892 freelist, new.counters,
1893 "drain percpu freelist"));
1895 freelist = nextfree;
1899 * Stage two: Ensure that the page is unfrozen while the
1900 * list presence reflects the actual number of objects
1903 * We setup the list membership and then perform a cmpxchg
1904 * with the count. If there is a mismatch then the page
1905 * is not unfrozen but the page is on the wrong list.
1907 * Then we restart the process which may have to remove
1908 * the page from the list that we just put it on again
1909 * because the number of objects in the slab may have
1914 old.freelist = page->freelist;
1915 old.counters = page->counters;
1916 VM_BUG_ON(!old.frozen);
1918 /* Determine target state of the slab */
1919 new.counters = old.counters;
1922 set_freepointer(s, freelist, old.freelist);
1923 new.freelist = freelist;
1925 new.freelist = old.freelist;
1929 if (!new.inuse && n->nr_partial >= s->min_partial)
1931 else if (new.freelist) {
1936 * Taking the spinlock removes the possiblity
1937 * that acquire_slab() will see a slab page that
1940 spin_lock(&n->list_lock);
1944 if (kmem_cache_debug(s) && !lock) {
1947 * This also ensures that the scanning of full
1948 * slabs from diagnostic functions will not see
1951 spin_lock(&n->list_lock);
1959 remove_partial(n, page);
1961 else if (l == M_FULL)
1963 remove_full(s, n, page);
1965 if (m == M_PARTIAL) {
1967 add_partial(n, page, tail);
1970 } else if (m == M_FULL) {
1972 stat(s, DEACTIVATE_FULL);
1973 add_full(s, n, page);
1979 if (!__cmpxchg_double_slab(s, page,
1980 old.freelist, old.counters,
1981 new.freelist, new.counters,
1986 spin_unlock(&n->list_lock);
1989 stat(s, DEACTIVATE_EMPTY);
1990 discard_slab(s, page);
1996 * Unfreeze all the cpu partial slabs.
1998 * This function must be called with interrupts disabled
1999 * for the cpu using c (or some other guarantee must be there
2000 * to guarantee no concurrent accesses).
2002 static void unfreeze_partials(struct kmem_cache *s,
2003 struct kmem_cache_cpu *c)
2005 #ifdef CONFIG_SLUB_CPU_PARTIAL
2006 struct kmem_cache_node *n = NULL, *n2 = NULL;
2007 struct page *page, *discard_page = NULL;
2009 while ((page = c->partial)) {
2013 c->partial = page->next;
2015 n2 = get_node(s, page_to_nid(page));
2018 spin_unlock(&n->list_lock);
2021 spin_lock(&n->list_lock);
2026 old.freelist = page->freelist;
2027 old.counters = page->counters;
2028 VM_BUG_ON(!old.frozen);
2030 new.counters = old.counters;
2031 new.freelist = old.freelist;
2035 } while (!__cmpxchg_double_slab(s, page,
2036 old.freelist, old.counters,
2037 new.freelist, new.counters,
2038 "unfreezing slab"));
2040 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2041 page->next = discard_page;
2042 discard_page = page;
2044 add_partial(n, page, DEACTIVATE_TO_TAIL);
2045 stat(s, FREE_ADD_PARTIAL);
2050 spin_unlock(&n->list_lock);
2052 while (discard_page) {
2053 page = discard_page;
2054 discard_page = discard_page->next;
2056 stat(s, DEACTIVATE_EMPTY);
2057 discard_slab(s, page);
2064 * Put a page that was just frozen (in __slab_free) into a partial page
2065 * slot if available. This is done without interrupts disabled and without
2066 * preemption disabled. The cmpxchg is racy and may put the partial page
2067 * onto a random cpus partial slot.
2069 * If we did not find a slot then simply move all the partials to the
2070 * per node partial list.
2072 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2074 #ifdef CONFIG_SLUB_CPU_PARTIAL
2075 struct page *oldpage;
2083 oldpage = this_cpu_read(s->cpu_slab->partial);
2086 pobjects = oldpage->pobjects;
2087 pages = oldpage->pages;
2088 if (drain && pobjects > s->cpu_partial) {
2089 unsigned long flags;
2091 * partial array is full. Move the existing
2092 * set to the per node partial list.
2094 local_irq_save(flags);
2095 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2096 local_irq_restore(flags);
2100 stat(s, CPU_PARTIAL_DRAIN);
2105 pobjects += page->objects - page->inuse;
2107 page->pages = pages;
2108 page->pobjects = pobjects;
2109 page->next = oldpage;
2111 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2113 if (unlikely(!s->cpu_partial)) {
2114 unsigned long flags;
2116 local_irq_save(flags);
2117 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2118 local_irq_restore(flags);
2124 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2126 stat(s, CPUSLAB_FLUSH);
2127 deactivate_slab(s, c->page, c->freelist);
2129 c->tid = next_tid(c->tid);
2137 * Called from IPI handler with interrupts disabled.
2139 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2141 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2147 unfreeze_partials(s, c);
2151 static void flush_cpu_slab(void *d)
2153 struct kmem_cache *s = d;
2155 __flush_cpu_slab(s, smp_processor_id());
2158 static bool has_cpu_slab(int cpu, void *info)
2160 struct kmem_cache *s = info;
2161 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2163 return c->page || c->partial;
2166 static void flush_all(struct kmem_cache *s)
2168 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2172 * Check if the objects in a per cpu structure fit numa
2173 * locality expectations.
2175 static inline int node_match(struct page *page, int node)
2178 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2184 #ifdef CONFIG_SLUB_DEBUG
2185 static int count_free(struct page *page)
2187 return page->objects - page->inuse;
2190 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2192 return atomic_long_read(&n->total_objects);
2194 #endif /* CONFIG_SLUB_DEBUG */
2196 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2197 static unsigned long count_partial(struct kmem_cache_node *n,
2198 int (*get_count)(struct page *))
2200 unsigned long flags;
2201 unsigned long x = 0;
2204 spin_lock_irqsave(&n->list_lock, flags);
2205 list_for_each_entry(page, &n->partial, lru)
2206 x += get_count(page);
2207 spin_unlock_irqrestore(&n->list_lock, flags);
2210 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2212 static noinline void
2213 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2215 #ifdef CONFIG_SLUB_DEBUG
2216 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2217 DEFAULT_RATELIMIT_BURST);
2219 struct kmem_cache_node *n;
2221 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2224 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2226 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2227 s->name, s->object_size, s->size, oo_order(s->oo),
2230 if (oo_order(s->min) > get_order(s->object_size))
2231 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2234 for_each_kmem_cache_node(s, node, n) {
2235 unsigned long nr_slabs;
2236 unsigned long nr_objs;
2237 unsigned long nr_free;
2239 nr_free = count_partial(n, count_free);
2240 nr_slabs = node_nr_slabs(n);
2241 nr_objs = node_nr_objs(n);
2243 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2244 node, nr_slabs, nr_objs, nr_free);
2249 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2250 int node, struct kmem_cache_cpu **pc)
2253 struct kmem_cache_cpu *c = *pc;
2256 freelist = get_partial(s, flags, node, c);
2261 page = new_slab(s, flags, node);
2263 c = raw_cpu_ptr(s->cpu_slab);
2268 * No other reference to the page yet so we can
2269 * muck around with it freely without cmpxchg
2271 freelist = page->freelist;
2272 page->freelist = NULL;
2274 stat(s, ALLOC_SLAB);
2283 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2285 if (unlikely(PageSlabPfmemalloc(page)))
2286 return gfp_pfmemalloc_allowed(gfpflags);
2292 * Check the page->freelist of a page and either transfer the freelist to the
2293 * per cpu freelist or deactivate the page.
2295 * The page is still frozen if the return value is not NULL.
2297 * If this function returns NULL then the page has been unfrozen.
2299 * This function must be called with interrupt disabled.
2301 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2304 unsigned long counters;
2308 freelist = page->freelist;
2309 counters = page->counters;
2311 new.counters = counters;
2312 VM_BUG_ON(!new.frozen);
2314 new.inuse = page->objects;
2315 new.frozen = freelist != NULL;
2317 } while (!__cmpxchg_double_slab(s, page,
2326 * Slow path. The lockless freelist is empty or we need to perform
2329 * Processing is still very fast if new objects have been freed to the
2330 * regular freelist. In that case we simply take over the regular freelist
2331 * as the lockless freelist and zap the regular freelist.
2333 * If that is not working then we fall back to the partial lists. We take the
2334 * first element of the freelist as the object to allocate now and move the
2335 * rest of the freelist to the lockless freelist.
2337 * And if we were unable to get a new slab from the partial slab lists then
2338 * we need to allocate a new slab. This is the slowest path since it involves
2339 * a call to the page allocator and the setup of a new slab.
2341 * Version of __slab_alloc to use when we know that interrupts are
2342 * already disabled (which is the case for bulk allocation).
2344 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2345 unsigned long addr, struct kmem_cache_cpu *c)
2355 if (unlikely(!node_match(page, node))) {
2356 int searchnode = node;
2358 if (node != NUMA_NO_NODE && !node_present_pages(node))
2359 searchnode = node_to_mem_node(node);
2361 if (unlikely(!node_match(page, searchnode))) {
2362 stat(s, ALLOC_NODE_MISMATCH);
2363 deactivate_slab(s, page, c->freelist);
2371 * By rights, we should be searching for a slab page that was
2372 * PFMEMALLOC but right now, we are losing the pfmemalloc
2373 * information when the page leaves the per-cpu allocator
2375 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2376 deactivate_slab(s, page, c->freelist);
2382 /* must check again c->freelist in case of cpu migration or IRQ */
2383 freelist = c->freelist;
2387 freelist = get_freelist(s, page);
2391 stat(s, DEACTIVATE_BYPASS);
2395 stat(s, ALLOC_REFILL);
2399 * freelist is pointing to the list of objects to be used.
2400 * page is pointing to the page from which the objects are obtained.
2401 * That page must be frozen for per cpu allocations to work.
2403 VM_BUG_ON(!c->page->frozen);
2404 c->freelist = get_freepointer(s, freelist);
2405 c->tid = next_tid(c->tid);
2411 page = c->page = c->partial;
2412 c->partial = page->next;
2413 stat(s, CPU_PARTIAL_ALLOC);
2418 freelist = new_slab_objects(s, gfpflags, node, &c);
2420 if (unlikely(!freelist)) {
2421 slab_out_of_memory(s, gfpflags, node);
2426 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2429 /* Only entered in the debug case */
2430 if (kmem_cache_debug(s) &&
2431 !alloc_debug_processing(s, page, freelist, addr))
2432 goto new_slab; /* Slab failed checks. Next slab needed */
2434 deactivate_slab(s, page, get_freepointer(s, freelist));
2441 * Another one that disabled interrupt and compensates for possible
2442 * cpu changes by refetching the per cpu area pointer.
2444 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2445 unsigned long addr, struct kmem_cache_cpu *c)
2448 unsigned long flags;
2450 local_irq_save(flags);
2451 #ifdef CONFIG_PREEMPT
2453 * We may have been preempted and rescheduled on a different
2454 * cpu before disabling interrupts. Need to reload cpu area
2457 c = this_cpu_ptr(s->cpu_slab);
2460 p = ___slab_alloc(s, gfpflags, node, addr, c);
2461 local_irq_restore(flags);
2466 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2467 * have the fastpath folded into their functions. So no function call
2468 * overhead for requests that can be satisfied on the fastpath.
2470 * The fastpath works by first checking if the lockless freelist can be used.
2471 * If not then __slab_alloc is called for slow processing.
2473 * Otherwise we can simply pick the next object from the lockless free list.
2475 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2476 gfp_t gfpflags, int node, unsigned long addr)
2479 struct kmem_cache_cpu *c;
2483 s = slab_pre_alloc_hook(s, gfpflags);
2488 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2489 * enabled. We may switch back and forth between cpus while
2490 * reading from one cpu area. That does not matter as long
2491 * as we end up on the original cpu again when doing the cmpxchg.
2493 * We should guarantee that tid and kmem_cache are retrieved on
2494 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2495 * to check if it is matched or not.
2498 tid = this_cpu_read(s->cpu_slab->tid);
2499 c = raw_cpu_ptr(s->cpu_slab);
2500 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2501 unlikely(tid != READ_ONCE(c->tid)));
2504 * Irqless object alloc/free algorithm used here depends on sequence
2505 * of fetching cpu_slab's data. tid should be fetched before anything
2506 * on c to guarantee that object and page associated with previous tid
2507 * won't be used with current tid. If we fetch tid first, object and
2508 * page could be one associated with next tid and our alloc/free
2509 * request will be failed. In this case, we will retry. So, no problem.
2514 * The transaction ids are globally unique per cpu and per operation on
2515 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2516 * occurs on the right processor and that there was no operation on the
2517 * linked list in between.
2520 object = c->freelist;
2522 if (unlikely(!object || !node_match(page, node))) {
2523 object = __slab_alloc(s, gfpflags, node, addr, c);
2524 stat(s, ALLOC_SLOWPATH);
2526 void *next_object = get_freepointer_safe(s, object);
2529 * The cmpxchg will only match if there was no additional
2530 * operation and if we are on the right processor.
2532 * The cmpxchg does the following atomically (without lock
2534 * 1. Relocate first pointer to the current per cpu area.
2535 * 2. Verify that tid and freelist have not been changed
2536 * 3. If they were not changed replace tid and freelist
2538 * Since this is without lock semantics the protection is only
2539 * against code executing on this cpu *not* from access by
2542 if (unlikely(!this_cpu_cmpxchg_double(
2543 s->cpu_slab->freelist, s->cpu_slab->tid,
2545 next_object, next_tid(tid)))) {
2547 note_cmpxchg_failure("slab_alloc", s, tid);
2550 prefetch_freepointer(s, next_object);
2551 stat(s, ALLOC_FASTPATH);
2554 if (unlikely(gfpflags & __GFP_ZERO) && object)
2555 memset(object, 0, s->object_size);
2557 slab_post_alloc_hook(s, gfpflags, object);
2562 static __always_inline void *slab_alloc(struct kmem_cache *s,
2563 gfp_t gfpflags, unsigned long addr)
2565 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2568 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2570 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2572 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2577 EXPORT_SYMBOL(kmem_cache_alloc);
2579 #ifdef CONFIG_TRACING
2580 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2582 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2583 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2584 kasan_kmalloc(s, ret, size);
2587 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2591 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2593 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2595 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2596 s->object_size, s->size, gfpflags, node);
2600 EXPORT_SYMBOL(kmem_cache_alloc_node);
2602 #ifdef CONFIG_TRACING
2603 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2605 int node, size_t size)
2607 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2609 trace_kmalloc_node(_RET_IP_, ret,
2610 size, s->size, gfpflags, node);
2612 kasan_kmalloc(s, ret, size);
2615 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2620 * Slow path handling. This may still be called frequently since objects
2621 * have a longer lifetime than the cpu slabs in most processing loads.
2623 * So we still attempt to reduce cache line usage. Just take the slab
2624 * lock and free the item. If there is no additional partial page
2625 * handling required then we can return immediately.
2627 static void __slab_free(struct kmem_cache *s, struct page *page,
2628 void *head, void *tail, int cnt,
2635 unsigned long counters;
2636 struct kmem_cache_node *n = NULL;
2637 unsigned long uninitialized_var(flags);
2639 stat(s, FREE_SLOWPATH);
2641 if (kmem_cache_debug(s) &&
2642 !(n = free_debug_processing(s, page, head, tail, cnt,
2648 spin_unlock_irqrestore(&n->list_lock, flags);
2651 prior = page->freelist;
2652 counters = page->counters;
2653 set_freepointer(s, tail, prior);
2654 new.counters = counters;
2655 was_frozen = new.frozen;
2657 if ((!new.inuse || !prior) && !was_frozen) {
2659 if (kmem_cache_has_cpu_partial(s) && !prior) {
2662 * Slab was on no list before and will be
2664 * We can defer the list move and instead
2669 } else { /* Needs to be taken off a list */
2671 n = get_node(s, page_to_nid(page));
2673 * Speculatively acquire the list_lock.
2674 * If the cmpxchg does not succeed then we may
2675 * drop the list_lock without any processing.
2677 * Otherwise the list_lock will synchronize with
2678 * other processors updating the list of slabs.
2680 spin_lock_irqsave(&n->list_lock, flags);
2685 } while (!cmpxchg_double_slab(s, page,
2693 * If we just froze the page then put it onto the
2694 * per cpu partial list.
2696 if (new.frozen && !was_frozen) {
2697 put_cpu_partial(s, page, 1);
2698 stat(s, CPU_PARTIAL_FREE);
2701 * The list lock was not taken therefore no list
2702 * activity can be necessary.
2705 stat(s, FREE_FROZEN);
2709 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2713 * Objects left in the slab. If it was not on the partial list before
2716 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2717 if (kmem_cache_debug(s))
2718 remove_full(s, n, page);
2719 add_partial(n, page, DEACTIVATE_TO_TAIL);
2720 stat(s, FREE_ADD_PARTIAL);
2722 spin_unlock_irqrestore(&n->list_lock, flags);
2728 * Slab on the partial list.
2730 remove_partial(n, page);
2731 stat(s, FREE_REMOVE_PARTIAL);
2733 /* Slab must be on the full list */
2734 remove_full(s, n, page);
2737 spin_unlock_irqrestore(&n->list_lock, flags);
2739 discard_slab(s, page);
2743 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2744 * can perform fastpath freeing without additional function calls.
2746 * The fastpath is only possible if we are freeing to the current cpu slab
2747 * of this processor. This typically the case if we have just allocated
2750 * If fastpath is not possible then fall back to __slab_free where we deal
2751 * with all sorts of special processing.
2753 * Bulk free of a freelist with several objects (all pointing to the
2754 * same page) possible by specifying head and tail ptr, plus objects
2755 * count (cnt). Bulk free indicated by tail pointer being set.
2757 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2758 void *head, void *tail, int cnt,
2761 void *tail_obj = tail ? : head;
2762 struct kmem_cache_cpu *c;
2765 slab_free_freelist_hook(s, head, tail);
2769 * Determine the currently cpus per cpu slab.
2770 * The cpu may change afterward. However that does not matter since
2771 * data is retrieved via this pointer. If we are on the same cpu
2772 * during the cmpxchg then the free will succeed.
2775 tid = this_cpu_read(s->cpu_slab->tid);
2776 c = raw_cpu_ptr(s->cpu_slab);
2777 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2778 unlikely(tid != READ_ONCE(c->tid)));
2780 /* Same with comment on barrier() in slab_alloc_node() */
2783 if (likely(page == c->page)) {
2784 set_freepointer(s, tail_obj, c->freelist);
2786 if (unlikely(!this_cpu_cmpxchg_double(
2787 s->cpu_slab->freelist, s->cpu_slab->tid,
2789 head, next_tid(tid)))) {
2791 note_cmpxchg_failure("slab_free", s, tid);
2794 stat(s, FREE_FASTPATH);
2796 __slab_free(s, page, head, tail_obj, cnt, addr);
2800 void kmem_cache_free(struct kmem_cache *s, void *x)
2802 s = cache_from_obj(s, x);
2805 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2806 trace_kmem_cache_free(_RET_IP_, x);
2808 EXPORT_SYMBOL(kmem_cache_free);
2810 struct detached_freelist {
2818 * This function progressively scans the array with free objects (with
2819 * a limited look ahead) and extract objects belonging to the same
2820 * page. It builds a detached freelist directly within the given
2821 * page/objects. This can happen without any need for
2822 * synchronization, because the objects are owned by running process.
2823 * The freelist is build up as a single linked list in the objects.
2824 * The idea is, that this detached freelist can then be bulk
2825 * transferred to the real freelist(s), but only requiring a single
2826 * synchronization primitive. Look ahead in the array is limited due
2827 * to performance reasons.
2829 static int build_detached_freelist(struct kmem_cache *s, size_t size,
2830 void **p, struct detached_freelist *df)
2832 size_t first_skipped_index = 0;
2836 /* Always re-init detached_freelist */
2841 } while (!object && size);
2846 /* Start new detached freelist */
2847 set_freepointer(s, object, NULL);
2848 df->page = virt_to_head_page(object);
2850 df->freelist = object;
2851 p[size] = NULL; /* mark object processed */
2857 continue; /* Skip processed objects */
2859 /* df->page is always set at this point */
2860 if (df->page == virt_to_head_page(object)) {
2861 /* Opportunity build freelist */
2862 set_freepointer(s, object, df->freelist);
2863 df->freelist = object;
2865 p[size] = NULL; /* mark object processed */
2870 /* Limit look ahead search */
2874 if (!first_skipped_index)
2875 first_skipped_index = size + 1;
2878 return first_skipped_index;
2882 /* Note that interrupts must be enabled when calling this function. */
2883 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2889 struct detached_freelist df;
2891 size = build_detached_freelist(s, size, p, &df);
2892 if (unlikely(!df.page))
2895 slab_free(s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
2896 } while (likely(size));
2898 EXPORT_SYMBOL(kmem_cache_free_bulk);
2900 /* Note that interrupts must be enabled when calling this function. */
2901 bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2904 struct kmem_cache_cpu *c;
2908 * Drain objects in the per cpu slab, while disabling local
2909 * IRQs, which protects against PREEMPT and interrupts
2910 * handlers invoking normal fastpath.
2912 local_irq_disable();
2913 c = this_cpu_ptr(s->cpu_slab);
2915 for (i = 0; i < size; i++) {
2916 void *object = c->freelist;
2918 if (unlikely(!object)) {
2920 * Invoking slow path likely have side-effect
2921 * of re-populating per CPU c->freelist
2923 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2925 if (unlikely(!p[i]))
2928 c = this_cpu_ptr(s->cpu_slab);
2929 continue; /* goto for-loop */
2932 /* kmem_cache debug support */
2933 s = slab_pre_alloc_hook(s, flags);
2937 c->freelist = get_freepointer(s, object);
2940 /* kmem_cache debug support */
2941 slab_post_alloc_hook(s, flags, object);
2943 c->tid = next_tid(c->tid);
2946 /* Clear memory outside IRQ disabled fastpath loop */
2947 if (unlikely(flags & __GFP_ZERO)) {
2950 for (j = 0; j < i; j++)
2951 memset(p[j], 0, s->object_size);
2957 __kmem_cache_free_bulk(s, i, p);
2961 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2965 * Object placement in a slab is made very easy because we always start at
2966 * offset 0. If we tune the size of the object to the alignment then we can
2967 * get the required alignment by putting one properly sized object after
2970 * Notice that the allocation order determines the sizes of the per cpu
2971 * caches. Each processor has always one slab available for allocations.
2972 * Increasing the allocation order reduces the number of times that slabs
2973 * must be moved on and off the partial lists and is therefore a factor in
2978 * Mininum / Maximum order of slab pages. This influences locking overhead
2979 * and slab fragmentation. A higher order reduces the number of partial slabs
2980 * and increases the number of allocations possible without having to
2981 * take the list_lock.
2983 static int slub_min_order;
2984 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2985 static int slub_min_objects;
2988 * Calculate the order of allocation given an slab object size.
2990 * The order of allocation has significant impact on performance and other
2991 * system components. Generally order 0 allocations should be preferred since
2992 * order 0 does not cause fragmentation in the page allocator. Larger objects
2993 * be problematic to put into order 0 slabs because there may be too much
2994 * unused space left. We go to a higher order if more than 1/16th of the slab
2997 * In order to reach satisfactory performance we must ensure that a minimum
2998 * number of objects is in one slab. Otherwise we may generate too much
2999 * activity on the partial lists which requires taking the list_lock. This is
3000 * less a concern for large slabs though which are rarely used.
3002 * slub_max_order specifies the order where we begin to stop considering the
3003 * number of objects in a slab as critical. If we reach slub_max_order then
3004 * we try to keep the page order as low as possible. So we accept more waste
3005 * of space in favor of a small page order.
3007 * Higher order allocations also allow the placement of more objects in a
3008 * slab and thereby reduce object handling overhead. If the user has
3009 * requested a higher mininum order then we start with that one instead of
3010 * the smallest order which will fit the object.
3012 static inline int slab_order(int size, int min_objects,
3013 int max_order, int fract_leftover, int reserved)
3017 int min_order = slub_min_order;
3019 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3020 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3022 for (order = max(min_order, get_order(min_objects * size + reserved));
3023 order <= max_order; order++) {
3025 unsigned long slab_size = PAGE_SIZE << order;
3027 rem = (slab_size - reserved) % size;
3029 if (rem <= slab_size / fract_leftover)
3036 static inline int calculate_order(int size, int reserved)
3044 * Attempt to find best configuration for a slab. This
3045 * works by first attempting to generate a layout with
3046 * the best configuration and backing off gradually.
3048 * First we increase the acceptable waste in a slab. Then
3049 * we reduce the minimum objects required in a slab.
3051 min_objects = slub_min_objects;
3053 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3054 max_objects = order_objects(slub_max_order, size, reserved);
3055 min_objects = min(min_objects, max_objects);
3057 while (min_objects > 1) {
3059 while (fraction >= 4) {
3060 order = slab_order(size, min_objects,
3061 slub_max_order, fraction, reserved);
3062 if (order <= slub_max_order)
3070 * We were unable to place multiple objects in a slab. Now
3071 * lets see if we can place a single object there.
3073 order = slab_order(size, 1, slub_max_order, 1, reserved);
3074 if (order <= slub_max_order)
3078 * Doh this slab cannot be placed using slub_max_order.
3080 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3081 if (order < MAX_ORDER)
3087 init_kmem_cache_node(struct kmem_cache_node *n)
3090 spin_lock_init(&n->list_lock);
3091 INIT_LIST_HEAD(&n->partial);
3092 #ifdef CONFIG_SLUB_DEBUG
3093 atomic_long_set(&n->nr_slabs, 0);
3094 atomic_long_set(&n->total_objects, 0);
3095 INIT_LIST_HEAD(&n->full);
3099 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3101 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3102 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3105 * Must align to double word boundary for the double cmpxchg
3106 * instructions to work; see __pcpu_double_call_return_bool().
3108 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3109 2 * sizeof(void *));
3114 init_kmem_cache_cpus(s);
3119 static struct kmem_cache *kmem_cache_node;
3122 * No kmalloc_node yet so do it by hand. We know that this is the first
3123 * slab on the node for this slabcache. There are no concurrent accesses
3126 * Note that this function only works on the kmem_cache_node
3127 * when allocating for the kmem_cache_node. This is used for bootstrapping
3128 * memory on a fresh node that has no slab structures yet.
3130 static void early_kmem_cache_node_alloc(int node)
3133 struct kmem_cache_node *n;
3135 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3137 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3140 if (page_to_nid(page) != node) {
3141 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3142 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3147 page->freelist = get_freepointer(kmem_cache_node, n);
3150 kmem_cache_node->node[node] = n;
3151 #ifdef CONFIG_SLUB_DEBUG
3152 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3153 init_tracking(kmem_cache_node, n);
3155 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3156 init_kmem_cache_node(n);
3157 inc_slabs_node(kmem_cache_node, node, page->objects);
3160 * No locks need to be taken here as it has just been
3161 * initialized and there is no concurrent access.
3163 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3166 static void free_kmem_cache_nodes(struct kmem_cache *s)
3169 struct kmem_cache_node *n;
3171 for_each_kmem_cache_node(s, node, n) {
3172 kmem_cache_free(kmem_cache_node, n);
3173 s->node[node] = NULL;
3177 static int init_kmem_cache_nodes(struct kmem_cache *s)
3181 for_each_node_state(node, N_NORMAL_MEMORY) {
3182 struct kmem_cache_node *n;
3184 if (slab_state == DOWN) {
3185 early_kmem_cache_node_alloc(node);
3188 n = kmem_cache_alloc_node(kmem_cache_node,
3192 free_kmem_cache_nodes(s);
3197 init_kmem_cache_node(n);
3202 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3204 if (min < MIN_PARTIAL)
3206 else if (min > MAX_PARTIAL)
3208 s->min_partial = min;
3212 * calculate_sizes() determines the order and the distribution of data within
3215 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3217 unsigned long flags = s->flags;
3218 unsigned long size = s->object_size;
3222 * Round up object size to the next word boundary. We can only
3223 * place the free pointer at word boundaries and this determines
3224 * the possible location of the free pointer.
3226 size = ALIGN(size, sizeof(void *));
3228 #ifdef CONFIG_SLUB_DEBUG
3230 * Determine if we can poison the object itself. If the user of
3231 * the slab may touch the object after free or before allocation
3232 * then we should never poison the object itself.
3234 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3236 s->flags |= __OBJECT_POISON;
3238 s->flags &= ~__OBJECT_POISON;
3242 * If we are Redzoning then check if there is some space between the
3243 * end of the object and the free pointer. If not then add an
3244 * additional word to have some bytes to store Redzone information.
3246 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3247 size += sizeof(void *);
3251 * With that we have determined the number of bytes in actual use
3252 * by the object. This is the potential offset to the free pointer.
3256 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3259 * Relocate free pointer after the object if it is not
3260 * permitted to overwrite the first word of the object on
3263 * This is the case if we do RCU, have a constructor or
3264 * destructor or are poisoning the objects.
3267 size += sizeof(void *);
3270 #ifdef CONFIG_SLUB_DEBUG
3271 if (flags & SLAB_STORE_USER)
3273 * Need to store information about allocs and frees after
3276 size += 2 * sizeof(struct track);
3278 if (flags & SLAB_RED_ZONE)
3280 * Add some empty padding so that we can catch
3281 * overwrites from earlier objects rather than let
3282 * tracking information or the free pointer be
3283 * corrupted if a user writes before the start
3286 size += sizeof(void *);
3290 * SLUB stores one object immediately after another beginning from
3291 * offset 0. In order to align the objects we have to simply size
3292 * each object to conform to the alignment.
3294 size = ALIGN(size, s->align);
3296 if (forced_order >= 0)
3297 order = forced_order;
3299 order = calculate_order(size, s->reserved);
3306 s->allocflags |= __GFP_COMP;
3308 if (s->flags & SLAB_CACHE_DMA)
3309 s->allocflags |= GFP_DMA;
3311 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3312 s->allocflags |= __GFP_RECLAIMABLE;
3315 * Determine the number of objects per slab
3317 s->oo = oo_make(order, size, s->reserved);
3318 s->min = oo_make(get_order(size), size, s->reserved);
3319 if (oo_objects(s->oo) > oo_objects(s->max))
3322 return !!oo_objects(s->oo);
3325 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3327 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3330 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3331 s->reserved = sizeof(struct rcu_head);
3333 if (!calculate_sizes(s, -1))
3335 if (disable_higher_order_debug) {
3337 * Disable debugging flags that store metadata if the min slab
3340 if (get_order(s->size) > get_order(s->object_size)) {
3341 s->flags &= ~DEBUG_METADATA_FLAGS;
3343 if (!calculate_sizes(s, -1))
3348 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3349 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3350 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3351 /* Enable fast mode */
3352 s->flags |= __CMPXCHG_DOUBLE;
3356 * The larger the object size is, the more pages we want on the partial
3357 * list to avoid pounding the page allocator excessively.
3359 set_min_partial(s, ilog2(s->size) / 2);
3362 * cpu_partial determined the maximum number of objects kept in the
3363 * per cpu partial lists of a processor.
3365 * Per cpu partial lists mainly contain slabs that just have one
3366 * object freed. If they are used for allocation then they can be
3367 * filled up again with minimal effort. The slab will never hit the
3368 * per node partial lists and therefore no locking will be required.
3370 * This setting also determines
3372 * A) The number of objects from per cpu partial slabs dumped to the
3373 * per node list when we reach the limit.
3374 * B) The number of objects in cpu partial slabs to extract from the
3375 * per node list when we run out of per cpu objects. We only fetch
3376 * 50% to keep some capacity around for frees.
3378 if (!kmem_cache_has_cpu_partial(s))
3380 else if (s->size >= PAGE_SIZE)
3382 else if (s->size >= 1024)
3384 else if (s->size >= 256)
3385 s->cpu_partial = 13;
3387 s->cpu_partial = 30;
3390 s->remote_node_defrag_ratio = 1000;
3392 if (!init_kmem_cache_nodes(s))
3395 if (alloc_kmem_cache_cpus(s))
3398 free_kmem_cache_nodes(s);
3400 if (flags & SLAB_PANIC)
3401 panic("Cannot create slab %s size=%lu realsize=%u "
3402 "order=%u offset=%u flags=%lx\n",
3403 s->name, (unsigned long)s->size, s->size,
3404 oo_order(s->oo), s->offset, flags);
3408 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3411 #ifdef CONFIG_SLUB_DEBUG
3412 void *addr = page_address(page);
3414 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3415 sizeof(long), GFP_ATOMIC);
3418 slab_err(s, page, text, s->name);
3421 get_map(s, page, map);
3422 for_each_object(p, s, addr, page->objects) {
3424 if (!test_bit(slab_index(p, s, addr), map)) {
3425 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3426 print_tracking(s, p);
3435 * Attempt to free all partial slabs on a node.
3436 * This is called from kmem_cache_close(). We must be the last thread
3437 * using the cache and therefore we do not need to lock anymore.
3439 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3441 struct page *page, *h;
3443 list_for_each_entry_safe(page, h, &n->partial, lru) {
3445 __remove_partial(n, page);
3446 discard_slab(s, page);
3448 list_slab_objects(s, page,
3449 "Objects remaining in %s on kmem_cache_close()");
3455 * Release all resources used by a slab cache.
3457 static inline int kmem_cache_close(struct kmem_cache *s)
3460 struct kmem_cache_node *n;
3463 /* Attempt to free all objects */
3464 for_each_kmem_cache_node(s, node, n) {
3466 if (n->nr_partial || slabs_node(s, node))
3469 free_percpu(s->cpu_slab);
3470 free_kmem_cache_nodes(s);
3474 int __kmem_cache_shutdown(struct kmem_cache *s)
3476 return kmem_cache_close(s);
3479 /********************************************************************
3481 *******************************************************************/
3483 static int __init setup_slub_min_order(char *str)
3485 get_option(&str, &slub_min_order);
3490 __setup("slub_min_order=", setup_slub_min_order);
3492 static int __init setup_slub_max_order(char *str)
3494 get_option(&str, &slub_max_order);
3495 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3500 __setup("slub_max_order=", setup_slub_max_order);
3502 static int __init setup_slub_min_objects(char *str)
3504 get_option(&str, &slub_min_objects);
3509 __setup("slub_min_objects=", setup_slub_min_objects);
3511 void *__kmalloc(size_t size, gfp_t flags)
3513 struct kmem_cache *s;
3516 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3517 return kmalloc_large(size, flags);
3519 s = kmalloc_slab(size, flags);
3521 if (unlikely(ZERO_OR_NULL_PTR(s)))
3524 ret = slab_alloc(s, flags, _RET_IP_);
3526 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3528 kasan_kmalloc(s, ret, size);
3532 EXPORT_SYMBOL(__kmalloc);
3535 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3540 flags |= __GFP_COMP | __GFP_NOTRACK;
3541 page = alloc_kmem_pages_node(node, flags, get_order(size));
3543 ptr = page_address(page);
3545 kmalloc_large_node_hook(ptr, size, flags);
3549 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3551 struct kmem_cache *s;
3554 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3555 ret = kmalloc_large_node(size, flags, node);
3557 trace_kmalloc_node(_RET_IP_, ret,
3558 size, PAGE_SIZE << get_order(size),
3564 s = kmalloc_slab(size, flags);
3566 if (unlikely(ZERO_OR_NULL_PTR(s)))
3569 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3571 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3573 kasan_kmalloc(s, ret, size);
3577 EXPORT_SYMBOL(__kmalloc_node);
3580 static size_t __ksize(const void *object)
3584 if (unlikely(object == ZERO_SIZE_PTR))
3587 page = virt_to_head_page(object);
3589 if (unlikely(!PageSlab(page))) {
3590 WARN_ON(!PageCompound(page));
3591 return PAGE_SIZE << compound_order(page);
3594 return slab_ksize(page->slab_cache);
3597 size_t ksize(const void *object)
3599 size_t size = __ksize(object);
3600 /* We assume that ksize callers could use whole allocated area,
3601 so we need unpoison this area. */
3602 kasan_krealloc(object, size);
3605 EXPORT_SYMBOL(ksize);
3607 void kfree(const void *x)
3610 void *object = (void *)x;
3612 trace_kfree(_RET_IP_, x);
3614 if (unlikely(ZERO_OR_NULL_PTR(x)))
3617 page = virt_to_head_page(x);
3618 if (unlikely(!PageSlab(page))) {
3619 BUG_ON(!PageCompound(page));
3621 __free_kmem_pages(page, compound_order(page));
3624 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3626 EXPORT_SYMBOL(kfree);
3628 #define SHRINK_PROMOTE_MAX 32
3631 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3632 * up most to the head of the partial lists. New allocations will then
3633 * fill those up and thus they can be removed from the partial lists.
3635 * The slabs with the least items are placed last. This results in them
3636 * being allocated from last increasing the chance that the last objects
3637 * are freed in them.
3639 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3643 struct kmem_cache_node *n;
3646 struct list_head discard;
3647 struct list_head promote[SHRINK_PROMOTE_MAX];
3648 unsigned long flags;
3653 * Disable empty slabs caching. Used to avoid pinning offline
3654 * memory cgroups by kmem pages that can be freed.
3660 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3661 * so we have to make sure the change is visible.
3663 kick_all_cpus_sync();
3667 for_each_kmem_cache_node(s, node, n) {
3668 INIT_LIST_HEAD(&discard);
3669 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3670 INIT_LIST_HEAD(promote + i);
3672 spin_lock_irqsave(&n->list_lock, flags);
3675 * Build lists of slabs to discard or promote.
3677 * Note that concurrent frees may occur while we hold the
3678 * list_lock. page->inuse here is the upper limit.
3680 list_for_each_entry_safe(page, t, &n->partial, lru) {
3681 int free = page->objects - page->inuse;
3683 /* Do not reread page->inuse */
3686 /* We do not keep full slabs on the list */
3689 if (free == page->objects) {
3690 list_move(&page->lru, &discard);
3692 } else if (free <= SHRINK_PROMOTE_MAX)
3693 list_move(&page->lru, promote + free - 1);
3697 * Promote the slabs filled up most to the head of the
3700 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3701 list_splice(promote + i, &n->partial);
3703 spin_unlock_irqrestore(&n->list_lock, flags);
3705 /* Release empty slabs */
3706 list_for_each_entry_safe(page, t, &discard, lru)
3707 discard_slab(s, page);
3709 if (slabs_node(s, node))
3716 static int slab_mem_going_offline_callback(void *arg)
3718 struct kmem_cache *s;
3720 mutex_lock(&slab_mutex);
3721 list_for_each_entry(s, &slab_caches, list)
3722 __kmem_cache_shrink(s, false);
3723 mutex_unlock(&slab_mutex);
3728 static void slab_mem_offline_callback(void *arg)
3730 struct kmem_cache_node *n;
3731 struct kmem_cache *s;
3732 struct memory_notify *marg = arg;
3735 offline_node = marg->status_change_nid_normal;
3738 * If the node still has available memory. we need kmem_cache_node
3741 if (offline_node < 0)
3744 mutex_lock(&slab_mutex);
3745 list_for_each_entry(s, &slab_caches, list) {
3746 n = get_node(s, offline_node);
3749 * if n->nr_slabs > 0, slabs still exist on the node
3750 * that is going down. We were unable to free them,
3751 * and offline_pages() function shouldn't call this
3752 * callback. So, we must fail.
3754 BUG_ON(slabs_node(s, offline_node));
3756 s->node[offline_node] = NULL;
3757 kmem_cache_free(kmem_cache_node, n);
3760 mutex_unlock(&slab_mutex);
3763 static int slab_mem_going_online_callback(void *arg)
3765 struct kmem_cache_node *n;
3766 struct kmem_cache *s;
3767 struct memory_notify *marg = arg;
3768 int nid = marg->status_change_nid_normal;
3772 * If the node's memory is already available, then kmem_cache_node is
3773 * already created. Nothing to do.
3779 * We are bringing a node online. No memory is available yet. We must
3780 * allocate a kmem_cache_node structure in order to bring the node
3783 mutex_lock(&slab_mutex);
3784 list_for_each_entry(s, &slab_caches, list) {
3786 * XXX: kmem_cache_alloc_node will fallback to other nodes
3787 * since memory is not yet available from the node that
3790 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3795 init_kmem_cache_node(n);
3799 mutex_unlock(&slab_mutex);
3803 static int slab_memory_callback(struct notifier_block *self,
3804 unsigned long action, void *arg)
3809 case MEM_GOING_ONLINE:
3810 ret = slab_mem_going_online_callback(arg);
3812 case MEM_GOING_OFFLINE:
3813 ret = slab_mem_going_offline_callback(arg);
3816 case MEM_CANCEL_ONLINE:
3817 slab_mem_offline_callback(arg);
3820 case MEM_CANCEL_OFFLINE:
3824 ret = notifier_from_errno(ret);
3830 static struct notifier_block slab_memory_callback_nb = {
3831 .notifier_call = slab_memory_callback,
3832 .priority = SLAB_CALLBACK_PRI,
3835 /********************************************************************
3836 * Basic setup of slabs
3837 *******************************************************************/
3840 * Used for early kmem_cache structures that were allocated using
3841 * the page allocator. Allocate them properly then fix up the pointers
3842 * that may be pointing to the wrong kmem_cache structure.
3845 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3848 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3849 struct kmem_cache_node *n;
3851 memcpy(s, static_cache, kmem_cache->object_size);
3854 * This runs very early, and only the boot processor is supposed to be
3855 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3858 __flush_cpu_slab(s, smp_processor_id());
3859 for_each_kmem_cache_node(s, node, n) {
3862 list_for_each_entry(p, &n->partial, lru)
3865 #ifdef CONFIG_SLUB_DEBUG
3866 list_for_each_entry(p, &n->full, lru)
3870 slab_init_memcg_params(s);
3871 list_add(&s->list, &slab_caches);
3875 void __init kmem_cache_init(void)
3877 static __initdata struct kmem_cache boot_kmem_cache,
3878 boot_kmem_cache_node;
3880 if (debug_guardpage_minorder())
3883 kmem_cache_node = &boot_kmem_cache_node;
3884 kmem_cache = &boot_kmem_cache;
3886 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3887 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3889 register_hotmemory_notifier(&slab_memory_callback_nb);
3891 /* Able to allocate the per node structures */
3892 slab_state = PARTIAL;
3894 create_boot_cache(kmem_cache, "kmem_cache",
3895 offsetof(struct kmem_cache, node) +
3896 nr_node_ids * sizeof(struct kmem_cache_node *),
3897 SLAB_HWCACHE_ALIGN);
3899 kmem_cache = bootstrap(&boot_kmem_cache);
3902 * Allocate kmem_cache_node properly from the kmem_cache slab.
3903 * kmem_cache_node is separately allocated so no need to
3904 * update any list pointers.
3906 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3908 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3909 setup_kmalloc_cache_index_table();
3910 create_kmalloc_caches(0);
3913 register_cpu_notifier(&slab_notifier);
3916 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3918 slub_min_order, slub_max_order, slub_min_objects,
3919 nr_cpu_ids, nr_node_ids);
3922 void __init kmem_cache_init_late(void)
3927 __kmem_cache_alias(const char *name, size_t size, size_t align,
3928 unsigned long flags, void (*ctor)(void *))
3930 struct kmem_cache *s, *c;
3932 s = find_mergeable(size, align, flags, name, ctor);
3937 * Adjust the object sizes so that we clear
3938 * the complete object on kzalloc.
3940 s->object_size = max(s->object_size, (int)size);
3941 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3943 for_each_memcg_cache(c, s) {
3944 c->object_size = s->object_size;
3945 c->inuse = max_t(int, c->inuse,
3946 ALIGN(size, sizeof(void *)));
3949 if (sysfs_slab_alias(s, name)) {
3958 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3962 err = kmem_cache_open(s, flags);
3966 /* Mutex is not taken during early boot */
3967 if (slab_state <= UP)
3970 memcg_propagate_slab_attrs(s);
3971 err = sysfs_slab_add(s);
3973 kmem_cache_close(s);
3980 * Use the cpu notifier to insure that the cpu slabs are flushed when
3983 static int slab_cpuup_callback(struct notifier_block *nfb,
3984 unsigned long action, void *hcpu)
3986 long cpu = (long)hcpu;
3987 struct kmem_cache *s;
3988 unsigned long flags;
3991 case CPU_UP_CANCELED:
3992 case CPU_UP_CANCELED_FROZEN:
3994 case CPU_DEAD_FROZEN:
3995 mutex_lock(&slab_mutex);
3996 list_for_each_entry(s, &slab_caches, list) {
3997 local_irq_save(flags);
3998 __flush_cpu_slab(s, cpu);
3999 local_irq_restore(flags);
4001 mutex_unlock(&slab_mutex);
4009 static struct notifier_block slab_notifier = {
4010 .notifier_call = slab_cpuup_callback
4015 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4017 struct kmem_cache *s;
4020 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4021 return kmalloc_large(size, gfpflags);
4023 s = kmalloc_slab(size, gfpflags);
4025 if (unlikely(ZERO_OR_NULL_PTR(s)))
4028 ret = slab_alloc(s, gfpflags, caller);
4030 /* Honor the call site pointer we received. */
4031 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4037 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4038 int node, unsigned long caller)
4040 struct kmem_cache *s;
4043 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4044 ret = kmalloc_large_node(size, gfpflags, node);
4046 trace_kmalloc_node(caller, ret,
4047 size, PAGE_SIZE << get_order(size),
4053 s = kmalloc_slab(size, gfpflags);
4055 if (unlikely(ZERO_OR_NULL_PTR(s)))
4058 ret = slab_alloc_node(s, gfpflags, node, caller);
4060 /* Honor the call site pointer we received. */
4061 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4068 static int count_inuse(struct page *page)
4073 static int count_total(struct page *page)
4075 return page->objects;
4079 #ifdef CONFIG_SLUB_DEBUG
4080 static int validate_slab(struct kmem_cache *s, struct page *page,
4084 void *addr = page_address(page);
4086 if (!check_slab(s, page) ||
4087 !on_freelist(s, page, NULL))
4090 /* Now we know that a valid freelist exists */
4091 bitmap_zero(map, page->objects);
4093 get_map(s, page, map);
4094 for_each_object(p, s, addr, page->objects) {
4095 if (test_bit(slab_index(p, s, addr), map))
4096 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4100 for_each_object(p, s, addr, page->objects)
4101 if (!test_bit(slab_index(p, s, addr), map))
4102 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4107 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4111 validate_slab(s, page, map);
4115 static int validate_slab_node(struct kmem_cache *s,
4116 struct kmem_cache_node *n, unsigned long *map)
4118 unsigned long count = 0;
4120 unsigned long flags;
4122 spin_lock_irqsave(&n->list_lock, flags);
4124 list_for_each_entry(page, &n->partial, lru) {
4125 validate_slab_slab(s, page, map);
4128 if (count != n->nr_partial)
4129 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4130 s->name, count, n->nr_partial);
4132 if (!(s->flags & SLAB_STORE_USER))
4135 list_for_each_entry(page, &n->full, lru) {
4136 validate_slab_slab(s, page, map);
4139 if (count != atomic_long_read(&n->nr_slabs))
4140 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4141 s->name, count, atomic_long_read(&n->nr_slabs));
4144 spin_unlock_irqrestore(&n->list_lock, flags);
4148 static long validate_slab_cache(struct kmem_cache *s)
4151 unsigned long count = 0;
4152 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4153 sizeof(unsigned long), GFP_KERNEL);
4154 struct kmem_cache_node *n;
4160 for_each_kmem_cache_node(s, node, n)
4161 count += validate_slab_node(s, n, map);
4166 * Generate lists of code addresses where slabcache objects are allocated
4171 unsigned long count;
4178 DECLARE_BITMAP(cpus, NR_CPUS);
4184 unsigned long count;
4185 struct location *loc;
4188 static void free_loc_track(struct loc_track *t)
4191 free_pages((unsigned long)t->loc,
4192 get_order(sizeof(struct location) * t->max));
4195 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4200 order = get_order(sizeof(struct location) * max);
4202 l = (void *)__get_free_pages(flags, order);
4207 memcpy(l, t->loc, sizeof(struct location) * t->count);
4215 static int add_location(struct loc_track *t, struct kmem_cache *s,
4216 const struct track *track)
4218 long start, end, pos;
4220 unsigned long caddr;
4221 unsigned long age = jiffies - track->when;
4227 pos = start + (end - start + 1) / 2;
4230 * There is nothing at "end". If we end up there
4231 * we need to add something to before end.
4236 caddr = t->loc[pos].addr;
4237 if (track->addr == caddr) {
4243 if (age < l->min_time)
4245 if (age > l->max_time)
4248 if (track->pid < l->min_pid)
4249 l->min_pid = track->pid;
4250 if (track->pid > l->max_pid)
4251 l->max_pid = track->pid;
4253 cpumask_set_cpu(track->cpu,
4254 to_cpumask(l->cpus));
4256 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4260 if (track->addr < caddr)
4267 * Not found. Insert new tracking element.
4269 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4275 (t->count - pos) * sizeof(struct location));
4278 l->addr = track->addr;
4282 l->min_pid = track->pid;
4283 l->max_pid = track->pid;
4284 cpumask_clear(to_cpumask(l->cpus));
4285 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4286 nodes_clear(l->nodes);
4287 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4291 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4292 struct page *page, enum track_item alloc,
4295 void *addr = page_address(page);
4298 bitmap_zero(map, page->objects);
4299 get_map(s, page, map);
4301 for_each_object(p, s, addr, page->objects)
4302 if (!test_bit(slab_index(p, s, addr), map))
4303 add_location(t, s, get_track(s, p, alloc));
4306 static int list_locations(struct kmem_cache *s, char *buf,
4307 enum track_item alloc)
4311 struct loc_track t = { 0, 0, NULL };
4313 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4314 sizeof(unsigned long), GFP_KERNEL);
4315 struct kmem_cache_node *n;
4317 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4320 return sprintf(buf, "Out of memory\n");
4322 /* Push back cpu slabs */
4325 for_each_kmem_cache_node(s, node, n) {
4326 unsigned long flags;
4329 if (!atomic_long_read(&n->nr_slabs))
4332 spin_lock_irqsave(&n->list_lock, flags);
4333 list_for_each_entry(page, &n->partial, lru)
4334 process_slab(&t, s, page, alloc, map);
4335 list_for_each_entry(page, &n->full, lru)
4336 process_slab(&t, s, page, alloc, map);
4337 spin_unlock_irqrestore(&n->list_lock, flags);
4340 for (i = 0; i < t.count; i++) {
4341 struct location *l = &t.loc[i];
4343 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4345 len += sprintf(buf + len, "%7ld ", l->count);
4348 len += sprintf(buf + len, "%pS", (void *)l->addr);
4350 len += sprintf(buf + len, "<not-available>");
4352 if (l->sum_time != l->min_time) {
4353 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4355 (long)div_u64(l->sum_time, l->count),
4358 len += sprintf(buf + len, " age=%ld",
4361 if (l->min_pid != l->max_pid)
4362 len += sprintf(buf + len, " pid=%ld-%ld",
4363 l->min_pid, l->max_pid);
4365 len += sprintf(buf + len, " pid=%ld",
4368 if (num_online_cpus() > 1 &&
4369 !cpumask_empty(to_cpumask(l->cpus)) &&
4370 len < PAGE_SIZE - 60)
4371 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4373 cpumask_pr_args(to_cpumask(l->cpus)));
4375 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4376 len < PAGE_SIZE - 60)
4377 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4379 nodemask_pr_args(&l->nodes));
4381 len += sprintf(buf + len, "\n");
4387 len += sprintf(buf, "No data\n");
4392 #ifdef SLUB_RESILIENCY_TEST
4393 static void __init resiliency_test(void)
4397 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4399 pr_err("SLUB resiliency testing\n");
4400 pr_err("-----------------------\n");
4401 pr_err("A. Corruption after allocation\n");
4403 p = kzalloc(16, GFP_KERNEL);
4405 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4408 validate_slab_cache(kmalloc_caches[4]);
4410 /* Hmmm... The next two are dangerous */
4411 p = kzalloc(32, GFP_KERNEL);
4412 p[32 + sizeof(void *)] = 0x34;
4413 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4415 pr_err("If allocated object is overwritten then not detectable\n\n");
4417 validate_slab_cache(kmalloc_caches[5]);
4418 p = kzalloc(64, GFP_KERNEL);
4419 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4421 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4423 pr_err("If allocated object is overwritten then not detectable\n\n");
4424 validate_slab_cache(kmalloc_caches[6]);
4426 pr_err("\nB. Corruption after free\n");
4427 p = kzalloc(128, GFP_KERNEL);
4430 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4431 validate_slab_cache(kmalloc_caches[7]);
4433 p = kzalloc(256, GFP_KERNEL);
4436 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4437 validate_slab_cache(kmalloc_caches[8]);
4439 p = kzalloc(512, GFP_KERNEL);
4442 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4443 validate_slab_cache(kmalloc_caches[9]);
4447 static void resiliency_test(void) {};
4452 enum slab_stat_type {
4453 SL_ALL, /* All slabs */
4454 SL_PARTIAL, /* Only partially allocated slabs */
4455 SL_CPU, /* Only slabs used for cpu caches */
4456 SL_OBJECTS, /* Determine allocated objects not slabs */
4457 SL_TOTAL /* Determine object capacity not slabs */
4460 #define SO_ALL (1 << SL_ALL)
4461 #define SO_PARTIAL (1 << SL_PARTIAL)
4462 #define SO_CPU (1 << SL_CPU)
4463 #define SO_OBJECTS (1 << SL_OBJECTS)
4464 #define SO_TOTAL (1 << SL_TOTAL)
4466 static ssize_t show_slab_objects(struct kmem_cache *s,
4467 char *buf, unsigned long flags)
4469 unsigned long total = 0;
4472 unsigned long *nodes;
4474 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4478 if (flags & SO_CPU) {
4481 for_each_possible_cpu(cpu) {
4482 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4487 page = READ_ONCE(c->page);
4491 node = page_to_nid(page);
4492 if (flags & SO_TOTAL)
4494 else if (flags & SO_OBJECTS)
4502 page = READ_ONCE(c->partial);
4504 node = page_to_nid(page);
4505 if (flags & SO_TOTAL)
4507 else if (flags & SO_OBJECTS)
4518 #ifdef CONFIG_SLUB_DEBUG
4519 if (flags & SO_ALL) {
4520 struct kmem_cache_node *n;
4522 for_each_kmem_cache_node(s, node, n) {
4524 if (flags & SO_TOTAL)
4525 x = atomic_long_read(&n->total_objects);
4526 else if (flags & SO_OBJECTS)
4527 x = atomic_long_read(&n->total_objects) -
4528 count_partial(n, count_free);
4530 x = atomic_long_read(&n->nr_slabs);
4537 if (flags & SO_PARTIAL) {
4538 struct kmem_cache_node *n;
4540 for_each_kmem_cache_node(s, node, n) {
4541 if (flags & SO_TOTAL)
4542 x = count_partial(n, count_total);
4543 else if (flags & SO_OBJECTS)
4544 x = count_partial(n, count_inuse);
4551 x = sprintf(buf, "%lu", total);
4553 for (node = 0; node < nr_node_ids; node++)
4555 x += sprintf(buf + x, " N%d=%lu",
4560 return x + sprintf(buf + x, "\n");
4563 #ifdef CONFIG_SLUB_DEBUG
4564 static int any_slab_objects(struct kmem_cache *s)
4567 struct kmem_cache_node *n;
4569 for_each_kmem_cache_node(s, node, n)
4570 if (atomic_long_read(&n->total_objects))
4577 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4578 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4580 struct slab_attribute {
4581 struct attribute attr;
4582 ssize_t (*show)(struct kmem_cache *s, char *buf);
4583 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4586 #define SLAB_ATTR_RO(_name) \
4587 static struct slab_attribute _name##_attr = \
4588 __ATTR(_name, 0400, _name##_show, NULL)
4590 #define SLAB_ATTR(_name) \
4591 static struct slab_attribute _name##_attr = \
4592 __ATTR(_name, 0600, _name##_show, _name##_store)
4594 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4596 return sprintf(buf, "%d\n", s->size);
4598 SLAB_ATTR_RO(slab_size);
4600 static ssize_t align_show(struct kmem_cache *s, char *buf)
4602 return sprintf(buf, "%d\n", s->align);
4604 SLAB_ATTR_RO(align);
4606 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4608 return sprintf(buf, "%d\n", s->object_size);
4610 SLAB_ATTR_RO(object_size);
4612 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4614 return sprintf(buf, "%d\n", oo_objects(s->oo));
4616 SLAB_ATTR_RO(objs_per_slab);
4618 static ssize_t order_store(struct kmem_cache *s,
4619 const char *buf, size_t length)
4621 unsigned long order;
4624 err = kstrtoul(buf, 10, &order);
4628 if (order > slub_max_order || order < slub_min_order)
4631 calculate_sizes(s, order);
4635 static ssize_t order_show(struct kmem_cache *s, char *buf)
4637 return sprintf(buf, "%d\n", oo_order(s->oo));
4641 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4643 return sprintf(buf, "%lu\n", s->min_partial);
4646 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4652 err = kstrtoul(buf, 10, &min);
4656 set_min_partial(s, min);
4659 SLAB_ATTR(min_partial);
4661 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4663 return sprintf(buf, "%u\n", s->cpu_partial);
4666 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4669 unsigned long objects;
4672 err = kstrtoul(buf, 10, &objects);
4675 if (objects && !kmem_cache_has_cpu_partial(s))
4678 s->cpu_partial = objects;
4682 SLAB_ATTR(cpu_partial);
4684 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4688 return sprintf(buf, "%pS\n", s->ctor);
4692 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4694 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4696 SLAB_ATTR_RO(aliases);
4698 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4700 return show_slab_objects(s, buf, SO_PARTIAL);
4702 SLAB_ATTR_RO(partial);
4704 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4706 return show_slab_objects(s, buf, SO_CPU);
4708 SLAB_ATTR_RO(cpu_slabs);
4710 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4712 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4714 SLAB_ATTR_RO(objects);
4716 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4718 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4720 SLAB_ATTR_RO(objects_partial);
4722 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4729 for_each_online_cpu(cpu) {
4730 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4733 pages += page->pages;
4734 objects += page->pobjects;
4738 len = sprintf(buf, "%d(%d)", objects, pages);
4741 for_each_online_cpu(cpu) {
4742 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4744 if (page && len < PAGE_SIZE - 20)
4745 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4746 page->pobjects, page->pages);
4749 return len + sprintf(buf + len, "\n");
4751 SLAB_ATTR_RO(slabs_cpu_partial);
4753 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4755 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4758 static ssize_t reclaim_account_store(struct kmem_cache *s,
4759 const char *buf, size_t length)
4761 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4763 s->flags |= SLAB_RECLAIM_ACCOUNT;
4766 SLAB_ATTR(reclaim_account);
4768 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4770 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4772 SLAB_ATTR_RO(hwcache_align);
4774 #ifdef CONFIG_ZONE_DMA
4775 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4777 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4779 SLAB_ATTR_RO(cache_dma);
4782 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4784 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4786 SLAB_ATTR_RO(destroy_by_rcu);
4788 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4790 return sprintf(buf, "%d\n", s->reserved);
4792 SLAB_ATTR_RO(reserved);
4794 #ifdef CONFIG_SLUB_DEBUG
4795 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4797 return show_slab_objects(s, buf, SO_ALL);
4799 SLAB_ATTR_RO(slabs);
4801 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4803 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4805 SLAB_ATTR_RO(total_objects);
4807 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4809 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4812 static ssize_t sanity_checks_store(struct kmem_cache *s,
4813 const char *buf, size_t length)
4815 s->flags &= ~SLAB_DEBUG_FREE;
4816 if (buf[0] == '1') {
4817 s->flags &= ~__CMPXCHG_DOUBLE;
4818 s->flags |= SLAB_DEBUG_FREE;
4822 SLAB_ATTR(sanity_checks);
4824 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4826 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4829 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4833 * Tracing a merged cache is going to give confusing results
4834 * as well as cause other issues like converting a mergeable
4835 * cache into an umergeable one.
4837 if (s->refcount > 1)
4840 s->flags &= ~SLAB_TRACE;
4841 if (buf[0] == '1') {
4842 s->flags &= ~__CMPXCHG_DOUBLE;
4843 s->flags |= SLAB_TRACE;
4849 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4851 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4854 static ssize_t red_zone_store(struct kmem_cache *s,
4855 const char *buf, size_t length)
4857 if (any_slab_objects(s))
4860 s->flags &= ~SLAB_RED_ZONE;
4861 if (buf[0] == '1') {
4862 s->flags &= ~__CMPXCHG_DOUBLE;
4863 s->flags |= SLAB_RED_ZONE;
4865 calculate_sizes(s, -1);
4868 SLAB_ATTR(red_zone);
4870 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4872 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4875 static ssize_t poison_store(struct kmem_cache *s,
4876 const char *buf, size_t length)
4878 if (any_slab_objects(s))
4881 s->flags &= ~SLAB_POISON;
4882 if (buf[0] == '1') {
4883 s->flags &= ~__CMPXCHG_DOUBLE;
4884 s->flags |= SLAB_POISON;
4886 calculate_sizes(s, -1);
4891 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4893 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4896 static ssize_t store_user_store(struct kmem_cache *s,
4897 const char *buf, size_t length)
4899 if (any_slab_objects(s))
4902 s->flags &= ~SLAB_STORE_USER;
4903 if (buf[0] == '1') {
4904 s->flags &= ~__CMPXCHG_DOUBLE;
4905 s->flags |= SLAB_STORE_USER;
4907 calculate_sizes(s, -1);
4910 SLAB_ATTR(store_user);
4912 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4917 static ssize_t validate_store(struct kmem_cache *s,
4918 const char *buf, size_t length)
4922 if (buf[0] == '1') {
4923 ret = validate_slab_cache(s);
4929 SLAB_ATTR(validate);
4931 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4933 if (!(s->flags & SLAB_STORE_USER))
4935 return list_locations(s, buf, TRACK_ALLOC);
4937 SLAB_ATTR_RO(alloc_calls);
4939 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4941 if (!(s->flags & SLAB_STORE_USER))
4943 return list_locations(s, buf, TRACK_FREE);
4945 SLAB_ATTR_RO(free_calls);
4946 #endif /* CONFIG_SLUB_DEBUG */
4948 #ifdef CONFIG_FAILSLAB
4949 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4951 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4954 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4957 if (s->refcount > 1)
4960 s->flags &= ~SLAB_FAILSLAB;
4962 s->flags |= SLAB_FAILSLAB;
4965 SLAB_ATTR(failslab);
4968 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4973 static ssize_t shrink_store(struct kmem_cache *s,
4974 const char *buf, size_t length)
4977 kmem_cache_shrink(s);
4985 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4987 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4990 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4991 const char *buf, size_t length)
4993 unsigned long ratio;
4996 err = kstrtoul(buf, 10, &ratio);
5001 s->remote_node_defrag_ratio = ratio * 10;
5005 SLAB_ATTR(remote_node_defrag_ratio);
5008 #ifdef CONFIG_SLUB_STATS
5009 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5011 unsigned long sum = 0;
5014 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5019 for_each_online_cpu(cpu) {
5020 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5026 len = sprintf(buf, "%lu", sum);
5029 for_each_online_cpu(cpu) {
5030 if (data[cpu] && len < PAGE_SIZE - 20)
5031 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5035 return len + sprintf(buf + len, "\n");
5038 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5042 for_each_online_cpu(cpu)
5043 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5046 #define STAT_ATTR(si, text) \
5047 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5049 return show_stat(s, buf, si); \
5051 static ssize_t text##_store(struct kmem_cache *s, \
5052 const char *buf, size_t length) \
5054 if (buf[0] != '0') \
5056 clear_stat(s, si); \
5061 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5062 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5063 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5064 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5065 STAT_ATTR(FREE_FROZEN, free_frozen);
5066 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5067 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5068 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5069 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5070 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5071 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5072 STAT_ATTR(FREE_SLAB, free_slab);
5073 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5074 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5075 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5076 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5077 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5078 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5079 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5080 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5081 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5082 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5083 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5084 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5085 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5086 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5089 static struct attribute *slab_attrs[] = {
5090 &slab_size_attr.attr,
5091 &object_size_attr.attr,
5092 &objs_per_slab_attr.attr,
5094 &min_partial_attr.attr,
5095 &cpu_partial_attr.attr,
5097 &objects_partial_attr.attr,
5099 &cpu_slabs_attr.attr,
5103 &hwcache_align_attr.attr,
5104 &reclaim_account_attr.attr,
5105 &destroy_by_rcu_attr.attr,
5107 &reserved_attr.attr,
5108 &slabs_cpu_partial_attr.attr,
5109 #ifdef CONFIG_SLUB_DEBUG
5110 &total_objects_attr.attr,
5112 &sanity_checks_attr.attr,
5114 &red_zone_attr.attr,
5116 &store_user_attr.attr,
5117 &validate_attr.attr,
5118 &alloc_calls_attr.attr,
5119 &free_calls_attr.attr,
5121 #ifdef CONFIG_ZONE_DMA
5122 &cache_dma_attr.attr,
5125 &remote_node_defrag_ratio_attr.attr,
5127 #ifdef CONFIG_SLUB_STATS
5128 &alloc_fastpath_attr.attr,
5129 &alloc_slowpath_attr.attr,
5130 &free_fastpath_attr.attr,
5131 &free_slowpath_attr.attr,
5132 &free_frozen_attr.attr,
5133 &free_add_partial_attr.attr,
5134 &free_remove_partial_attr.attr,
5135 &alloc_from_partial_attr.attr,
5136 &alloc_slab_attr.attr,
5137 &alloc_refill_attr.attr,
5138 &alloc_node_mismatch_attr.attr,
5139 &free_slab_attr.attr,
5140 &cpuslab_flush_attr.attr,
5141 &deactivate_full_attr.attr,
5142 &deactivate_empty_attr.attr,
5143 &deactivate_to_head_attr.attr,
5144 &deactivate_to_tail_attr.attr,
5145 &deactivate_remote_frees_attr.attr,
5146 &deactivate_bypass_attr.attr,
5147 &order_fallback_attr.attr,
5148 &cmpxchg_double_fail_attr.attr,
5149 &cmpxchg_double_cpu_fail_attr.attr,
5150 &cpu_partial_alloc_attr.attr,
5151 &cpu_partial_free_attr.attr,
5152 &cpu_partial_node_attr.attr,
5153 &cpu_partial_drain_attr.attr,
5155 #ifdef CONFIG_FAILSLAB
5156 &failslab_attr.attr,
5162 static struct attribute_group slab_attr_group = {
5163 .attrs = slab_attrs,
5166 static ssize_t slab_attr_show(struct kobject *kobj,
5167 struct attribute *attr,
5170 struct slab_attribute *attribute;
5171 struct kmem_cache *s;
5174 attribute = to_slab_attr(attr);
5177 if (!attribute->show)
5180 err = attribute->show(s, buf);
5185 static ssize_t slab_attr_store(struct kobject *kobj,
5186 struct attribute *attr,
5187 const char *buf, size_t len)
5189 struct slab_attribute *attribute;
5190 struct kmem_cache *s;
5193 attribute = to_slab_attr(attr);
5196 if (!attribute->store)
5199 err = attribute->store(s, buf, len);
5200 #ifdef CONFIG_MEMCG_KMEM
5201 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5202 struct kmem_cache *c;
5204 mutex_lock(&slab_mutex);
5205 if (s->max_attr_size < len)
5206 s->max_attr_size = len;
5209 * This is a best effort propagation, so this function's return
5210 * value will be determined by the parent cache only. This is
5211 * basically because not all attributes will have a well
5212 * defined semantics for rollbacks - most of the actions will
5213 * have permanent effects.
5215 * Returning the error value of any of the children that fail
5216 * is not 100 % defined, in the sense that users seeing the
5217 * error code won't be able to know anything about the state of
5220 * Only returning the error code for the parent cache at least
5221 * has well defined semantics. The cache being written to
5222 * directly either failed or succeeded, in which case we loop
5223 * through the descendants with best-effort propagation.
5225 for_each_memcg_cache(c, s)
5226 attribute->store(c, buf, len);
5227 mutex_unlock(&slab_mutex);
5233 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5235 #ifdef CONFIG_MEMCG_KMEM
5237 char *buffer = NULL;
5238 struct kmem_cache *root_cache;
5240 if (is_root_cache(s))
5243 root_cache = s->memcg_params.root_cache;
5246 * This mean this cache had no attribute written. Therefore, no point
5247 * in copying default values around
5249 if (!root_cache->max_attr_size)
5252 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5255 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5257 if (!attr || !attr->store || !attr->show)
5261 * It is really bad that we have to allocate here, so we will
5262 * do it only as a fallback. If we actually allocate, though,
5263 * we can just use the allocated buffer until the end.
5265 * Most of the slub attributes will tend to be very small in
5266 * size, but sysfs allows buffers up to a page, so they can
5267 * theoretically happen.
5271 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5274 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5275 if (WARN_ON(!buffer))
5280 attr->show(root_cache, buf);
5281 attr->store(s, buf, strlen(buf));
5285 free_page((unsigned long)buffer);
5289 static void kmem_cache_release(struct kobject *k)
5291 slab_kmem_cache_release(to_slab(k));
5294 static const struct sysfs_ops slab_sysfs_ops = {
5295 .show = slab_attr_show,
5296 .store = slab_attr_store,
5299 static struct kobj_type slab_ktype = {
5300 .sysfs_ops = &slab_sysfs_ops,
5301 .release = kmem_cache_release,
5304 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5306 struct kobj_type *ktype = get_ktype(kobj);
5308 if (ktype == &slab_ktype)
5313 static const struct kset_uevent_ops slab_uevent_ops = {
5314 .filter = uevent_filter,
5317 static struct kset *slab_kset;
5319 static inline struct kset *cache_kset(struct kmem_cache *s)
5321 #ifdef CONFIG_MEMCG_KMEM
5322 if (!is_root_cache(s))
5323 return s->memcg_params.root_cache->memcg_kset;
5328 #define ID_STR_LENGTH 64
5330 /* Create a unique string id for a slab cache:
5332 * Format :[flags-]size
5334 static char *create_unique_id(struct kmem_cache *s)
5336 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5343 * First flags affecting slabcache operations. We will only
5344 * get here for aliasable slabs so we do not need to support
5345 * too many flags. The flags here must cover all flags that
5346 * are matched during merging to guarantee that the id is
5349 if (s->flags & SLAB_CACHE_DMA)
5351 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5353 if (s->flags & SLAB_DEBUG_FREE)
5355 if (!(s->flags & SLAB_NOTRACK))
5359 p += sprintf(p, "%07d", s->size);
5361 BUG_ON(p > name + ID_STR_LENGTH - 1);
5365 static int sysfs_slab_add(struct kmem_cache *s)
5369 int unmergeable = slab_unmergeable(s);
5373 * Slabcache can never be merged so we can use the name proper.
5374 * This is typically the case for debug situations. In that
5375 * case we can catch duplicate names easily.
5377 sysfs_remove_link(&slab_kset->kobj, s->name);
5381 * Create a unique name for the slab as a target
5384 name = create_unique_id(s);
5387 s->kobj.kset = cache_kset(s);
5388 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5392 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5396 #ifdef CONFIG_MEMCG_KMEM
5397 if (is_root_cache(s)) {
5398 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5399 if (!s->memcg_kset) {
5406 kobject_uevent(&s->kobj, KOBJ_ADD);
5408 /* Setup first alias */
5409 sysfs_slab_alias(s, s->name);
5416 kobject_del(&s->kobj);
5420 void sysfs_slab_remove(struct kmem_cache *s)
5422 if (slab_state < FULL)
5424 * Sysfs has not been setup yet so no need to remove the
5429 #ifdef CONFIG_MEMCG_KMEM
5430 kset_unregister(s->memcg_kset);
5432 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5433 kobject_del(&s->kobj);
5434 kobject_put(&s->kobj);
5438 * Need to buffer aliases during bootup until sysfs becomes
5439 * available lest we lose that information.
5441 struct saved_alias {
5442 struct kmem_cache *s;
5444 struct saved_alias *next;
5447 static struct saved_alias *alias_list;
5449 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5451 struct saved_alias *al;
5453 if (slab_state == FULL) {
5455 * If we have a leftover link then remove it.
5457 sysfs_remove_link(&slab_kset->kobj, name);
5458 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5461 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5467 al->next = alias_list;
5472 static int __init slab_sysfs_init(void)
5474 struct kmem_cache *s;
5477 mutex_lock(&slab_mutex);
5479 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5481 mutex_unlock(&slab_mutex);
5482 pr_err("Cannot register slab subsystem.\n");
5488 list_for_each_entry(s, &slab_caches, list) {
5489 err = sysfs_slab_add(s);
5491 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5495 while (alias_list) {
5496 struct saved_alias *al = alias_list;
5498 alias_list = alias_list->next;
5499 err = sysfs_slab_alias(al->s, al->name);
5501 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5506 mutex_unlock(&slab_mutex);
5511 __initcall(slab_sysfs_init);
5512 #endif /* CONFIG_SYSFS */
5515 * The /proc/slabinfo ABI
5517 #ifdef CONFIG_SLABINFO
5518 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5520 unsigned long nr_slabs = 0;
5521 unsigned long nr_objs = 0;
5522 unsigned long nr_free = 0;
5524 struct kmem_cache_node *n;
5526 for_each_kmem_cache_node(s, node, n) {
5527 nr_slabs += node_nr_slabs(n);
5528 nr_objs += node_nr_objs(n);
5529 nr_free += count_partial(n, count_free);
5532 sinfo->active_objs = nr_objs - nr_free;
5533 sinfo->num_objs = nr_objs;
5534 sinfo->active_slabs = nr_slabs;
5535 sinfo->num_slabs = nr_slabs;
5536 sinfo->objects_per_slab = oo_objects(s->oo);
5537 sinfo->cache_order = oo_order(s->oo);
5540 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5544 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5545 size_t count, loff_t *ppos)
5549 #endif /* CONFIG_SLABINFO */