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 __printf(3, 4) 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, gfp_t flags,
1296 size_t size, void **p)
1300 flags &= gfp_allowed_mask;
1301 for (i = 0; i < size; i++) {
1302 void *object = p[i];
1304 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1305 kmemleak_alloc_recursive(object, s->object_size, 1,
1307 kasan_slab_alloc(s, object);
1309 memcg_kmem_put_cache(s);
1312 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1314 kmemleak_free_recursive(x, s->flags);
1317 * Trouble is that we may no longer disable interrupts in the fast path
1318 * So in order to make the debug calls that expect irqs to be
1319 * disabled we need to disable interrupts temporarily.
1321 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1323 unsigned long flags;
1325 local_irq_save(flags);
1326 kmemcheck_slab_free(s, x, s->object_size);
1327 debug_check_no_locks_freed(x, s->object_size);
1328 local_irq_restore(flags);
1331 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1332 debug_check_no_obj_freed(x, s->object_size);
1334 kasan_slab_free(s, x);
1337 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1338 void *head, void *tail)
1341 * Compiler cannot detect this function can be removed if slab_free_hook()
1342 * evaluates to nothing. Thus, catch all relevant config debug options here.
1344 #if defined(CONFIG_KMEMCHECK) || \
1345 defined(CONFIG_LOCKDEP) || \
1346 defined(CONFIG_DEBUG_KMEMLEAK) || \
1347 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1348 defined(CONFIG_KASAN)
1350 void *object = head;
1351 void *tail_obj = tail ? : head;
1354 slab_free_hook(s, object);
1355 } while ((object != tail_obj) &&
1356 (object = get_freepointer(s, object)));
1360 static void setup_object(struct kmem_cache *s, struct page *page,
1363 setup_object_debug(s, page, object);
1364 if (unlikely(s->ctor)) {
1365 kasan_unpoison_object_data(s, object);
1367 kasan_poison_object_data(s, object);
1372 * Slab allocation and freeing
1374 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1375 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1378 int order = oo_order(oo);
1380 flags |= __GFP_NOTRACK;
1382 if (node == NUMA_NO_NODE)
1383 page = alloc_pages(flags, order);
1385 page = __alloc_pages_node(node, flags, order);
1387 if (page && memcg_charge_slab(page, flags, order, s)) {
1388 __free_pages(page, order);
1395 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1398 struct kmem_cache_order_objects oo = s->oo;
1403 flags &= gfp_allowed_mask;
1405 if (gfpflags_allow_blocking(flags))
1408 flags |= s->allocflags;
1411 * Let the initial higher-order allocation fail under memory pressure
1412 * so we fall-back to the minimum order allocation.
1414 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1415 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1416 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1418 page = alloc_slab_page(s, alloc_gfp, node, oo);
1419 if (unlikely(!page)) {
1423 * Allocation may have failed due to fragmentation.
1424 * Try a lower order alloc if possible
1426 page = alloc_slab_page(s, alloc_gfp, node, oo);
1427 if (unlikely(!page))
1429 stat(s, ORDER_FALLBACK);
1432 if (kmemcheck_enabled &&
1433 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1434 int pages = 1 << oo_order(oo);
1436 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1439 * Objects from caches that have a constructor don't get
1440 * cleared when they're allocated, so we need to do it here.
1443 kmemcheck_mark_uninitialized_pages(page, pages);
1445 kmemcheck_mark_unallocated_pages(page, pages);
1448 page->objects = oo_objects(oo);
1450 order = compound_order(page);
1451 page->slab_cache = s;
1452 __SetPageSlab(page);
1453 if (page_is_pfmemalloc(page))
1454 SetPageSlabPfmemalloc(page);
1456 start = page_address(page);
1458 if (unlikely(s->flags & SLAB_POISON))
1459 memset(start, POISON_INUSE, PAGE_SIZE << order);
1461 kasan_poison_slab(page);
1463 for_each_object_idx(p, idx, s, start, page->objects) {
1464 setup_object(s, page, p);
1465 if (likely(idx < page->objects))
1466 set_freepointer(s, p, p + s->size);
1468 set_freepointer(s, p, NULL);
1471 page->freelist = start;
1472 page->inuse = page->objects;
1476 if (gfpflags_allow_blocking(flags))
1477 local_irq_disable();
1481 mod_zone_page_state(page_zone(page),
1482 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1483 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1486 inc_slabs_node(s, page_to_nid(page), page->objects);
1491 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1493 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1494 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1498 return allocate_slab(s,
1499 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1502 static void __free_slab(struct kmem_cache *s, struct page *page)
1504 int order = compound_order(page);
1505 int pages = 1 << order;
1507 if (kmem_cache_debug(s)) {
1510 slab_pad_check(s, page);
1511 for_each_object(p, s, page_address(page),
1513 check_object(s, page, p, SLUB_RED_INACTIVE);
1516 kmemcheck_free_shadow(page, compound_order(page));
1518 mod_zone_page_state(page_zone(page),
1519 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1520 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1523 __ClearPageSlabPfmemalloc(page);
1524 __ClearPageSlab(page);
1526 page_mapcount_reset(page);
1527 if (current->reclaim_state)
1528 current->reclaim_state->reclaimed_slab += pages;
1529 __free_kmem_pages(page, order);
1532 #define need_reserve_slab_rcu \
1533 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1535 static void rcu_free_slab(struct rcu_head *h)
1539 if (need_reserve_slab_rcu)
1540 page = virt_to_head_page(h);
1542 page = container_of((struct list_head *)h, struct page, lru);
1544 __free_slab(page->slab_cache, page);
1547 static void free_slab(struct kmem_cache *s, struct page *page)
1549 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1550 struct rcu_head *head;
1552 if (need_reserve_slab_rcu) {
1553 int order = compound_order(page);
1554 int offset = (PAGE_SIZE << order) - s->reserved;
1556 VM_BUG_ON(s->reserved != sizeof(*head));
1557 head = page_address(page) + offset;
1559 head = &page->rcu_head;
1562 call_rcu(head, rcu_free_slab);
1564 __free_slab(s, page);
1567 static void discard_slab(struct kmem_cache *s, struct page *page)
1569 dec_slabs_node(s, page_to_nid(page), page->objects);
1574 * Management of partially allocated slabs.
1577 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1580 if (tail == DEACTIVATE_TO_TAIL)
1581 list_add_tail(&page->lru, &n->partial);
1583 list_add(&page->lru, &n->partial);
1586 static inline void add_partial(struct kmem_cache_node *n,
1587 struct page *page, int tail)
1589 lockdep_assert_held(&n->list_lock);
1590 __add_partial(n, page, tail);
1594 __remove_partial(struct kmem_cache_node *n, struct page *page)
1596 list_del(&page->lru);
1600 static inline void remove_partial(struct kmem_cache_node *n,
1603 lockdep_assert_held(&n->list_lock);
1604 __remove_partial(n, page);
1608 * Remove slab from the partial list, freeze it and
1609 * return the pointer to the freelist.
1611 * Returns a list of objects or NULL if it fails.
1613 static inline void *acquire_slab(struct kmem_cache *s,
1614 struct kmem_cache_node *n, struct page *page,
1615 int mode, int *objects)
1618 unsigned long counters;
1621 lockdep_assert_held(&n->list_lock);
1624 * Zap the freelist and set the frozen bit.
1625 * The old freelist is the list of objects for the
1626 * per cpu allocation list.
1628 freelist = page->freelist;
1629 counters = page->counters;
1630 new.counters = counters;
1631 *objects = new.objects - new.inuse;
1633 new.inuse = page->objects;
1634 new.freelist = NULL;
1636 new.freelist = freelist;
1639 VM_BUG_ON(new.frozen);
1642 if (!__cmpxchg_double_slab(s, page,
1644 new.freelist, new.counters,
1648 remove_partial(n, page);
1653 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1654 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1657 * Try to allocate a partial slab from a specific node.
1659 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1660 struct kmem_cache_cpu *c, gfp_t flags)
1662 struct page *page, *page2;
1663 void *object = NULL;
1664 unsigned int available = 0;
1668 * Racy check. If we mistakenly see no partial slabs then we
1669 * just allocate an empty slab. If we mistakenly try to get a
1670 * partial slab and there is none available then get_partials()
1673 if (!n || !n->nr_partial)
1676 spin_lock(&n->list_lock);
1677 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1680 if (!pfmemalloc_match(page, flags))
1683 t = acquire_slab(s, n, page, object == NULL, &objects);
1687 available += objects;
1690 stat(s, ALLOC_FROM_PARTIAL);
1693 put_cpu_partial(s, page, 0);
1694 stat(s, CPU_PARTIAL_NODE);
1696 if (!kmem_cache_has_cpu_partial(s)
1697 || available > s->cpu_partial / 2)
1701 spin_unlock(&n->list_lock);
1706 * Get a page from somewhere. Search in increasing NUMA distances.
1708 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1709 struct kmem_cache_cpu *c)
1712 struct zonelist *zonelist;
1715 enum zone_type high_zoneidx = gfp_zone(flags);
1717 unsigned int cpuset_mems_cookie;
1720 * The defrag ratio allows a configuration of the tradeoffs between
1721 * inter node defragmentation and node local allocations. A lower
1722 * defrag_ratio increases the tendency to do local allocations
1723 * instead of attempting to obtain partial slabs from other nodes.
1725 * If the defrag_ratio is set to 0 then kmalloc() always
1726 * returns node local objects. If the ratio is higher then kmalloc()
1727 * may return off node objects because partial slabs are obtained
1728 * from other nodes and filled up.
1730 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1731 * defrag_ratio = 1000) then every (well almost) allocation will
1732 * first attempt to defrag slab caches on other nodes. This means
1733 * scanning over all nodes to look for partial slabs which may be
1734 * expensive if we do it every time we are trying to find a slab
1735 * with available objects.
1737 if (!s->remote_node_defrag_ratio ||
1738 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1742 cpuset_mems_cookie = read_mems_allowed_begin();
1743 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1744 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1745 struct kmem_cache_node *n;
1747 n = get_node(s, zone_to_nid(zone));
1749 if (n && cpuset_zone_allowed(zone, flags) &&
1750 n->nr_partial > s->min_partial) {
1751 object = get_partial_node(s, n, c, flags);
1754 * Don't check read_mems_allowed_retry()
1755 * here - if mems_allowed was updated in
1756 * parallel, that was a harmless race
1757 * between allocation and the cpuset
1764 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1770 * Get a partial page, lock it and return it.
1772 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1773 struct kmem_cache_cpu *c)
1776 int searchnode = node;
1778 if (node == NUMA_NO_NODE)
1779 searchnode = numa_mem_id();
1781 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1782 if (object || node != NUMA_NO_NODE)
1785 return get_any_partial(s, flags, c);
1788 #ifdef CONFIG_PREEMPT
1790 * Calculate the next globally unique transaction for disambiguiation
1791 * during cmpxchg. The transactions start with the cpu number and are then
1792 * incremented by CONFIG_NR_CPUS.
1794 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1797 * No preemption supported therefore also no need to check for
1803 static inline unsigned long next_tid(unsigned long tid)
1805 return tid + TID_STEP;
1808 static inline unsigned int tid_to_cpu(unsigned long tid)
1810 return tid % TID_STEP;
1813 static inline unsigned long tid_to_event(unsigned long tid)
1815 return tid / TID_STEP;
1818 static inline unsigned int init_tid(int cpu)
1823 static inline void note_cmpxchg_failure(const char *n,
1824 const struct kmem_cache *s, unsigned long tid)
1826 #ifdef SLUB_DEBUG_CMPXCHG
1827 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1829 pr_info("%s %s: cmpxchg redo ", n, s->name);
1831 #ifdef CONFIG_PREEMPT
1832 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1833 pr_warn("due to cpu change %d -> %d\n",
1834 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1837 if (tid_to_event(tid) != tid_to_event(actual_tid))
1838 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1839 tid_to_event(tid), tid_to_event(actual_tid));
1841 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1842 actual_tid, tid, next_tid(tid));
1844 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1847 static void init_kmem_cache_cpus(struct kmem_cache *s)
1851 for_each_possible_cpu(cpu)
1852 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1856 * Remove the cpu slab
1858 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1861 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1862 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1864 enum slab_modes l = M_NONE, m = M_NONE;
1866 int tail = DEACTIVATE_TO_HEAD;
1870 if (page->freelist) {
1871 stat(s, DEACTIVATE_REMOTE_FREES);
1872 tail = DEACTIVATE_TO_TAIL;
1876 * Stage one: Free all available per cpu objects back
1877 * to the page freelist while it is still frozen. Leave the
1880 * There is no need to take the list->lock because the page
1883 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1885 unsigned long counters;
1888 prior = page->freelist;
1889 counters = page->counters;
1890 set_freepointer(s, freelist, prior);
1891 new.counters = counters;
1893 VM_BUG_ON(!new.frozen);
1895 } while (!__cmpxchg_double_slab(s, page,
1897 freelist, new.counters,
1898 "drain percpu freelist"));
1900 freelist = nextfree;
1904 * Stage two: Ensure that the page is unfrozen while the
1905 * list presence reflects the actual number of objects
1908 * We setup the list membership and then perform a cmpxchg
1909 * with the count. If there is a mismatch then the page
1910 * is not unfrozen but the page is on the wrong list.
1912 * Then we restart the process which may have to remove
1913 * the page from the list that we just put it on again
1914 * because the number of objects in the slab may have
1919 old.freelist = page->freelist;
1920 old.counters = page->counters;
1921 VM_BUG_ON(!old.frozen);
1923 /* Determine target state of the slab */
1924 new.counters = old.counters;
1927 set_freepointer(s, freelist, old.freelist);
1928 new.freelist = freelist;
1930 new.freelist = old.freelist;
1934 if (!new.inuse && n->nr_partial >= s->min_partial)
1936 else if (new.freelist) {
1941 * Taking the spinlock removes the possiblity
1942 * that acquire_slab() will see a slab page that
1945 spin_lock(&n->list_lock);
1949 if (kmem_cache_debug(s) && !lock) {
1952 * This also ensures that the scanning of full
1953 * slabs from diagnostic functions will not see
1956 spin_lock(&n->list_lock);
1964 remove_partial(n, page);
1966 else if (l == M_FULL)
1968 remove_full(s, n, page);
1970 if (m == M_PARTIAL) {
1972 add_partial(n, page, tail);
1975 } else if (m == M_FULL) {
1977 stat(s, DEACTIVATE_FULL);
1978 add_full(s, n, page);
1984 if (!__cmpxchg_double_slab(s, page,
1985 old.freelist, old.counters,
1986 new.freelist, new.counters,
1991 spin_unlock(&n->list_lock);
1994 stat(s, DEACTIVATE_EMPTY);
1995 discard_slab(s, page);
2001 * Unfreeze all the cpu partial slabs.
2003 * This function must be called with interrupts disabled
2004 * for the cpu using c (or some other guarantee must be there
2005 * to guarantee no concurrent accesses).
2007 static void unfreeze_partials(struct kmem_cache *s,
2008 struct kmem_cache_cpu *c)
2010 #ifdef CONFIG_SLUB_CPU_PARTIAL
2011 struct kmem_cache_node *n = NULL, *n2 = NULL;
2012 struct page *page, *discard_page = NULL;
2014 while ((page = c->partial)) {
2018 c->partial = page->next;
2020 n2 = get_node(s, page_to_nid(page));
2023 spin_unlock(&n->list_lock);
2026 spin_lock(&n->list_lock);
2031 old.freelist = page->freelist;
2032 old.counters = page->counters;
2033 VM_BUG_ON(!old.frozen);
2035 new.counters = old.counters;
2036 new.freelist = old.freelist;
2040 } while (!__cmpxchg_double_slab(s, page,
2041 old.freelist, old.counters,
2042 new.freelist, new.counters,
2043 "unfreezing slab"));
2045 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2046 page->next = discard_page;
2047 discard_page = page;
2049 add_partial(n, page, DEACTIVATE_TO_TAIL);
2050 stat(s, FREE_ADD_PARTIAL);
2055 spin_unlock(&n->list_lock);
2057 while (discard_page) {
2058 page = discard_page;
2059 discard_page = discard_page->next;
2061 stat(s, DEACTIVATE_EMPTY);
2062 discard_slab(s, page);
2069 * Put a page that was just frozen (in __slab_free) into a partial page
2070 * slot if available. This is done without interrupts disabled and without
2071 * preemption disabled. The cmpxchg is racy and may put the partial page
2072 * onto a random cpus partial slot.
2074 * If we did not find a slot then simply move all the partials to the
2075 * per node partial list.
2077 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2079 #ifdef CONFIG_SLUB_CPU_PARTIAL
2080 struct page *oldpage;
2088 oldpage = this_cpu_read(s->cpu_slab->partial);
2091 pobjects = oldpage->pobjects;
2092 pages = oldpage->pages;
2093 if (drain && pobjects > s->cpu_partial) {
2094 unsigned long flags;
2096 * partial array is full. Move the existing
2097 * set to the per node partial list.
2099 local_irq_save(flags);
2100 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2101 local_irq_restore(flags);
2105 stat(s, CPU_PARTIAL_DRAIN);
2110 pobjects += page->objects - page->inuse;
2112 page->pages = pages;
2113 page->pobjects = pobjects;
2114 page->next = oldpage;
2116 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2118 if (unlikely(!s->cpu_partial)) {
2119 unsigned long flags;
2121 local_irq_save(flags);
2122 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2123 local_irq_restore(flags);
2129 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2131 stat(s, CPUSLAB_FLUSH);
2132 deactivate_slab(s, c->page, c->freelist);
2134 c->tid = next_tid(c->tid);
2142 * Called from IPI handler with interrupts disabled.
2144 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2146 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2152 unfreeze_partials(s, c);
2156 static void flush_cpu_slab(void *d)
2158 struct kmem_cache *s = d;
2160 __flush_cpu_slab(s, smp_processor_id());
2163 static bool has_cpu_slab(int cpu, void *info)
2165 struct kmem_cache *s = info;
2166 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2168 return c->page || c->partial;
2171 static void flush_all(struct kmem_cache *s)
2173 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2177 * Check if the objects in a per cpu structure fit numa
2178 * locality expectations.
2180 static inline int node_match(struct page *page, int node)
2183 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2189 #ifdef CONFIG_SLUB_DEBUG
2190 static int count_free(struct page *page)
2192 return page->objects - page->inuse;
2195 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2197 return atomic_long_read(&n->total_objects);
2199 #endif /* CONFIG_SLUB_DEBUG */
2201 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2202 static unsigned long count_partial(struct kmem_cache_node *n,
2203 int (*get_count)(struct page *))
2205 unsigned long flags;
2206 unsigned long x = 0;
2209 spin_lock_irqsave(&n->list_lock, flags);
2210 list_for_each_entry(page, &n->partial, lru)
2211 x += get_count(page);
2212 spin_unlock_irqrestore(&n->list_lock, flags);
2215 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2217 static noinline void
2218 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2220 #ifdef CONFIG_SLUB_DEBUG
2221 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2222 DEFAULT_RATELIMIT_BURST);
2224 struct kmem_cache_node *n;
2226 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2229 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2231 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2232 s->name, s->object_size, s->size, oo_order(s->oo),
2235 if (oo_order(s->min) > get_order(s->object_size))
2236 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2239 for_each_kmem_cache_node(s, node, n) {
2240 unsigned long nr_slabs;
2241 unsigned long nr_objs;
2242 unsigned long nr_free;
2244 nr_free = count_partial(n, count_free);
2245 nr_slabs = node_nr_slabs(n);
2246 nr_objs = node_nr_objs(n);
2248 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2249 node, nr_slabs, nr_objs, nr_free);
2254 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2255 int node, struct kmem_cache_cpu **pc)
2258 struct kmem_cache_cpu *c = *pc;
2261 freelist = get_partial(s, flags, node, c);
2266 page = new_slab(s, flags, node);
2268 c = raw_cpu_ptr(s->cpu_slab);
2273 * No other reference to the page yet so we can
2274 * muck around with it freely without cmpxchg
2276 freelist = page->freelist;
2277 page->freelist = NULL;
2279 stat(s, ALLOC_SLAB);
2288 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2290 if (unlikely(PageSlabPfmemalloc(page)))
2291 return gfp_pfmemalloc_allowed(gfpflags);
2297 * Check the page->freelist of a page and either transfer the freelist to the
2298 * per cpu freelist or deactivate the page.
2300 * The page is still frozen if the return value is not NULL.
2302 * If this function returns NULL then the page has been unfrozen.
2304 * This function must be called with interrupt disabled.
2306 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2309 unsigned long counters;
2313 freelist = page->freelist;
2314 counters = page->counters;
2316 new.counters = counters;
2317 VM_BUG_ON(!new.frozen);
2319 new.inuse = page->objects;
2320 new.frozen = freelist != NULL;
2322 } while (!__cmpxchg_double_slab(s, page,
2331 * Slow path. The lockless freelist is empty or we need to perform
2334 * Processing is still very fast if new objects have been freed to the
2335 * regular freelist. In that case we simply take over the regular freelist
2336 * as the lockless freelist and zap the regular freelist.
2338 * If that is not working then we fall back to the partial lists. We take the
2339 * first element of the freelist as the object to allocate now and move the
2340 * rest of the freelist to the lockless freelist.
2342 * And if we were unable to get a new slab from the partial slab lists then
2343 * we need to allocate a new slab. This is the slowest path since it involves
2344 * a call to the page allocator and the setup of a new slab.
2346 * Version of __slab_alloc to use when we know that interrupts are
2347 * already disabled (which is the case for bulk allocation).
2349 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2350 unsigned long addr, struct kmem_cache_cpu *c)
2358 * if the node is not online or has no normal memory, just
2359 * ignore the node constraint
2361 if (unlikely(node != NUMA_NO_NODE &&
2362 !node_state(node, N_NORMAL_MEMORY)))
2363 node = NUMA_NO_NODE;
2368 if (unlikely(!node_match(page, node))) {
2370 * same as above but node_match() being false already
2371 * implies node != NUMA_NO_NODE
2373 if (!node_state(node, N_NORMAL_MEMORY)) {
2374 node = NUMA_NO_NODE;
2377 stat(s, ALLOC_NODE_MISMATCH);
2378 deactivate_slab(s, page, c->freelist);
2386 * By rights, we should be searching for a slab page that was
2387 * PFMEMALLOC but right now, we are losing the pfmemalloc
2388 * information when the page leaves the per-cpu allocator
2390 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2391 deactivate_slab(s, page, c->freelist);
2397 /* must check again c->freelist in case of cpu migration or IRQ */
2398 freelist = c->freelist;
2402 freelist = get_freelist(s, page);
2406 stat(s, DEACTIVATE_BYPASS);
2410 stat(s, ALLOC_REFILL);
2414 * freelist is pointing to the list of objects to be used.
2415 * page is pointing to the page from which the objects are obtained.
2416 * That page must be frozen for per cpu allocations to work.
2418 VM_BUG_ON(!c->page->frozen);
2419 c->freelist = get_freepointer(s, freelist);
2420 c->tid = next_tid(c->tid);
2426 page = c->page = c->partial;
2427 c->partial = page->next;
2428 stat(s, CPU_PARTIAL_ALLOC);
2433 freelist = new_slab_objects(s, gfpflags, node, &c);
2435 if (unlikely(!freelist)) {
2436 slab_out_of_memory(s, gfpflags, node);
2441 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2444 /* Only entered in the debug case */
2445 if (kmem_cache_debug(s) &&
2446 !alloc_debug_processing(s, page, freelist, addr))
2447 goto new_slab; /* Slab failed checks. Next slab needed */
2449 deactivate_slab(s, page, get_freepointer(s, freelist));
2456 * Another one that disabled interrupt and compensates for possible
2457 * cpu changes by refetching the per cpu area pointer.
2459 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2460 unsigned long addr, struct kmem_cache_cpu *c)
2463 unsigned long flags;
2465 local_irq_save(flags);
2466 #ifdef CONFIG_PREEMPT
2468 * We may have been preempted and rescheduled on a different
2469 * cpu before disabling interrupts. Need to reload cpu area
2472 c = this_cpu_ptr(s->cpu_slab);
2475 p = ___slab_alloc(s, gfpflags, node, addr, c);
2476 local_irq_restore(flags);
2481 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2482 * have the fastpath folded into their functions. So no function call
2483 * overhead for requests that can be satisfied on the fastpath.
2485 * The fastpath works by first checking if the lockless freelist can be used.
2486 * If not then __slab_alloc is called for slow processing.
2488 * Otherwise we can simply pick the next object from the lockless free list.
2490 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2491 gfp_t gfpflags, int node, unsigned long addr)
2494 struct kmem_cache_cpu *c;
2498 s = slab_pre_alloc_hook(s, gfpflags);
2503 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2504 * enabled. We may switch back and forth between cpus while
2505 * reading from one cpu area. That does not matter as long
2506 * as we end up on the original cpu again when doing the cmpxchg.
2508 * We should guarantee that tid and kmem_cache are retrieved on
2509 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2510 * to check if it is matched or not.
2513 tid = this_cpu_read(s->cpu_slab->tid);
2514 c = raw_cpu_ptr(s->cpu_slab);
2515 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2516 unlikely(tid != READ_ONCE(c->tid)));
2519 * Irqless object alloc/free algorithm used here depends on sequence
2520 * of fetching cpu_slab's data. tid should be fetched before anything
2521 * on c to guarantee that object and page associated with previous tid
2522 * won't be used with current tid. If we fetch tid first, object and
2523 * page could be one associated with next tid and our alloc/free
2524 * request will be failed. In this case, we will retry. So, no problem.
2529 * The transaction ids are globally unique per cpu and per operation on
2530 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2531 * occurs on the right processor and that there was no operation on the
2532 * linked list in between.
2535 object = c->freelist;
2537 if (unlikely(!object || !node_match(page, node))) {
2538 object = __slab_alloc(s, gfpflags, node, addr, c);
2539 stat(s, ALLOC_SLOWPATH);
2541 void *next_object = get_freepointer_safe(s, object);
2544 * The cmpxchg will only match if there was no additional
2545 * operation and if we are on the right processor.
2547 * The cmpxchg does the following atomically (without lock
2549 * 1. Relocate first pointer to the current per cpu area.
2550 * 2. Verify that tid and freelist have not been changed
2551 * 3. If they were not changed replace tid and freelist
2553 * Since this is without lock semantics the protection is only
2554 * against code executing on this cpu *not* from access by
2557 if (unlikely(!this_cpu_cmpxchg_double(
2558 s->cpu_slab->freelist, s->cpu_slab->tid,
2560 next_object, next_tid(tid)))) {
2562 note_cmpxchg_failure("slab_alloc", s, tid);
2565 prefetch_freepointer(s, next_object);
2566 stat(s, ALLOC_FASTPATH);
2569 if (unlikely(gfpflags & __GFP_ZERO) && object)
2570 memset(object, 0, s->object_size);
2572 slab_post_alloc_hook(s, gfpflags, 1, &object);
2577 static __always_inline void *slab_alloc(struct kmem_cache *s,
2578 gfp_t gfpflags, unsigned long addr)
2580 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2583 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2585 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2587 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2592 EXPORT_SYMBOL(kmem_cache_alloc);
2594 #ifdef CONFIG_TRACING
2595 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2597 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2598 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2599 kasan_kmalloc(s, ret, size);
2602 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2606 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2608 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2610 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2611 s->object_size, s->size, gfpflags, node);
2615 EXPORT_SYMBOL(kmem_cache_alloc_node);
2617 #ifdef CONFIG_TRACING
2618 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2620 int node, size_t size)
2622 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2624 trace_kmalloc_node(_RET_IP_, ret,
2625 size, s->size, gfpflags, node);
2627 kasan_kmalloc(s, ret, size);
2630 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2635 * Slow path handling. This may still be called frequently since objects
2636 * have a longer lifetime than the cpu slabs in most processing loads.
2638 * So we still attempt to reduce cache line usage. Just take the slab
2639 * lock and free the item. If there is no additional partial page
2640 * handling required then we can return immediately.
2642 static void __slab_free(struct kmem_cache *s, struct page *page,
2643 void *head, void *tail, int cnt,
2650 unsigned long counters;
2651 struct kmem_cache_node *n = NULL;
2652 unsigned long uninitialized_var(flags);
2654 stat(s, FREE_SLOWPATH);
2656 if (kmem_cache_debug(s) &&
2657 !(n = free_debug_processing(s, page, head, tail, cnt,
2663 spin_unlock_irqrestore(&n->list_lock, flags);
2666 prior = page->freelist;
2667 counters = page->counters;
2668 set_freepointer(s, tail, prior);
2669 new.counters = counters;
2670 was_frozen = new.frozen;
2672 if ((!new.inuse || !prior) && !was_frozen) {
2674 if (kmem_cache_has_cpu_partial(s) && !prior) {
2677 * Slab was on no list before and will be
2679 * We can defer the list move and instead
2684 } else { /* Needs to be taken off a list */
2686 n = get_node(s, page_to_nid(page));
2688 * Speculatively acquire the list_lock.
2689 * If the cmpxchg does not succeed then we may
2690 * drop the list_lock without any processing.
2692 * Otherwise the list_lock will synchronize with
2693 * other processors updating the list of slabs.
2695 spin_lock_irqsave(&n->list_lock, flags);
2700 } while (!cmpxchg_double_slab(s, page,
2708 * If we just froze the page then put it onto the
2709 * per cpu partial list.
2711 if (new.frozen && !was_frozen) {
2712 put_cpu_partial(s, page, 1);
2713 stat(s, CPU_PARTIAL_FREE);
2716 * The list lock was not taken therefore no list
2717 * activity can be necessary.
2720 stat(s, FREE_FROZEN);
2724 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2728 * Objects left in the slab. If it was not on the partial list before
2731 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2732 if (kmem_cache_debug(s))
2733 remove_full(s, n, page);
2734 add_partial(n, page, DEACTIVATE_TO_TAIL);
2735 stat(s, FREE_ADD_PARTIAL);
2737 spin_unlock_irqrestore(&n->list_lock, flags);
2743 * Slab on the partial list.
2745 remove_partial(n, page);
2746 stat(s, FREE_REMOVE_PARTIAL);
2748 /* Slab must be on the full list */
2749 remove_full(s, n, page);
2752 spin_unlock_irqrestore(&n->list_lock, flags);
2754 discard_slab(s, page);
2758 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2759 * can perform fastpath freeing without additional function calls.
2761 * The fastpath is only possible if we are freeing to the current cpu slab
2762 * of this processor. This typically the case if we have just allocated
2765 * If fastpath is not possible then fall back to __slab_free where we deal
2766 * with all sorts of special processing.
2768 * Bulk free of a freelist with several objects (all pointing to the
2769 * same page) possible by specifying head and tail ptr, plus objects
2770 * count (cnt). Bulk free indicated by tail pointer being set.
2772 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2773 void *head, void *tail, int cnt,
2776 void *tail_obj = tail ? : head;
2777 struct kmem_cache_cpu *c;
2780 slab_free_freelist_hook(s, head, tail);
2784 * Determine the currently cpus per cpu slab.
2785 * The cpu may change afterward. However that does not matter since
2786 * data is retrieved via this pointer. If we are on the same cpu
2787 * during the cmpxchg then the free will succeed.
2790 tid = this_cpu_read(s->cpu_slab->tid);
2791 c = raw_cpu_ptr(s->cpu_slab);
2792 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2793 unlikely(tid != READ_ONCE(c->tid)));
2795 /* Same with comment on barrier() in slab_alloc_node() */
2798 if (likely(page == c->page)) {
2799 void **freelist = READ_ONCE(c->freelist);
2801 set_freepointer(s, tail_obj, freelist);
2803 if (unlikely(!this_cpu_cmpxchg_double(
2804 s->cpu_slab->freelist, s->cpu_slab->tid,
2806 head, next_tid(tid)))) {
2808 note_cmpxchg_failure("slab_free", s, tid);
2811 stat(s, FREE_FASTPATH);
2813 __slab_free(s, page, head, tail_obj, cnt, addr);
2817 void kmem_cache_free(struct kmem_cache *s, void *x)
2819 s = cache_from_obj(s, x);
2822 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2823 trace_kmem_cache_free(_RET_IP_, x);
2825 EXPORT_SYMBOL(kmem_cache_free);
2827 struct detached_freelist {
2832 struct kmem_cache *s;
2836 * This function progressively scans the array with free objects (with
2837 * a limited look ahead) and extract objects belonging to the same
2838 * page. It builds a detached freelist directly within the given
2839 * page/objects. This can happen without any need for
2840 * synchronization, because the objects are owned by running process.
2841 * The freelist is build up as a single linked list in the objects.
2842 * The idea is, that this detached freelist can then be bulk
2843 * transferred to the real freelist(s), but only requiring a single
2844 * synchronization primitive. Look ahead in the array is limited due
2845 * to performance reasons.
2848 int build_detached_freelist(struct kmem_cache *s, size_t size,
2849 void **p, struct detached_freelist *df)
2851 size_t first_skipped_index = 0;
2855 /* Always re-init detached_freelist */
2860 } while (!object && size);
2865 /* Support for memcg, compiler can optimize this out */
2866 df->s = cache_from_obj(s, object);
2868 /* Start new detached freelist */
2869 set_freepointer(df->s, object, NULL);
2870 df->page = virt_to_head_page(object);
2872 df->freelist = object;
2873 p[size] = NULL; /* mark object processed */
2879 continue; /* Skip processed objects */
2881 /* df->page is always set at this point */
2882 if (df->page == virt_to_head_page(object)) {
2883 /* Opportunity build freelist */
2884 set_freepointer(df->s, object, df->freelist);
2885 df->freelist = object;
2887 p[size] = NULL; /* mark object processed */
2892 /* Limit look ahead search */
2896 if (!first_skipped_index)
2897 first_skipped_index = size + 1;
2900 return first_skipped_index;
2903 /* Note that interrupts must be enabled when calling this function. */
2904 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2910 struct detached_freelist df;
2912 size = build_detached_freelist(s, size, p, &df);
2913 if (unlikely(!df.page))
2916 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2917 } while (likely(size));
2919 EXPORT_SYMBOL(kmem_cache_free_bulk);
2921 /* Note that interrupts must be enabled when calling this function. */
2922 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2925 struct kmem_cache_cpu *c;
2928 /* memcg and kmem_cache debug support */
2929 s = slab_pre_alloc_hook(s, flags);
2933 * Drain objects in the per cpu slab, while disabling local
2934 * IRQs, which protects against PREEMPT and interrupts
2935 * handlers invoking normal fastpath.
2937 local_irq_disable();
2938 c = this_cpu_ptr(s->cpu_slab);
2940 for (i = 0; i < size; i++) {
2941 void *object = c->freelist;
2943 if (unlikely(!object)) {
2945 * We may have removed an object from c->freelist using
2946 * the fastpath in the previous iteration; in that case,
2947 * c->tid has not been bumped yet.
2948 * Since ___slab_alloc() may reenable interrupts while
2949 * allocating memory, we should bump c->tid now.
2951 c->tid = next_tid(c->tid);
2954 * Invoking slow path likely have side-effect
2955 * of re-populating per CPU c->freelist
2957 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2959 if (unlikely(!p[i]))
2962 c = this_cpu_ptr(s->cpu_slab);
2963 continue; /* goto for-loop */
2965 c->freelist = get_freepointer(s, object);
2968 c->tid = next_tid(c->tid);
2971 /* Clear memory outside IRQ disabled fastpath loop */
2972 if (unlikely(flags & __GFP_ZERO)) {
2975 for (j = 0; j < i; j++)
2976 memset(p[j], 0, s->object_size);
2979 /* memcg and kmem_cache debug support */
2980 slab_post_alloc_hook(s, flags, size, p);
2984 slab_post_alloc_hook(s, flags, i, p);
2985 __kmem_cache_free_bulk(s, i, p);
2988 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2992 * Object placement in a slab is made very easy because we always start at
2993 * offset 0. If we tune the size of the object to the alignment then we can
2994 * get the required alignment by putting one properly sized object after
2997 * Notice that the allocation order determines the sizes of the per cpu
2998 * caches. Each processor has always one slab available for allocations.
2999 * Increasing the allocation order reduces the number of times that slabs
3000 * must be moved on and off the partial lists and is therefore a factor in
3005 * Mininum / Maximum order of slab pages. This influences locking overhead
3006 * and slab fragmentation. A higher order reduces the number of partial slabs
3007 * and increases the number of allocations possible without having to
3008 * take the list_lock.
3010 static int slub_min_order;
3011 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3012 static int slub_min_objects;
3015 * Calculate the order of allocation given an slab object size.
3017 * The order of allocation has significant impact on performance and other
3018 * system components. Generally order 0 allocations should be preferred since
3019 * order 0 does not cause fragmentation in the page allocator. Larger objects
3020 * be problematic to put into order 0 slabs because there may be too much
3021 * unused space left. We go to a higher order if more than 1/16th of the slab
3024 * In order to reach satisfactory performance we must ensure that a minimum
3025 * number of objects is in one slab. Otherwise we may generate too much
3026 * activity on the partial lists which requires taking the list_lock. This is
3027 * less a concern for large slabs though which are rarely used.
3029 * slub_max_order specifies the order where we begin to stop considering the
3030 * number of objects in a slab as critical. If we reach slub_max_order then
3031 * we try to keep the page order as low as possible. So we accept more waste
3032 * of space in favor of a small page order.
3034 * Higher order allocations also allow the placement of more objects in a
3035 * slab and thereby reduce object handling overhead. If the user has
3036 * requested a higher mininum order then we start with that one instead of
3037 * the smallest order which will fit the object.
3039 static inline int slab_order(int size, int min_objects,
3040 int max_order, int fract_leftover, int reserved)
3044 int min_order = slub_min_order;
3046 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3047 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3049 for (order = max(min_order, get_order(min_objects * size + reserved));
3050 order <= max_order; order++) {
3052 unsigned long slab_size = PAGE_SIZE << order;
3054 rem = (slab_size - reserved) % size;
3056 if (rem <= slab_size / fract_leftover)
3063 static inline int calculate_order(int size, int reserved)
3071 * Attempt to find best configuration for a slab. This
3072 * works by first attempting to generate a layout with
3073 * the best configuration and backing off gradually.
3075 * First we increase the acceptable waste in a slab. Then
3076 * we reduce the minimum objects required in a slab.
3078 min_objects = slub_min_objects;
3080 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3081 max_objects = order_objects(slub_max_order, size, reserved);
3082 min_objects = min(min_objects, max_objects);
3084 while (min_objects > 1) {
3086 while (fraction >= 4) {
3087 order = slab_order(size, min_objects,
3088 slub_max_order, fraction, reserved);
3089 if (order <= slub_max_order)
3097 * We were unable to place multiple objects in a slab. Now
3098 * lets see if we can place a single object there.
3100 order = slab_order(size, 1, slub_max_order, 1, reserved);
3101 if (order <= slub_max_order)
3105 * Doh this slab cannot be placed using slub_max_order.
3107 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3108 if (order < MAX_ORDER)
3114 init_kmem_cache_node(struct kmem_cache_node *n)
3117 spin_lock_init(&n->list_lock);
3118 INIT_LIST_HEAD(&n->partial);
3119 #ifdef CONFIG_SLUB_DEBUG
3120 atomic_long_set(&n->nr_slabs, 0);
3121 atomic_long_set(&n->total_objects, 0);
3122 INIT_LIST_HEAD(&n->full);
3126 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3128 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3129 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3132 * Must align to double word boundary for the double cmpxchg
3133 * instructions to work; see __pcpu_double_call_return_bool().
3135 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3136 2 * sizeof(void *));
3141 init_kmem_cache_cpus(s);
3146 static struct kmem_cache *kmem_cache_node;
3149 * No kmalloc_node yet so do it by hand. We know that this is the first
3150 * slab on the node for this slabcache. There are no concurrent accesses
3153 * Note that this function only works on the kmem_cache_node
3154 * when allocating for the kmem_cache_node. This is used for bootstrapping
3155 * memory on a fresh node that has no slab structures yet.
3157 static void early_kmem_cache_node_alloc(int node)
3160 struct kmem_cache_node *n;
3162 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3164 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3167 if (page_to_nid(page) != node) {
3168 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3169 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3174 page->freelist = get_freepointer(kmem_cache_node, n);
3177 kmem_cache_node->node[node] = n;
3178 #ifdef CONFIG_SLUB_DEBUG
3179 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3180 init_tracking(kmem_cache_node, n);
3182 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3183 init_kmem_cache_node(n);
3184 inc_slabs_node(kmem_cache_node, node, page->objects);
3187 * No locks need to be taken here as it has just been
3188 * initialized and there is no concurrent access.
3190 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3193 static void free_kmem_cache_nodes(struct kmem_cache *s)
3196 struct kmem_cache_node *n;
3198 for_each_kmem_cache_node(s, node, n) {
3199 kmem_cache_free(kmem_cache_node, n);
3200 s->node[node] = NULL;
3204 static int init_kmem_cache_nodes(struct kmem_cache *s)
3208 for_each_node_state(node, N_NORMAL_MEMORY) {
3209 struct kmem_cache_node *n;
3211 if (slab_state == DOWN) {
3212 early_kmem_cache_node_alloc(node);
3215 n = kmem_cache_alloc_node(kmem_cache_node,
3219 free_kmem_cache_nodes(s);
3224 init_kmem_cache_node(n);
3229 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3231 if (min < MIN_PARTIAL)
3233 else if (min > MAX_PARTIAL)
3235 s->min_partial = min;
3239 * calculate_sizes() determines the order and the distribution of data within
3242 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3244 unsigned long flags = s->flags;
3245 unsigned long size = s->object_size;
3249 * Round up object size to the next word boundary. We can only
3250 * place the free pointer at word boundaries and this determines
3251 * the possible location of the free pointer.
3253 size = ALIGN(size, sizeof(void *));
3255 #ifdef CONFIG_SLUB_DEBUG
3257 * Determine if we can poison the object itself. If the user of
3258 * the slab may touch the object after free or before allocation
3259 * then we should never poison the object itself.
3261 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3263 s->flags |= __OBJECT_POISON;
3265 s->flags &= ~__OBJECT_POISON;
3269 * If we are Redzoning then check if there is some space between the
3270 * end of the object and the free pointer. If not then add an
3271 * additional word to have some bytes to store Redzone information.
3273 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3274 size += sizeof(void *);
3278 * With that we have determined the number of bytes in actual use
3279 * by the object. This is the potential offset to the free pointer.
3283 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3286 * Relocate free pointer after the object if it is not
3287 * permitted to overwrite the first word of the object on
3290 * This is the case if we do RCU, have a constructor or
3291 * destructor or are poisoning the objects.
3294 size += sizeof(void *);
3297 #ifdef CONFIG_SLUB_DEBUG
3298 if (flags & SLAB_STORE_USER)
3300 * Need to store information about allocs and frees after
3303 size += 2 * sizeof(struct track);
3305 if (flags & SLAB_RED_ZONE)
3307 * Add some empty padding so that we can catch
3308 * overwrites from earlier objects rather than let
3309 * tracking information or the free pointer be
3310 * corrupted if a user writes before the start
3313 size += sizeof(void *);
3317 * SLUB stores one object immediately after another beginning from
3318 * offset 0. In order to align the objects we have to simply size
3319 * each object to conform to the alignment.
3321 size = ALIGN(size, s->align);
3323 if (forced_order >= 0)
3324 order = forced_order;
3326 order = calculate_order(size, s->reserved);
3333 s->allocflags |= __GFP_COMP;
3335 if (s->flags & SLAB_CACHE_DMA)
3336 s->allocflags |= GFP_DMA;
3338 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3339 s->allocflags |= __GFP_RECLAIMABLE;
3342 * Determine the number of objects per slab
3344 s->oo = oo_make(order, size, s->reserved);
3345 s->min = oo_make(get_order(size), size, s->reserved);
3346 if (oo_objects(s->oo) > oo_objects(s->max))
3349 return !!oo_objects(s->oo);
3352 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3354 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3357 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3358 s->reserved = sizeof(struct rcu_head);
3360 if (!calculate_sizes(s, -1))
3362 if (disable_higher_order_debug) {
3364 * Disable debugging flags that store metadata if the min slab
3367 if (get_order(s->size) > get_order(s->object_size)) {
3368 s->flags &= ~DEBUG_METADATA_FLAGS;
3370 if (!calculate_sizes(s, -1))
3375 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3377 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3378 /* Enable fast mode */
3379 s->flags |= __CMPXCHG_DOUBLE;
3383 * The larger the object size is, the more pages we want on the partial
3384 * list to avoid pounding the page allocator excessively.
3386 set_min_partial(s, ilog2(s->size) / 2);
3389 * cpu_partial determined the maximum number of objects kept in the
3390 * per cpu partial lists of a processor.
3392 * Per cpu partial lists mainly contain slabs that just have one
3393 * object freed. If they are used for allocation then they can be
3394 * filled up again with minimal effort. The slab will never hit the
3395 * per node partial lists and therefore no locking will be required.
3397 * This setting also determines
3399 * A) The number of objects from per cpu partial slabs dumped to the
3400 * per node list when we reach the limit.
3401 * B) The number of objects in cpu partial slabs to extract from the
3402 * per node list when we run out of per cpu objects. We only fetch
3403 * 50% to keep some capacity around for frees.
3405 if (!kmem_cache_has_cpu_partial(s))
3407 else if (s->size >= PAGE_SIZE)
3409 else if (s->size >= 1024)
3411 else if (s->size >= 256)
3412 s->cpu_partial = 13;
3414 s->cpu_partial = 30;
3417 s->remote_node_defrag_ratio = 1000;
3419 if (!init_kmem_cache_nodes(s))
3422 if (alloc_kmem_cache_cpus(s))
3425 free_kmem_cache_nodes(s);
3427 if (flags & SLAB_PANIC)
3428 panic("Cannot create slab %s size=%lu realsize=%u "
3429 "order=%u offset=%u flags=%lx\n",
3430 s->name, (unsigned long)s->size, s->size,
3431 oo_order(s->oo), s->offset, flags);
3435 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3438 #ifdef CONFIG_SLUB_DEBUG
3439 void *addr = page_address(page);
3441 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3442 sizeof(long), GFP_ATOMIC);
3445 slab_err(s, page, text, s->name);
3448 get_map(s, page, map);
3449 for_each_object(p, s, addr, page->objects) {
3451 if (!test_bit(slab_index(p, s, addr), map)) {
3452 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3453 print_tracking(s, p);
3462 * Attempt to free all partial slabs on a node.
3463 * This is called from kmem_cache_close(). We must be the last thread
3464 * using the cache and therefore we do not need to lock anymore.
3466 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3468 struct page *page, *h;
3470 list_for_each_entry_safe(page, h, &n->partial, lru) {
3472 __remove_partial(n, page);
3473 discard_slab(s, page);
3475 list_slab_objects(s, page,
3476 "Objects remaining in %s on kmem_cache_close()");
3482 * Release all resources used by a slab cache.
3484 static inline int kmem_cache_close(struct kmem_cache *s)
3487 struct kmem_cache_node *n;
3490 /* Attempt to free all objects */
3491 for_each_kmem_cache_node(s, node, n) {
3493 if (n->nr_partial || slabs_node(s, node))
3496 free_percpu(s->cpu_slab);
3497 free_kmem_cache_nodes(s);
3501 int __kmem_cache_shutdown(struct kmem_cache *s)
3503 return kmem_cache_close(s);
3506 /********************************************************************
3508 *******************************************************************/
3510 static int __init setup_slub_min_order(char *str)
3512 get_option(&str, &slub_min_order);
3517 __setup("slub_min_order=", setup_slub_min_order);
3519 static int __init setup_slub_max_order(char *str)
3521 get_option(&str, &slub_max_order);
3522 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3527 __setup("slub_max_order=", setup_slub_max_order);
3529 static int __init setup_slub_min_objects(char *str)
3531 get_option(&str, &slub_min_objects);
3536 __setup("slub_min_objects=", setup_slub_min_objects);
3538 void *__kmalloc(size_t size, gfp_t flags)
3540 struct kmem_cache *s;
3543 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3544 return kmalloc_large(size, flags);
3546 s = kmalloc_slab(size, flags);
3548 if (unlikely(ZERO_OR_NULL_PTR(s)))
3551 ret = slab_alloc(s, flags, _RET_IP_);
3553 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3555 kasan_kmalloc(s, ret, size);
3559 EXPORT_SYMBOL(__kmalloc);
3562 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3567 flags |= __GFP_COMP | __GFP_NOTRACK;
3568 page = alloc_kmem_pages_node(node, flags, get_order(size));
3570 ptr = page_address(page);
3572 kmalloc_large_node_hook(ptr, size, flags);
3576 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3578 struct kmem_cache *s;
3581 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3582 ret = kmalloc_large_node(size, flags, node);
3584 trace_kmalloc_node(_RET_IP_, ret,
3585 size, PAGE_SIZE << get_order(size),
3591 s = kmalloc_slab(size, flags);
3593 if (unlikely(ZERO_OR_NULL_PTR(s)))
3596 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3598 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3600 kasan_kmalloc(s, ret, size);
3604 EXPORT_SYMBOL(__kmalloc_node);
3607 static size_t __ksize(const void *object)
3611 if (unlikely(object == ZERO_SIZE_PTR))
3614 page = virt_to_head_page(object);
3616 if (unlikely(!PageSlab(page))) {
3617 WARN_ON(!PageCompound(page));
3618 return PAGE_SIZE << compound_order(page);
3621 return slab_ksize(page->slab_cache);
3624 size_t ksize(const void *object)
3626 size_t size = __ksize(object);
3627 /* We assume that ksize callers could use whole allocated area,
3628 so we need unpoison this area. */
3629 kasan_krealloc(object, size);
3632 EXPORT_SYMBOL(ksize);
3634 void kfree(const void *x)
3637 void *object = (void *)x;
3639 trace_kfree(_RET_IP_, x);
3641 if (unlikely(ZERO_OR_NULL_PTR(x)))
3644 page = virt_to_head_page(x);
3645 if (unlikely(!PageSlab(page))) {
3646 BUG_ON(!PageCompound(page));
3648 __free_kmem_pages(page, compound_order(page));
3651 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3653 EXPORT_SYMBOL(kfree);
3655 #define SHRINK_PROMOTE_MAX 32
3658 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3659 * up most to the head of the partial lists. New allocations will then
3660 * fill those up and thus they can be removed from the partial lists.
3662 * The slabs with the least items are placed last. This results in them
3663 * being allocated from last increasing the chance that the last objects
3664 * are freed in them.
3666 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3670 struct kmem_cache_node *n;
3673 struct list_head discard;
3674 struct list_head promote[SHRINK_PROMOTE_MAX];
3675 unsigned long flags;
3680 * Disable empty slabs caching. Used to avoid pinning offline
3681 * memory cgroups by kmem pages that can be freed.
3687 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3688 * so we have to make sure the change is visible.
3690 kick_all_cpus_sync();
3694 for_each_kmem_cache_node(s, node, n) {
3695 INIT_LIST_HEAD(&discard);
3696 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3697 INIT_LIST_HEAD(promote + i);
3699 spin_lock_irqsave(&n->list_lock, flags);
3702 * Build lists of slabs to discard or promote.
3704 * Note that concurrent frees may occur while we hold the
3705 * list_lock. page->inuse here is the upper limit.
3707 list_for_each_entry_safe(page, t, &n->partial, lru) {
3708 int free = page->objects - page->inuse;
3710 /* Do not reread page->inuse */
3713 /* We do not keep full slabs on the list */
3716 if (free == page->objects) {
3717 list_move(&page->lru, &discard);
3719 } else if (free <= SHRINK_PROMOTE_MAX)
3720 list_move(&page->lru, promote + free - 1);
3724 * Promote the slabs filled up most to the head of the
3727 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3728 list_splice(promote + i, &n->partial);
3730 spin_unlock_irqrestore(&n->list_lock, flags);
3732 /* Release empty slabs */
3733 list_for_each_entry_safe(page, t, &discard, lru)
3734 discard_slab(s, page);
3736 if (slabs_node(s, node))
3743 static int slab_mem_going_offline_callback(void *arg)
3745 struct kmem_cache *s;
3747 mutex_lock(&slab_mutex);
3748 list_for_each_entry(s, &slab_caches, list)
3749 __kmem_cache_shrink(s, false);
3750 mutex_unlock(&slab_mutex);
3755 static void slab_mem_offline_callback(void *arg)
3757 struct kmem_cache_node *n;
3758 struct kmem_cache *s;
3759 struct memory_notify *marg = arg;
3762 offline_node = marg->status_change_nid_normal;
3765 * If the node still has available memory. we need kmem_cache_node
3768 if (offline_node < 0)
3771 mutex_lock(&slab_mutex);
3772 list_for_each_entry(s, &slab_caches, list) {
3773 n = get_node(s, offline_node);
3776 * if n->nr_slabs > 0, slabs still exist on the node
3777 * that is going down. We were unable to free them,
3778 * and offline_pages() function shouldn't call this
3779 * callback. So, we must fail.
3781 BUG_ON(slabs_node(s, offline_node));
3783 s->node[offline_node] = NULL;
3784 kmem_cache_free(kmem_cache_node, n);
3787 mutex_unlock(&slab_mutex);
3790 static int slab_mem_going_online_callback(void *arg)
3792 struct kmem_cache_node *n;
3793 struct kmem_cache *s;
3794 struct memory_notify *marg = arg;
3795 int nid = marg->status_change_nid_normal;
3799 * If the node's memory is already available, then kmem_cache_node is
3800 * already created. Nothing to do.
3806 * We are bringing a node online. No memory is available yet. We must
3807 * allocate a kmem_cache_node structure in order to bring the node
3810 mutex_lock(&slab_mutex);
3811 list_for_each_entry(s, &slab_caches, list) {
3813 * XXX: kmem_cache_alloc_node will fallback to other nodes
3814 * since memory is not yet available from the node that
3817 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3822 init_kmem_cache_node(n);
3826 mutex_unlock(&slab_mutex);
3830 static int slab_memory_callback(struct notifier_block *self,
3831 unsigned long action, void *arg)
3836 case MEM_GOING_ONLINE:
3837 ret = slab_mem_going_online_callback(arg);
3839 case MEM_GOING_OFFLINE:
3840 ret = slab_mem_going_offline_callback(arg);
3843 case MEM_CANCEL_ONLINE:
3844 slab_mem_offline_callback(arg);
3847 case MEM_CANCEL_OFFLINE:
3851 ret = notifier_from_errno(ret);
3857 static struct notifier_block slab_memory_callback_nb = {
3858 .notifier_call = slab_memory_callback,
3859 .priority = SLAB_CALLBACK_PRI,
3862 /********************************************************************
3863 * Basic setup of slabs
3864 *******************************************************************/
3867 * Used for early kmem_cache structures that were allocated using
3868 * the page allocator. Allocate them properly then fix up the pointers
3869 * that may be pointing to the wrong kmem_cache structure.
3872 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3875 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3876 struct kmem_cache_node *n;
3878 memcpy(s, static_cache, kmem_cache->object_size);
3881 * This runs very early, and only the boot processor is supposed to be
3882 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3885 __flush_cpu_slab(s, smp_processor_id());
3886 for_each_kmem_cache_node(s, node, n) {
3889 list_for_each_entry(p, &n->partial, lru)
3892 #ifdef CONFIG_SLUB_DEBUG
3893 list_for_each_entry(p, &n->full, lru)
3897 slab_init_memcg_params(s);
3898 list_add(&s->list, &slab_caches);
3902 void __init kmem_cache_init(void)
3904 static __initdata struct kmem_cache boot_kmem_cache,
3905 boot_kmem_cache_node;
3907 if (debug_guardpage_minorder())
3910 kmem_cache_node = &boot_kmem_cache_node;
3911 kmem_cache = &boot_kmem_cache;
3913 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3914 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3916 register_hotmemory_notifier(&slab_memory_callback_nb);
3918 /* Able to allocate the per node structures */
3919 slab_state = PARTIAL;
3921 create_boot_cache(kmem_cache, "kmem_cache",
3922 offsetof(struct kmem_cache, node) +
3923 nr_node_ids * sizeof(struct kmem_cache_node *),
3924 SLAB_HWCACHE_ALIGN);
3926 kmem_cache = bootstrap(&boot_kmem_cache);
3929 * Allocate kmem_cache_node properly from the kmem_cache slab.
3930 * kmem_cache_node is separately allocated so no need to
3931 * update any list pointers.
3933 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3935 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3936 setup_kmalloc_cache_index_table();
3937 create_kmalloc_caches(0);
3940 register_cpu_notifier(&slab_notifier);
3943 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3945 slub_min_order, slub_max_order, slub_min_objects,
3946 nr_cpu_ids, nr_node_ids);
3949 void __init kmem_cache_init_late(void)
3954 __kmem_cache_alias(const char *name, size_t size, size_t align,
3955 unsigned long flags, void (*ctor)(void *))
3957 struct kmem_cache *s, *c;
3959 s = find_mergeable(size, align, flags, name, ctor);
3964 * Adjust the object sizes so that we clear
3965 * the complete object on kzalloc.
3967 s->object_size = max(s->object_size, (int)size);
3968 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3970 for_each_memcg_cache(c, s) {
3971 c->object_size = s->object_size;
3972 c->inuse = max_t(int, c->inuse,
3973 ALIGN(size, sizeof(void *)));
3976 if (sysfs_slab_alias(s, name)) {
3985 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3989 err = kmem_cache_open(s, flags);
3993 /* Mutex is not taken during early boot */
3994 if (slab_state <= UP)
3997 memcg_propagate_slab_attrs(s);
3998 err = sysfs_slab_add(s);
4000 kmem_cache_close(s);
4007 * Use the cpu notifier to insure that the cpu slabs are flushed when
4010 static int slab_cpuup_callback(struct notifier_block *nfb,
4011 unsigned long action, void *hcpu)
4013 long cpu = (long)hcpu;
4014 struct kmem_cache *s;
4015 unsigned long flags;
4018 case CPU_UP_CANCELED:
4019 case CPU_UP_CANCELED_FROZEN:
4021 case CPU_DEAD_FROZEN:
4022 mutex_lock(&slab_mutex);
4023 list_for_each_entry(s, &slab_caches, list) {
4024 local_irq_save(flags);
4025 __flush_cpu_slab(s, cpu);
4026 local_irq_restore(flags);
4028 mutex_unlock(&slab_mutex);
4036 static struct notifier_block slab_notifier = {
4037 .notifier_call = slab_cpuup_callback
4042 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4044 struct kmem_cache *s;
4047 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4048 return kmalloc_large(size, gfpflags);
4050 s = kmalloc_slab(size, gfpflags);
4052 if (unlikely(ZERO_OR_NULL_PTR(s)))
4055 ret = slab_alloc(s, gfpflags, caller);
4057 /* Honor the call site pointer we received. */
4058 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4064 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4065 int node, unsigned long caller)
4067 struct kmem_cache *s;
4070 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4071 ret = kmalloc_large_node(size, gfpflags, node);
4073 trace_kmalloc_node(caller, ret,
4074 size, PAGE_SIZE << get_order(size),
4080 s = kmalloc_slab(size, gfpflags);
4082 if (unlikely(ZERO_OR_NULL_PTR(s)))
4085 ret = slab_alloc_node(s, gfpflags, node, caller);
4087 /* Honor the call site pointer we received. */
4088 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4095 static int count_inuse(struct page *page)
4100 static int count_total(struct page *page)
4102 return page->objects;
4106 #ifdef CONFIG_SLUB_DEBUG
4107 static int validate_slab(struct kmem_cache *s, struct page *page,
4111 void *addr = page_address(page);
4113 if (!check_slab(s, page) ||
4114 !on_freelist(s, page, NULL))
4117 /* Now we know that a valid freelist exists */
4118 bitmap_zero(map, page->objects);
4120 get_map(s, page, map);
4121 for_each_object(p, s, addr, page->objects) {
4122 if (test_bit(slab_index(p, s, addr), map))
4123 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4127 for_each_object(p, s, addr, page->objects)
4128 if (!test_bit(slab_index(p, s, addr), map))
4129 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4134 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4138 validate_slab(s, page, map);
4142 static int validate_slab_node(struct kmem_cache *s,
4143 struct kmem_cache_node *n, unsigned long *map)
4145 unsigned long count = 0;
4147 unsigned long flags;
4149 spin_lock_irqsave(&n->list_lock, flags);
4151 list_for_each_entry(page, &n->partial, lru) {
4152 validate_slab_slab(s, page, map);
4155 if (count != n->nr_partial)
4156 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4157 s->name, count, n->nr_partial);
4159 if (!(s->flags & SLAB_STORE_USER))
4162 list_for_each_entry(page, &n->full, lru) {
4163 validate_slab_slab(s, page, map);
4166 if (count != atomic_long_read(&n->nr_slabs))
4167 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4168 s->name, count, atomic_long_read(&n->nr_slabs));
4171 spin_unlock_irqrestore(&n->list_lock, flags);
4175 static long validate_slab_cache(struct kmem_cache *s)
4178 unsigned long count = 0;
4179 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4180 sizeof(unsigned long), GFP_KERNEL);
4181 struct kmem_cache_node *n;
4187 for_each_kmem_cache_node(s, node, n)
4188 count += validate_slab_node(s, n, map);
4193 * Generate lists of code addresses where slabcache objects are allocated
4198 unsigned long count;
4205 DECLARE_BITMAP(cpus, NR_CPUS);
4211 unsigned long count;
4212 struct location *loc;
4215 static void free_loc_track(struct loc_track *t)
4218 free_pages((unsigned long)t->loc,
4219 get_order(sizeof(struct location) * t->max));
4222 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4227 order = get_order(sizeof(struct location) * max);
4229 l = (void *)__get_free_pages(flags, order);
4234 memcpy(l, t->loc, sizeof(struct location) * t->count);
4242 static int add_location(struct loc_track *t, struct kmem_cache *s,
4243 const struct track *track)
4245 long start, end, pos;
4247 unsigned long caddr;
4248 unsigned long age = jiffies - track->when;
4254 pos = start + (end - start + 1) / 2;
4257 * There is nothing at "end". If we end up there
4258 * we need to add something to before end.
4263 caddr = t->loc[pos].addr;
4264 if (track->addr == caddr) {
4270 if (age < l->min_time)
4272 if (age > l->max_time)
4275 if (track->pid < l->min_pid)
4276 l->min_pid = track->pid;
4277 if (track->pid > l->max_pid)
4278 l->max_pid = track->pid;
4280 cpumask_set_cpu(track->cpu,
4281 to_cpumask(l->cpus));
4283 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4287 if (track->addr < caddr)
4294 * Not found. Insert new tracking element.
4296 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4302 (t->count - pos) * sizeof(struct location));
4305 l->addr = track->addr;
4309 l->min_pid = track->pid;
4310 l->max_pid = track->pid;
4311 cpumask_clear(to_cpumask(l->cpus));
4312 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4313 nodes_clear(l->nodes);
4314 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4318 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4319 struct page *page, enum track_item alloc,
4322 void *addr = page_address(page);
4325 bitmap_zero(map, page->objects);
4326 get_map(s, page, map);
4328 for_each_object(p, s, addr, page->objects)
4329 if (!test_bit(slab_index(p, s, addr), map))
4330 add_location(t, s, get_track(s, p, alloc));
4333 static int list_locations(struct kmem_cache *s, char *buf,
4334 enum track_item alloc)
4338 struct loc_track t = { 0, 0, NULL };
4340 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4341 sizeof(unsigned long), GFP_KERNEL);
4342 struct kmem_cache_node *n;
4344 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4347 return sprintf(buf, "Out of memory\n");
4349 /* Push back cpu slabs */
4352 for_each_kmem_cache_node(s, node, n) {
4353 unsigned long flags;
4356 if (!atomic_long_read(&n->nr_slabs))
4359 spin_lock_irqsave(&n->list_lock, flags);
4360 list_for_each_entry(page, &n->partial, lru)
4361 process_slab(&t, s, page, alloc, map);
4362 list_for_each_entry(page, &n->full, lru)
4363 process_slab(&t, s, page, alloc, map);
4364 spin_unlock_irqrestore(&n->list_lock, flags);
4367 for (i = 0; i < t.count; i++) {
4368 struct location *l = &t.loc[i];
4370 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4372 len += sprintf(buf + len, "%7ld ", l->count);
4375 len += sprintf(buf + len, "%pS", (void *)l->addr);
4377 len += sprintf(buf + len, "<not-available>");
4379 if (l->sum_time != l->min_time) {
4380 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4382 (long)div_u64(l->sum_time, l->count),
4385 len += sprintf(buf + len, " age=%ld",
4388 if (l->min_pid != l->max_pid)
4389 len += sprintf(buf + len, " pid=%ld-%ld",
4390 l->min_pid, l->max_pid);
4392 len += sprintf(buf + len, " pid=%ld",
4395 if (num_online_cpus() > 1 &&
4396 !cpumask_empty(to_cpumask(l->cpus)) &&
4397 len < PAGE_SIZE - 60)
4398 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4400 cpumask_pr_args(to_cpumask(l->cpus)));
4402 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4403 len < PAGE_SIZE - 60)
4404 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4406 nodemask_pr_args(&l->nodes));
4408 len += sprintf(buf + len, "\n");
4414 len += sprintf(buf, "No data\n");
4419 #ifdef SLUB_RESILIENCY_TEST
4420 static void __init resiliency_test(void)
4424 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4426 pr_err("SLUB resiliency testing\n");
4427 pr_err("-----------------------\n");
4428 pr_err("A. Corruption after allocation\n");
4430 p = kzalloc(16, GFP_KERNEL);
4432 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4435 validate_slab_cache(kmalloc_caches[4]);
4437 /* Hmmm... The next two are dangerous */
4438 p = kzalloc(32, GFP_KERNEL);
4439 p[32 + sizeof(void *)] = 0x34;
4440 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4442 pr_err("If allocated object is overwritten then not detectable\n\n");
4444 validate_slab_cache(kmalloc_caches[5]);
4445 p = kzalloc(64, GFP_KERNEL);
4446 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4448 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4450 pr_err("If allocated object is overwritten then not detectable\n\n");
4451 validate_slab_cache(kmalloc_caches[6]);
4453 pr_err("\nB. Corruption after free\n");
4454 p = kzalloc(128, GFP_KERNEL);
4457 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4458 validate_slab_cache(kmalloc_caches[7]);
4460 p = kzalloc(256, GFP_KERNEL);
4463 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4464 validate_slab_cache(kmalloc_caches[8]);
4466 p = kzalloc(512, GFP_KERNEL);
4469 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4470 validate_slab_cache(kmalloc_caches[9]);
4474 static void resiliency_test(void) {};
4479 enum slab_stat_type {
4480 SL_ALL, /* All slabs */
4481 SL_PARTIAL, /* Only partially allocated slabs */
4482 SL_CPU, /* Only slabs used for cpu caches */
4483 SL_OBJECTS, /* Determine allocated objects not slabs */
4484 SL_TOTAL /* Determine object capacity not slabs */
4487 #define SO_ALL (1 << SL_ALL)
4488 #define SO_PARTIAL (1 << SL_PARTIAL)
4489 #define SO_CPU (1 << SL_CPU)
4490 #define SO_OBJECTS (1 << SL_OBJECTS)
4491 #define SO_TOTAL (1 << SL_TOTAL)
4493 static ssize_t show_slab_objects(struct kmem_cache *s,
4494 char *buf, unsigned long flags)
4496 unsigned long total = 0;
4499 unsigned long *nodes;
4501 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4505 if (flags & SO_CPU) {
4508 for_each_possible_cpu(cpu) {
4509 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4514 page = READ_ONCE(c->page);
4518 node = page_to_nid(page);
4519 if (flags & SO_TOTAL)
4521 else if (flags & SO_OBJECTS)
4529 page = READ_ONCE(c->partial);
4531 node = page_to_nid(page);
4532 if (flags & SO_TOTAL)
4534 else if (flags & SO_OBJECTS)
4545 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4546 * already held which will conflict with an existing lock order:
4548 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4550 * We don't really need mem_hotplug_lock (to hold off
4551 * slab_mem_going_offline_callback) here because slab's memory hot
4552 * unplug code doesn't destroy the kmem_cache->node[] data.
4555 #ifdef CONFIG_SLUB_DEBUG
4556 if (flags & SO_ALL) {
4557 struct kmem_cache_node *n;
4559 for_each_kmem_cache_node(s, node, n) {
4561 if (flags & SO_TOTAL)
4562 x = atomic_long_read(&n->total_objects);
4563 else if (flags & SO_OBJECTS)
4564 x = atomic_long_read(&n->total_objects) -
4565 count_partial(n, count_free);
4567 x = atomic_long_read(&n->nr_slabs);
4574 if (flags & SO_PARTIAL) {
4575 struct kmem_cache_node *n;
4577 for_each_kmem_cache_node(s, node, n) {
4578 if (flags & SO_TOTAL)
4579 x = count_partial(n, count_total);
4580 else if (flags & SO_OBJECTS)
4581 x = count_partial(n, count_inuse);
4588 x = sprintf(buf, "%lu", total);
4590 for (node = 0; node < nr_node_ids; node++)
4592 x += sprintf(buf + x, " N%d=%lu",
4596 return x + sprintf(buf + x, "\n");
4599 #ifdef CONFIG_SLUB_DEBUG
4600 static int any_slab_objects(struct kmem_cache *s)
4603 struct kmem_cache_node *n;
4605 for_each_kmem_cache_node(s, node, n)
4606 if (atomic_long_read(&n->total_objects))
4613 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4614 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4616 struct slab_attribute {
4617 struct attribute attr;
4618 ssize_t (*show)(struct kmem_cache *s, char *buf);
4619 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4622 #define SLAB_ATTR_RO(_name) \
4623 static struct slab_attribute _name##_attr = \
4624 __ATTR(_name, 0400, _name##_show, NULL)
4626 #define SLAB_ATTR(_name) \
4627 static struct slab_attribute _name##_attr = \
4628 __ATTR(_name, 0600, _name##_show, _name##_store)
4630 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4632 return sprintf(buf, "%d\n", s->size);
4634 SLAB_ATTR_RO(slab_size);
4636 static ssize_t align_show(struct kmem_cache *s, char *buf)
4638 return sprintf(buf, "%d\n", s->align);
4640 SLAB_ATTR_RO(align);
4642 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4644 return sprintf(buf, "%d\n", s->object_size);
4646 SLAB_ATTR_RO(object_size);
4648 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4650 return sprintf(buf, "%d\n", oo_objects(s->oo));
4652 SLAB_ATTR_RO(objs_per_slab);
4654 static ssize_t order_store(struct kmem_cache *s,
4655 const char *buf, size_t length)
4657 unsigned long order;
4660 err = kstrtoul(buf, 10, &order);
4664 if (order > slub_max_order || order < slub_min_order)
4667 calculate_sizes(s, order);
4671 static ssize_t order_show(struct kmem_cache *s, char *buf)
4673 return sprintf(buf, "%d\n", oo_order(s->oo));
4677 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4679 return sprintf(buf, "%lu\n", s->min_partial);
4682 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4688 err = kstrtoul(buf, 10, &min);
4692 set_min_partial(s, min);
4695 SLAB_ATTR(min_partial);
4697 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4699 return sprintf(buf, "%u\n", s->cpu_partial);
4702 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4705 unsigned int objects;
4708 err = kstrtouint(buf, 10, &objects);
4711 if (objects && !kmem_cache_has_cpu_partial(s))
4714 s->cpu_partial = objects;
4718 SLAB_ATTR(cpu_partial);
4720 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4724 return sprintf(buf, "%pS\n", s->ctor);
4728 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4730 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4732 SLAB_ATTR_RO(aliases);
4734 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4736 return show_slab_objects(s, buf, SO_PARTIAL);
4738 SLAB_ATTR_RO(partial);
4740 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4742 return show_slab_objects(s, buf, SO_CPU);
4744 SLAB_ATTR_RO(cpu_slabs);
4746 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4748 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4750 SLAB_ATTR_RO(objects);
4752 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4754 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4756 SLAB_ATTR_RO(objects_partial);
4758 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4765 for_each_online_cpu(cpu) {
4766 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4769 pages += page->pages;
4770 objects += page->pobjects;
4774 len = sprintf(buf, "%d(%d)", objects, pages);
4777 for_each_online_cpu(cpu) {
4778 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4780 if (page && len < PAGE_SIZE - 20)
4781 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4782 page->pobjects, page->pages);
4785 return len + sprintf(buf + len, "\n");
4787 SLAB_ATTR_RO(slabs_cpu_partial);
4789 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4791 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4794 static ssize_t reclaim_account_store(struct kmem_cache *s,
4795 const char *buf, size_t length)
4797 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4799 s->flags |= SLAB_RECLAIM_ACCOUNT;
4802 SLAB_ATTR(reclaim_account);
4804 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4806 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4808 SLAB_ATTR_RO(hwcache_align);
4810 #ifdef CONFIG_ZONE_DMA
4811 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4813 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4815 SLAB_ATTR_RO(cache_dma);
4818 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4820 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4822 SLAB_ATTR_RO(destroy_by_rcu);
4824 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4826 return sprintf(buf, "%d\n", s->reserved);
4828 SLAB_ATTR_RO(reserved);
4830 #ifdef CONFIG_SLUB_DEBUG
4831 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4833 return show_slab_objects(s, buf, SO_ALL);
4835 SLAB_ATTR_RO(slabs);
4837 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4839 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4841 SLAB_ATTR_RO(total_objects);
4843 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4845 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4848 static ssize_t sanity_checks_store(struct kmem_cache *s,
4849 const char *buf, size_t length)
4851 s->flags &= ~SLAB_DEBUG_FREE;
4852 if (buf[0] == '1') {
4853 s->flags &= ~__CMPXCHG_DOUBLE;
4854 s->flags |= SLAB_DEBUG_FREE;
4858 SLAB_ATTR(sanity_checks);
4860 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4862 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4865 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4869 * Tracing a merged cache is going to give confusing results
4870 * as well as cause other issues like converting a mergeable
4871 * cache into an umergeable one.
4873 if (s->refcount > 1)
4876 s->flags &= ~SLAB_TRACE;
4877 if (buf[0] == '1') {
4878 s->flags &= ~__CMPXCHG_DOUBLE;
4879 s->flags |= SLAB_TRACE;
4885 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4887 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4890 static ssize_t red_zone_store(struct kmem_cache *s,
4891 const char *buf, size_t length)
4893 if (any_slab_objects(s))
4896 s->flags &= ~SLAB_RED_ZONE;
4897 if (buf[0] == '1') {
4898 s->flags &= ~__CMPXCHG_DOUBLE;
4899 s->flags |= SLAB_RED_ZONE;
4901 calculate_sizes(s, -1);
4904 SLAB_ATTR(red_zone);
4906 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4908 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4911 static ssize_t poison_store(struct kmem_cache *s,
4912 const char *buf, size_t length)
4914 if (any_slab_objects(s))
4917 s->flags &= ~SLAB_POISON;
4918 if (buf[0] == '1') {
4919 s->flags &= ~__CMPXCHG_DOUBLE;
4920 s->flags |= SLAB_POISON;
4922 calculate_sizes(s, -1);
4927 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4929 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4932 static ssize_t store_user_store(struct kmem_cache *s,
4933 const char *buf, size_t length)
4935 if (any_slab_objects(s))
4938 s->flags &= ~SLAB_STORE_USER;
4939 if (buf[0] == '1') {
4940 s->flags &= ~__CMPXCHG_DOUBLE;
4941 s->flags |= SLAB_STORE_USER;
4943 calculate_sizes(s, -1);
4946 SLAB_ATTR(store_user);
4948 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4953 static ssize_t validate_store(struct kmem_cache *s,
4954 const char *buf, size_t length)
4958 if (buf[0] == '1') {
4959 ret = validate_slab_cache(s);
4965 SLAB_ATTR(validate);
4967 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4969 if (!(s->flags & SLAB_STORE_USER))
4971 return list_locations(s, buf, TRACK_ALLOC);
4973 SLAB_ATTR_RO(alloc_calls);
4975 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4977 if (!(s->flags & SLAB_STORE_USER))
4979 return list_locations(s, buf, TRACK_FREE);
4981 SLAB_ATTR_RO(free_calls);
4982 #endif /* CONFIG_SLUB_DEBUG */
4984 #ifdef CONFIG_FAILSLAB
4985 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4987 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4990 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4993 if (s->refcount > 1)
4996 s->flags &= ~SLAB_FAILSLAB;
4998 s->flags |= SLAB_FAILSLAB;
5001 SLAB_ATTR(failslab);
5004 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5009 static ssize_t shrink_store(struct kmem_cache *s,
5010 const char *buf, size_t length)
5013 kmem_cache_shrink(s);
5021 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5023 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5026 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5027 const char *buf, size_t length)
5029 unsigned long ratio;
5032 err = kstrtoul(buf, 10, &ratio);
5037 s->remote_node_defrag_ratio = ratio * 10;
5041 SLAB_ATTR(remote_node_defrag_ratio);
5044 #ifdef CONFIG_SLUB_STATS
5045 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5047 unsigned long sum = 0;
5050 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5055 for_each_online_cpu(cpu) {
5056 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5062 len = sprintf(buf, "%lu", sum);
5065 for_each_online_cpu(cpu) {
5066 if (data[cpu] && len < PAGE_SIZE - 20)
5067 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5071 return len + sprintf(buf + len, "\n");
5074 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5078 for_each_online_cpu(cpu)
5079 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5082 #define STAT_ATTR(si, text) \
5083 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5085 return show_stat(s, buf, si); \
5087 static ssize_t text##_store(struct kmem_cache *s, \
5088 const char *buf, size_t length) \
5090 if (buf[0] != '0') \
5092 clear_stat(s, si); \
5097 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5098 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5099 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5100 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5101 STAT_ATTR(FREE_FROZEN, free_frozen);
5102 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5103 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5104 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5105 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5106 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5107 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5108 STAT_ATTR(FREE_SLAB, free_slab);
5109 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5110 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5111 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5112 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5113 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5114 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5115 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5116 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5117 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5118 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5119 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5120 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5121 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5122 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5125 static struct attribute *slab_attrs[] = {
5126 &slab_size_attr.attr,
5127 &object_size_attr.attr,
5128 &objs_per_slab_attr.attr,
5130 &min_partial_attr.attr,
5131 &cpu_partial_attr.attr,
5133 &objects_partial_attr.attr,
5135 &cpu_slabs_attr.attr,
5139 &hwcache_align_attr.attr,
5140 &reclaim_account_attr.attr,
5141 &destroy_by_rcu_attr.attr,
5143 &reserved_attr.attr,
5144 &slabs_cpu_partial_attr.attr,
5145 #ifdef CONFIG_SLUB_DEBUG
5146 &total_objects_attr.attr,
5148 &sanity_checks_attr.attr,
5150 &red_zone_attr.attr,
5152 &store_user_attr.attr,
5153 &validate_attr.attr,
5154 &alloc_calls_attr.attr,
5155 &free_calls_attr.attr,
5157 #ifdef CONFIG_ZONE_DMA
5158 &cache_dma_attr.attr,
5161 &remote_node_defrag_ratio_attr.attr,
5163 #ifdef CONFIG_SLUB_STATS
5164 &alloc_fastpath_attr.attr,
5165 &alloc_slowpath_attr.attr,
5166 &free_fastpath_attr.attr,
5167 &free_slowpath_attr.attr,
5168 &free_frozen_attr.attr,
5169 &free_add_partial_attr.attr,
5170 &free_remove_partial_attr.attr,
5171 &alloc_from_partial_attr.attr,
5172 &alloc_slab_attr.attr,
5173 &alloc_refill_attr.attr,
5174 &alloc_node_mismatch_attr.attr,
5175 &free_slab_attr.attr,
5176 &cpuslab_flush_attr.attr,
5177 &deactivate_full_attr.attr,
5178 &deactivate_empty_attr.attr,
5179 &deactivate_to_head_attr.attr,
5180 &deactivate_to_tail_attr.attr,
5181 &deactivate_remote_frees_attr.attr,
5182 &deactivate_bypass_attr.attr,
5183 &order_fallback_attr.attr,
5184 &cmpxchg_double_fail_attr.attr,
5185 &cmpxchg_double_cpu_fail_attr.attr,
5186 &cpu_partial_alloc_attr.attr,
5187 &cpu_partial_free_attr.attr,
5188 &cpu_partial_node_attr.attr,
5189 &cpu_partial_drain_attr.attr,
5191 #ifdef CONFIG_FAILSLAB
5192 &failslab_attr.attr,
5198 static struct attribute_group slab_attr_group = {
5199 .attrs = slab_attrs,
5202 static ssize_t slab_attr_show(struct kobject *kobj,
5203 struct attribute *attr,
5206 struct slab_attribute *attribute;
5207 struct kmem_cache *s;
5210 attribute = to_slab_attr(attr);
5213 if (!attribute->show)
5216 err = attribute->show(s, buf);
5221 static ssize_t slab_attr_store(struct kobject *kobj,
5222 struct attribute *attr,
5223 const char *buf, size_t len)
5225 struct slab_attribute *attribute;
5226 struct kmem_cache *s;
5229 attribute = to_slab_attr(attr);
5232 if (!attribute->store)
5235 err = attribute->store(s, buf, len);
5236 #ifdef CONFIG_MEMCG_KMEM
5237 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5238 struct kmem_cache *c;
5240 mutex_lock(&slab_mutex);
5241 if (s->max_attr_size < len)
5242 s->max_attr_size = len;
5245 * This is a best effort propagation, so this function's return
5246 * value will be determined by the parent cache only. This is
5247 * basically because not all attributes will have a well
5248 * defined semantics for rollbacks - most of the actions will
5249 * have permanent effects.
5251 * Returning the error value of any of the children that fail
5252 * is not 100 % defined, in the sense that users seeing the
5253 * error code won't be able to know anything about the state of
5256 * Only returning the error code for the parent cache at least
5257 * has well defined semantics. The cache being written to
5258 * directly either failed or succeeded, in which case we loop
5259 * through the descendants with best-effort propagation.
5261 for_each_memcg_cache(c, s)
5262 attribute->store(c, buf, len);
5263 mutex_unlock(&slab_mutex);
5269 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5271 #ifdef CONFIG_MEMCG_KMEM
5273 char *buffer = NULL;
5274 struct kmem_cache *root_cache;
5276 if (is_root_cache(s))
5279 root_cache = s->memcg_params.root_cache;
5282 * This mean this cache had no attribute written. Therefore, no point
5283 * in copying default values around
5285 if (!root_cache->max_attr_size)
5288 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5291 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5294 if (!attr || !attr->store || !attr->show)
5298 * It is really bad that we have to allocate here, so we will
5299 * do it only as a fallback. If we actually allocate, though,
5300 * we can just use the allocated buffer until the end.
5302 * Most of the slub attributes will tend to be very small in
5303 * size, but sysfs allows buffers up to a page, so they can
5304 * theoretically happen.
5308 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5309 !IS_ENABLED(CONFIG_SLUB_STATS))
5312 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5313 if (WARN_ON(!buffer))
5318 len = attr->show(root_cache, buf);
5320 attr->store(s, buf, len);
5324 free_page((unsigned long)buffer);
5328 static void kmem_cache_release(struct kobject *k)
5330 slab_kmem_cache_release(to_slab(k));
5333 static const struct sysfs_ops slab_sysfs_ops = {
5334 .show = slab_attr_show,
5335 .store = slab_attr_store,
5338 static struct kobj_type slab_ktype = {
5339 .sysfs_ops = &slab_sysfs_ops,
5340 .release = kmem_cache_release,
5343 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5345 struct kobj_type *ktype = get_ktype(kobj);
5347 if (ktype == &slab_ktype)
5352 static const struct kset_uevent_ops slab_uevent_ops = {
5353 .filter = uevent_filter,
5356 static struct kset *slab_kset;
5358 static inline struct kset *cache_kset(struct kmem_cache *s)
5360 #ifdef CONFIG_MEMCG_KMEM
5361 if (!is_root_cache(s))
5362 return s->memcg_params.root_cache->memcg_kset;
5367 #define ID_STR_LENGTH 64
5369 /* Create a unique string id for a slab cache:
5371 * Format :[flags-]size
5373 static char *create_unique_id(struct kmem_cache *s)
5375 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5382 * First flags affecting slabcache operations. We will only
5383 * get here for aliasable slabs so we do not need to support
5384 * too many flags. The flags here must cover all flags that
5385 * are matched during merging to guarantee that the id is
5388 if (s->flags & SLAB_CACHE_DMA)
5390 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5392 if (s->flags & SLAB_DEBUG_FREE)
5394 if (!(s->flags & SLAB_NOTRACK))
5398 p += sprintf(p, "%07d", s->size);
5400 BUG_ON(p > name + ID_STR_LENGTH - 1);
5404 static int sysfs_slab_add(struct kmem_cache *s)
5408 int unmergeable = slab_unmergeable(s);
5412 * Slabcache can never be merged so we can use the name proper.
5413 * This is typically the case for debug situations. In that
5414 * case we can catch duplicate names easily.
5416 sysfs_remove_link(&slab_kset->kobj, s->name);
5420 * Create a unique name for the slab as a target
5423 name = create_unique_id(s);
5426 s->kobj.kset = cache_kset(s);
5427 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5431 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5435 #ifdef CONFIG_MEMCG_KMEM
5436 if (is_root_cache(s)) {
5437 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5438 if (!s->memcg_kset) {
5445 kobject_uevent(&s->kobj, KOBJ_ADD);
5447 /* Setup first alias */
5448 sysfs_slab_alias(s, s->name);
5455 kobject_del(&s->kobj);
5459 void sysfs_slab_remove(struct kmem_cache *s)
5461 if (slab_state < FULL)
5463 * Sysfs has not been setup yet so no need to remove the
5468 #ifdef CONFIG_MEMCG_KMEM
5469 kset_unregister(s->memcg_kset);
5471 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5472 kobject_del(&s->kobj);
5473 kobject_put(&s->kobj);
5477 * Need to buffer aliases during bootup until sysfs becomes
5478 * available lest we lose that information.
5480 struct saved_alias {
5481 struct kmem_cache *s;
5483 struct saved_alias *next;
5486 static struct saved_alias *alias_list;
5488 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5490 struct saved_alias *al;
5492 if (slab_state == FULL) {
5494 * If we have a leftover link then remove it.
5496 sysfs_remove_link(&slab_kset->kobj, name);
5497 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5500 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5506 al->next = alias_list;
5511 static int __init slab_sysfs_init(void)
5513 struct kmem_cache *s;
5516 mutex_lock(&slab_mutex);
5518 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5520 mutex_unlock(&slab_mutex);
5521 pr_err("Cannot register slab subsystem.\n");
5527 list_for_each_entry(s, &slab_caches, list) {
5528 err = sysfs_slab_add(s);
5530 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5534 while (alias_list) {
5535 struct saved_alias *al = alias_list;
5537 alias_list = alias_list->next;
5538 err = sysfs_slab_alias(al->s, al->name);
5540 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5545 mutex_unlock(&slab_mutex);
5550 __initcall(slab_sysfs_init);
5551 #endif /* CONFIG_SYSFS */
5554 * The /proc/slabinfo ABI
5556 #ifdef CONFIG_SLABINFO
5557 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5559 unsigned long nr_slabs = 0;
5560 unsigned long nr_objs = 0;
5561 unsigned long nr_free = 0;
5563 struct kmem_cache_node *n;
5565 for_each_kmem_cache_node(s, node, n) {
5566 nr_slabs += node_nr_slabs(n);
5567 nr_objs += node_nr_objs(n);
5568 nr_free += count_partial(n, count_free);
5571 sinfo->active_objs = nr_objs - nr_free;
5572 sinfo->num_objs = nr_objs;
5573 sinfo->active_slabs = nr_slabs;
5574 sinfo->num_slabs = nr_slabs;
5575 sinfo->objects_per_slab = oo_objects(s->oo);
5576 sinfo->cache_order = oo_order(s->oo);
5579 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5583 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5584 size_t count, loff_t *ppos)
5588 #endif /* CONFIG_SLABINFO */