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 int order_objects(int order, unsigned long size, int reserved)
289 return ((PAGE_SIZE << order) - reserved) / size;
292 static inline struct kmem_cache_order_objects oo_make(int order,
293 unsigned long size, int reserved)
295 struct kmem_cache_order_objects x = {
296 (order << OO_SHIFT) + order_objects(order, size, reserved)
302 static inline int oo_order(struct kmem_cache_order_objects x)
304 return x.x >> OO_SHIFT;
307 static inline int oo_objects(struct kmem_cache_order_objects x)
309 return x.x & OO_MASK;
313 * Per slab locking using the pagelock
315 static __always_inline void slab_lock(struct page *page)
317 VM_BUG_ON_PAGE(PageTail(page), page);
318 bit_spin_lock(PG_locked, &page->flags);
321 static __always_inline void slab_unlock(struct page *page)
323 VM_BUG_ON_PAGE(PageTail(page), page);
324 __bit_spin_unlock(PG_locked, &page->flags);
327 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
330 tmp.counters = counters_new;
332 * page->counters can cover frozen/inuse/objects as well
333 * as page->_count. If we assign to ->counters directly
334 * we run the risk of losing updates to page->_count, so
335 * be careful and only assign to the fields we need.
337 page->frozen = tmp.frozen;
338 page->inuse = tmp.inuse;
339 page->objects = tmp.objects;
342 /* Interrupts must be disabled (for the fallback code to work right) */
343 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
344 void *freelist_old, unsigned long counters_old,
345 void *freelist_new, unsigned long counters_new,
348 VM_BUG_ON(!irqs_disabled());
349 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
350 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
351 if (s->flags & __CMPXCHG_DOUBLE) {
352 if (cmpxchg_double(&page->freelist, &page->counters,
353 freelist_old, counters_old,
354 freelist_new, counters_new))
360 if (page->freelist == freelist_old &&
361 page->counters == counters_old) {
362 page->freelist = freelist_new;
363 set_page_slub_counters(page, counters_new);
371 stat(s, CMPXCHG_DOUBLE_FAIL);
373 #ifdef SLUB_DEBUG_CMPXCHG
374 pr_info("%s %s: cmpxchg double redo ", n, s->name);
380 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
381 void *freelist_old, unsigned long counters_old,
382 void *freelist_new, unsigned long counters_new,
385 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
386 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
387 if (s->flags & __CMPXCHG_DOUBLE) {
388 if (cmpxchg_double(&page->freelist, &page->counters,
389 freelist_old, counters_old,
390 freelist_new, counters_new))
397 local_irq_save(flags);
399 if (page->freelist == freelist_old &&
400 page->counters == counters_old) {
401 page->freelist = freelist_new;
402 set_page_slub_counters(page, counters_new);
404 local_irq_restore(flags);
408 local_irq_restore(flags);
412 stat(s, CMPXCHG_DOUBLE_FAIL);
414 #ifdef SLUB_DEBUG_CMPXCHG
415 pr_info("%s %s: cmpxchg double redo ", n, s->name);
421 #ifdef CONFIG_SLUB_DEBUG
423 * Determine a map of object in use on a page.
425 * Node listlock must be held to guarantee that the page does
426 * not vanish from under us.
428 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
431 void *addr = page_address(page);
433 for (p = page->freelist; p; p = get_freepointer(s, p))
434 set_bit(slab_index(p, s, addr), map);
440 #if defined(CONFIG_SLUB_DEBUG_ON)
441 static int slub_debug = DEBUG_DEFAULT_FLAGS;
442 #elif defined(CONFIG_KASAN)
443 static int slub_debug = SLAB_STORE_USER;
445 static int slub_debug;
448 static char *slub_debug_slabs;
449 static int disable_higher_order_debug;
452 * slub is about to manipulate internal object metadata. This memory lies
453 * outside the range of the allocated object, so accessing it would normally
454 * be reported by kasan as a bounds error. metadata_access_enable() is used
455 * to tell kasan that these accesses are OK.
457 static inline void metadata_access_enable(void)
459 kasan_disable_current();
462 static inline void metadata_access_disable(void)
464 kasan_enable_current();
470 static void print_section(char *text, u8 *addr, unsigned int length)
472 metadata_access_enable();
473 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
475 metadata_access_disable();
478 static struct track *get_track(struct kmem_cache *s, void *object,
479 enum track_item alloc)
484 p = object + s->offset + sizeof(void *);
486 p = object + s->inuse;
491 static void set_track(struct kmem_cache *s, void *object,
492 enum track_item alloc, unsigned long addr)
494 struct track *p = get_track(s, object, alloc);
497 #ifdef CONFIG_STACKTRACE
498 struct stack_trace trace;
501 trace.nr_entries = 0;
502 trace.max_entries = TRACK_ADDRS_COUNT;
503 trace.entries = p->addrs;
505 metadata_access_enable();
506 save_stack_trace(&trace);
507 metadata_access_disable();
509 /* See rant in lockdep.c */
510 if (trace.nr_entries != 0 &&
511 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
514 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
518 p->cpu = smp_processor_id();
519 p->pid = current->pid;
522 memset(p, 0, sizeof(struct track));
525 static void init_tracking(struct kmem_cache *s, void *object)
527 if (!(s->flags & SLAB_STORE_USER))
530 set_track(s, object, TRACK_FREE, 0UL);
531 set_track(s, object, TRACK_ALLOC, 0UL);
534 static void print_track(const char *s, struct track *t)
539 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
540 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
541 #ifdef CONFIG_STACKTRACE
544 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
546 pr_err("\t%pS\n", (void *)t->addrs[i]);
553 static void print_tracking(struct kmem_cache *s, void *object)
555 if (!(s->flags & SLAB_STORE_USER))
558 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
559 print_track("Freed", get_track(s, object, TRACK_FREE));
562 static void print_page_info(struct page *page)
564 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
565 page, page->objects, page->inuse, page->freelist, page->flags);
569 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
571 struct va_format vaf;
577 pr_err("=============================================================================\n");
578 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
579 pr_err("-----------------------------------------------------------------------------\n\n");
581 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
585 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
587 struct va_format vaf;
593 pr_err("FIX %s: %pV\n", s->name, &vaf);
597 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
599 unsigned int off; /* Offset of last byte */
600 u8 *addr = page_address(page);
602 print_tracking(s, p);
604 print_page_info(page);
606 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
610 print_section("Bytes b4 ", p - 16, 16);
612 print_section("Object ", p, min_t(unsigned long, s->object_size,
614 if (s->flags & SLAB_RED_ZONE)
615 print_section("Redzone ", p + s->object_size,
616 s->inuse - s->object_size);
619 off = s->offset + sizeof(void *);
623 if (s->flags & SLAB_STORE_USER)
624 off += 2 * sizeof(struct track);
627 /* Beginning of the filler is the free pointer */
628 print_section("Padding ", p + off, s->size - off);
633 void object_err(struct kmem_cache *s, struct page *page,
634 u8 *object, char *reason)
636 slab_bug(s, "%s", reason);
637 print_trailer(s, page, object);
640 static void slab_err(struct kmem_cache *s, struct page *page,
641 const char *fmt, ...)
647 vsnprintf(buf, sizeof(buf), fmt, args);
649 slab_bug(s, "%s", buf);
650 print_page_info(page);
654 static void init_object(struct kmem_cache *s, void *object, u8 val)
658 if (s->flags & __OBJECT_POISON) {
659 memset(p, POISON_FREE, s->object_size - 1);
660 p[s->object_size - 1] = POISON_END;
663 if (s->flags & SLAB_RED_ZONE)
664 memset(p + s->object_size, val, s->inuse - s->object_size);
667 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
668 void *from, void *to)
670 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
671 memset(from, data, to - from);
674 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
675 u8 *object, char *what,
676 u8 *start, unsigned int value, unsigned int bytes)
681 metadata_access_enable();
682 fault = memchr_inv(start, value, bytes);
683 metadata_access_disable();
688 while (end > fault && end[-1] == value)
691 slab_bug(s, "%s overwritten", what);
692 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
693 fault, end - 1, fault[0], value);
694 print_trailer(s, page, object);
696 restore_bytes(s, what, value, fault, end);
704 * Bytes of the object to be managed.
705 * If the freepointer may overlay the object then the free
706 * pointer is the first word of the object.
708 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
711 * object + s->object_size
712 * Padding to reach word boundary. This is also used for Redzoning.
713 * Padding is extended by another word if Redzoning is enabled and
714 * object_size == inuse.
716 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
717 * 0xcc (RED_ACTIVE) for objects in use.
720 * Meta data starts here.
722 * A. Free pointer (if we cannot overwrite object on free)
723 * B. Tracking data for SLAB_STORE_USER
724 * C. Padding to reach required alignment boundary or at mininum
725 * one word if debugging is on to be able to detect writes
726 * before the word boundary.
728 * Padding is done using 0x5a (POISON_INUSE)
731 * Nothing is used beyond s->size.
733 * If slabcaches are merged then the object_size and inuse boundaries are mostly
734 * ignored. And therefore no slab options that rely on these boundaries
735 * may be used with merged slabcaches.
738 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
740 unsigned long off = s->inuse; /* The end of info */
743 /* Freepointer is placed after the object. */
744 off += sizeof(void *);
746 if (s->flags & SLAB_STORE_USER)
747 /* We also have user information there */
748 off += 2 * sizeof(struct track);
753 return check_bytes_and_report(s, page, p, "Object padding",
754 p + off, POISON_INUSE, s->size - off);
757 /* Check the pad bytes at the end of a slab page */
758 static int slab_pad_check(struct kmem_cache *s, struct page *page)
766 if (!(s->flags & SLAB_POISON))
769 start = page_address(page);
770 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
771 end = start + length;
772 remainder = length % s->size;
776 metadata_access_enable();
777 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
778 metadata_access_disable();
781 while (end > fault && end[-1] == POISON_INUSE)
784 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
785 print_section("Padding ", end - remainder, remainder);
787 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
791 static int check_object(struct kmem_cache *s, struct page *page,
792 void *object, u8 val)
795 u8 *endobject = object + s->object_size;
797 if (s->flags & SLAB_RED_ZONE) {
798 if (!check_bytes_and_report(s, page, object, "Redzone",
799 endobject, val, s->inuse - s->object_size))
802 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
803 check_bytes_and_report(s, page, p, "Alignment padding",
804 endobject, POISON_INUSE,
805 s->inuse - s->object_size);
809 if (s->flags & SLAB_POISON) {
810 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
811 (!check_bytes_and_report(s, page, p, "Poison", p,
812 POISON_FREE, s->object_size - 1) ||
813 !check_bytes_and_report(s, page, p, "Poison",
814 p + s->object_size - 1, POISON_END, 1)))
817 * check_pad_bytes cleans up on its own.
819 check_pad_bytes(s, page, p);
822 if (!s->offset && val == SLUB_RED_ACTIVE)
824 * Object and freepointer overlap. Cannot check
825 * freepointer while object is allocated.
829 /* Check free pointer validity */
830 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
831 object_err(s, page, p, "Freepointer corrupt");
833 * No choice but to zap it and thus lose the remainder
834 * of the free objects in this slab. May cause
835 * another error because the object count is now wrong.
837 set_freepointer(s, p, NULL);
843 static int check_slab(struct kmem_cache *s, struct page *page)
847 VM_BUG_ON(!irqs_disabled());
849 if (!PageSlab(page)) {
850 slab_err(s, page, "Not a valid slab page");
854 maxobj = order_objects(compound_order(page), s->size, s->reserved);
855 if (page->objects > maxobj) {
856 slab_err(s, page, "objects %u > max %u",
857 page->objects, maxobj);
860 if (page->inuse > page->objects) {
861 slab_err(s, page, "inuse %u > max %u",
862 page->inuse, page->objects);
865 /* Slab_pad_check fixes things up after itself */
866 slab_pad_check(s, page);
871 * Determine if a certain object on a page is on the freelist. Must hold the
872 * slab lock to guarantee that the chains are in a consistent state.
874 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
882 while (fp && nr <= page->objects) {
885 if (!check_valid_pointer(s, page, fp)) {
887 object_err(s, page, object,
888 "Freechain corrupt");
889 set_freepointer(s, object, NULL);
891 slab_err(s, page, "Freepointer corrupt");
892 page->freelist = NULL;
893 page->inuse = page->objects;
894 slab_fix(s, "Freelist cleared");
900 fp = get_freepointer(s, object);
904 max_objects = order_objects(compound_order(page), s->size, s->reserved);
905 if (max_objects > MAX_OBJS_PER_PAGE)
906 max_objects = MAX_OBJS_PER_PAGE;
908 if (page->objects != max_objects) {
909 slab_err(s, page, "Wrong number of objects. Found %d but "
910 "should be %d", page->objects, max_objects);
911 page->objects = max_objects;
912 slab_fix(s, "Number of objects adjusted.");
914 if (page->inuse != page->objects - nr) {
915 slab_err(s, page, "Wrong object count. Counter is %d but "
916 "counted were %d", page->inuse, page->objects - nr);
917 page->inuse = page->objects - nr;
918 slab_fix(s, "Object count adjusted.");
920 return search == NULL;
923 static void trace(struct kmem_cache *s, struct page *page, void *object,
926 if (s->flags & SLAB_TRACE) {
927 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
929 alloc ? "alloc" : "free",
934 print_section("Object ", (void *)object,
942 * Tracking of fully allocated slabs for debugging purposes.
944 static void add_full(struct kmem_cache *s,
945 struct kmem_cache_node *n, struct page *page)
947 if (!(s->flags & SLAB_STORE_USER))
950 lockdep_assert_held(&n->list_lock);
951 list_add(&page->lru, &n->full);
954 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
956 if (!(s->flags & SLAB_STORE_USER))
959 lockdep_assert_held(&n->list_lock);
960 list_del(&page->lru);
963 /* Tracking of the number of slabs for debugging purposes */
964 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
966 struct kmem_cache_node *n = get_node(s, node);
968 return atomic_long_read(&n->nr_slabs);
971 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
973 return atomic_long_read(&n->nr_slabs);
976 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
978 struct kmem_cache_node *n = get_node(s, node);
981 * May be called early in order to allocate a slab for the
982 * kmem_cache_node structure. Solve the chicken-egg
983 * dilemma by deferring the increment of the count during
984 * bootstrap (see early_kmem_cache_node_alloc).
987 atomic_long_inc(&n->nr_slabs);
988 atomic_long_add(objects, &n->total_objects);
991 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
993 struct kmem_cache_node *n = get_node(s, node);
995 atomic_long_dec(&n->nr_slabs);
996 atomic_long_sub(objects, &n->total_objects);
999 /* Object debug checks for alloc/free paths */
1000 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1003 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1006 init_object(s, object, SLUB_RED_INACTIVE);
1007 init_tracking(s, object);
1010 static noinline int alloc_debug_processing(struct kmem_cache *s,
1012 void *object, unsigned long addr)
1014 if (!check_slab(s, page))
1017 if (!check_valid_pointer(s, page, object)) {
1018 object_err(s, page, object, "Freelist Pointer check fails");
1022 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1025 /* Success perform special debug activities for allocs */
1026 if (s->flags & SLAB_STORE_USER)
1027 set_track(s, object, TRACK_ALLOC, addr);
1028 trace(s, page, object, 1);
1029 init_object(s, object, SLUB_RED_ACTIVE);
1033 if (PageSlab(page)) {
1035 * If this is a slab page then lets do the best we can
1036 * to avoid issues in the future. Marking all objects
1037 * as used avoids touching the remaining objects.
1039 slab_fix(s, "Marking all objects used");
1040 page->inuse = page->objects;
1041 page->freelist = NULL;
1046 /* Supports checking bulk free of a constructed freelist */
1047 static noinline struct kmem_cache_node *free_debug_processing(
1048 struct kmem_cache *s, struct page *page,
1049 void *head, void *tail, int bulk_cnt,
1050 unsigned long addr, unsigned long *flags)
1052 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1053 void *object = head;
1056 spin_lock_irqsave(&n->list_lock, *flags);
1059 if (!check_slab(s, page))
1065 if (!check_valid_pointer(s, page, object)) {
1066 slab_err(s, page, "Invalid object pointer 0x%p", object);
1070 if (on_freelist(s, page, object)) {
1071 object_err(s, page, object, "Object already free");
1075 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1078 if (unlikely(s != page->slab_cache)) {
1079 if (!PageSlab(page)) {
1080 slab_err(s, page, "Attempt to free object(0x%p) "
1081 "outside of slab", object);
1082 } else if (!page->slab_cache) {
1083 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1087 object_err(s, page, object,
1088 "page slab pointer corrupt.");
1092 if (s->flags & SLAB_STORE_USER)
1093 set_track(s, object, TRACK_FREE, addr);
1094 trace(s, page, object, 0);
1095 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1096 init_object(s, object, SLUB_RED_INACTIVE);
1098 /* Reached end of constructed freelist yet? */
1099 if (object != tail) {
1100 object = get_freepointer(s, object);
1104 if (cnt != bulk_cnt)
1105 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1110 * Keep node_lock to preserve integrity
1111 * until the object is actually freed
1117 spin_unlock_irqrestore(&n->list_lock, *flags);
1118 slab_fix(s, "Object at 0x%p not freed", object);
1122 static int __init setup_slub_debug(char *str)
1124 slub_debug = DEBUG_DEFAULT_FLAGS;
1125 if (*str++ != '=' || !*str)
1127 * No options specified. Switch on full debugging.
1133 * No options but restriction on slabs. This means full
1134 * debugging for slabs matching a pattern.
1141 * Switch off all debugging measures.
1146 * Determine which debug features should be switched on
1148 for (; *str && *str != ','; str++) {
1149 switch (tolower(*str)) {
1151 slub_debug |= SLAB_DEBUG_FREE;
1154 slub_debug |= SLAB_RED_ZONE;
1157 slub_debug |= SLAB_POISON;
1160 slub_debug |= SLAB_STORE_USER;
1163 slub_debug |= SLAB_TRACE;
1166 slub_debug |= SLAB_FAILSLAB;
1170 * Avoid enabling debugging on caches if its minimum
1171 * order would increase as a result.
1173 disable_higher_order_debug = 1;
1176 pr_err("slub_debug option '%c' unknown. skipped\n",
1183 slub_debug_slabs = str + 1;
1188 __setup("slub_debug", setup_slub_debug);
1190 unsigned long kmem_cache_flags(unsigned long object_size,
1191 unsigned long flags, const char *name,
1192 void (*ctor)(void *))
1195 * Enable debugging if selected on the kernel commandline.
1197 if (slub_debug && (!slub_debug_slabs || (name &&
1198 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1199 flags |= slub_debug;
1203 #else /* !CONFIG_SLUB_DEBUG */
1204 static inline void setup_object_debug(struct kmem_cache *s,
1205 struct page *page, void *object) {}
1207 static inline int alloc_debug_processing(struct kmem_cache *s,
1208 struct page *page, void *object, unsigned long addr) { return 0; }
1210 static inline struct kmem_cache_node *free_debug_processing(
1211 struct kmem_cache *s, struct page *page,
1212 void *head, void *tail, int bulk_cnt,
1213 unsigned long addr, unsigned long *flags) { return NULL; }
1215 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1217 static inline int check_object(struct kmem_cache *s, struct page *page,
1218 void *object, u8 val) { return 1; }
1219 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1220 struct page *page) {}
1221 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1222 struct page *page) {}
1223 unsigned long kmem_cache_flags(unsigned long object_size,
1224 unsigned long flags, const char *name,
1225 void (*ctor)(void *))
1229 #define slub_debug 0
1231 #define disable_higher_order_debug 0
1233 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1235 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1237 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1239 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1242 #endif /* CONFIG_SLUB_DEBUG */
1245 * Hooks for other subsystems that check memory allocations. In a typical
1246 * production configuration these hooks all should produce no code at all.
1248 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1250 kmemleak_alloc(ptr, size, 1, flags);
1251 kasan_kmalloc_large(ptr, size);
1254 static inline void kfree_hook(const void *x)
1257 kasan_kfree_large(x);
1260 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1262 kmemleak_free_recursive(x, s->flags);
1265 * Trouble is that we may no longer disable interrupts in the fast path
1266 * So in order to make the debug calls that expect irqs to be
1267 * disabled we need to disable interrupts temporarily.
1269 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1271 unsigned long flags;
1273 local_irq_save(flags);
1274 kmemcheck_slab_free(s, x, s->object_size);
1275 debug_check_no_locks_freed(x, s->object_size);
1276 local_irq_restore(flags);
1279 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1280 debug_check_no_obj_freed(x, s->object_size);
1282 kasan_slab_free(s, x);
1285 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1286 void *head, void *tail)
1289 * Compiler cannot detect this function can be removed if slab_free_hook()
1290 * evaluates to nothing. Thus, catch all relevant config debug options here.
1292 #if defined(CONFIG_KMEMCHECK) || \
1293 defined(CONFIG_LOCKDEP) || \
1294 defined(CONFIG_DEBUG_KMEMLEAK) || \
1295 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1296 defined(CONFIG_KASAN)
1298 void *object = head;
1299 void *tail_obj = tail ? : head;
1302 slab_free_hook(s, object);
1303 } while ((object != tail_obj) &&
1304 (object = get_freepointer(s, object)));
1308 static void setup_object(struct kmem_cache *s, struct page *page,
1311 setup_object_debug(s, page, object);
1312 if (unlikely(s->ctor)) {
1313 kasan_unpoison_object_data(s, object);
1315 kasan_poison_object_data(s, object);
1320 * Slab allocation and freeing
1322 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1323 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1326 int order = oo_order(oo);
1328 flags |= __GFP_NOTRACK;
1330 if (node == NUMA_NO_NODE)
1331 page = alloc_pages(flags, order);
1333 page = __alloc_pages_node(node, flags, order);
1335 if (page && memcg_charge_slab(page, flags, order, s)) {
1336 __free_pages(page, order);
1343 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1346 struct kmem_cache_order_objects oo = s->oo;
1351 flags &= gfp_allowed_mask;
1353 if (gfpflags_allow_blocking(flags))
1356 flags |= s->allocflags;
1359 * Let the initial higher-order allocation fail under memory pressure
1360 * so we fall-back to the minimum order allocation.
1362 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1363 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1364 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1366 page = alloc_slab_page(s, alloc_gfp, node, oo);
1367 if (unlikely(!page)) {
1371 * Allocation may have failed due to fragmentation.
1372 * Try a lower order alloc if possible
1374 page = alloc_slab_page(s, alloc_gfp, node, oo);
1375 if (unlikely(!page))
1377 stat(s, ORDER_FALLBACK);
1380 if (kmemcheck_enabled &&
1381 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1382 int pages = 1 << oo_order(oo);
1384 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1387 * Objects from caches that have a constructor don't get
1388 * cleared when they're allocated, so we need to do it here.
1391 kmemcheck_mark_uninitialized_pages(page, pages);
1393 kmemcheck_mark_unallocated_pages(page, pages);
1396 page->objects = oo_objects(oo);
1398 order = compound_order(page);
1399 page->slab_cache = s;
1400 __SetPageSlab(page);
1401 if (page_is_pfmemalloc(page))
1402 SetPageSlabPfmemalloc(page);
1404 start = page_address(page);
1406 if (unlikely(s->flags & SLAB_POISON))
1407 memset(start, POISON_INUSE, PAGE_SIZE << order);
1409 kasan_poison_slab(page);
1411 for_each_object_idx(p, idx, s, start, page->objects) {
1412 setup_object(s, page, p);
1413 if (likely(idx < page->objects))
1414 set_freepointer(s, p, p + s->size);
1416 set_freepointer(s, p, NULL);
1419 page->freelist = start;
1420 page->inuse = page->objects;
1424 if (gfpflags_allow_blocking(flags))
1425 local_irq_disable();
1429 mod_zone_page_state(page_zone(page),
1430 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1431 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1434 inc_slabs_node(s, page_to_nid(page), page->objects);
1439 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1441 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1442 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1446 return allocate_slab(s,
1447 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1450 static void __free_slab(struct kmem_cache *s, struct page *page)
1452 int order = compound_order(page);
1453 int pages = 1 << order;
1455 if (kmem_cache_debug(s)) {
1458 slab_pad_check(s, page);
1459 for_each_object(p, s, page_address(page),
1461 check_object(s, page, p, SLUB_RED_INACTIVE);
1464 kmemcheck_free_shadow(page, compound_order(page));
1466 mod_zone_page_state(page_zone(page),
1467 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1468 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1471 __ClearPageSlabPfmemalloc(page);
1472 __ClearPageSlab(page);
1474 page_mapcount_reset(page);
1475 if (current->reclaim_state)
1476 current->reclaim_state->reclaimed_slab += pages;
1477 __free_kmem_pages(page, order);
1480 #define need_reserve_slab_rcu \
1481 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1483 static void rcu_free_slab(struct rcu_head *h)
1487 if (need_reserve_slab_rcu)
1488 page = virt_to_head_page(h);
1490 page = container_of((struct list_head *)h, struct page, lru);
1492 __free_slab(page->slab_cache, page);
1495 static void free_slab(struct kmem_cache *s, struct page *page)
1497 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1498 struct rcu_head *head;
1500 if (need_reserve_slab_rcu) {
1501 int order = compound_order(page);
1502 int offset = (PAGE_SIZE << order) - s->reserved;
1504 VM_BUG_ON(s->reserved != sizeof(*head));
1505 head = page_address(page) + offset;
1507 head = &page->rcu_head;
1510 call_rcu(head, rcu_free_slab);
1512 __free_slab(s, page);
1515 static void discard_slab(struct kmem_cache *s, struct page *page)
1517 dec_slabs_node(s, page_to_nid(page), page->objects);
1522 * Management of partially allocated slabs.
1525 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1528 if (tail == DEACTIVATE_TO_TAIL)
1529 list_add_tail(&page->lru, &n->partial);
1531 list_add(&page->lru, &n->partial);
1534 static inline void add_partial(struct kmem_cache_node *n,
1535 struct page *page, int tail)
1537 lockdep_assert_held(&n->list_lock);
1538 __add_partial(n, page, tail);
1541 static inline void remove_partial(struct kmem_cache_node *n,
1544 lockdep_assert_held(&n->list_lock);
1545 list_del(&page->lru);
1550 * Remove slab from the partial list, freeze it and
1551 * return the pointer to the freelist.
1553 * Returns a list of objects or NULL if it fails.
1555 static inline void *acquire_slab(struct kmem_cache *s,
1556 struct kmem_cache_node *n, struct page *page,
1557 int mode, int *objects)
1560 unsigned long counters;
1563 lockdep_assert_held(&n->list_lock);
1566 * Zap the freelist and set the frozen bit.
1567 * The old freelist is the list of objects for the
1568 * per cpu allocation list.
1570 freelist = page->freelist;
1571 counters = page->counters;
1572 new.counters = counters;
1573 *objects = new.objects - new.inuse;
1575 new.inuse = page->objects;
1576 new.freelist = NULL;
1578 new.freelist = freelist;
1581 VM_BUG_ON(new.frozen);
1584 if (!__cmpxchg_double_slab(s, page,
1586 new.freelist, new.counters,
1590 remove_partial(n, page);
1595 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1596 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1599 * Try to allocate a partial slab from a specific node.
1601 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1602 struct kmem_cache_cpu *c, gfp_t flags)
1604 struct page *page, *page2;
1605 void *object = NULL;
1610 * Racy check. If we mistakenly see no partial slabs then we
1611 * just allocate an empty slab. If we mistakenly try to get a
1612 * partial slab and there is none available then get_partials()
1615 if (!n || !n->nr_partial)
1618 spin_lock(&n->list_lock);
1619 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1622 if (!pfmemalloc_match(page, flags))
1625 t = acquire_slab(s, n, page, object == NULL, &objects);
1629 available += objects;
1632 stat(s, ALLOC_FROM_PARTIAL);
1635 put_cpu_partial(s, page, 0);
1636 stat(s, CPU_PARTIAL_NODE);
1638 if (!kmem_cache_has_cpu_partial(s)
1639 || available > s->cpu_partial / 2)
1643 spin_unlock(&n->list_lock);
1648 * Get a page from somewhere. Search in increasing NUMA distances.
1650 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1651 struct kmem_cache_cpu *c)
1654 struct zonelist *zonelist;
1657 enum zone_type high_zoneidx = gfp_zone(flags);
1659 unsigned int cpuset_mems_cookie;
1662 * The defrag ratio allows a configuration of the tradeoffs between
1663 * inter node defragmentation and node local allocations. A lower
1664 * defrag_ratio increases the tendency to do local allocations
1665 * instead of attempting to obtain partial slabs from other nodes.
1667 * If the defrag_ratio is set to 0 then kmalloc() always
1668 * returns node local objects. If the ratio is higher then kmalloc()
1669 * may return off node objects because partial slabs are obtained
1670 * from other nodes and filled up.
1672 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1673 * defrag_ratio = 1000) then every (well almost) allocation will
1674 * first attempt to defrag slab caches on other nodes. This means
1675 * scanning over all nodes to look for partial slabs which may be
1676 * expensive if we do it every time we are trying to find a slab
1677 * with available objects.
1679 if (!s->remote_node_defrag_ratio ||
1680 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1684 cpuset_mems_cookie = read_mems_allowed_begin();
1685 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1686 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1687 struct kmem_cache_node *n;
1689 n = get_node(s, zone_to_nid(zone));
1691 if (n && cpuset_zone_allowed(zone, flags) &&
1692 n->nr_partial > s->min_partial) {
1693 object = get_partial_node(s, n, c, flags);
1696 * Don't check read_mems_allowed_retry()
1697 * here - if mems_allowed was updated in
1698 * parallel, that was a harmless race
1699 * between allocation and the cpuset
1706 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1712 * Get a partial page, lock it and return it.
1714 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1715 struct kmem_cache_cpu *c)
1718 int searchnode = node;
1720 if (node == NUMA_NO_NODE)
1721 searchnode = numa_mem_id();
1722 else if (!node_present_pages(node))
1723 searchnode = node_to_mem_node(node);
1725 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1726 if (object || node != NUMA_NO_NODE)
1729 return get_any_partial(s, flags, c);
1732 #ifdef CONFIG_PREEMPT
1734 * Calculate the next globally unique transaction for disambiguiation
1735 * during cmpxchg. The transactions start with the cpu number and are then
1736 * incremented by CONFIG_NR_CPUS.
1738 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1741 * No preemption supported therefore also no need to check for
1747 static inline unsigned long next_tid(unsigned long tid)
1749 return tid + TID_STEP;
1752 static inline unsigned int tid_to_cpu(unsigned long tid)
1754 return tid % TID_STEP;
1757 static inline unsigned long tid_to_event(unsigned long tid)
1759 return tid / TID_STEP;
1762 static inline unsigned int init_tid(int cpu)
1767 static inline void note_cmpxchg_failure(const char *n,
1768 const struct kmem_cache *s, unsigned long tid)
1770 #ifdef SLUB_DEBUG_CMPXCHG
1771 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1773 pr_info("%s %s: cmpxchg redo ", n, s->name);
1775 #ifdef CONFIG_PREEMPT
1776 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1777 pr_warn("due to cpu change %d -> %d\n",
1778 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1781 if (tid_to_event(tid) != tid_to_event(actual_tid))
1782 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1783 tid_to_event(tid), tid_to_event(actual_tid));
1785 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1786 actual_tid, tid, next_tid(tid));
1788 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1791 static void init_kmem_cache_cpus(struct kmem_cache *s)
1795 for_each_possible_cpu(cpu)
1796 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1800 * Remove the cpu slab
1802 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1805 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1806 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1808 enum slab_modes l = M_NONE, m = M_NONE;
1810 int tail = DEACTIVATE_TO_HEAD;
1814 if (page->freelist) {
1815 stat(s, DEACTIVATE_REMOTE_FREES);
1816 tail = DEACTIVATE_TO_TAIL;
1820 * Stage one: Free all available per cpu objects back
1821 * to the page freelist while it is still frozen. Leave the
1824 * There is no need to take the list->lock because the page
1827 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1829 unsigned long counters;
1832 prior = page->freelist;
1833 counters = page->counters;
1834 set_freepointer(s, freelist, prior);
1835 new.counters = counters;
1837 VM_BUG_ON(!new.frozen);
1839 } while (!__cmpxchg_double_slab(s, page,
1841 freelist, new.counters,
1842 "drain percpu freelist"));
1844 freelist = nextfree;
1848 * Stage two: Ensure that the page is unfrozen while the
1849 * list presence reflects the actual number of objects
1852 * We setup the list membership and then perform a cmpxchg
1853 * with the count. If there is a mismatch then the page
1854 * is not unfrozen but the page is on the wrong list.
1856 * Then we restart the process which may have to remove
1857 * the page from the list that we just put it on again
1858 * because the number of objects in the slab may have
1863 old.freelist = page->freelist;
1864 old.counters = page->counters;
1865 VM_BUG_ON(!old.frozen);
1867 /* Determine target state of the slab */
1868 new.counters = old.counters;
1871 set_freepointer(s, freelist, old.freelist);
1872 new.freelist = freelist;
1874 new.freelist = old.freelist;
1878 if (!new.inuse && n->nr_partial >= s->min_partial)
1880 else if (new.freelist) {
1885 * Taking the spinlock removes the possiblity
1886 * that acquire_slab() will see a slab page that
1889 spin_lock(&n->list_lock);
1893 if (kmem_cache_debug(s) && !lock) {
1896 * This also ensures that the scanning of full
1897 * slabs from diagnostic functions will not see
1900 spin_lock(&n->list_lock);
1908 remove_partial(n, page);
1910 else if (l == M_FULL)
1912 remove_full(s, n, page);
1914 if (m == M_PARTIAL) {
1916 add_partial(n, page, tail);
1919 } else if (m == M_FULL) {
1921 stat(s, DEACTIVATE_FULL);
1922 add_full(s, n, page);
1928 if (!__cmpxchg_double_slab(s, page,
1929 old.freelist, old.counters,
1930 new.freelist, new.counters,
1935 spin_unlock(&n->list_lock);
1938 stat(s, DEACTIVATE_EMPTY);
1939 discard_slab(s, page);
1945 * Unfreeze all the cpu partial slabs.
1947 * This function must be called with interrupts disabled
1948 * for the cpu using c (or some other guarantee must be there
1949 * to guarantee no concurrent accesses).
1951 static void unfreeze_partials(struct kmem_cache *s,
1952 struct kmem_cache_cpu *c)
1954 #ifdef CONFIG_SLUB_CPU_PARTIAL
1955 struct kmem_cache_node *n = NULL, *n2 = NULL;
1956 struct page *page, *discard_page = NULL;
1958 while ((page = c->partial)) {
1962 c->partial = page->next;
1964 n2 = get_node(s, page_to_nid(page));
1967 spin_unlock(&n->list_lock);
1970 spin_lock(&n->list_lock);
1975 old.freelist = page->freelist;
1976 old.counters = page->counters;
1977 VM_BUG_ON(!old.frozen);
1979 new.counters = old.counters;
1980 new.freelist = old.freelist;
1984 } while (!__cmpxchg_double_slab(s, page,
1985 old.freelist, old.counters,
1986 new.freelist, new.counters,
1987 "unfreezing slab"));
1989 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1990 page->next = discard_page;
1991 discard_page = page;
1993 add_partial(n, page, DEACTIVATE_TO_TAIL);
1994 stat(s, FREE_ADD_PARTIAL);
1999 spin_unlock(&n->list_lock);
2001 while (discard_page) {
2002 page = discard_page;
2003 discard_page = discard_page->next;
2005 stat(s, DEACTIVATE_EMPTY);
2006 discard_slab(s, page);
2013 * Put a page that was just frozen (in __slab_free) into a partial page
2014 * slot if available. This is done without interrupts disabled and without
2015 * preemption disabled. The cmpxchg is racy and may put the partial page
2016 * onto a random cpus partial slot.
2018 * If we did not find a slot then simply move all the partials to the
2019 * per node partial list.
2021 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2023 #ifdef CONFIG_SLUB_CPU_PARTIAL
2024 struct page *oldpage;
2032 oldpage = this_cpu_read(s->cpu_slab->partial);
2035 pobjects = oldpage->pobjects;
2036 pages = oldpage->pages;
2037 if (drain && pobjects > s->cpu_partial) {
2038 unsigned long flags;
2040 * partial array is full. Move the existing
2041 * set to the per node partial list.
2043 local_irq_save(flags);
2044 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2045 local_irq_restore(flags);
2049 stat(s, CPU_PARTIAL_DRAIN);
2054 pobjects += page->objects - page->inuse;
2056 page->pages = pages;
2057 page->pobjects = pobjects;
2058 page->next = oldpage;
2060 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2062 if (unlikely(!s->cpu_partial)) {
2063 unsigned long flags;
2065 local_irq_save(flags);
2066 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2067 local_irq_restore(flags);
2073 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2075 stat(s, CPUSLAB_FLUSH);
2076 deactivate_slab(s, c->page, c->freelist);
2078 c->tid = next_tid(c->tid);
2086 * Called from IPI handler with interrupts disabled.
2088 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2090 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2096 unfreeze_partials(s, c);
2100 static void flush_cpu_slab(void *d)
2102 struct kmem_cache *s = d;
2104 __flush_cpu_slab(s, smp_processor_id());
2107 static bool has_cpu_slab(int cpu, void *info)
2109 struct kmem_cache *s = info;
2110 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2112 return c->page || c->partial;
2115 static void flush_all(struct kmem_cache *s)
2117 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2121 * Check if the objects in a per cpu structure fit numa
2122 * locality expectations.
2124 static inline int node_match(struct page *page, int node)
2127 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2133 #ifdef CONFIG_SLUB_DEBUG
2134 static int count_free(struct page *page)
2136 return page->objects - page->inuse;
2139 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2141 return atomic_long_read(&n->total_objects);
2143 #endif /* CONFIG_SLUB_DEBUG */
2145 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2146 static unsigned long count_partial(struct kmem_cache_node *n,
2147 int (*get_count)(struct page *))
2149 unsigned long flags;
2150 unsigned long x = 0;
2153 spin_lock_irqsave(&n->list_lock, flags);
2154 list_for_each_entry(page, &n->partial, lru)
2155 x += get_count(page);
2156 spin_unlock_irqrestore(&n->list_lock, flags);
2159 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2161 static noinline void
2162 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2164 #ifdef CONFIG_SLUB_DEBUG
2165 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2166 DEFAULT_RATELIMIT_BURST);
2168 struct kmem_cache_node *n;
2170 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2173 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2175 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2176 s->name, s->object_size, s->size, oo_order(s->oo),
2179 if (oo_order(s->min) > get_order(s->object_size))
2180 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2183 for_each_kmem_cache_node(s, node, n) {
2184 unsigned long nr_slabs;
2185 unsigned long nr_objs;
2186 unsigned long nr_free;
2188 nr_free = count_partial(n, count_free);
2189 nr_slabs = node_nr_slabs(n);
2190 nr_objs = node_nr_objs(n);
2192 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2193 node, nr_slabs, nr_objs, nr_free);
2198 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2199 int node, struct kmem_cache_cpu **pc)
2202 struct kmem_cache_cpu *c = *pc;
2205 freelist = get_partial(s, flags, node, c);
2210 page = new_slab(s, flags, node);
2212 c = raw_cpu_ptr(s->cpu_slab);
2217 * No other reference to the page yet so we can
2218 * muck around with it freely without cmpxchg
2220 freelist = page->freelist;
2221 page->freelist = NULL;
2223 stat(s, ALLOC_SLAB);
2232 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2234 if (unlikely(PageSlabPfmemalloc(page)))
2235 return gfp_pfmemalloc_allowed(gfpflags);
2241 * Check the page->freelist of a page and either transfer the freelist to the
2242 * per cpu freelist or deactivate the page.
2244 * The page is still frozen if the return value is not NULL.
2246 * If this function returns NULL then the page has been unfrozen.
2248 * This function must be called with interrupt disabled.
2250 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2253 unsigned long counters;
2257 freelist = page->freelist;
2258 counters = page->counters;
2260 new.counters = counters;
2261 VM_BUG_ON(!new.frozen);
2263 new.inuse = page->objects;
2264 new.frozen = freelist != NULL;
2266 } while (!__cmpxchg_double_slab(s, page,
2275 * Slow path. The lockless freelist is empty or we need to perform
2278 * Processing is still very fast if new objects have been freed to the
2279 * regular freelist. In that case we simply take over the regular freelist
2280 * as the lockless freelist and zap the regular freelist.
2282 * If that is not working then we fall back to the partial lists. We take the
2283 * first element of the freelist as the object to allocate now and move the
2284 * rest of the freelist to the lockless freelist.
2286 * And if we were unable to get a new slab from the partial slab lists then
2287 * we need to allocate a new slab. This is the slowest path since it involves
2288 * a call to the page allocator and the setup of a new slab.
2290 * Version of __slab_alloc to use when we know that interrupts are
2291 * already disabled (which is the case for bulk allocation).
2293 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2294 unsigned long addr, struct kmem_cache_cpu *c)
2304 if (unlikely(!node_match(page, node))) {
2305 int searchnode = node;
2307 if (node != NUMA_NO_NODE && !node_present_pages(node))
2308 searchnode = node_to_mem_node(node);
2310 if (unlikely(!node_match(page, searchnode))) {
2311 stat(s, ALLOC_NODE_MISMATCH);
2312 deactivate_slab(s, page, c->freelist);
2320 * By rights, we should be searching for a slab page that was
2321 * PFMEMALLOC but right now, we are losing the pfmemalloc
2322 * information when the page leaves the per-cpu allocator
2324 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2325 deactivate_slab(s, page, c->freelist);
2331 /* must check again c->freelist in case of cpu migration or IRQ */
2332 freelist = c->freelist;
2336 freelist = get_freelist(s, page);
2340 stat(s, DEACTIVATE_BYPASS);
2344 stat(s, ALLOC_REFILL);
2348 * freelist is pointing to the list of objects to be used.
2349 * page is pointing to the page from which the objects are obtained.
2350 * That page must be frozen for per cpu allocations to work.
2352 VM_BUG_ON(!c->page->frozen);
2353 c->freelist = get_freepointer(s, freelist);
2354 c->tid = next_tid(c->tid);
2360 page = c->page = c->partial;
2361 c->partial = page->next;
2362 stat(s, CPU_PARTIAL_ALLOC);
2367 freelist = new_slab_objects(s, gfpflags, node, &c);
2369 if (unlikely(!freelist)) {
2370 slab_out_of_memory(s, gfpflags, node);
2375 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2378 /* Only entered in the debug case */
2379 if (kmem_cache_debug(s) &&
2380 !alloc_debug_processing(s, page, freelist, addr))
2381 goto new_slab; /* Slab failed checks. Next slab needed */
2383 deactivate_slab(s, page, get_freepointer(s, freelist));
2390 * Another one that disabled interrupt and compensates for possible
2391 * cpu changes by refetching the per cpu area pointer.
2393 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2394 unsigned long addr, struct kmem_cache_cpu *c)
2397 unsigned long flags;
2399 local_irq_save(flags);
2400 #ifdef CONFIG_PREEMPT
2402 * We may have been preempted and rescheduled on a different
2403 * cpu before disabling interrupts. Need to reload cpu area
2406 c = this_cpu_ptr(s->cpu_slab);
2409 p = ___slab_alloc(s, gfpflags, node, addr, c);
2410 local_irq_restore(flags);
2415 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2416 * have the fastpath folded into their functions. So no function call
2417 * overhead for requests that can be satisfied on the fastpath.
2419 * The fastpath works by first checking if the lockless freelist can be used.
2420 * If not then __slab_alloc is called for slow processing.
2422 * Otherwise we can simply pick the next object from the lockless free list.
2424 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2425 gfp_t gfpflags, int node, unsigned long addr)
2428 struct kmem_cache_cpu *c;
2432 s = slab_pre_alloc_hook(s, gfpflags);
2437 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2438 * enabled. We may switch back and forth between cpus while
2439 * reading from one cpu area. That does not matter as long
2440 * as we end up on the original cpu again when doing the cmpxchg.
2442 * We should guarantee that tid and kmem_cache are retrieved on
2443 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2444 * to check if it is matched or not.
2447 tid = this_cpu_read(s->cpu_slab->tid);
2448 c = raw_cpu_ptr(s->cpu_slab);
2449 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2450 unlikely(tid != READ_ONCE(c->tid)));
2453 * Irqless object alloc/free algorithm used here depends on sequence
2454 * of fetching cpu_slab's data. tid should be fetched before anything
2455 * on c to guarantee that object and page associated with previous tid
2456 * won't be used with current tid. If we fetch tid first, object and
2457 * page could be one associated with next tid and our alloc/free
2458 * request will be failed. In this case, we will retry. So, no problem.
2463 * The transaction ids are globally unique per cpu and per operation on
2464 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2465 * occurs on the right processor and that there was no operation on the
2466 * linked list in between.
2469 object = c->freelist;
2471 if (unlikely(!object || !node_match(page, node))) {
2472 object = __slab_alloc(s, gfpflags, node, addr, c);
2473 stat(s, ALLOC_SLOWPATH);
2475 void *next_object = get_freepointer_safe(s, object);
2478 * The cmpxchg will only match if there was no additional
2479 * operation and if we are on the right processor.
2481 * The cmpxchg does the following atomically (without lock
2483 * 1. Relocate first pointer to the current per cpu area.
2484 * 2. Verify that tid and freelist have not been changed
2485 * 3. If they were not changed replace tid and freelist
2487 * Since this is without lock semantics the protection is only
2488 * against code executing on this cpu *not* from access by
2491 if (unlikely(!this_cpu_cmpxchg_double(
2492 s->cpu_slab->freelist, s->cpu_slab->tid,
2494 next_object, next_tid(tid)))) {
2496 note_cmpxchg_failure("slab_alloc", s, tid);
2499 prefetch_freepointer(s, next_object);
2500 stat(s, ALLOC_FASTPATH);
2503 if (unlikely(gfpflags & __GFP_ZERO) && object)
2504 memset(object, 0, s->object_size);
2506 slab_post_alloc_hook(s, gfpflags, 1, &object);
2511 static __always_inline void *slab_alloc(struct kmem_cache *s,
2512 gfp_t gfpflags, unsigned long addr)
2514 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2517 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2519 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2521 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2526 EXPORT_SYMBOL(kmem_cache_alloc);
2528 #ifdef CONFIG_TRACING
2529 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2531 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2532 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2533 kasan_kmalloc(s, ret, size);
2536 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2540 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2542 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2544 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2545 s->object_size, s->size, gfpflags, node);
2549 EXPORT_SYMBOL(kmem_cache_alloc_node);
2551 #ifdef CONFIG_TRACING
2552 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2554 int node, size_t size)
2556 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2558 trace_kmalloc_node(_RET_IP_, ret,
2559 size, s->size, gfpflags, node);
2561 kasan_kmalloc(s, ret, size);
2564 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2569 * Slow path handling. This may still be called frequently since objects
2570 * have a longer lifetime than the cpu slabs in most processing loads.
2572 * So we still attempt to reduce cache line usage. Just take the slab
2573 * lock and free the item. If there is no additional partial page
2574 * handling required then we can return immediately.
2576 static void __slab_free(struct kmem_cache *s, struct page *page,
2577 void *head, void *tail, int cnt,
2584 unsigned long counters;
2585 struct kmem_cache_node *n = NULL;
2586 unsigned long uninitialized_var(flags);
2588 stat(s, FREE_SLOWPATH);
2590 if (kmem_cache_debug(s) &&
2591 !(n = free_debug_processing(s, page, head, tail, cnt,
2597 spin_unlock_irqrestore(&n->list_lock, flags);
2600 prior = page->freelist;
2601 counters = page->counters;
2602 set_freepointer(s, tail, prior);
2603 new.counters = counters;
2604 was_frozen = new.frozen;
2606 if ((!new.inuse || !prior) && !was_frozen) {
2608 if (kmem_cache_has_cpu_partial(s) && !prior) {
2611 * Slab was on no list before and will be
2613 * We can defer the list move and instead
2618 } else { /* Needs to be taken off a list */
2620 n = get_node(s, page_to_nid(page));
2622 * Speculatively acquire the list_lock.
2623 * If the cmpxchg does not succeed then we may
2624 * drop the list_lock without any processing.
2626 * Otherwise the list_lock will synchronize with
2627 * other processors updating the list of slabs.
2629 spin_lock_irqsave(&n->list_lock, flags);
2634 } while (!cmpxchg_double_slab(s, page,
2642 * If we just froze the page then put it onto the
2643 * per cpu partial list.
2645 if (new.frozen && !was_frozen) {
2646 put_cpu_partial(s, page, 1);
2647 stat(s, CPU_PARTIAL_FREE);
2650 * The list lock was not taken therefore no list
2651 * activity can be necessary.
2654 stat(s, FREE_FROZEN);
2658 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2662 * Objects left in the slab. If it was not on the partial list before
2665 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2666 if (kmem_cache_debug(s))
2667 remove_full(s, n, page);
2668 add_partial(n, page, DEACTIVATE_TO_TAIL);
2669 stat(s, FREE_ADD_PARTIAL);
2671 spin_unlock_irqrestore(&n->list_lock, flags);
2677 * Slab on the partial list.
2679 remove_partial(n, page);
2680 stat(s, FREE_REMOVE_PARTIAL);
2682 /* Slab must be on the full list */
2683 remove_full(s, n, page);
2686 spin_unlock_irqrestore(&n->list_lock, flags);
2688 discard_slab(s, page);
2692 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2693 * can perform fastpath freeing without additional function calls.
2695 * The fastpath is only possible if we are freeing to the current cpu slab
2696 * of this processor. This typically the case if we have just allocated
2699 * If fastpath is not possible then fall back to __slab_free where we deal
2700 * with all sorts of special processing.
2702 * Bulk free of a freelist with several objects (all pointing to the
2703 * same page) possible by specifying head and tail ptr, plus objects
2704 * count (cnt). Bulk free indicated by tail pointer being set.
2706 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2707 void *head, void *tail, int cnt,
2710 void *tail_obj = tail ? : head;
2711 struct kmem_cache_cpu *c;
2714 slab_free_freelist_hook(s, head, tail);
2718 * Determine the currently cpus per cpu slab.
2719 * The cpu may change afterward. However that does not matter since
2720 * data is retrieved via this pointer. If we are on the same cpu
2721 * during the cmpxchg then the free will succeed.
2724 tid = this_cpu_read(s->cpu_slab->tid);
2725 c = raw_cpu_ptr(s->cpu_slab);
2726 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2727 unlikely(tid != READ_ONCE(c->tid)));
2729 /* Same with comment on barrier() in slab_alloc_node() */
2732 if (likely(page == c->page)) {
2733 set_freepointer(s, tail_obj, c->freelist);
2735 if (unlikely(!this_cpu_cmpxchg_double(
2736 s->cpu_slab->freelist, s->cpu_slab->tid,
2738 head, next_tid(tid)))) {
2740 note_cmpxchg_failure("slab_free", s, tid);
2743 stat(s, FREE_FASTPATH);
2745 __slab_free(s, page, head, tail_obj, cnt, addr);
2749 void kmem_cache_free(struct kmem_cache *s, void *x)
2751 s = cache_from_obj(s, x);
2754 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2755 trace_kmem_cache_free(_RET_IP_, x);
2757 EXPORT_SYMBOL(kmem_cache_free);
2759 struct detached_freelist {
2764 struct kmem_cache *s;
2768 * This function progressively scans the array with free objects (with
2769 * a limited look ahead) and extract objects belonging to the same
2770 * page. It builds a detached freelist directly within the given
2771 * page/objects. This can happen without any need for
2772 * synchronization, because the objects are owned by running process.
2773 * The freelist is build up as a single linked list in the objects.
2774 * The idea is, that this detached freelist can then be bulk
2775 * transferred to the real freelist(s), but only requiring a single
2776 * synchronization primitive. Look ahead in the array is limited due
2777 * to performance reasons.
2780 int build_detached_freelist(struct kmem_cache *s, size_t size,
2781 void **p, struct detached_freelist *df)
2783 size_t first_skipped_index = 0;
2788 /* Always re-init detached_freelist */
2793 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2794 } while (!object && size);
2799 page = virt_to_head_page(object);
2801 /* Handle kalloc'ed objects */
2802 if (unlikely(!PageSlab(page))) {
2803 BUG_ON(!PageCompound(page));
2805 __free_kmem_pages(page, compound_order(page));
2806 p[size] = NULL; /* mark object processed */
2809 /* Derive kmem_cache from object */
2810 df->s = page->slab_cache;
2812 df->s = cache_from_obj(s, object); /* Support for memcg */
2815 /* Start new detached freelist */
2817 set_freepointer(df->s, object, NULL);
2819 df->freelist = object;
2820 p[size] = NULL; /* mark object processed */
2826 continue; /* Skip processed objects */
2828 /* df->page is always set at this point */
2829 if (df->page == virt_to_head_page(object)) {
2830 /* Opportunity build freelist */
2831 set_freepointer(df->s, object, df->freelist);
2832 df->freelist = object;
2834 p[size] = NULL; /* mark object processed */
2839 /* Limit look ahead search */
2843 if (!first_skipped_index)
2844 first_skipped_index = size + 1;
2847 return first_skipped_index;
2850 /* Note that interrupts must be enabled when calling this function. */
2851 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2857 struct detached_freelist df;
2859 size = build_detached_freelist(s, size, p, &df);
2860 if (unlikely(!df.page))
2863 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2864 } while (likely(size));
2866 EXPORT_SYMBOL(kmem_cache_free_bulk);
2868 /* Note that interrupts must be enabled when calling this function. */
2869 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2872 struct kmem_cache_cpu *c;
2875 /* memcg and kmem_cache debug support */
2876 s = slab_pre_alloc_hook(s, flags);
2880 * Drain objects in the per cpu slab, while disabling local
2881 * IRQs, which protects against PREEMPT and interrupts
2882 * handlers invoking normal fastpath.
2884 local_irq_disable();
2885 c = this_cpu_ptr(s->cpu_slab);
2887 for (i = 0; i < size; i++) {
2888 void *object = c->freelist;
2890 if (unlikely(!object)) {
2892 * Invoking slow path likely have side-effect
2893 * of re-populating per CPU c->freelist
2895 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2897 if (unlikely(!p[i]))
2900 c = this_cpu_ptr(s->cpu_slab);
2901 continue; /* goto for-loop */
2903 c->freelist = get_freepointer(s, object);
2906 c->tid = next_tid(c->tid);
2909 /* Clear memory outside IRQ disabled fastpath loop */
2910 if (unlikely(flags & __GFP_ZERO)) {
2913 for (j = 0; j < i; j++)
2914 memset(p[j], 0, s->object_size);
2917 /* memcg and kmem_cache debug support */
2918 slab_post_alloc_hook(s, flags, size, p);
2922 slab_post_alloc_hook(s, flags, i, p);
2923 __kmem_cache_free_bulk(s, i, p);
2926 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2930 * Object placement in a slab is made very easy because we always start at
2931 * offset 0. If we tune the size of the object to the alignment then we can
2932 * get the required alignment by putting one properly sized object after
2935 * Notice that the allocation order determines the sizes of the per cpu
2936 * caches. Each processor has always one slab available for allocations.
2937 * Increasing the allocation order reduces the number of times that slabs
2938 * must be moved on and off the partial lists and is therefore a factor in
2943 * Mininum / Maximum order of slab pages. This influences locking overhead
2944 * and slab fragmentation. A higher order reduces the number of partial slabs
2945 * and increases the number of allocations possible without having to
2946 * take the list_lock.
2948 static int slub_min_order;
2949 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2950 static int slub_min_objects;
2953 * Calculate the order of allocation given an slab object size.
2955 * The order of allocation has significant impact on performance and other
2956 * system components. Generally order 0 allocations should be preferred since
2957 * order 0 does not cause fragmentation in the page allocator. Larger objects
2958 * be problematic to put into order 0 slabs because there may be too much
2959 * unused space left. We go to a higher order if more than 1/16th of the slab
2962 * In order to reach satisfactory performance we must ensure that a minimum
2963 * number of objects is in one slab. Otherwise we may generate too much
2964 * activity on the partial lists which requires taking the list_lock. This is
2965 * less a concern for large slabs though which are rarely used.
2967 * slub_max_order specifies the order where we begin to stop considering the
2968 * number of objects in a slab as critical. If we reach slub_max_order then
2969 * we try to keep the page order as low as possible. So we accept more waste
2970 * of space in favor of a small page order.
2972 * Higher order allocations also allow the placement of more objects in a
2973 * slab and thereby reduce object handling overhead. If the user has
2974 * requested a higher mininum order then we start with that one instead of
2975 * the smallest order which will fit the object.
2977 static inline int slab_order(int size, int min_objects,
2978 int max_order, int fract_leftover, int reserved)
2982 int min_order = slub_min_order;
2984 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2985 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2987 for (order = max(min_order, get_order(min_objects * size + reserved));
2988 order <= max_order; order++) {
2990 unsigned long slab_size = PAGE_SIZE << order;
2992 rem = (slab_size - reserved) % size;
2994 if (rem <= slab_size / fract_leftover)
3001 static inline int calculate_order(int size, int reserved)
3009 * Attempt to find best configuration for a slab. This
3010 * works by first attempting to generate a layout with
3011 * the best configuration and backing off gradually.
3013 * First we increase the acceptable waste in a slab. Then
3014 * we reduce the minimum objects required in a slab.
3016 min_objects = slub_min_objects;
3018 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3019 max_objects = order_objects(slub_max_order, size, reserved);
3020 min_objects = min(min_objects, max_objects);
3022 while (min_objects > 1) {
3024 while (fraction >= 4) {
3025 order = slab_order(size, min_objects,
3026 slub_max_order, fraction, reserved);
3027 if (order <= slub_max_order)
3035 * We were unable to place multiple objects in a slab. Now
3036 * lets see if we can place a single object there.
3038 order = slab_order(size, 1, slub_max_order, 1, reserved);
3039 if (order <= slub_max_order)
3043 * Doh this slab cannot be placed using slub_max_order.
3045 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3046 if (order < MAX_ORDER)
3052 init_kmem_cache_node(struct kmem_cache_node *n)
3055 spin_lock_init(&n->list_lock);
3056 INIT_LIST_HEAD(&n->partial);
3057 #ifdef CONFIG_SLUB_DEBUG
3058 atomic_long_set(&n->nr_slabs, 0);
3059 atomic_long_set(&n->total_objects, 0);
3060 INIT_LIST_HEAD(&n->full);
3064 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3066 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3067 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3070 * Must align to double word boundary for the double cmpxchg
3071 * instructions to work; see __pcpu_double_call_return_bool().
3073 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3074 2 * sizeof(void *));
3079 init_kmem_cache_cpus(s);
3084 static struct kmem_cache *kmem_cache_node;
3087 * No kmalloc_node yet so do it by hand. We know that this is the first
3088 * slab on the node for this slabcache. There are no concurrent accesses
3091 * Note that this function only works on the kmem_cache_node
3092 * when allocating for the kmem_cache_node. This is used for bootstrapping
3093 * memory on a fresh node that has no slab structures yet.
3095 static void early_kmem_cache_node_alloc(int node)
3098 struct kmem_cache_node *n;
3100 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3102 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3105 if (page_to_nid(page) != node) {
3106 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3107 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3112 page->freelist = get_freepointer(kmem_cache_node, n);
3115 kmem_cache_node->node[node] = n;
3116 #ifdef CONFIG_SLUB_DEBUG
3117 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3118 init_tracking(kmem_cache_node, n);
3120 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3121 init_kmem_cache_node(n);
3122 inc_slabs_node(kmem_cache_node, node, page->objects);
3125 * No locks need to be taken here as it has just been
3126 * initialized and there is no concurrent access.
3128 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3131 static void free_kmem_cache_nodes(struct kmem_cache *s)
3134 struct kmem_cache_node *n;
3136 for_each_kmem_cache_node(s, node, n) {
3137 kmem_cache_free(kmem_cache_node, n);
3138 s->node[node] = NULL;
3142 void __kmem_cache_release(struct kmem_cache *s)
3144 free_percpu(s->cpu_slab);
3145 free_kmem_cache_nodes(s);
3148 static int init_kmem_cache_nodes(struct kmem_cache *s)
3152 for_each_node_state(node, N_NORMAL_MEMORY) {
3153 struct kmem_cache_node *n;
3155 if (slab_state == DOWN) {
3156 early_kmem_cache_node_alloc(node);
3159 n = kmem_cache_alloc_node(kmem_cache_node,
3163 free_kmem_cache_nodes(s);
3168 init_kmem_cache_node(n);
3173 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3175 if (min < MIN_PARTIAL)
3177 else if (min > MAX_PARTIAL)
3179 s->min_partial = min;
3183 * calculate_sizes() determines the order and the distribution of data within
3186 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3188 unsigned long flags = s->flags;
3189 unsigned long size = s->object_size;
3193 * Round up object size to the next word boundary. We can only
3194 * place the free pointer at word boundaries and this determines
3195 * the possible location of the free pointer.
3197 size = ALIGN(size, sizeof(void *));
3199 #ifdef CONFIG_SLUB_DEBUG
3201 * Determine if we can poison the object itself. If the user of
3202 * the slab may touch the object after free or before allocation
3203 * then we should never poison the object itself.
3205 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3207 s->flags |= __OBJECT_POISON;
3209 s->flags &= ~__OBJECT_POISON;
3213 * If we are Redzoning then check if there is some space between the
3214 * end of the object and the free pointer. If not then add an
3215 * additional word to have some bytes to store Redzone information.
3217 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3218 size += sizeof(void *);
3222 * With that we have determined the number of bytes in actual use
3223 * by the object. This is the potential offset to the free pointer.
3227 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3230 * Relocate free pointer after the object if it is not
3231 * permitted to overwrite the first word of the object on
3234 * This is the case if we do RCU, have a constructor or
3235 * destructor or are poisoning the objects.
3238 size += sizeof(void *);
3241 #ifdef CONFIG_SLUB_DEBUG
3242 if (flags & SLAB_STORE_USER)
3244 * Need to store information about allocs and frees after
3247 size += 2 * sizeof(struct track);
3249 if (flags & SLAB_RED_ZONE)
3251 * Add some empty padding so that we can catch
3252 * overwrites from earlier objects rather than let
3253 * tracking information or the free pointer be
3254 * corrupted if a user writes before the start
3257 size += sizeof(void *);
3261 * SLUB stores one object immediately after another beginning from
3262 * offset 0. In order to align the objects we have to simply size
3263 * each object to conform to the alignment.
3265 size = ALIGN(size, s->align);
3267 if (forced_order >= 0)
3268 order = forced_order;
3270 order = calculate_order(size, s->reserved);
3277 s->allocflags |= __GFP_COMP;
3279 if (s->flags & SLAB_CACHE_DMA)
3280 s->allocflags |= GFP_DMA;
3282 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3283 s->allocflags |= __GFP_RECLAIMABLE;
3286 * Determine the number of objects per slab
3288 s->oo = oo_make(order, size, s->reserved);
3289 s->min = oo_make(get_order(size), size, s->reserved);
3290 if (oo_objects(s->oo) > oo_objects(s->max))
3293 return !!oo_objects(s->oo);
3296 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3298 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3301 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3302 s->reserved = sizeof(struct rcu_head);
3304 if (!calculate_sizes(s, -1))
3306 if (disable_higher_order_debug) {
3308 * Disable debugging flags that store metadata if the min slab
3311 if (get_order(s->size) > get_order(s->object_size)) {
3312 s->flags &= ~DEBUG_METADATA_FLAGS;
3314 if (!calculate_sizes(s, -1))
3319 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3320 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3321 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3322 /* Enable fast mode */
3323 s->flags |= __CMPXCHG_DOUBLE;
3327 * The larger the object size is, the more pages we want on the partial
3328 * list to avoid pounding the page allocator excessively.
3330 set_min_partial(s, ilog2(s->size) / 2);
3333 * cpu_partial determined the maximum number of objects kept in the
3334 * per cpu partial lists of a processor.
3336 * Per cpu partial lists mainly contain slabs that just have one
3337 * object freed. If they are used for allocation then they can be
3338 * filled up again with minimal effort. The slab will never hit the
3339 * per node partial lists and therefore no locking will be required.
3341 * This setting also determines
3343 * A) The number of objects from per cpu partial slabs dumped to the
3344 * per node list when we reach the limit.
3345 * B) The number of objects in cpu partial slabs to extract from the
3346 * per node list when we run out of per cpu objects. We only fetch
3347 * 50% to keep some capacity around for frees.
3349 if (!kmem_cache_has_cpu_partial(s))
3351 else if (s->size >= PAGE_SIZE)
3353 else if (s->size >= 1024)
3355 else if (s->size >= 256)
3356 s->cpu_partial = 13;
3358 s->cpu_partial = 30;
3361 s->remote_node_defrag_ratio = 1000;
3363 if (!init_kmem_cache_nodes(s))
3366 if (alloc_kmem_cache_cpus(s))
3369 free_kmem_cache_nodes(s);
3371 if (flags & SLAB_PANIC)
3372 panic("Cannot create slab %s size=%lu realsize=%u "
3373 "order=%u offset=%u flags=%lx\n",
3374 s->name, (unsigned long)s->size, s->size,
3375 oo_order(s->oo), s->offset, flags);
3379 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3382 #ifdef CONFIG_SLUB_DEBUG
3383 void *addr = page_address(page);
3385 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3386 sizeof(long), GFP_ATOMIC);
3389 slab_err(s, page, text, s->name);
3392 get_map(s, page, map);
3393 for_each_object(p, s, addr, page->objects) {
3395 if (!test_bit(slab_index(p, s, addr), map)) {
3396 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3397 print_tracking(s, p);
3406 * Attempt to free all partial slabs on a node.
3407 * This is called from __kmem_cache_shutdown(). We must take list_lock
3408 * because sysfs file might still access partial list after the shutdowning.
3410 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3412 struct page *page, *h;
3414 BUG_ON(irqs_disabled());
3415 spin_lock_irq(&n->list_lock);
3416 list_for_each_entry_safe(page, h, &n->partial, lru) {
3418 remove_partial(n, page);
3419 discard_slab(s, page);
3421 list_slab_objects(s, page,
3422 "Objects remaining in %s on __kmem_cache_shutdown()");
3425 spin_unlock_irq(&n->list_lock);
3429 * Release all resources used by a slab cache.
3431 int __kmem_cache_shutdown(struct kmem_cache *s)
3434 struct kmem_cache_node *n;
3437 /* Attempt to free all objects */
3438 for_each_kmem_cache_node(s, node, n) {
3440 if (n->nr_partial || slabs_node(s, node))
3446 /********************************************************************
3448 *******************************************************************/
3450 static int __init setup_slub_min_order(char *str)
3452 get_option(&str, &slub_min_order);
3457 __setup("slub_min_order=", setup_slub_min_order);
3459 static int __init setup_slub_max_order(char *str)
3461 get_option(&str, &slub_max_order);
3462 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3467 __setup("slub_max_order=", setup_slub_max_order);
3469 static int __init setup_slub_min_objects(char *str)
3471 get_option(&str, &slub_min_objects);
3476 __setup("slub_min_objects=", setup_slub_min_objects);
3478 void *__kmalloc(size_t size, gfp_t flags)
3480 struct kmem_cache *s;
3483 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3484 return kmalloc_large(size, flags);
3486 s = kmalloc_slab(size, flags);
3488 if (unlikely(ZERO_OR_NULL_PTR(s)))
3491 ret = slab_alloc(s, flags, _RET_IP_);
3493 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3495 kasan_kmalloc(s, ret, size);
3499 EXPORT_SYMBOL(__kmalloc);
3502 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3507 flags |= __GFP_COMP | __GFP_NOTRACK;
3508 page = alloc_kmem_pages_node(node, flags, get_order(size));
3510 ptr = page_address(page);
3512 kmalloc_large_node_hook(ptr, size, flags);
3516 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3518 struct kmem_cache *s;
3521 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3522 ret = kmalloc_large_node(size, flags, node);
3524 trace_kmalloc_node(_RET_IP_, ret,
3525 size, PAGE_SIZE << get_order(size),
3531 s = kmalloc_slab(size, flags);
3533 if (unlikely(ZERO_OR_NULL_PTR(s)))
3536 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3538 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3540 kasan_kmalloc(s, ret, size);
3544 EXPORT_SYMBOL(__kmalloc_node);
3547 static size_t __ksize(const void *object)
3551 if (unlikely(object == ZERO_SIZE_PTR))
3554 page = virt_to_head_page(object);
3556 if (unlikely(!PageSlab(page))) {
3557 WARN_ON(!PageCompound(page));
3558 return PAGE_SIZE << compound_order(page);
3561 return slab_ksize(page->slab_cache);
3564 size_t ksize(const void *object)
3566 size_t size = __ksize(object);
3567 /* We assume that ksize callers could use whole allocated area,
3568 so we need unpoison this area. */
3569 kasan_krealloc(object, size);
3572 EXPORT_SYMBOL(ksize);
3574 void kfree(const void *x)
3577 void *object = (void *)x;
3579 trace_kfree(_RET_IP_, x);
3581 if (unlikely(ZERO_OR_NULL_PTR(x)))
3584 page = virt_to_head_page(x);
3585 if (unlikely(!PageSlab(page))) {
3586 BUG_ON(!PageCompound(page));
3588 __free_kmem_pages(page, compound_order(page));
3591 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3593 EXPORT_SYMBOL(kfree);
3595 #define SHRINK_PROMOTE_MAX 32
3598 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3599 * up most to the head of the partial lists. New allocations will then
3600 * fill those up and thus they can be removed from the partial lists.
3602 * The slabs with the least items are placed last. This results in them
3603 * being allocated from last increasing the chance that the last objects
3604 * are freed in them.
3606 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3610 struct kmem_cache_node *n;
3613 struct list_head discard;
3614 struct list_head promote[SHRINK_PROMOTE_MAX];
3615 unsigned long flags;
3620 * Disable empty slabs caching. Used to avoid pinning offline
3621 * memory cgroups by kmem pages that can be freed.
3627 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3628 * so we have to make sure the change is visible.
3630 kick_all_cpus_sync();
3634 for_each_kmem_cache_node(s, node, n) {
3635 INIT_LIST_HEAD(&discard);
3636 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3637 INIT_LIST_HEAD(promote + i);
3639 spin_lock_irqsave(&n->list_lock, flags);
3642 * Build lists of slabs to discard or promote.
3644 * Note that concurrent frees may occur while we hold the
3645 * list_lock. page->inuse here is the upper limit.
3647 list_for_each_entry_safe(page, t, &n->partial, lru) {
3648 int free = page->objects - page->inuse;
3650 /* Do not reread page->inuse */
3653 /* We do not keep full slabs on the list */
3656 if (free == page->objects) {
3657 list_move(&page->lru, &discard);
3659 } else if (free <= SHRINK_PROMOTE_MAX)
3660 list_move(&page->lru, promote + free - 1);
3664 * Promote the slabs filled up most to the head of the
3667 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3668 list_splice(promote + i, &n->partial);
3670 spin_unlock_irqrestore(&n->list_lock, flags);
3672 /* Release empty slabs */
3673 list_for_each_entry_safe(page, t, &discard, lru)
3674 discard_slab(s, page);
3676 if (slabs_node(s, node))
3683 static int slab_mem_going_offline_callback(void *arg)
3685 struct kmem_cache *s;
3687 mutex_lock(&slab_mutex);
3688 list_for_each_entry(s, &slab_caches, list)
3689 __kmem_cache_shrink(s, false);
3690 mutex_unlock(&slab_mutex);
3695 static void slab_mem_offline_callback(void *arg)
3697 struct kmem_cache_node *n;
3698 struct kmem_cache *s;
3699 struct memory_notify *marg = arg;
3702 offline_node = marg->status_change_nid_normal;
3705 * If the node still has available memory. we need kmem_cache_node
3708 if (offline_node < 0)
3711 mutex_lock(&slab_mutex);
3712 list_for_each_entry(s, &slab_caches, list) {
3713 n = get_node(s, offline_node);
3716 * if n->nr_slabs > 0, slabs still exist on the node
3717 * that is going down. We were unable to free them,
3718 * and offline_pages() function shouldn't call this
3719 * callback. So, we must fail.
3721 BUG_ON(slabs_node(s, offline_node));
3723 s->node[offline_node] = NULL;
3724 kmem_cache_free(kmem_cache_node, n);
3727 mutex_unlock(&slab_mutex);
3730 static int slab_mem_going_online_callback(void *arg)
3732 struct kmem_cache_node *n;
3733 struct kmem_cache *s;
3734 struct memory_notify *marg = arg;
3735 int nid = marg->status_change_nid_normal;
3739 * If the node's memory is already available, then kmem_cache_node is
3740 * already created. Nothing to do.
3746 * We are bringing a node online. No memory is available yet. We must
3747 * allocate a kmem_cache_node structure in order to bring the node
3750 mutex_lock(&slab_mutex);
3751 list_for_each_entry(s, &slab_caches, list) {
3753 * XXX: kmem_cache_alloc_node will fallback to other nodes
3754 * since memory is not yet available from the node that
3757 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3762 init_kmem_cache_node(n);
3766 mutex_unlock(&slab_mutex);
3770 static int slab_memory_callback(struct notifier_block *self,
3771 unsigned long action, void *arg)
3776 case MEM_GOING_ONLINE:
3777 ret = slab_mem_going_online_callback(arg);
3779 case MEM_GOING_OFFLINE:
3780 ret = slab_mem_going_offline_callback(arg);
3783 case MEM_CANCEL_ONLINE:
3784 slab_mem_offline_callback(arg);
3787 case MEM_CANCEL_OFFLINE:
3791 ret = notifier_from_errno(ret);
3797 static struct notifier_block slab_memory_callback_nb = {
3798 .notifier_call = slab_memory_callback,
3799 .priority = SLAB_CALLBACK_PRI,
3802 /********************************************************************
3803 * Basic setup of slabs
3804 *******************************************************************/
3807 * Used for early kmem_cache structures that were allocated using
3808 * the page allocator. Allocate them properly then fix up the pointers
3809 * that may be pointing to the wrong kmem_cache structure.
3812 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3815 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3816 struct kmem_cache_node *n;
3818 memcpy(s, static_cache, kmem_cache->object_size);
3821 * This runs very early, and only the boot processor is supposed to be
3822 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3825 __flush_cpu_slab(s, smp_processor_id());
3826 for_each_kmem_cache_node(s, node, n) {
3829 list_for_each_entry(p, &n->partial, lru)
3832 #ifdef CONFIG_SLUB_DEBUG
3833 list_for_each_entry(p, &n->full, lru)
3837 slab_init_memcg_params(s);
3838 list_add(&s->list, &slab_caches);
3842 void __init kmem_cache_init(void)
3844 static __initdata struct kmem_cache boot_kmem_cache,
3845 boot_kmem_cache_node;
3847 if (debug_guardpage_minorder())
3850 kmem_cache_node = &boot_kmem_cache_node;
3851 kmem_cache = &boot_kmem_cache;
3853 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3854 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3856 register_hotmemory_notifier(&slab_memory_callback_nb);
3858 /* Able to allocate the per node structures */
3859 slab_state = PARTIAL;
3861 create_boot_cache(kmem_cache, "kmem_cache",
3862 offsetof(struct kmem_cache, node) +
3863 nr_node_ids * sizeof(struct kmem_cache_node *),
3864 SLAB_HWCACHE_ALIGN);
3866 kmem_cache = bootstrap(&boot_kmem_cache);
3869 * Allocate kmem_cache_node properly from the kmem_cache slab.
3870 * kmem_cache_node is separately allocated so no need to
3871 * update any list pointers.
3873 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3875 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3876 setup_kmalloc_cache_index_table();
3877 create_kmalloc_caches(0);
3880 register_cpu_notifier(&slab_notifier);
3883 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3885 slub_min_order, slub_max_order, slub_min_objects,
3886 nr_cpu_ids, nr_node_ids);
3889 void __init kmem_cache_init_late(void)
3894 __kmem_cache_alias(const char *name, size_t size, size_t align,
3895 unsigned long flags, void (*ctor)(void *))
3897 struct kmem_cache *s, *c;
3899 s = find_mergeable(size, align, flags, name, ctor);
3904 * Adjust the object sizes so that we clear
3905 * the complete object on kzalloc.
3907 s->object_size = max(s->object_size, (int)size);
3908 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3910 for_each_memcg_cache(c, s) {
3911 c->object_size = s->object_size;
3912 c->inuse = max_t(int, c->inuse,
3913 ALIGN(size, sizeof(void *)));
3916 if (sysfs_slab_alias(s, name)) {
3925 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3929 err = kmem_cache_open(s, flags);
3933 /* Mutex is not taken during early boot */
3934 if (slab_state <= UP)
3937 memcg_propagate_slab_attrs(s);
3938 err = sysfs_slab_add(s);
3940 __kmem_cache_release(s);
3947 * Use the cpu notifier to insure that the cpu slabs are flushed when
3950 static int slab_cpuup_callback(struct notifier_block *nfb,
3951 unsigned long action, void *hcpu)
3953 long cpu = (long)hcpu;
3954 struct kmem_cache *s;
3955 unsigned long flags;
3958 case CPU_UP_CANCELED:
3959 case CPU_UP_CANCELED_FROZEN:
3961 case CPU_DEAD_FROZEN:
3962 mutex_lock(&slab_mutex);
3963 list_for_each_entry(s, &slab_caches, list) {
3964 local_irq_save(flags);
3965 __flush_cpu_slab(s, cpu);
3966 local_irq_restore(flags);
3968 mutex_unlock(&slab_mutex);
3976 static struct notifier_block slab_notifier = {
3977 .notifier_call = slab_cpuup_callback
3982 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3984 struct kmem_cache *s;
3987 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3988 return kmalloc_large(size, gfpflags);
3990 s = kmalloc_slab(size, gfpflags);
3992 if (unlikely(ZERO_OR_NULL_PTR(s)))
3995 ret = slab_alloc(s, gfpflags, caller);
3997 /* Honor the call site pointer we received. */
3998 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4004 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4005 int node, unsigned long caller)
4007 struct kmem_cache *s;
4010 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4011 ret = kmalloc_large_node(size, gfpflags, node);
4013 trace_kmalloc_node(caller, ret,
4014 size, PAGE_SIZE << get_order(size),
4020 s = kmalloc_slab(size, gfpflags);
4022 if (unlikely(ZERO_OR_NULL_PTR(s)))
4025 ret = slab_alloc_node(s, gfpflags, node, caller);
4027 /* Honor the call site pointer we received. */
4028 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4035 static int count_inuse(struct page *page)
4040 static int count_total(struct page *page)
4042 return page->objects;
4046 #ifdef CONFIG_SLUB_DEBUG
4047 static int validate_slab(struct kmem_cache *s, struct page *page,
4051 void *addr = page_address(page);
4053 if (!check_slab(s, page) ||
4054 !on_freelist(s, page, NULL))
4057 /* Now we know that a valid freelist exists */
4058 bitmap_zero(map, page->objects);
4060 get_map(s, page, map);
4061 for_each_object(p, s, addr, page->objects) {
4062 if (test_bit(slab_index(p, s, addr), map))
4063 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4067 for_each_object(p, s, addr, page->objects)
4068 if (!test_bit(slab_index(p, s, addr), map))
4069 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4074 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4078 validate_slab(s, page, map);
4082 static int validate_slab_node(struct kmem_cache *s,
4083 struct kmem_cache_node *n, unsigned long *map)
4085 unsigned long count = 0;
4087 unsigned long flags;
4089 spin_lock_irqsave(&n->list_lock, flags);
4091 list_for_each_entry(page, &n->partial, lru) {
4092 validate_slab_slab(s, page, map);
4095 if (count != n->nr_partial)
4096 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4097 s->name, count, n->nr_partial);
4099 if (!(s->flags & SLAB_STORE_USER))
4102 list_for_each_entry(page, &n->full, lru) {
4103 validate_slab_slab(s, page, map);
4106 if (count != atomic_long_read(&n->nr_slabs))
4107 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4108 s->name, count, atomic_long_read(&n->nr_slabs));
4111 spin_unlock_irqrestore(&n->list_lock, flags);
4115 static long validate_slab_cache(struct kmem_cache *s)
4118 unsigned long count = 0;
4119 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4120 sizeof(unsigned long), GFP_KERNEL);
4121 struct kmem_cache_node *n;
4127 for_each_kmem_cache_node(s, node, n)
4128 count += validate_slab_node(s, n, map);
4133 * Generate lists of code addresses where slabcache objects are allocated
4138 unsigned long count;
4145 DECLARE_BITMAP(cpus, NR_CPUS);
4151 unsigned long count;
4152 struct location *loc;
4155 static void free_loc_track(struct loc_track *t)
4158 free_pages((unsigned long)t->loc,
4159 get_order(sizeof(struct location) * t->max));
4162 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4167 order = get_order(sizeof(struct location) * max);
4169 l = (void *)__get_free_pages(flags, order);
4174 memcpy(l, t->loc, sizeof(struct location) * t->count);
4182 static int add_location(struct loc_track *t, struct kmem_cache *s,
4183 const struct track *track)
4185 long start, end, pos;
4187 unsigned long caddr;
4188 unsigned long age = jiffies - track->when;
4194 pos = start + (end - start + 1) / 2;
4197 * There is nothing at "end". If we end up there
4198 * we need to add something to before end.
4203 caddr = t->loc[pos].addr;
4204 if (track->addr == caddr) {
4210 if (age < l->min_time)
4212 if (age > l->max_time)
4215 if (track->pid < l->min_pid)
4216 l->min_pid = track->pid;
4217 if (track->pid > l->max_pid)
4218 l->max_pid = track->pid;
4220 cpumask_set_cpu(track->cpu,
4221 to_cpumask(l->cpus));
4223 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4227 if (track->addr < caddr)
4234 * Not found. Insert new tracking element.
4236 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4242 (t->count - pos) * sizeof(struct location));
4245 l->addr = track->addr;
4249 l->min_pid = track->pid;
4250 l->max_pid = track->pid;
4251 cpumask_clear(to_cpumask(l->cpus));
4252 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4253 nodes_clear(l->nodes);
4254 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4258 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4259 struct page *page, enum track_item alloc,
4262 void *addr = page_address(page);
4265 bitmap_zero(map, page->objects);
4266 get_map(s, page, map);
4268 for_each_object(p, s, addr, page->objects)
4269 if (!test_bit(slab_index(p, s, addr), map))
4270 add_location(t, s, get_track(s, p, alloc));
4273 static int list_locations(struct kmem_cache *s, char *buf,
4274 enum track_item alloc)
4278 struct loc_track t = { 0, 0, NULL };
4280 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4281 sizeof(unsigned long), GFP_KERNEL);
4282 struct kmem_cache_node *n;
4284 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4287 return sprintf(buf, "Out of memory\n");
4289 /* Push back cpu slabs */
4292 for_each_kmem_cache_node(s, node, n) {
4293 unsigned long flags;
4296 if (!atomic_long_read(&n->nr_slabs))
4299 spin_lock_irqsave(&n->list_lock, flags);
4300 list_for_each_entry(page, &n->partial, lru)
4301 process_slab(&t, s, page, alloc, map);
4302 list_for_each_entry(page, &n->full, lru)
4303 process_slab(&t, s, page, alloc, map);
4304 spin_unlock_irqrestore(&n->list_lock, flags);
4307 for (i = 0; i < t.count; i++) {
4308 struct location *l = &t.loc[i];
4310 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4312 len += sprintf(buf + len, "%7ld ", l->count);
4315 len += sprintf(buf + len, "%pS", (void *)l->addr);
4317 len += sprintf(buf + len, "<not-available>");
4319 if (l->sum_time != l->min_time) {
4320 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4322 (long)div_u64(l->sum_time, l->count),
4325 len += sprintf(buf + len, " age=%ld",
4328 if (l->min_pid != l->max_pid)
4329 len += sprintf(buf + len, " pid=%ld-%ld",
4330 l->min_pid, l->max_pid);
4332 len += sprintf(buf + len, " pid=%ld",
4335 if (num_online_cpus() > 1 &&
4336 !cpumask_empty(to_cpumask(l->cpus)) &&
4337 len < PAGE_SIZE - 60)
4338 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4340 cpumask_pr_args(to_cpumask(l->cpus)));
4342 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4343 len < PAGE_SIZE - 60)
4344 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4346 nodemask_pr_args(&l->nodes));
4348 len += sprintf(buf + len, "\n");
4354 len += sprintf(buf, "No data\n");
4359 #ifdef SLUB_RESILIENCY_TEST
4360 static void __init resiliency_test(void)
4364 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4366 pr_err("SLUB resiliency testing\n");
4367 pr_err("-----------------------\n");
4368 pr_err("A. Corruption after allocation\n");
4370 p = kzalloc(16, GFP_KERNEL);
4372 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4375 validate_slab_cache(kmalloc_caches[4]);
4377 /* Hmmm... The next two are dangerous */
4378 p = kzalloc(32, GFP_KERNEL);
4379 p[32 + sizeof(void *)] = 0x34;
4380 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4382 pr_err("If allocated object is overwritten then not detectable\n\n");
4384 validate_slab_cache(kmalloc_caches[5]);
4385 p = kzalloc(64, GFP_KERNEL);
4386 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4388 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4390 pr_err("If allocated object is overwritten then not detectable\n\n");
4391 validate_slab_cache(kmalloc_caches[6]);
4393 pr_err("\nB. Corruption after free\n");
4394 p = kzalloc(128, GFP_KERNEL);
4397 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4398 validate_slab_cache(kmalloc_caches[7]);
4400 p = kzalloc(256, GFP_KERNEL);
4403 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4404 validate_slab_cache(kmalloc_caches[8]);
4406 p = kzalloc(512, GFP_KERNEL);
4409 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4410 validate_slab_cache(kmalloc_caches[9]);
4414 static void resiliency_test(void) {};
4419 enum slab_stat_type {
4420 SL_ALL, /* All slabs */
4421 SL_PARTIAL, /* Only partially allocated slabs */
4422 SL_CPU, /* Only slabs used for cpu caches */
4423 SL_OBJECTS, /* Determine allocated objects not slabs */
4424 SL_TOTAL /* Determine object capacity not slabs */
4427 #define SO_ALL (1 << SL_ALL)
4428 #define SO_PARTIAL (1 << SL_PARTIAL)
4429 #define SO_CPU (1 << SL_CPU)
4430 #define SO_OBJECTS (1 << SL_OBJECTS)
4431 #define SO_TOTAL (1 << SL_TOTAL)
4433 static ssize_t show_slab_objects(struct kmem_cache *s,
4434 char *buf, unsigned long flags)
4436 unsigned long total = 0;
4439 unsigned long *nodes;
4441 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4445 if (flags & SO_CPU) {
4448 for_each_possible_cpu(cpu) {
4449 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4454 page = READ_ONCE(c->page);
4458 node = page_to_nid(page);
4459 if (flags & SO_TOTAL)
4461 else if (flags & SO_OBJECTS)
4469 page = READ_ONCE(c->partial);
4471 node = page_to_nid(page);
4472 if (flags & SO_TOTAL)
4474 else if (flags & SO_OBJECTS)
4485 #ifdef CONFIG_SLUB_DEBUG
4486 if (flags & SO_ALL) {
4487 struct kmem_cache_node *n;
4489 for_each_kmem_cache_node(s, node, n) {
4491 if (flags & SO_TOTAL)
4492 x = atomic_long_read(&n->total_objects);
4493 else if (flags & SO_OBJECTS)
4494 x = atomic_long_read(&n->total_objects) -
4495 count_partial(n, count_free);
4497 x = atomic_long_read(&n->nr_slabs);
4504 if (flags & SO_PARTIAL) {
4505 struct kmem_cache_node *n;
4507 for_each_kmem_cache_node(s, node, n) {
4508 if (flags & SO_TOTAL)
4509 x = count_partial(n, count_total);
4510 else if (flags & SO_OBJECTS)
4511 x = count_partial(n, count_inuse);
4518 x = sprintf(buf, "%lu", total);
4520 for (node = 0; node < nr_node_ids; node++)
4522 x += sprintf(buf + x, " N%d=%lu",
4527 return x + sprintf(buf + x, "\n");
4530 #ifdef CONFIG_SLUB_DEBUG
4531 static int any_slab_objects(struct kmem_cache *s)
4534 struct kmem_cache_node *n;
4536 for_each_kmem_cache_node(s, node, n)
4537 if (atomic_long_read(&n->total_objects))
4544 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4545 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4547 struct slab_attribute {
4548 struct attribute attr;
4549 ssize_t (*show)(struct kmem_cache *s, char *buf);
4550 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4553 #define SLAB_ATTR_RO(_name) \
4554 static struct slab_attribute _name##_attr = \
4555 __ATTR(_name, 0400, _name##_show, NULL)
4557 #define SLAB_ATTR(_name) \
4558 static struct slab_attribute _name##_attr = \
4559 __ATTR(_name, 0600, _name##_show, _name##_store)
4561 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4563 return sprintf(buf, "%d\n", s->size);
4565 SLAB_ATTR_RO(slab_size);
4567 static ssize_t align_show(struct kmem_cache *s, char *buf)
4569 return sprintf(buf, "%d\n", s->align);
4571 SLAB_ATTR_RO(align);
4573 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4575 return sprintf(buf, "%d\n", s->object_size);
4577 SLAB_ATTR_RO(object_size);
4579 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4581 return sprintf(buf, "%d\n", oo_objects(s->oo));
4583 SLAB_ATTR_RO(objs_per_slab);
4585 static ssize_t order_store(struct kmem_cache *s,
4586 const char *buf, size_t length)
4588 unsigned long order;
4591 err = kstrtoul(buf, 10, &order);
4595 if (order > slub_max_order || order < slub_min_order)
4598 calculate_sizes(s, order);
4602 static ssize_t order_show(struct kmem_cache *s, char *buf)
4604 return sprintf(buf, "%d\n", oo_order(s->oo));
4608 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4610 return sprintf(buf, "%lu\n", s->min_partial);
4613 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4619 err = kstrtoul(buf, 10, &min);
4623 set_min_partial(s, min);
4626 SLAB_ATTR(min_partial);
4628 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4630 return sprintf(buf, "%u\n", s->cpu_partial);
4633 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4636 unsigned long objects;
4639 err = kstrtoul(buf, 10, &objects);
4642 if (objects && !kmem_cache_has_cpu_partial(s))
4645 s->cpu_partial = objects;
4649 SLAB_ATTR(cpu_partial);
4651 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4655 return sprintf(buf, "%pS\n", s->ctor);
4659 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4661 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4663 SLAB_ATTR_RO(aliases);
4665 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4667 return show_slab_objects(s, buf, SO_PARTIAL);
4669 SLAB_ATTR_RO(partial);
4671 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4673 return show_slab_objects(s, buf, SO_CPU);
4675 SLAB_ATTR_RO(cpu_slabs);
4677 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4679 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4681 SLAB_ATTR_RO(objects);
4683 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4685 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4687 SLAB_ATTR_RO(objects_partial);
4689 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4696 for_each_online_cpu(cpu) {
4697 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4700 pages += page->pages;
4701 objects += page->pobjects;
4705 len = sprintf(buf, "%d(%d)", objects, pages);
4708 for_each_online_cpu(cpu) {
4709 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4711 if (page && len < PAGE_SIZE - 20)
4712 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4713 page->pobjects, page->pages);
4716 return len + sprintf(buf + len, "\n");
4718 SLAB_ATTR_RO(slabs_cpu_partial);
4720 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4722 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4725 static ssize_t reclaim_account_store(struct kmem_cache *s,
4726 const char *buf, size_t length)
4728 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4730 s->flags |= SLAB_RECLAIM_ACCOUNT;
4733 SLAB_ATTR(reclaim_account);
4735 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4737 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4739 SLAB_ATTR_RO(hwcache_align);
4741 #ifdef CONFIG_ZONE_DMA
4742 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4744 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4746 SLAB_ATTR_RO(cache_dma);
4749 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4751 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4753 SLAB_ATTR_RO(destroy_by_rcu);
4755 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4757 return sprintf(buf, "%d\n", s->reserved);
4759 SLAB_ATTR_RO(reserved);
4761 #ifdef CONFIG_SLUB_DEBUG
4762 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4764 return show_slab_objects(s, buf, SO_ALL);
4766 SLAB_ATTR_RO(slabs);
4768 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4770 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4772 SLAB_ATTR_RO(total_objects);
4774 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4776 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4779 static ssize_t sanity_checks_store(struct kmem_cache *s,
4780 const char *buf, size_t length)
4782 s->flags &= ~SLAB_DEBUG_FREE;
4783 if (buf[0] == '1') {
4784 s->flags &= ~__CMPXCHG_DOUBLE;
4785 s->flags |= SLAB_DEBUG_FREE;
4789 SLAB_ATTR(sanity_checks);
4791 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4793 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4796 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4800 * Tracing a merged cache is going to give confusing results
4801 * as well as cause other issues like converting a mergeable
4802 * cache into an umergeable one.
4804 if (s->refcount > 1)
4807 s->flags &= ~SLAB_TRACE;
4808 if (buf[0] == '1') {
4809 s->flags &= ~__CMPXCHG_DOUBLE;
4810 s->flags |= SLAB_TRACE;
4816 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4818 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4821 static ssize_t red_zone_store(struct kmem_cache *s,
4822 const char *buf, size_t length)
4824 if (any_slab_objects(s))
4827 s->flags &= ~SLAB_RED_ZONE;
4828 if (buf[0] == '1') {
4829 s->flags &= ~__CMPXCHG_DOUBLE;
4830 s->flags |= SLAB_RED_ZONE;
4832 calculate_sizes(s, -1);
4835 SLAB_ATTR(red_zone);
4837 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4839 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4842 static ssize_t poison_store(struct kmem_cache *s,
4843 const char *buf, size_t length)
4845 if (any_slab_objects(s))
4848 s->flags &= ~SLAB_POISON;
4849 if (buf[0] == '1') {
4850 s->flags &= ~__CMPXCHG_DOUBLE;
4851 s->flags |= SLAB_POISON;
4853 calculate_sizes(s, -1);
4858 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4860 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4863 static ssize_t store_user_store(struct kmem_cache *s,
4864 const char *buf, size_t length)
4866 if (any_slab_objects(s))
4869 s->flags &= ~SLAB_STORE_USER;
4870 if (buf[0] == '1') {
4871 s->flags &= ~__CMPXCHG_DOUBLE;
4872 s->flags |= SLAB_STORE_USER;
4874 calculate_sizes(s, -1);
4877 SLAB_ATTR(store_user);
4879 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4884 static ssize_t validate_store(struct kmem_cache *s,
4885 const char *buf, size_t length)
4889 if (buf[0] == '1') {
4890 ret = validate_slab_cache(s);
4896 SLAB_ATTR(validate);
4898 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4900 if (!(s->flags & SLAB_STORE_USER))
4902 return list_locations(s, buf, TRACK_ALLOC);
4904 SLAB_ATTR_RO(alloc_calls);
4906 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4908 if (!(s->flags & SLAB_STORE_USER))
4910 return list_locations(s, buf, TRACK_FREE);
4912 SLAB_ATTR_RO(free_calls);
4913 #endif /* CONFIG_SLUB_DEBUG */
4915 #ifdef CONFIG_FAILSLAB
4916 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4918 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4921 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4924 if (s->refcount > 1)
4927 s->flags &= ~SLAB_FAILSLAB;
4929 s->flags |= SLAB_FAILSLAB;
4932 SLAB_ATTR(failslab);
4935 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4940 static ssize_t shrink_store(struct kmem_cache *s,
4941 const char *buf, size_t length)
4944 kmem_cache_shrink(s);
4952 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4954 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4957 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4958 const char *buf, size_t length)
4960 unsigned long ratio;
4963 err = kstrtoul(buf, 10, &ratio);
4968 s->remote_node_defrag_ratio = ratio * 10;
4972 SLAB_ATTR(remote_node_defrag_ratio);
4975 #ifdef CONFIG_SLUB_STATS
4976 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4978 unsigned long sum = 0;
4981 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4986 for_each_online_cpu(cpu) {
4987 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4993 len = sprintf(buf, "%lu", sum);
4996 for_each_online_cpu(cpu) {
4997 if (data[cpu] && len < PAGE_SIZE - 20)
4998 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5002 return len + sprintf(buf + len, "\n");
5005 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5009 for_each_online_cpu(cpu)
5010 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5013 #define STAT_ATTR(si, text) \
5014 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5016 return show_stat(s, buf, si); \
5018 static ssize_t text##_store(struct kmem_cache *s, \
5019 const char *buf, size_t length) \
5021 if (buf[0] != '0') \
5023 clear_stat(s, si); \
5028 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5029 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5030 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5031 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5032 STAT_ATTR(FREE_FROZEN, free_frozen);
5033 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5034 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5035 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5036 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5037 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5038 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5039 STAT_ATTR(FREE_SLAB, free_slab);
5040 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5041 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5042 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5043 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5044 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5045 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5046 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5047 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5048 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5049 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5050 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5051 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5052 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5053 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5056 static struct attribute *slab_attrs[] = {
5057 &slab_size_attr.attr,
5058 &object_size_attr.attr,
5059 &objs_per_slab_attr.attr,
5061 &min_partial_attr.attr,
5062 &cpu_partial_attr.attr,
5064 &objects_partial_attr.attr,
5066 &cpu_slabs_attr.attr,
5070 &hwcache_align_attr.attr,
5071 &reclaim_account_attr.attr,
5072 &destroy_by_rcu_attr.attr,
5074 &reserved_attr.attr,
5075 &slabs_cpu_partial_attr.attr,
5076 #ifdef CONFIG_SLUB_DEBUG
5077 &total_objects_attr.attr,
5079 &sanity_checks_attr.attr,
5081 &red_zone_attr.attr,
5083 &store_user_attr.attr,
5084 &validate_attr.attr,
5085 &alloc_calls_attr.attr,
5086 &free_calls_attr.attr,
5088 #ifdef CONFIG_ZONE_DMA
5089 &cache_dma_attr.attr,
5092 &remote_node_defrag_ratio_attr.attr,
5094 #ifdef CONFIG_SLUB_STATS
5095 &alloc_fastpath_attr.attr,
5096 &alloc_slowpath_attr.attr,
5097 &free_fastpath_attr.attr,
5098 &free_slowpath_attr.attr,
5099 &free_frozen_attr.attr,
5100 &free_add_partial_attr.attr,
5101 &free_remove_partial_attr.attr,
5102 &alloc_from_partial_attr.attr,
5103 &alloc_slab_attr.attr,
5104 &alloc_refill_attr.attr,
5105 &alloc_node_mismatch_attr.attr,
5106 &free_slab_attr.attr,
5107 &cpuslab_flush_attr.attr,
5108 &deactivate_full_attr.attr,
5109 &deactivate_empty_attr.attr,
5110 &deactivate_to_head_attr.attr,
5111 &deactivate_to_tail_attr.attr,
5112 &deactivate_remote_frees_attr.attr,
5113 &deactivate_bypass_attr.attr,
5114 &order_fallback_attr.attr,
5115 &cmpxchg_double_fail_attr.attr,
5116 &cmpxchg_double_cpu_fail_attr.attr,
5117 &cpu_partial_alloc_attr.attr,
5118 &cpu_partial_free_attr.attr,
5119 &cpu_partial_node_attr.attr,
5120 &cpu_partial_drain_attr.attr,
5122 #ifdef CONFIG_FAILSLAB
5123 &failslab_attr.attr,
5129 static struct attribute_group slab_attr_group = {
5130 .attrs = slab_attrs,
5133 static ssize_t slab_attr_show(struct kobject *kobj,
5134 struct attribute *attr,
5137 struct slab_attribute *attribute;
5138 struct kmem_cache *s;
5141 attribute = to_slab_attr(attr);
5144 if (!attribute->show)
5147 err = attribute->show(s, buf);
5152 static ssize_t slab_attr_store(struct kobject *kobj,
5153 struct attribute *attr,
5154 const char *buf, size_t len)
5156 struct slab_attribute *attribute;
5157 struct kmem_cache *s;
5160 attribute = to_slab_attr(attr);
5163 if (!attribute->store)
5166 err = attribute->store(s, buf, len);
5168 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5169 struct kmem_cache *c;
5171 mutex_lock(&slab_mutex);
5172 if (s->max_attr_size < len)
5173 s->max_attr_size = len;
5176 * This is a best effort propagation, so this function's return
5177 * value will be determined by the parent cache only. This is
5178 * basically because not all attributes will have a well
5179 * defined semantics for rollbacks - most of the actions will
5180 * have permanent effects.
5182 * Returning the error value of any of the children that fail
5183 * is not 100 % defined, in the sense that users seeing the
5184 * error code won't be able to know anything about the state of
5187 * Only returning the error code for the parent cache at least
5188 * has well defined semantics. The cache being written to
5189 * directly either failed or succeeded, in which case we loop
5190 * through the descendants with best-effort propagation.
5192 for_each_memcg_cache(c, s)
5193 attribute->store(c, buf, len);
5194 mutex_unlock(&slab_mutex);
5200 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5204 char *buffer = NULL;
5205 struct kmem_cache *root_cache;
5207 if (is_root_cache(s))
5210 root_cache = s->memcg_params.root_cache;
5213 * This mean this cache had no attribute written. Therefore, no point
5214 * in copying default values around
5216 if (!root_cache->max_attr_size)
5219 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5222 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5224 if (!attr || !attr->store || !attr->show)
5228 * It is really bad that we have to allocate here, so we will
5229 * do it only as a fallback. If we actually allocate, though,
5230 * we can just use the allocated buffer until the end.
5232 * Most of the slub attributes will tend to be very small in
5233 * size, but sysfs allows buffers up to a page, so they can
5234 * theoretically happen.
5238 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5241 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5242 if (WARN_ON(!buffer))
5247 attr->show(root_cache, buf);
5248 attr->store(s, buf, strlen(buf));
5252 free_page((unsigned long)buffer);
5256 static void kmem_cache_release(struct kobject *k)
5258 slab_kmem_cache_release(to_slab(k));
5261 static const struct sysfs_ops slab_sysfs_ops = {
5262 .show = slab_attr_show,
5263 .store = slab_attr_store,
5266 static struct kobj_type slab_ktype = {
5267 .sysfs_ops = &slab_sysfs_ops,
5268 .release = kmem_cache_release,
5271 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5273 struct kobj_type *ktype = get_ktype(kobj);
5275 if (ktype == &slab_ktype)
5280 static const struct kset_uevent_ops slab_uevent_ops = {
5281 .filter = uevent_filter,
5284 static struct kset *slab_kset;
5286 static inline struct kset *cache_kset(struct kmem_cache *s)
5289 if (!is_root_cache(s))
5290 return s->memcg_params.root_cache->memcg_kset;
5295 #define ID_STR_LENGTH 64
5297 /* Create a unique string id for a slab cache:
5299 * Format :[flags-]size
5301 static char *create_unique_id(struct kmem_cache *s)
5303 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5310 * First flags affecting slabcache operations. We will only
5311 * get here for aliasable slabs so we do not need to support
5312 * too many flags. The flags here must cover all flags that
5313 * are matched during merging to guarantee that the id is
5316 if (s->flags & SLAB_CACHE_DMA)
5318 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5320 if (s->flags & SLAB_DEBUG_FREE)
5322 if (!(s->flags & SLAB_NOTRACK))
5324 if (s->flags & SLAB_ACCOUNT)
5328 p += sprintf(p, "%07d", s->size);
5330 BUG_ON(p > name + ID_STR_LENGTH - 1);
5334 static int sysfs_slab_add(struct kmem_cache *s)
5338 int unmergeable = slab_unmergeable(s);
5342 * Slabcache can never be merged so we can use the name proper.
5343 * This is typically the case for debug situations. In that
5344 * case we can catch duplicate names easily.
5346 sysfs_remove_link(&slab_kset->kobj, s->name);
5350 * Create a unique name for the slab as a target
5353 name = create_unique_id(s);
5356 s->kobj.kset = cache_kset(s);
5357 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5361 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5366 if (is_root_cache(s)) {
5367 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5368 if (!s->memcg_kset) {
5375 kobject_uevent(&s->kobj, KOBJ_ADD);
5377 /* Setup first alias */
5378 sysfs_slab_alias(s, s->name);
5385 kobject_del(&s->kobj);
5389 void sysfs_slab_remove(struct kmem_cache *s)
5391 if (slab_state < FULL)
5393 * Sysfs has not been setup yet so no need to remove the
5399 kset_unregister(s->memcg_kset);
5401 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5402 kobject_del(&s->kobj);
5403 kobject_put(&s->kobj);
5407 * Need to buffer aliases during bootup until sysfs becomes
5408 * available lest we lose that information.
5410 struct saved_alias {
5411 struct kmem_cache *s;
5413 struct saved_alias *next;
5416 static struct saved_alias *alias_list;
5418 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5420 struct saved_alias *al;
5422 if (slab_state == FULL) {
5424 * If we have a leftover link then remove it.
5426 sysfs_remove_link(&slab_kset->kobj, name);
5427 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5430 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5436 al->next = alias_list;
5441 static int __init slab_sysfs_init(void)
5443 struct kmem_cache *s;
5446 mutex_lock(&slab_mutex);
5448 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5450 mutex_unlock(&slab_mutex);
5451 pr_err("Cannot register slab subsystem.\n");
5457 list_for_each_entry(s, &slab_caches, list) {
5458 err = sysfs_slab_add(s);
5460 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5464 while (alias_list) {
5465 struct saved_alias *al = alias_list;
5467 alias_list = alias_list->next;
5468 err = sysfs_slab_alias(al->s, al->name);
5470 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5475 mutex_unlock(&slab_mutex);
5480 __initcall(slab_sysfs_init);
5481 #endif /* CONFIG_SYSFS */
5484 * The /proc/slabinfo ABI
5486 #ifdef CONFIG_SLABINFO
5487 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5489 unsigned long nr_slabs = 0;
5490 unsigned long nr_objs = 0;
5491 unsigned long nr_free = 0;
5493 struct kmem_cache_node *n;
5495 for_each_kmem_cache_node(s, node, n) {
5496 nr_slabs += node_nr_slabs(n);
5497 nr_objs += node_nr_objs(n);
5498 nr_free += count_partial(n, count_free);
5501 sinfo->active_objs = nr_objs - nr_free;
5502 sinfo->num_objs = nr_objs;
5503 sinfo->active_slabs = nr_slabs;
5504 sinfo->num_slabs = nr_slabs;
5505 sinfo->objects_per_slab = oo_objects(s->oo);
5506 sinfo->cache_order = oo_order(s->oo);
5509 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5513 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5514 size_t count, loff_t *ppos)
5518 #endif /* CONFIG_SLABINFO */