2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
219 raw_cpu_inc(s->cpu_slab->stat[si]);
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
236 base = page_address(page);
237 if (object < base || object >= base + page->objects * s->size ||
238 (object - base) % s->size) {
245 static inline void *get_freepointer(struct kmem_cache *s, void *object)
247 return *(void **)(object + s->offset);
250 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
252 prefetch(object + s->offset);
255 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
262 p = get_freepointer(s, object);
267 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
269 *(void **)(object + s->offset) = fp;
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 bit_spin_lock(PG_locked, &page->flags);
344 static __always_inline void slab_unlock(struct page *page)
346 __bit_spin_unlock(PG_locked, &page->flags);
349 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
352 tmp.counters = counters_new;
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
359 page->frozen = tmp.frozen;
360 page->inuse = tmp.inuse;
361 page->objects = tmp.objects;
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 set_page_slub_counters(page, counters_new);
393 stat(s, CMPXCHG_DOUBLE_FAIL);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
419 local_irq_save(flags);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 set_page_slub_counters(page, counters_new);
426 local_irq_restore(flags);
430 local_irq_restore(flags);
434 stat(s, CMPXCHG_DOUBLE_FAIL);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 void *addr = page_address(page);
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
462 #if defined(CONFIG_SLUB_DEBUG_ON)
463 static int slub_debug = DEBUG_DEFAULT_FLAGS;
464 #elif defined(CONFIG_KASAN)
465 static int slub_debug = SLAB_STORE_USER;
467 static int slub_debug;
470 static char *slub_debug_slabs;
471 static int disable_higher_order_debug;
474 * slub is about to manipulate internal object metadata. This memory lies
475 * outside the range of the allocated object, so accessing it would normally
476 * be reported by kasan as a bounds error. metadata_access_enable() is used
477 * to tell kasan that these accesses are OK.
479 static inline void metadata_access_enable(void)
481 kasan_disable_current();
484 static inline void metadata_access_disable(void)
486 kasan_enable_current();
492 static void print_section(char *text, u8 *addr, unsigned int length)
494 metadata_access_enable();
495 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
497 metadata_access_disable();
500 static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
506 p = object + s->offset + sizeof(void *);
508 p = object + s->inuse;
513 static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
516 struct track *p = get_track(s, object, alloc);
519 #ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
527 metadata_access_enable();
528 save_stack_trace(&trace);
529 metadata_access_disable();
531 /* See rant in lockdep.c */
532 if (trace.nr_entries != 0 &&
533 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
536 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
540 p->cpu = smp_processor_id();
541 p->pid = current->pid;
544 memset(p, 0, sizeof(struct track));
547 static void init_tracking(struct kmem_cache *s, void *object)
549 if (!(s->flags & SLAB_STORE_USER))
552 set_track(s, object, TRACK_FREE, 0UL);
553 set_track(s, object, TRACK_ALLOC, 0UL);
556 static void print_track(const char *s, struct track *t)
561 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
562 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
563 #ifdef CONFIG_STACKTRACE
566 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
568 pr_err("\t%pS\n", (void *)t->addrs[i]);
575 static void print_tracking(struct kmem_cache *s, void *object)
577 if (!(s->flags & SLAB_STORE_USER))
580 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
581 print_track("Freed", get_track(s, object, TRACK_FREE));
584 static void print_page_info(struct page *page)
586 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
587 page, page->objects, page->inuse, page->freelist, page->flags);
591 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
593 struct va_format vaf;
599 pr_err("=============================================================================\n");
600 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
601 pr_err("-----------------------------------------------------------------------------\n\n");
603 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
607 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
609 struct va_format vaf;
615 pr_err("FIX %s: %pV\n", s->name, &vaf);
619 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
621 unsigned int off; /* Offset of last byte */
622 u8 *addr = page_address(page);
624 print_tracking(s, p);
626 print_page_info(page);
628 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
629 p, p - addr, get_freepointer(s, p));
632 print_section("Bytes b4 ", p - 16, 16);
634 print_section("Object ", p, min_t(unsigned long, s->object_size,
636 if (s->flags & SLAB_RED_ZONE)
637 print_section("Redzone ", p + s->object_size,
638 s->inuse - s->object_size);
641 off = s->offset + sizeof(void *);
645 if (s->flags & SLAB_STORE_USER)
646 off += 2 * sizeof(struct track);
649 /* Beginning of the filler is the free pointer */
650 print_section("Padding ", p + off, s->size - off);
655 void object_err(struct kmem_cache *s, struct page *page,
656 u8 *object, char *reason)
658 slab_bug(s, "%s", reason);
659 print_trailer(s, page, object);
662 static void slab_err(struct kmem_cache *s, struct page *page,
663 const char *fmt, ...)
669 vsnprintf(buf, sizeof(buf), fmt, args);
671 slab_bug(s, "%s", buf);
672 print_page_info(page);
676 static void init_object(struct kmem_cache *s, void *object, u8 val)
680 if (s->flags & __OBJECT_POISON) {
681 memset(p, POISON_FREE, s->object_size - 1);
682 p[s->object_size - 1] = POISON_END;
685 if (s->flags & SLAB_RED_ZONE)
686 memset(p + s->object_size, val, s->inuse - s->object_size);
689 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
690 void *from, void *to)
692 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
693 memset(from, data, to - from);
696 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
697 u8 *object, char *what,
698 u8 *start, unsigned int value, unsigned int bytes)
703 metadata_access_enable();
704 fault = memchr_inv(start, value, bytes);
705 metadata_access_disable();
710 while (end > fault && end[-1] == value)
713 slab_bug(s, "%s overwritten", what);
714 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
715 fault, end - 1, fault[0], value);
716 print_trailer(s, page, object);
718 restore_bytes(s, what, value, fault, end);
726 * Bytes of the object to be managed.
727 * If the freepointer may overlay the object then the free
728 * pointer is the first word of the object.
730 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
733 * object + s->object_size
734 * Padding to reach word boundary. This is also used for Redzoning.
735 * Padding is extended by another word if Redzoning is enabled and
736 * object_size == inuse.
738 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
739 * 0xcc (RED_ACTIVE) for objects in use.
742 * Meta data starts here.
744 * A. Free pointer (if we cannot overwrite object on free)
745 * B. Tracking data for SLAB_STORE_USER
746 * C. Padding to reach required alignment boundary or at mininum
747 * one word if debugging is on to be able to detect writes
748 * before the word boundary.
750 * Padding is done using 0x5a (POISON_INUSE)
753 * Nothing is used beyond s->size.
755 * If slabcaches are merged then the object_size and inuse boundaries are mostly
756 * ignored. And therefore no slab options that rely on these boundaries
757 * may be used with merged slabcaches.
760 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
762 unsigned long off = s->inuse; /* The end of info */
765 /* Freepointer is placed after the object. */
766 off += sizeof(void *);
768 if (s->flags & SLAB_STORE_USER)
769 /* We also have user information there */
770 off += 2 * sizeof(struct track);
775 return check_bytes_and_report(s, page, p, "Object padding",
776 p + off, POISON_INUSE, s->size - off);
779 /* Check the pad bytes at the end of a slab page */
780 static int slab_pad_check(struct kmem_cache *s, struct page *page)
788 if (!(s->flags & SLAB_POISON))
791 start = page_address(page);
792 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
793 end = start + length;
794 remainder = length % s->size;
798 metadata_access_enable();
799 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
800 metadata_access_disable();
803 while (end > fault && end[-1] == POISON_INUSE)
806 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
807 print_section("Padding ", end - remainder, remainder);
809 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
813 static int check_object(struct kmem_cache *s, struct page *page,
814 void *object, u8 val)
817 u8 *endobject = object + s->object_size;
819 if (s->flags & SLAB_RED_ZONE) {
820 if (!check_bytes_and_report(s, page, object, "Redzone",
821 endobject, val, s->inuse - s->object_size))
824 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
825 check_bytes_and_report(s, page, p, "Alignment padding",
826 endobject, POISON_INUSE,
827 s->inuse - s->object_size);
831 if (s->flags & SLAB_POISON) {
832 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
833 (!check_bytes_and_report(s, page, p, "Poison", p,
834 POISON_FREE, s->object_size - 1) ||
835 !check_bytes_and_report(s, page, p, "Poison",
836 p + s->object_size - 1, POISON_END, 1)))
839 * check_pad_bytes cleans up on its own.
841 check_pad_bytes(s, page, p);
844 if (!s->offset && val == SLUB_RED_ACTIVE)
846 * Object and freepointer overlap. Cannot check
847 * freepointer while object is allocated.
851 /* Check free pointer validity */
852 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
853 object_err(s, page, p, "Freepointer corrupt");
855 * No choice but to zap it and thus lose the remainder
856 * of the free objects in this slab. May cause
857 * another error because the object count is now wrong.
859 set_freepointer(s, p, NULL);
865 static int check_slab(struct kmem_cache *s, struct page *page)
869 VM_BUG_ON(!irqs_disabled());
871 if (!PageSlab(page)) {
872 slab_err(s, page, "Not a valid slab page");
876 maxobj = order_objects(compound_order(page), s->size, s->reserved);
877 if (page->objects > maxobj) {
878 slab_err(s, page, "objects %u > max %u",
879 page->objects, maxobj);
882 if (page->inuse > page->objects) {
883 slab_err(s, page, "inuse %u > max %u",
884 page->inuse, page->objects);
887 /* Slab_pad_check fixes things up after itself */
888 slab_pad_check(s, page);
893 * Determine if a certain object on a page is on the freelist. Must hold the
894 * slab lock to guarantee that the chains are in a consistent state.
896 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
904 while (fp && nr <= page->objects) {
907 if (!check_valid_pointer(s, page, fp)) {
909 object_err(s, page, object,
910 "Freechain corrupt");
911 set_freepointer(s, object, NULL);
913 slab_err(s, page, "Freepointer corrupt");
914 page->freelist = NULL;
915 page->inuse = page->objects;
916 slab_fix(s, "Freelist cleared");
922 fp = get_freepointer(s, object);
926 max_objects = order_objects(compound_order(page), s->size, s->reserved);
927 if (max_objects > MAX_OBJS_PER_PAGE)
928 max_objects = MAX_OBJS_PER_PAGE;
930 if (page->objects != max_objects) {
931 slab_err(s, page, "Wrong number of objects. Found %d but "
932 "should be %d", page->objects, max_objects);
933 page->objects = max_objects;
934 slab_fix(s, "Number of objects adjusted.");
936 if (page->inuse != page->objects - nr) {
937 slab_err(s, page, "Wrong object count. Counter is %d but "
938 "counted were %d", page->inuse, page->objects - nr);
939 page->inuse = page->objects - nr;
940 slab_fix(s, "Object count adjusted.");
942 return search == NULL;
945 static void trace(struct kmem_cache *s, struct page *page, void *object,
948 if (s->flags & SLAB_TRACE) {
949 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
951 alloc ? "alloc" : "free",
956 print_section("Object ", (void *)object,
964 * Tracking of fully allocated slabs for debugging purposes.
966 static void add_full(struct kmem_cache *s,
967 struct kmem_cache_node *n, struct page *page)
969 if (!(s->flags & SLAB_STORE_USER))
972 lockdep_assert_held(&n->list_lock);
973 list_add(&page->lru, &n->full);
976 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
978 if (!(s->flags & SLAB_STORE_USER))
981 lockdep_assert_held(&n->list_lock);
982 list_del(&page->lru);
985 /* Tracking of the number of slabs for debugging purposes */
986 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
988 struct kmem_cache_node *n = get_node(s, node);
990 return atomic_long_read(&n->nr_slabs);
993 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
995 return atomic_long_read(&n->nr_slabs);
998 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1000 struct kmem_cache_node *n = get_node(s, node);
1003 * May be called early in order to allocate a slab for the
1004 * kmem_cache_node structure. Solve the chicken-egg
1005 * dilemma by deferring the increment of the count during
1006 * bootstrap (see early_kmem_cache_node_alloc).
1009 atomic_long_inc(&n->nr_slabs);
1010 atomic_long_add(objects, &n->total_objects);
1013 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1015 struct kmem_cache_node *n = get_node(s, node);
1017 atomic_long_dec(&n->nr_slabs);
1018 atomic_long_sub(objects, &n->total_objects);
1021 /* Object debug checks for alloc/free paths */
1022 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1025 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1028 init_object(s, object, SLUB_RED_INACTIVE);
1029 init_tracking(s, object);
1032 static noinline int alloc_debug_processing(struct kmem_cache *s,
1034 void *object, unsigned long addr)
1036 if (!check_slab(s, page))
1039 if (!check_valid_pointer(s, page, object)) {
1040 object_err(s, page, object, "Freelist Pointer check fails");
1044 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1047 /* Success perform special debug activities for allocs */
1048 if (s->flags & SLAB_STORE_USER)
1049 set_track(s, object, TRACK_ALLOC, addr);
1050 trace(s, page, object, 1);
1051 init_object(s, object, SLUB_RED_ACTIVE);
1055 if (PageSlab(page)) {
1057 * If this is a slab page then lets do the best we can
1058 * to avoid issues in the future. Marking all objects
1059 * as used avoids touching the remaining objects.
1061 slab_fix(s, "Marking all objects used");
1062 page->inuse = page->objects;
1063 page->freelist = NULL;
1068 /* Supports checking bulk free of a constructed freelist */
1069 static noinline struct kmem_cache_node *free_debug_processing(
1070 struct kmem_cache *s, struct page *page,
1071 void *head, void *tail, int bulk_cnt,
1072 unsigned long addr, unsigned long *flags)
1074 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1075 void *object = head;
1078 raw_spin_lock_irqsave(&n->list_lock, *flags);
1081 if (!check_slab(s, page))
1087 if (!check_valid_pointer(s, page, object)) {
1088 slab_err(s, page, "Invalid object pointer 0x%p", object);
1092 if (on_freelist(s, page, object)) {
1093 object_err(s, page, object, "Object already free");
1097 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1100 if (unlikely(s != page->slab_cache)) {
1101 if (!PageSlab(page)) {
1102 slab_err(s, page, "Attempt to free object(0x%p) "
1103 "outside of slab", object);
1104 } else if (!page->slab_cache) {
1105 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1109 object_err(s, page, object,
1110 "page slab pointer corrupt.");
1114 if (s->flags & SLAB_STORE_USER)
1115 set_track(s, object, TRACK_FREE, addr);
1116 trace(s, page, object, 0);
1117 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1118 init_object(s, object, SLUB_RED_INACTIVE);
1120 /* Reached end of constructed freelist yet? */
1121 if (object != tail) {
1122 object = get_freepointer(s, object);
1126 if (cnt != bulk_cnt)
1127 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1132 * Keep node_lock to preserve integrity
1133 * until the object is actually freed
1139 raw_spin_unlock_irqrestore(&n->list_lock, *flags);
1140 slab_fix(s, "Object at 0x%p not freed", object);
1144 static int __init setup_slub_debug(char *str)
1146 slub_debug = DEBUG_DEFAULT_FLAGS;
1147 if (*str++ != '=' || !*str)
1149 * No options specified. Switch on full debugging.
1155 * No options but restriction on slabs. This means full
1156 * debugging for slabs matching a pattern.
1163 * Switch off all debugging measures.
1168 * Determine which debug features should be switched on
1170 for (; *str && *str != ','; str++) {
1171 switch (tolower(*str)) {
1173 slub_debug |= SLAB_DEBUG_FREE;
1176 slub_debug |= SLAB_RED_ZONE;
1179 slub_debug |= SLAB_POISON;
1182 slub_debug |= SLAB_STORE_USER;
1185 slub_debug |= SLAB_TRACE;
1188 slub_debug |= SLAB_FAILSLAB;
1192 * Avoid enabling debugging on caches if its minimum
1193 * order would increase as a result.
1195 disable_higher_order_debug = 1;
1198 pr_err("slub_debug option '%c' unknown. skipped\n",
1205 slub_debug_slabs = str + 1;
1210 __setup("slub_debug", setup_slub_debug);
1212 unsigned long kmem_cache_flags(unsigned long object_size,
1213 unsigned long flags, const char *name,
1214 void (*ctor)(void *))
1217 * Enable debugging if selected on the kernel commandline.
1219 if (slub_debug && (!slub_debug_slabs || (name &&
1220 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1221 flags |= slub_debug;
1225 #else /* !CONFIG_SLUB_DEBUG */
1226 static inline void setup_object_debug(struct kmem_cache *s,
1227 struct page *page, void *object) {}
1229 static inline int alloc_debug_processing(struct kmem_cache *s,
1230 struct page *page, void *object, unsigned long addr) { return 0; }
1232 static inline struct kmem_cache_node *free_debug_processing(
1233 struct kmem_cache *s, struct page *page,
1234 void *head, void *tail, int bulk_cnt,
1235 unsigned long addr, unsigned long *flags) { return NULL; }
1237 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1239 static inline int check_object(struct kmem_cache *s, struct page *page,
1240 void *object, u8 val) { return 1; }
1241 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1242 struct page *page) {}
1243 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1244 struct page *page) {}
1245 unsigned long kmem_cache_flags(unsigned long object_size,
1246 unsigned long flags, const char *name,
1247 void (*ctor)(void *))
1251 #define slub_debug 0
1253 #define disable_higher_order_debug 0
1255 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1257 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1259 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1261 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1264 #endif /* CONFIG_SLUB_DEBUG */
1266 struct slub_free_list {
1267 raw_spinlock_t lock;
1268 struct list_head list;
1270 static DEFINE_PER_CPU(struct slub_free_list, slub_free_list);
1273 * Hooks for other subsystems that check memory allocations. In a typical
1274 * production configuration these hooks all should produce no code at all.
1276 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1278 kmemleak_alloc(ptr, size, 1, flags);
1279 kasan_kmalloc_large(ptr, size);
1282 static inline void kfree_hook(const void *x)
1285 kasan_kfree_large(x);
1288 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1291 flags &= gfp_allowed_mask;
1292 lockdep_trace_alloc(flags);
1293 might_sleep_if(gfpflags_allow_blocking(flags));
1295 if (should_failslab(s->object_size, flags, s->flags))
1298 return memcg_kmem_get_cache(s, flags);
1301 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1302 size_t size, void **p)
1306 flags &= gfp_allowed_mask;
1307 for (i = 0; i < size; i++) {
1308 void *object = p[i];
1310 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1311 kmemleak_alloc_recursive(object, s->object_size, 1,
1313 kasan_slab_alloc(s, object);
1315 memcg_kmem_put_cache(s);
1318 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1320 kmemleak_free_recursive(x, s->flags);
1323 * Trouble is that we may no longer disable interrupts in the fast path
1324 * So in order to make the debug calls that expect irqs to be
1325 * disabled we need to disable interrupts temporarily.
1327 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1329 unsigned long flags;
1331 local_irq_save(flags);
1332 kmemcheck_slab_free(s, x, s->object_size);
1333 debug_check_no_locks_freed(x, s->object_size);
1334 local_irq_restore(flags);
1337 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1338 debug_check_no_obj_freed(x, s->object_size);
1340 kasan_slab_free(s, x);
1343 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1344 void *head, void *tail)
1347 * Compiler cannot detect this function can be removed if slab_free_hook()
1348 * evaluates to nothing. Thus, catch all relevant config debug options here.
1350 #if defined(CONFIG_KMEMCHECK) || \
1351 defined(CONFIG_LOCKDEP) || \
1352 defined(CONFIG_DEBUG_KMEMLEAK) || \
1353 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1354 defined(CONFIG_KASAN)
1356 void *object = head;
1357 void *tail_obj = tail ? : head;
1360 slab_free_hook(s, object);
1361 } while ((object != tail_obj) &&
1362 (object = get_freepointer(s, object)));
1366 static void setup_object(struct kmem_cache *s, struct page *page,
1369 setup_object_debug(s, page, object);
1370 if (unlikely(s->ctor)) {
1371 kasan_unpoison_object_data(s, object);
1373 kasan_poison_object_data(s, object);
1378 * Slab allocation and freeing
1380 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1381 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1384 int order = oo_order(oo);
1386 flags |= __GFP_NOTRACK;
1388 if (node == NUMA_NO_NODE)
1389 page = alloc_pages(flags, order);
1391 page = __alloc_pages_node(node, flags, order);
1393 if (page && memcg_charge_slab(page, flags, order, s)) {
1394 __free_pages(page, order);
1401 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1404 struct kmem_cache_order_objects oo = s->oo;
1408 bool enableirqs = false;
1410 flags &= gfp_allowed_mask;
1412 if (gfpflags_allow_blocking(flags))
1414 #ifdef CONFIG_PREEMPT_RT_FULL
1415 if (system_state == SYSTEM_RUNNING)
1421 flags |= s->allocflags;
1424 * Let the initial higher-order allocation fail under memory pressure
1425 * so we fall-back to the minimum order allocation.
1427 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1428 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1429 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1431 page = alloc_slab_page(s, alloc_gfp, node, oo);
1432 if (unlikely(!page)) {
1436 * Allocation may have failed due to fragmentation.
1437 * Try a lower order alloc if possible
1439 page = alloc_slab_page(s, alloc_gfp, node, oo);
1440 if (unlikely(!page))
1442 stat(s, ORDER_FALLBACK);
1445 if (kmemcheck_enabled &&
1446 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1447 int pages = 1 << oo_order(oo);
1449 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1452 * Objects from caches that have a constructor don't get
1453 * cleared when they're allocated, so we need to do it here.
1456 kmemcheck_mark_uninitialized_pages(page, pages);
1458 kmemcheck_mark_unallocated_pages(page, pages);
1461 page->objects = oo_objects(oo);
1463 order = compound_order(page);
1464 page->slab_cache = s;
1465 __SetPageSlab(page);
1466 if (page_is_pfmemalloc(page))
1467 SetPageSlabPfmemalloc(page);
1469 start = page_address(page);
1471 if (unlikely(s->flags & SLAB_POISON))
1472 memset(start, POISON_INUSE, PAGE_SIZE << order);
1474 kasan_poison_slab(page);
1476 for_each_object_idx(p, idx, s, start, page->objects) {
1477 setup_object(s, page, p);
1478 if (likely(idx < page->objects))
1479 set_freepointer(s, p, p + s->size);
1481 set_freepointer(s, p, NULL);
1484 page->freelist = start;
1485 page->inuse = page->objects;
1490 local_irq_disable();
1494 mod_zone_page_state(page_zone(page),
1495 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1496 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1499 inc_slabs_node(s, page_to_nid(page), page->objects);
1504 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1506 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1507 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1511 return allocate_slab(s,
1512 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1515 static void __free_slab(struct kmem_cache *s, struct page *page)
1517 int order = compound_order(page);
1518 int pages = 1 << order;
1520 if (kmem_cache_debug(s)) {
1523 slab_pad_check(s, page);
1524 for_each_object(p, s, page_address(page),
1526 check_object(s, page, p, SLUB_RED_INACTIVE);
1529 kmemcheck_free_shadow(page, compound_order(page));
1531 mod_zone_page_state(page_zone(page),
1532 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1533 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1536 __ClearPageSlabPfmemalloc(page);
1537 __ClearPageSlab(page);
1539 page_mapcount_reset(page);
1540 if (current->reclaim_state)
1541 current->reclaim_state->reclaimed_slab += pages;
1542 __free_kmem_pages(page, order);
1545 static void free_delayed(struct list_head *h)
1547 while(!list_empty(h)) {
1548 struct page *page = list_first_entry(h, struct page, lru);
1550 list_del(&page->lru);
1551 __free_slab(page->slab_cache, page);
1555 #define need_reserve_slab_rcu \
1556 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1558 static void rcu_free_slab(struct rcu_head *h)
1562 if (need_reserve_slab_rcu)
1563 page = virt_to_head_page(h);
1565 page = container_of((struct list_head *)h, struct page, lru);
1567 __free_slab(page->slab_cache, page);
1570 static void free_slab(struct kmem_cache *s, struct page *page)
1572 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1573 struct rcu_head *head;
1575 if (need_reserve_slab_rcu) {
1576 int order = compound_order(page);
1577 int offset = (PAGE_SIZE << order) - s->reserved;
1579 VM_BUG_ON(s->reserved != sizeof(*head));
1580 head = page_address(page) + offset;
1582 head = &page->rcu_head;
1585 call_rcu(head, rcu_free_slab);
1586 } else if (irqs_disabled()) {
1587 struct slub_free_list *f = this_cpu_ptr(&slub_free_list);
1589 raw_spin_lock(&f->lock);
1590 list_add(&page->lru, &f->list);
1591 raw_spin_unlock(&f->lock);
1593 __free_slab(s, page);
1596 static void discard_slab(struct kmem_cache *s, struct page *page)
1598 dec_slabs_node(s, page_to_nid(page), page->objects);
1603 * Management of partially allocated slabs.
1606 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1609 if (tail == DEACTIVATE_TO_TAIL)
1610 list_add_tail(&page->lru, &n->partial);
1612 list_add(&page->lru, &n->partial);
1615 static inline void add_partial(struct kmem_cache_node *n,
1616 struct page *page, int tail)
1618 lockdep_assert_held(&n->list_lock);
1619 __add_partial(n, page, tail);
1623 __remove_partial(struct kmem_cache_node *n, struct page *page)
1625 list_del(&page->lru);
1629 static inline void remove_partial(struct kmem_cache_node *n,
1632 lockdep_assert_held(&n->list_lock);
1633 __remove_partial(n, page);
1637 * Remove slab from the partial list, freeze it and
1638 * return the pointer to the freelist.
1640 * Returns a list of objects or NULL if it fails.
1642 static inline void *acquire_slab(struct kmem_cache *s,
1643 struct kmem_cache_node *n, struct page *page,
1644 int mode, int *objects)
1647 unsigned long counters;
1650 lockdep_assert_held(&n->list_lock);
1653 * Zap the freelist and set the frozen bit.
1654 * The old freelist is the list of objects for the
1655 * per cpu allocation list.
1657 freelist = page->freelist;
1658 counters = page->counters;
1659 new.counters = counters;
1660 *objects = new.objects - new.inuse;
1662 new.inuse = page->objects;
1663 new.freelist = NULL;
1665 new.freelist = freelist;
1668 VM_BUG_ON(new.frozen);
1671 if (!__cmpxchg_double_slab(s, page,
1673 new.freelist, new.counters,
1677 remove_partial(n, page);
1682 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1683 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1686 * Try to allocate a partial slab from a specific node.
1688 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1689 struct kmem_cache_cpu *c, gfp_t flags)
1691 struct page *page, *page2;
1692 void *object = NULL;
1697 * Racy check. If we mistakenly see no partial slabs then we
1698 * just allocate an empty slab. If we mistakenly try to get a
1699 * partial slab and there is none available then get_partials()
1702 if (!n || !n->nr_partial)
1705 raw_spin_lock(&n->list_lock);
1706 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1709 if (!pfmemalloc_match(page, flags))
1712 t = acquire_slab(s, n, page, object == NULL, &objects);
1716 available += objects;
1719 stat(s, ALLOC_FROM_PARTIAL);
1722 put_cpu_partial(s, page, 0);
1723 stat(s, CPU_PARTIAL_NODE);
1725 if (!kmem_cache_has_cpu_partial(s)
1726 || available > s->cpu_partial / 2)
1730 raw_spin_unlock(&n->list_lock);
1735 * Get a page from somewhere. Search in increasing NUMA distances.
1737 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1738 struct kmem_cache_cpu *c)
1741 struct zonelist *zonelist;
1744 enum zone_type high_zoneidx = gfp_zone(flags);
1746 unsigned int cpuset_mems_cookie;
1749 * The defrag ratio allows a configuration of the tradeoffs between
1750 * inter node defragmentation and node local allocations. A lower
1751 * defrag_ratio increases the tendency to do local allocations
1752 * instead of attempting to obtain partial slabs from other nodes.
1754 * If the defrag_ratio is set to 0 then kmalloc() always
1755 * returns node local objects. If the ratio is higher then kmalloc()
1756 * may return off node objects because partial slabs are obtained
1757 * from other nodes and filled up.
1759 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1760 * defrag_ratio = 1000) then every (well almost) allocation will
1761 * first attempt to defrag slab caches on other nodes. This means
1762 * scanning over all nodes to look for partial slabs which may be
1763 * expensive if we do it every time we are trying to find a slab
1764 * with available objects.
1766 if (!s->remote_node_defrag_ratio ||
1767 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1771 cpuset_mems_cookie = read_mems_allowed_begin();
1772 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1773 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1774 struct kmem_cache_node *n;
1776 n = get_node(s, zone_to_nid(zone));
1778 if (n && cpuset_zone_allowed(zone, flags) &&
1779 n->nr_partial > s->min_partial) {
1780 object = get_partial_node(s, n, c, flags);
1783 * Don't check read_mems_allowed_retry()
1784 * here - if mems_allowed was updated in
1785 * parallel, that was a harmless race
1786 * between allocation and the cpuset
1793 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1799 * Get a partial page, lock it and return it.
1801 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1802 struct kmem_cache_cpu *c)
1805 int searchnode = node;
1807 if (node == NUMA_NO_NODE)
1808 searchnode = numa_mem_id();
1809 else if (!node_present_pages(node))
1810 searchnode = node_to_mem_node(node);
1812 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1813 if (object || node != NUMA_NO_NODE)
1816 return get_any_partial(s, flags, c);
1819 #ifdef CONFIG_PREEMPT
1821 * Calculate the next globally unique transaction for disambiguiation
1822 * during cmpxchg. The transactions start with the cpu number and are then
1823 * incremented by CONFIG_NR_CPUS.
1825 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1828 * No preemption supported therefore also no need to check for
1834 static inline unsigned long next_tid(unsigned long tid)
1836 return tid + TID_STEP;
1839 static inline unsigned int tid_to_cpu(unsigned long tid)
1841 return tid % TID_STEP;
1844 static inline unsigned long tid_to_event(unsigned long tid)
1846 return tid / TID_STEP;
1849 static inline unsigned int init_tid(int cpu)
1854 static inline void note_cmpxchg_failure(const char *n,
1855 const struct kmem_cache *s, unsigned long tid)
1857 #ifdef SLUB_DEBUG_CMPXCHG
1858 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1860 pr_info("%s %s: cmpxchg redo ", n, s->name);
1862 #ifdef CONFIG_PREEMPT
1863 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1864 pr_warn("due to cpu change %d -> %d\n",
1865 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1868 if (tid_to_event(tid) != tid_to_event(actual_tid))
1869 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1870 tid_to_event(tid), tid_to_event(actual_tid));
1872 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1873 actual_tid, tid, next_tid(tid));
1875 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1878 static void init_kmem_cache_cpus(struct kmem_cache *s)
1882 for_each_possible_cpu(cpu)
1883 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1887 * Remove the cpu slab
1889 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1892 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1893 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1895 enum slab_modes l = M_NONE, m = M_NONE;
1897 int tail = DEACTIVATE_TO_HEAD;
1901 if (page->freelist) {
1902 stat(s, DEACTIVATE_REMOTE_FREES);
1903 tail = DEACTIVATE_TO_TAIL;
1907 * Stage one: Free all available per cpu objects back
1908 * to the page freelist while it is still frozen. Leave the
1911 * There is no need to take the list->lock because the page
1914 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1916 unsigned long counters;
1919 prior = page->freelist;
1920 counters = page->counters;
1921 set_freepointer(s, freelist, prior);
1922 new.counters = counters;
1924 VM_BUG_ON(!new.frozen);
1926 } while (!__cmpxchg_double_slab(s, page,
1928 freelist, new.counters,
1929 "drain percpu freelist"));
1931 freelist = nextfree;
1935 * Stage two: Ensure that the page is unfrozen while the
1936 * list presence reflects the actual number of objects
1939 * We setup the list membership and then perform a cmpxchg
1940 * with the count. If there is a mismatch then the page
1941 * is not unfrozen but the page is on the wrong list.
1943 * Then we restart the process which may have to remove
1944 * the page from the list that we just put it on again
1945 * because the number of objects in the slab may have
1950 old.freelist = page->freelist;
1951 old.counters = page->counters;
1952 VM_BUG_ON(!old.frozen);
1954 /* Determine target state of the slab */
1955 new.counters = old.counters;
1958 set_freepointer(s, freelist, old.freelist);
1959 new.freelist = freelist;
1961 new.freelist = old.freelist;
1965 if (!new.inuse && n->nr_partial >= s->min_partial)
1967 else if (new.freelist) {
1972 * Taking the spinlock removes the possiblity
1973 * that acquire_slab() will see a slab page that
1976 raw_spin_lock(&n->list_lock);
1980 if (kmem_cache_debug(s) && !lock) {
1983 * This also ensures that the scanning of full
1984 * slabs from diagnostic functions will not see
1987 raw_spin_lock(&n->list_lock);
1995 remove_partial(n, page);
1997 else if (l == M_FULL)
1999 remove_full(s, n, page);
2001 if (m == M_PARTIAL) {
2003 add_partial(n, page, tail);
2006 } else if (m == M_FULL) {
2008 stat(s, DEACTIVATE_FULL);
2009 add_full(s, n, page);
2015 if (!__cmpxchg_double_slab(s, page,
2016 old.freelist, old.counters,
2017 new.freelist, new.counters,
2022 raw_spin_unlock(&n->list_lock);
2025 stat(s, DEACTIVATE_EMPTY);
2026 discard_slab(s, page);
2032 * Unfreeze all the cpu partial slabs.
2034 * This function must be called with interrupts disabled
2035 * for the cpu using c (or some other guarantee must be there
2036 * to guarantee no concurrent accesses).
2038 static void unfreeze_partials(struct kmem_cache *s,
2039 struct kmem_cache_cpu *c)
2041 #ifdef CONFIG_SLUB_CPU_PARTIAL
2042 struct kmem_cache_node *n = NULL, *n2 = NULL;
2043 struct page *page, *discard_page = NULL;
2045 while ((page = c->partial)) {
2049 c->partial = page->next;
2051 n2 = get_node(s, page_to_nid(page));
2054 raw_spin_unlock(&n->list_lock);
2057 raw_spin_lock(&n->list_lock);
2062 old.freelist = page->freelist;
2063 old.counters = page->counters;
2064 VM_BUG_ON(!old.frozen);
2066 new.counters = old.counters;
2067 new.freelist = old.freelist;
2071 } while (!__cmpxchg_double_slab(s, page,
2072 old.freelist, old.counters,
2073 new.freelist, new.counters,
2074 "unfreezing slab"));
2076 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2077 page->next = discard_page;
2078 discard_page = page;
2080 add_partial(n, page, DEACTIVATE_TO_TAIL);
2081 stat(s, FREE_ADD_PARTIAL);
2086 raw_spin_unlock(&n->list_lock);
2088 while (discard_page) {
2089 page = discard_page;
2090 discard_page = discard_page->next;
2092 stat(s, DEACTIVATE_EMPTY);
2093 discard_slab(s, page);
2100 * Put a page that was just frozen (in __slab_free) into a partial page
2101 * slot if available. This is done without interrupts disabled and without
2102 * preemption disabled. The cmpxchg is racy and may put the partial page
2103 * onto a random cpus partial slot.
2105 * If we did not find a slot then simply move all the partials to the
2106 * per node partial list.
2108 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2110 #ifdef CONFIG_SLUB_CPU_PARTIAL
2111 struct page *oldpage;
2119 oldpage = this_cpu_read(s->cpu_slab->partial);
2122 pobjects = oldpage->pobjects;
2123 pages = oldpage->pages;
2124 if (drain && pobjects > s->cpu_partial) {
2125 struct slub_free_list *f;
2126 unsigned long flags;
2129 * partial array is full. Move the existing
2130 * set to the per node partial list.
2132 local_irq_save(flags);
2133 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2134 f = this_cpu_ptr(&slub_free_list);
2135 raw_spin_lock(&f->lock);
2136 list_splice_init(&f->list, &tofree);
2137 raw_spin_unlock(&f->lock);
2138 local_irq_restore(flags);
2139 free_delayed(&tofree);
2143 stat(s, CPU_PARTIAL_DRAIN);
2148 pobjects += page->objects - page->inuse;
2150 page->pages = pages;
2151 page->pobjects = pobjects;
2152 page->next = oldpage;
2154 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2156 if (unlikely(!s->cpu_partial)) {
2157 unsigned long flags;
2159 local_irq_save(flags);
2160 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2161 local_irq_restore(flags);
2167 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2169 stat(s, CPUSLAB_FLUSH);
2170 deactivate_slab(s, c->page, c->freelist);
2172 c->tid = next_tid(c->tid);
2180 * Called from IPI handler with interrupts disabled.
2182 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2184 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2190 unfreeze_partials(s, c);
2194 static void flush_cpu_slab(void *d)
2196 struct kmem_cache *s = d;
2198 __flush_cpu_slab(s, smp_processor_id());
2201 static bool has_cpu_slab(int cpu, void *info)
2203 struct kmem_cache *s = info;
2204 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2206 return c->page || c->partial;
2209 static void flush_all(struct kmem_cache *s)
2214 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2215 for_each_online_cpu(cpu) {
2216 struct slub_free_list *f;
2218 if (!has_cpu_slab(cpu, s))
2221 f = &per_cpu(slub_free_list, cpu);
2222 raw_spin_lock_irq(&f->lock);
2223 list_splice_init(&f->list, &tofree);
2224 raw_spin_unlock_irq(&f->lock);
2225 free_delayed(&tofree);
2230 * Check if the objects in a per cpu structure fit numa
2231 * locality expectations.
2233 static inline int node_match(struct page *page, int node)
2236 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2242 #ifdef CONFIG_SLUB_DEBUG
2243 static int count_free(struct page *page)
2245 return page->objects - page->inuse;
2248 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2250 return atomic_long_read(&n->total_objects);
2252 #endif /* CONFIG_SLUB_DEBUG */
2254 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2255 static unsigned long count_partial(struct kmem_cache_node *n,
2256 int (*get_count)(struct page *))
2258 unsigned long flags;
2259 unsigned long x = 0;
2262 raw_spin_lock_irqsave(&n->list_lock, flags);
2263 list_for_each_entry(page, &n->partial, lru)
2264 x += get_count(page);
2265 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2268 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2270 static noinline void
2271 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2273 #ifdef CONFIG_SLUB_DEBUG
2274 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2275 DEFAULT_RATELIMIT_BURST);
2277 struct kmem_cache_node *n;
2279 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2282 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2284 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2285 s->name, s->object_size, s->size, oo_order(s->oo),
2288 if (oo_order(s->min) > get_order(s->object_size))
2289 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2292 for_each_kmem_cache_node(s, node, n) {
2293 unsigned long nr_slabs;
2294 unsigned long nr_objs;
2295 unsigned long nr_free;
2297 nr_free = count_partial(n, count_free);
2298 nr_slabs = node_nr_slabs(n);
2299 nr_objs = node_nr_objs(n);
2301 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2302 node, nr_slabs, nr_objs, nr_free);
2307 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2308 int node, struct kmem_cache_cpu **pc)
2311 struct kmem_cache_cpu *c = *pc;
2314 freelist = get_partial(s, flags, node, c);
2319 page = new_slab(s, flags, node);
2321 c = raw_cpu_ptr(s->cpu_slab);
2326 * No other reference to the page yet so we can
2327 * muck around with it freely without cmpxchg
2329 freelist = page->freelist;
2330 page->freelist = NULL;
2332 stat(s, ALLOC_SLAB);
2341 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2343 if (unlikely(PageSlabPfmemalloc(page)))
2344 return gfp_pfmemalloc_allowed(gfpflags);
2350 * Check the page->freelist of a page and either transfer the freelist to the
2351 * per cpu freelist or deactivate the page.
2353 * The page is still frozen if the return value is not NULL.
2355 * If this function returns NULL then the page has been unfrozen.
2357 * This function must be called with interrupt disabled.
2359 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2362 unsigned long counters;
2366 freelist = page->freelist;
2367 counters = page->counters;
2369 new.counters = counters;
2370 VM_BUG_ON(!new.frozen);
2372 new.inuse = page->objects;
2373 new.frozen = freelist != NULL;
2375 } while (!__cmpxchg_double_slab(s, page,
2384 * Slow path. The lockless freelist is empty or we need to perform
2387 * Processing is still very fast if new objects have been freed to the
2388 * regular freelist. In that case we simply take over the regular freelist
2389 * as the lockless freelist and zap the regular freelist.
2391 * If that is not working then we fall back to the partial lists. We take the
2392 * first element of the freelist as the object to allocate now and move the
2393 * rest of the freelist to the lockless freelist.
2395 * And if we were unable to get a new slab from the partial slab lists then
2396 * we need to allocate a new slab. This is the slowest path since it involves
2397 * a call to the page allocator and the setup of a new slab.
2399 * Version of __slab_alloc to use when we know that interrupts are
2400 * already disabled (which is the case for bulk allocation).
2402 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2403 unsigned long addr, struct kmem_cache_cpu *c,
2404 struct list_head *to_free)
2406 struct slub_free_list *f;
2415 if (unlikely(!node_match(page, node))) {
2416 int searchnode = node;
2418 if (node != NUMA_NO_NODE && !node_present_pages(node))
2419 searchnode = node_to_mem_node(node);
2421 if (unlikely(!node_match(page, searchnode))) {
2422 stat(s, ALLOC_NODE_MISMATCH);
2423 deactivate_slab(s, page, c->freelist);
2431 * By rights, we should be searching for a slab page that was
2432 * PFMEMALLOC but right now, we are losing the pfmemalloc
2433 * information when the page leaves the per-cpu allocator
2435 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2436 deactivate_slab(s, page, c->freelist);
2442 /* must check again c->freelist in case of cpu migration or IRQ */
2443 freelist = c->freelist;
2447 freelist = get_freelist(s, page);
2451 stat(s, DEACTIVATE_BYPASS);
2455 stat(s, ALLOC_REFILL);
2459 * freelist is pointing to the list of objects to be used.
2460 * page is pointing to the page from which the objects are obtained.
2461 * That page must be frozen for per cpu allocations to work.
2463 VM_BUG_ON(!c->page->frozen);
2464 c->freelist = get_freepointer(s, freelist);
2465 c->tid = next_tid(c->tid);
2468 f = this_cpu_ptr(&slub_free_list);
2469 raw_spin_lock(&f->lock);
2470 list_splice_init(&f->list, to_free);
2471 raw_spin_unlock(&f->lock);
2478 page = c->page = c->partial;
2479 c->partial = page->next;
2480 stat(s, CPU_PARTIAL_ALLOC);
2485 freelist = new_slab_objects(s, gfpflags, node, &c);
2487 if (unlikely(!freelist)) {
2488 slab_out_of_memory(s, gfpflags, node);
2493 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2496 /* Only entered in the debug case */
2497 if (kmem_cache_debug(s) &&
2498 !alloc_debug_processing(s, page, freelist, addr))
2499 goto new_slab; /* Slab failed checks. Next slab needed */
2501 deactivate_slab(s, page, get_freepointer(s, freelist));
2508 * Another one that disabled interrupt and compensates for possible
2509 * cpu changes by refetching the per cpu area pointer.
2511 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2512 unsigned long addr, struct kmem_cache_cpu *c)
2515 unsigned long flags;
2518 local_irq_save(flags);
2519 #ifdef CONFIG_PREEMPT
2521 * We may have been preempted and rescheduled on a different
2522 * cpu before disabling interrupts. Need to reload cpu area
2525 c = this_cpu_ptr(s->cpu_slab);
2528 p = ___slab_alloc(s, gfpflags, node, addr, c, &tofree);
2529 local_irq_restore(flags);
2530 free_delayed(&tofree);
2535 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2536 * have the fastpath folded into their functions. So no function call
2537 * overhead for requests that can be satisfied on the fastpath.
2539 * The fastpath works by first checking if the lockless freelist can be used.
2540 * If not then __slab_alloc is called for slow processing.
2542 * Otherwise we can simply pick the next object from the lockless free list.
2544 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2545 gfp_t gfpflags, int node, unsigned long addr)
2548 struct kmem_cache_cpu *c;
2552 s = slab_pre_alloc_hook(s, gfpflags);
2557 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2558 * enabled. We may switch back and forth between cpus while
2559 * reading from one cpu area. That does not matter as long
2560 * as we end up on the original cpu again when doing the cmpxchg.
2562 * We should guarantee that tid and kmem_cache are retrieved on
2563 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2564 * to check if it is matched or not.
2567 tid = this_cpu_read(s->cpu_slab->tid);
2568 c = raw_cpu_ptr(s->cpu_slab);
2569 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2570 unlikely(tid != READ_ONCE(c->tid)));
2573 * Irqless object alloc/free algorithm used here depends on sequence
2574 * of fetching cpu_slab's data. tid should be fetched before anything
2575 * on c to guarantee that object and page associated with previous tid
2576 * won't be used with current tid. If we fetch tid first, object and
2577 * page could be one associated with next tid and our alloc/free
2578 * request will be failed. In this case, we will retry. So, no problem.
2583 * The transaction ids are globally unique per cpu and per operation on
2584 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2585 * occurs on the right processor and that there was no operation on the
2586 * linked list in between.
2589 object = c->freelist;
2591 if (unlikely(!object || !node_match(page, node))) {
2592 object = __slab_alloc(s, gfpflags, node, addr, c);
2593 stat(s, ALLOC_SLOWPATH);
2595 void *next_object = get_freepointer_safe(s, object);
2598 * The cmpxchg will only match if there was no additional
2599 * operation and if we are on the right processor.
2601 * The cmpxchg does the following atomically (without lock
2603 * 1. Relocate first pointer to the current per cpu area.
2604 * 2. Verify that tid and freelist have not been changed
2605 * 3. If they were not changed replace tid and freelist
2607 * Since this is without lock semantics the protection is only
2608 * against code executing on this cpu *not* from access by
2611 if (unlikely(!this_cpu_cmpxchg_double(
2612 s->cpu_slab->freelist, s->cpu_slab->tid,
2614 next_object, next_tid(tid)))) {
2616 note_cmpxchg_failure("slab_alloc", s, tid);
2619 prefetch_freepointer(s, next_object);
2620 stat(s, ALLOC_FASTPATH);
2623 if (unlikely(gfpflags & __GFP_ZERO) && object)
2624 memset(object, 0, s->object_size);
2626 slab_post_alloc_hook(s, gfpflags, 1, &object);
2631 static __always_inline void *slab_alloc(struct kmem_cache *s,
2632 gfp_t gfpflags, unsigned long addr)
2634 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2637 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2639 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2641 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2646 EXPORT_SYMBOL(kmem_cache_alloc);
2648 #ifdef CONFIG_TRACING
2649 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2651 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2652 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2653 kasan_kmalloc(s, ret, size);
2656 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2660 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2662 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2664 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2665 s->object_size, s->size, gfpflags, node);
2669 EXPORT_SYMBOL(kmem_cache_alloc_node);
2671 #ifdef CONFIG_TRACING
2672 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2674 int node, size_t size)
2676 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2678 trace_kmalloc_node(_RET_IP_, ret,
2679 size, s->size, gfpflags, node);
2681 kasan_kmalloc(s, ret, size);
2684 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2689 * Slow path handling. This may still be called frequently since objects
2690 * have a longer lifetime than the cpu slabs in most processing loads.
2692 * So we still attempt to reduce cache line usage. Just take the slab
2693 * lock and free the item. If there is no additional partial page
2694 * handling required then we can return immediately.
2696 static void __slab_free(struct kmem_cache *s, struct page *page,
2697 void *head, void *tail, int cnt,
2704 unsigned long counters;
2705 struct kmem_cache_node *n = NULL;
2706 unsigned long uninitialized_var(flags);
2708 stat(s, FREE_SLOWPATH);
2710 if (kmem_cache_debug(s) &&
2711 !(n = free_debug_processing(s, page, head, tail, cnt,
2717 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2720 prior = page->freelist;
2721 counters = page->counters;
2722 set_freepointer(s, tail, prior);
2723 new.counters = counters;
2724 was_frozen = new.frozen;
2726 if ((!new.inuse || !prior) && !was_frozen) {
2728 if (kmem_cache_has_cpu_partial(s) && !prior) {
2731 * Slab was on no list before and will be
2733 * We can defer the list move and instead
2738 } else { /* Needs to be taken off a list */
2740 n = get_node(s, page_to_nid(page));
2742 * Speculatively acquire the list_lock.
2743 * If the cmpxchg does not succeed then we may
2744 * drop the list_lock without any processing.
2746 * Otherwise the list_lock will synchronize with
2747 * other processors updating the list of slabs.
2749 raw_spin_lock_irqsave(&n->list_lock, flags);
2754 } while (!cmpxchg_double_slab(s, page,
2762 * If we just froze the page then put it onto the
2763 * per cpu partial list.
2765 if (new.frozen && !was_frozen) {
2766 put_cpu_partial(s, page, 1);
2767 stat(s, CPU_PARTIAL_FREE);
2770 * The list lock was not taken therefore no list
2771 * activity can be necessary.
2774 stat(s, FREE_FROZEN);
2778 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2782 * Objects left in the slab. If it was not on the partial list before
2785 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2786 if (kmem_cache_debug(s))
2787 remove_full(s, n, page);
2788 add_partial(n, page, DEACTIVATE_TO_TAIL);
2789 stat(s, FREE_ADD_PARTIAL);
2791 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2797 * Slab on the partial list.
2799 remove_partial(n, page);
2800 stat(s, FREE_REMOVE_PARTIAL);
2802 /* Slab must be on the full list */
2803 remove_full(s, n, page);
2806 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2808 discard_slab(s, page);
2812 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2813 * can perform fastpath freeing without additional function calls.
2815 * The fastpath is only possible if we are freeing to the current cpu slab
2816 * of this processor. This typically the case if we have just allocated
2819 * If fastpath is not possible then fall back to __slab_free where we deal
2820 * with all sorts of special processing.
2822 * Bulk free of a freelist with several objects (all pointing to the
2823 * same page) possible by specifying head and tail ptr, plus objects
2824 * count (cnt). Bulk free indicated by tail pointer being set.
2826 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2827 void *head, void *tail, int cnt,
2830 void *tail_obj = tail ? : head;
2831 struct kmem_cache_cpu *c;
2834 slab_free_freelist_hook(s, head, tail);
2838 * Determine the currently cpus per cpu slab.
2839 * The cpu may change afterward. However that does not matter since
2840 * data is retrieved via this pointer. If we are on the same cpu
2841 * during the cmpxchg then the free will succeed.
2844 tid = this_cpu_read(s->cpu_slab->tid);
2845 c = raw_cpu_ptr(s->cpu_slab);
2846 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2847 unlikely(tid != READ_ONCE(c->tid)));
2849 /* Same with comment on barrier() in slab_alloc_node() */
2852 if (likely(page == c->page)) {
2853 set_freepointer(s, tail_obj, c->freelist);
2855 if (unlikely(!this_cpu_cmpxchg_double(
2856 s->cpu_slab->freelist, s->cpu_slab->tid,
2858 head, next_tid(tid)))) {
2860 note_cmpxchg_failure("slab_free", s, tid);
2863 stat(s, FREE_FASTPATH);
2865 __slab_free(s, page, head, tail_obj, cnt, addr);
2869 void kmem_cache_free(struct kmem_cache *s, void *x)
2871 s = cache_from_obj(s, x);
2874 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2875 trace_kmem_cache_free(_RET_IP_, x);
2877 EXPORT_SYMBOL(kmem_cache_free);
2879 struct detached_freelist {
2887 * This function progressively scans the array with free objects (with
2888 * a limited look ahead) and extract objects belonging to the same
2889 * page. It builds a detached freelist directly within the given
2890 * page/objects. This can happen without any need for
2891 * synchronization, because the objects are owned by running process.
2892 * The freelist is build up as a single linked list in the objects.
2893 * The idea is, that this detached freelist can then be bulk
2894 * transferred to the real freelist(s), but only requiring a single
2895 * synchronization primitive. Look ahead in the array is limited due
2896 * to performance reasons.
2898 static int build_detached_freelist(struct kmem_cache *s, size_t size,
2899 void **p, struct detached_freelist *df)
2901 size_t first_skipped_index = 0;
2905 /* Always re-init detached_freelist */
2910 } while (!object && size);
2915 /* Start new detached freelist */
2916 set_freepointer(s, object, NULL);
2917 df->page = virt_to_head_page(object);
2919 df->freelist = object;
2920 p[size] = NULL; /* mark object processed */
2926 continue; /* Skip processed objects */
2928 /* df->page is always set at this point */
2929 if (df->page == virt_to_head_page(object)) {
2930 /* Opportunity build freelist */
2931 set_freepointer(s, object, df->freelist);
2932 df->freelist = object;
2934 p[size] = NULL; /* mark object processed */
2939 /* Limit look ahead search */
2943 if (!first_skipped_index)
2944 first_skipped_index = size + 1;
2947 return first_skipped_index;
2951 /* Note that interrupts must be enabled when calling this function. */
2952 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
2958 struct detached_freelist df;
2959 struct kmem_cache *s;
2961 /* Support for memcg */
2962 s = cache_from_obj(orig_s, p[size - 1]);
2964 size = build_detached_freelist(s, size, p, &df);
2965 if (unlikely(!df.page))
2968 slab_free(s, df.page, df.freelist, df.tail, df.cnt, _RET_IP_);
2969 } while (likely(size));
2971 EXPORT_SYMBOL(kmem_cache_free_bulk);
2973 /* Note that interrupts must be enabled when calling this function. */
2974 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2977 struct kmem_cache_cpu *c;
2981 /* memcg and kmem_cache debug support */
2982 s = slab_pre_alloc_hook(s, flags);
2986 * Drain objects in the per cpu slab, while disabling local
2987 * IRQs, which protects against PREEMPT and interrupts
2988 * handlers invoking normal fastpath.
2990 local_irq_disable();
2991 c = this_cpu_ptr(s->cpu_slab);
2993 for (i = 0; i < size; i++) {
2994 void *object = c->freelist;
2996 if (unlikely(!object)) {
2998 * Invoking slow path likely have side-effect
2999 * of re-populating per CPU c->freelist
3001 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3002 _RET_IP_, c, &to_free);
3003 if (unlikely(!p[i]))
3006 c = this_cpu_ptr(s->cpu_slab);
3007 continue; /* goto for-loop */
3009 c->freelist = get_freepointer(s, object);
3012 c->tid = next_tid(c->tid);
3014 free_delayed(&to_free);
3016 /* Clear memory outside IRQ disabled fastpath loop */
3017 if (unlikely(flags & __GFP_ZERO)) {
3020 for (j = 0; j < i; j++)
3021 memset(p[j], 0, s->object_size);
3024 /* memcg and kmem_cache debug support */
3025 slab_post_alloc_hook(s, flags, size, p);
3029 slab_post_alloc_hook(s, flags, i, p);
3030 __kmem_cache_free_bulk(s, i, p);
3033 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3037 * Object placement in a slab is made very easy because we always start at
3038 * offset 0. If we tune the size of the object to the alignment then we can
3039 * get the required alignment by putting one properly sized object after
3042 * Notice that the allocation order determines the sizes of the per cpu
3043 * caches. Each processor has always one slab available for allocations.
3044 * Increasing the allocation order reduces the number of times that slabs
3045 * must be moved on and off the partial lists and is therefore a factor in
3050 * Mininum / Maximum order of slab pages. This influences locking overhead
3051 * and slab fragmentation. A higher order reduces the number of partial slabs
3052 * and increases the number of allocations possible without having to
3053 * take the list_lock.
3055 static int slub_min_order;
3056 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3057 static int slub_min_objects;
3060 * Calculate the order of allocation given an slab object size.
3062 * The order of allocation has significant impact on performance and other
3063 * system components. Generally order 0 allocations should be preferred since
3064 * order 0 does not cause fragmentation in the page allocator. Larger objects
3065 * be problematic to put into order 0 slabs because there may be too much
3066 * unused space left. We go to a higher order if more than 1/16th of the slab
3069 * In order to reach satisfactory performance we must ensure that a minimum
3070 * number of objects is in one slab. Otherwise we may generate too much
3071 * activity on the partial lists which requires taking the list_lock. This is
3072 * less a concern for large slabs though which are rarely used.
3074 * slub_max_order specifies the order where we begin to stop considering the
3075 * number of objects in a slab as critical. If we reach slub_max_order then
3076 * we try to keep the page order as low as possible. So we accept more waste
3077 * of space in favor of a small page order.
3079 * Higher order allocations also allow the placement of more objects in a
3080 * slab and thereby reduce object handling overhead. If the user has
3081 * requested a higher mininum order then we start with that one instead of
3082 * the smallest order which will fit the object.
3084 static inline int slab_order(int size, int min_objects,
3085 int max_order, int fract_leftover, int reserved)
3089 int min_order = slub_min_order;
3091 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3092 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3094 for (order = max(min_order, get_order(min_objects * size + reserved));
3095 order <= max_order; order++) {
3097 unsigned long slab_size = PAGE_SIZE << order;
3099 rem = (slab_size - reserved) % size;
3101 if (rem <= slab_size / fract_leftover)
3108 static inline int calculate_order(int size, int reserved)
3116 * Attempt to find best configuration for a slab. This
3117 * works by first attempting to generate a layout with
3118 * the best configuration and backing off gradually.
3120 * First we increase the acceptable waste in a slab. Then
3121 * we reduce the minimum objects required in a slab.
3123 min_objects = slub_min_objects;
3125 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3126 max_objects = order_objects(slub_max_order, size, reserved);
3127 min_objects = min(min_objects, max_objects);
3129 while (min_objects > 1) {
3131 while (fraction >= 4) {
3132 order = slab_order(size, min_objects,
3133 slub_max_order, fraction, reserved);
3134 if (order <= slub_max_order)
3142 * We were unable to place multiple objects in a slab. Now
3143 * lets see if we can place a single object there.
3145 order = slab_order(size, 1, slub_max_order, 1, reserved);
3146 if (order <= slub_max_order)
3150 * Doh this slab cannot be placed using slub_max_order.
3152 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3153 if (order < MAX_ORDER)
3159 init_kmem_cache_node(struct kmem_cache_node *n)
3162 raw_spin_lock_init(&n->list_lock);
3163 INIT_LIST_HEAD(&n->partial);
3164 #ifdef CONFIG_SLUB_DEBUG
3165 atomic_long_set(&n->nr_slabs, 0);
3166 atomic_long_set(&n->total_objects, 0);
3167 INIT_LIST_HEAD(&n->full);
3171 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3173 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3174 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3177 * Must align to double word boundary for the double cmpxchg
3178 * instructions to work; see __pcpu_double_call_return_bool().
3180 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3181 2 * sizeof(void *));
3186 init_kmem_cache_cpus(s);
3191 static struct kmem_cache *kmem_cache_node;
3194 * No kmalloc_node yet so do it by hand. We know that this is the first
3195 * slab on the node for this slabcache. There are no concurrent accesses
3198 * Note that this function only works on the kmem_cache_node
3199 * when allocating for the kmem_cache_node. This is used for bootstrapping
3200 * memory on a fresh node that has no slab structures yet.
3202 static void early_kmem_cache_node_alloc(int node)
3205 struct kmem_cache_node *n;
3207 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3209 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3212 if (page_to_nid(page) != node) {
3213 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3214 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3219 page->freelist = get_freepointer(kmem_cache_node, n);
3222 kmem_cache_node->node[node] = n;
3223 #ifdef CONFIG_SLUB_DEBUG
3224 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3225 init_tracking(kmem_cache_node, n);
3227 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3228 init_kmem_cache_node(n);
3229 inc_slabs_node(kmem_cache_node, node, page->objects);
3232 * No locks need to be taken here as it has just been
3233 * initialized and there is no concurrent access.
3235 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3238 static void free_kmem_cache_nodes(struct kmem_cache *s)
3241 struct kmem_cache_node *n;
3243 for_each_kmem_cache_node(s, node, n) {
3244 kmem_cache_free(kmem_cache_node, n);
3245 s->node[node] = NULL;
3249 static int init_kmem_cache_nodes(struct kmem_cache *s)
3253 for_each_node_state(node, N_NORMAL_MEMORY) {
3254 struct kmem_cache_node *n;
3256 if (slab_state == DOWN) {
3257 early_kmem_cache_node_alloc(node);
3260 n = kmem_cache_alloc_node(kmem_cache_node,
3264 free_kmem_cache_nodes(s);
3269 init_kmem_cache_node(n);
3274 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3276 if (min < MIN_PARTIAL)
3278 else if (min > MAX_PARTIAL)
3280 s->min_partial = min;
3284 * calculate_sizes() determines the order and the distribution of data within
3287 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3289 unsigned long flags = s->flags;
3290 unsigned long size = s->object_size;
3294 * Round up object size to the next word boundary. We can only
3295 * place the free pointer at word boundaries and this determines
3296 * the possible location of the free pointer.
3298 size = ALIGN(size, sizeof(void *));
3300 #ifdef CONFIG_SLUB_DEBUG
3302 * Determine if we can poison the object itself. If the user of
3303 * the slab may touch the object after free or before allocation
3304 * then we should never poison the object itself.
3306 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3308 s->flags |= __OBJECT_POISON;
3310 s->flags &= ~__OBJECT_POISON;
3314 * If we are Redzoning then check if there is some space between the
3315 * end of the object and the free pointer. If not then add an
3316 * additional word to have some bytes to store Redzone information.
3318 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3319 size += sizeof(void *);
3323 * With that we have determined the number of bytes in actual use
3324 * by the object. This is the potential offset to the free pointer.
3328 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3331 * Relocate free pointer after the object if it is not
3332 * permitted to overwrite the first word of the object on
3335 * This is the case if we do RCU, have a constructor or
3336 * destructor or are poisoning the objects.
3339 size += sizeof(void *);
3342 #ifdef CONFIG_SLUB_DEBUG
3343 if (flags & SLAB_STORE_USER)
3345 * Need to store information about allocs and frees after
3348 size += 2 * sizeof(struct track);
3350 if (flags & SLAB_RED_ZONE)
3352 * Add some empty padding so that we can catch
3353 * overwrites from earlier objects rather than let
3354 * tracking information or the free pointer be
3355 * corrupted if a user writes before the start
3358 size += sizeof(void *);
3362 * SLUB stores one object immediately after another beginning from
3363 * offset 0. In order to align the objects we have to simply size
3364 * each object to conform to the alignment.
3366 size = ALIGN(size, s->align);
3368 if (forced_order >= 0)
3369 order = forced_order;
3371 order = calculate_order(size, s->reserved);
3378 s->allocflags |= __GFP_COMP;
3380 if (s->flags & SLAB_CACHE_DMA)
3381 s->allocflags |= GFP_DMA;
3383 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3384 s->allocflags |= __GFP_RECLAIMABLE;
3387 * Determine the number of objects per slab
3389 s->oo = oo_make(order, size, s->reserved);
3390 s->min = oo_make(get_order(size), size, s->reserved);
3391 if (oo_objects(s->oo) > oo_objects(s->max))
3394 return !!oo_objects(s->oo);
3397 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3399 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3402 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3403 s->reserved = sizeof(struct rcu_head);
3405 if (!calculate_sizes(s, -1))
3407 if (disable_higher_order_debug) {
3409 * Disable debugging flags that store metadata if the min slab
3412 if (get_order(s->size) > get_order(s->object_size)) {
3413 s->flags &= ~DEBUG_METADATA_FLAGS;
3415 if (!calculate_sizes(s, -1))
3420 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3421 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3422 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3423 /* Enable fast mode */
3424 s->flags |= __CMPXCHG_DOUBLE;
3428 * The larger the object size is, the more pages we want on the partial
3429 * list to avoid pounding the page allocator excessively.
3431 set_min_partial(s, ilog2(s->size) / 2);
3434 * cpu_partial determined the maximum number of objects kept in the
3435 * per cpu partial lists of a processor.
3437 * Per cpu partial lists mainly contain slabs that just have one
3438 * object freed. If they are used for allocation then they can be
3439 * filled up again with minimal effort. The slab will never hit the
3440 * per node partial lists and therefore no locking will be required.
3442 * This setting also determines
3444 * A) The number of objects from per cpu partial slabs dumped to the
3445 * per node list when we reach the limit.
3446 * B) The number of objects in cpu partial slabs to extract from the
3447 * per node list when we run out of per cpu objects. We only fetch
3448 * 50% to keep some capacity around for frees.
3450 if (!kmem_cache_has_cpu_partial(s))
3452 else if (s->size >= PAGE_SIZE)
3454 else if (s->size >= 1024)
3456 else if (s->size >= 256)
3457 s->cpu_partial = 13;
3459 s->cpu_partial = 30;
3462 s->remote_node_defrag_ratio = 1000;
3464 if (!init_kmem_cache_nodes(s))
3467 if (alloc_kmem_cache_cpus(s))
3470 free_kmem_cache_nodes(s);
3472 if (flags & SLAB_PANIC)
3473 panic("Cannot create slab %s size=%lu realsize=%u "
3474 "order=%u offset=%u flags=%lx\n",
3475 s->name, (unsigned long)s->size, s->size,
3476 oo_order(s->oo), s->offset, flags);
3480 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3483 #ifdef CONFIG_SLUB_DEBUG
3484 void *addr = page_address(page);
3486 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3487 sizeof(long), GFP_ATOMIC);
3490 slab_err(s, page, text, s->name);
3493 get_map(s, page, map);
3494 for_each_object(p, s, addr, page->objects) {
3496 if (!test_bit(slab_index(p, s, addr), map)) {
3497 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3498 print_tracking(s, p);
3507 * Attempt to free all partial slabs on a node.
3508 * This is called from kmem_cache_close(). We must be the last thread
3509 * using the cache and therefore we do not need to lock anymore.
3511 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3513 struct page *page, *h;
3515 list_for_each_entry_safe(page, h, &n->partial, lru) {
3517 __remove_partial(n, page);
3518 discard_slab(s, page);
3520 list_slab_objects(s, page,
3521 "Objects remaining in %s on kmem_cache_close()");
3527 * Release all resources used by a slab cache.
3529 static inline int kmem_cache_close(struct kmem_cache *s)
3532 struct kmem_cache_node *n;
3535 /* Attempt to free all objects */
3536 for_each_kmem_cache_node(s, node, n) {
3538 if (n->nr_partial || slabs_node(s, node))
3541 free_percpu(s->cpu_slab);
3542 free_kmem_cache_nodes(s);
3546 int __kmem_cache_shutdown(struct kmem_cache *s)
3548 return kmem_cache_close(s);
3551 /********************************************************************
3553 *******************************************************************/
3555 static int __init setup_slub_min_order(char *str)
3557 get_option(&str, &slub_min_order);
3562 __setup("slub_min_order=", setup_slub_min_order);
3564 static int __init setup_slub_max_order(char *str)
3566 get_option(&str, &slub_max_order);
3567 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3572 __setup("slub_max_order=", setup_slub_max_order);
3574 static int __init setup_slub_min_objects(char *str)
3576 get_option(&str, &slub_min_objects);
3581 __setup("slub_min_objects=", setup_slub_min_objects);
3583 void *__kmalloc(size_t size, gfp_t flags)
3585 struct kmem_cache *s;
3588 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3589 return kmalloc_large(size, flags);
3591 s = kmalloc_slab(size, flags);
3593 if (unlikely(ZERO_OR_NULL_PTR(s)))
3596 ret = slab_alloc(s, flags, _RET_IP_);
3598 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3600 kasan_kmalloc(s, ret, size);
3604 EXPORT_SYMBOL(__kmalloc);
3607 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3612 flags |= __GFP_COMP | __GFP_NOTRACK;
3613 page = alloc_kmem_pages_node(node, flags, get_order(size));
3615 ptr = page_address(page);
3617 kmalloc_large_node_hook(ptr, size, flags);
3621 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3623 struct kmem_cache *s;
3626 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3627 ret = kmalloc_large_node(size, flags, node);
3629 trace_kmalloc_node(_RET_IP_, ret,
3630 size, PAGE_SIZE << get_order(size),
3636 s = kmalloc_slab(size, flags);
3638 if (unlikely(ZERO_OR_NULL_PTR(s)))
3641 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3643 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3645 kasan_kmalloc(s, ret, size);
3649 EXPORT_SYMBOL(__kmalloc_node);
3652 static size_t __ksize(const void *object)
3656 if (unlikely(object == ZERO_SIZE_PTR))
3659 page = virt_to_head_page(object);
3661 if (unlikely(!PageSlab(page))) {
3662 WARN_ON(!PageCompound(page));
3663 return PAGE_SIZE << compound_order(page);
3666 return slab_ksize(page->slab_cache);
3669 size_t ksize(const void *object)
3671 size_t size = __ksize(object);
3672 /* We assume that ksize callers could use whole allocated area,
3673 so we need unpoison this area. */
3674 kasan_krealloc(object, size);
3677 EXPORT_SYMBOL(ksize);
3679 void kfree(const void *x)
3682 void *object = (void *)x;
3684 trace_kfree(_RET_IP_, x);
3686 if (unlikely(ZERO_OR_NULL_PTR(x)))
3689 page = virt_to_head_page(x);
3690 if (unlikely(!PageSlab(page))) {
3691 BUG_ON(!PageCompound(page));
3693 __free_kmem_pages(page, compound_order(page));
3696 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3698 EXPORT_SYMBOL(kfree);
3700 #define SHRINK_PROMOTE_MAX 32
3703 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3704 * up most to the head of the partial lists. New allocations will then
3705 * fill those up and thus they can be removed from the partial lists.
3707 * The slabs with the least items are placed last. This results in them
3708 * being allocated from last increasing the chance that the last objects
3709 * are freed in them.
3711 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3715 struct kmem_cache_node *n;
3718 struct list_head discard;
3719 struct list_head promote[SHRINK_PROMOTE_MAX];
3720 unsigned long flags;
3725 * Disable empty slabs caching. Used to avoid pinning offline
3726 * memory cgroups by kmem pages that can be freed.
3732 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3733 * so we have to make sure the change is visible.
3735 kick_all_cpus_sync();
3739 for_each_kmem_cache_node(s, node, n) {
3740 INIT_LIST_HEAD(&discard);
3741 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3742 INIT_LIST_HEAD(promote + i);
3744 raw_spin_lock_irqsave(&n->list_lock, flags);
3747 * Build lists of slabs to discard or promote.
3749 * Note that concurrent frees may occur while we hold the
3750 * list_lock. page->inuse here is the upper limit.
3752 list_for_each_entry_safe(page, t, &n->partial, lru) {
3753 int free = page->objects - page->inuse;
3755 /* Do not reread page->inuse */
3758 /* We do not keep full slabs on the list */
3761 if (free == page->objects) {
3762 list_move(&page->lru, &discard);
3764 } else if (free <= SHRINK_PROMOTE_MAX)
3765 list_move(&page->lru, promote + free - 1);
3769 * Promote the slabs filled up most to the head of the
3772 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3773 list_splice(promote + i, &n->partial);
3775 raw_spin_unlock_irqrestore(&n->list_lock, flags);
3777 /* Release empty slabs */
3778 list_for_each_entry_safe(page, t, &discard, lru)
3779 discard_slab(s, page);
3781 if (slabs_node(s, node))
3788 static int slab_mem_going_offline_callback(void *arg)
3790 struct kmem_cache *s;
3792 mutex_lock(&slab_mutex);
3793 list_for_each_entry(s, &slab_caches, list)
3794 __kmem_cache_shrink(s, false);
3795 mutex_unlock(&slab_mutex);
3800 static void slab_mem_offline_callback(void *arg)
3802 struct kmem_cache_node *n;
3803 struct kmem_cache *s;
3804 struct memory_notify *marg = arg;
3807 offline_node = marg->status_change_nid_normal;
3810 * If the node still has available memory. we need kmem_cache_node
3813 if (offline_node < 0)
3816 mutex_lock(&slab_mutex);
3817 list_for_each_entry(s, &slab_caches, list) {
3818 n = get_node(s, offline_node);
3821 * if n->nr_slabs > 0, slabs still exist on the node
3822 * that is going down. We were unable to free them,
3823 * and offline_pages() function shouldn't call this
3824 * callback. So, we must fail.
3826 BUG_ON(slabs_node(s, offline_node));
3828 s->node[offline_node] = NULL;
3829 kmem_cache_free(kmem_cache_node, n);
3832 mutex_unlock(&slab_mutex);
3835 static int slab_mem_going_online_callback(void *arg)
3837 struct kmem_cache_node *n;
3838 struct kmem_cache *s;
3839 struct memory_notify *marg = arg;
3840 int nid = marg->status_change_nid_normal;
3844 * If the node's memory is already available, then kmem_cache_node is
3845 * already created. Nothing to do.
3851 * We are bringing a node online. No memory is available yet. We must
3852 * allocate a kmem_cache_node structure in order to bring the node
3855 mutex_lock(&slab_mutex);
3856 list_for_each_entry(s, &slab_caches, list) {
3858 * XXX: kmem_cache_alloc_node will fallback to other nodes
3859 * since memory is not yet available from the node that
3862 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3867 init_kmem_cache_node(n);
3871 mutex_unlock(&slab_mutex);
3875 static int slab_memory_callback(struct notifier_block *self,
3876 unsigned long action, void *arg)
3881 case MEM_GOING_ONLINE:
3882 ret = slab_mem_going_online_callback(arg);
3884 case MEM_GOING_OFFLINE:
3885 ret = slab_mem_going_offline_callback(arg);
3888 case MEM_CANCEL_ONLINE:
3889 slab_mem_offline_callback(arg);
3892 case MEM_CANCEL_OFFLINE:
3896 ret = notifier_from_errno(ret);
3902 static struct notifier_block slab_memory_callback_nb = {
3903 .notifier_call = slab_memory_callback,
3904 .priority = SLAB_CALLBACK_PRI,
3907 /********************************************************************
3908 * Basic setup of slabs
3909 *******************************************************************/
3912 * Used for early kmem_cache structures that were allocated using
3913 * the page allocator. Allocate them properly then fix up the pointers
3914 * that may be pointing to the wrong kmem_cache structure.
3917 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3920 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3921 struct kmem_cache_node *n;
3923 memcpy(s, static_cache, kmem_cache->object_size);
3926 * This runs very early, and only the boot processor is supposed to be
3927 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3930 __flush_cpu_slab(s, smp_processor_id());
3931 for_each_kmem_cache_node(s, node, n) {
3934 list_for_each_entry(p, &n->partial, lru)
3937 #ifdef CONFIG_SLUB_DEBUG
3938 list_for_each_entry(p, &n->full, lru)
3942 slab_init_memcg_params(s);
3943 list_add(&s->list, &slab_caches);
3947 void __init kmem_cache_init(void)
3949 static __initdata struct kmem_cache boot_kmem_cache,
3950 boot_kmem_cache_node;
3953 for_each_possible_cpu(cpu) {
3954 raw_spin_lock_init(&per_cpu(slub_free_list, cpu).lock);
3955 INIT_LIST_HEAD(&per_cpu(slub_free_list, cpu).list);
3958 if (debug_guardpage_minorder())
3961 kmem_cache_node = &boot_kmem_cache_node;
3962 kmem_cache = &boot_kmem_cache;
3964 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3965 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3967 register_hotmemory_notifier(&slab_memory_callback_nb);
3969 /* Able to allocate the per node structures */
3970 slab_state = PARTIAL;
3972 create_boot_cache(kmem_cache, "kmem_cache",
3973 offsetof(struct kmem_cache, node) +
3974 nr_node_ids * sizeof(struct kmem_cache_node *),
3975 SLAB_HWCACHE_ALIGN);
3977 kmem_cache = bootstrap(&boot_kmem_cache);
3980 * Allocate kmem_cache_node properly from the kmem_cache slab.
3981 * kmem_cache_node is separately allocated so no need to
3982 * update any list pointers.
3984 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3986 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3987 setup_kmalloc_cache_index_table();
3988 create_kmalloc_caches(0);
3991 register_cpu_notifier(&slab_notifier);
3994 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3996 slub_min_order, slub_max_order, slub_min_objects,
3997 nr_cpu_ids, nr_node_ids);
4000 void __init kmem_cache_init_late(void)
4005 __kmem_cache_alias(const char *name, size_t size, size_t align,
4006 unsigned long flags, void (*ctor)(void *))
4008 struct kmem_cache *s, *c;
4010 s = find_mergeable(size, align, flags, name, ctor);
4015 * Adjust the object sizes so that we clear
4016 * the complete object on kzalloc.
4018 s->object_size = max(s->object_size, (int)size);
4019 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
4021 for_each_memcg_cache(c, s) {
4022 c->object_size = s->object_size;
4023 c->inuse = max_t(int, c->inuse,
4024 ALIGN(size, sizeof(void *)));
4027 if (sysfs_slab_alias(s, name)) {
4036 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4040 err = kmem_cache_open(s, flags);
4044 /* Mutex is not taken during early boot */
4045 if (slab_state <= UP)
4048 memcg_propagate_slab_attrs(s);
4049 err = sysfs_slab_add(s);
4051 kmem_cache_close(s);
4058 * Use the cpu notifier to insure that the cpu slabs are flushed when
4061 static int slab_cpuup_callback(struct notifier_block *nfb,
4062 unsigned long action, void *hcpu)
4064 long cpu = (long)hcpu;
4065 struct kmem_cache *s;
4066 unsigned long flags;
4069 case CPU_UP_CANCELED:
4070 case CPU_UP_CANCELED_FROZEN:
4072 case CPU_DEAD_FROZEN:
4073 mutex_lock(&slab_mutex);
4074 list_for_each_entry(s, &slab_caches, list) {
4075 local_irq_save(flags);
4076 __flush_cpu_slab(s, cpu);
4077 local_irq_restore(flags);
4079 mutex_unlock(&slab_mutex);
4087 static struct notifier_block slab_notifier = {
4088 .notifier_call = slab_cpuup_callback
4093 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4095 struct kmem_cache *s;
4098 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4099 return kmalloc_large(size, gfpflags);
4101 s = kmalloc_slab(size, gfpflags);
4103 if (unlikely(ZERO_OR_NULL_PTR(s)))
4106 ret = slab_alloc(s, gfpflags, caller);
4108 /* Honor the call site pointer we received. */
4109 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4115 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4116 int node, unsigned long caller)
4118 struct kmem_cache *s;
4121 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4122 ret = kmalloc_large_node(size, gfpflags, node);
4124 trace_kmalloc_node(caller, ret,
4125 size, PAGE_SIZE << get_order(size),
4131 s = kmalloc_slab(size, gfpflags);
4133 if (unlikely(ZERO_OR_NULL_PTR(s)))
4136 ret = slab_alloc_node(s, gfpflags, node, caller);
4138 /* Honor the call site pointer we received. */
4139 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4146 static int count_inuse(struct page *page)
4151 static int count_total(struct page *page)
4153 return page->objects;
4157 #ifdef CONFIG_SLUB_DEBUG
4158 static int validate_slab(struct kmem_cache *s, struct page *page,
4162 void *addr = page_address(page);
4164 if (!check_slab(s, page) ||
4165 !on_freelist(s, page, NULL))
4168 /* Now we know that a valid freelist exists */
4169 bitmap_zero(map, page->objects);
4171 get_map(s, page, map);
4172 for_each_object(p, s, addr, page->objects) {
4173 if (test_bit(slab_index(p, s, addr), map))
4174 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4178 for_each_object(p, s, addr, page->objects)
4179 if (!test_bit(slab_index(p, s, addr), map))
4180 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4185 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4189 validate_slab(s, page, map);
4193 static int validate_slab_node(struct kmem_cache *s,
4194 struct kmem_cache_node *n, unsigned long *map)
4196 unsigned long count = 0;
4198 unsigned long flags;
4200 raw_spin_lock_irqsave(&n->list_lock, flags);
4202 list_for_each_entry(page, &n->partial, lru) {
4203 validate_slab_slab(s, page, map);
4206 if (count != n->nr_partial)
4207 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4208 s->name, count, n->nr_partial);
4210 if (!(s->flags & SLAB_STORE_USER))
4213 list_for_each_entry(page, &n->full, lru) {
4214 validate_slab_slab(s, page, map);
4217 if (count != atomic_long_read(&n->nr_slabs))
4218 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4219 s->name, count, atomic_long_read(&n->nr_slabs));
4222 raw_spin_unlock_irqrestore(&n->list_lock, flags);
4226 static long validate_slab_cache(struct kmem_cache *s)
4229 unsigned long count = 0;
4230 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4231 sizeof(unsigned long), GFP_KERNEL);
4232 struct kmem_cache_node *n;
4238 for_each_kmem_cache_node(s, node, n)
4239 count += validate_slab_node(s, n, map);
4244 * Generate lists of code addresses where slabcache objects are allocated
4249 unsigned long count;
4256 DECLARE_BITMAP(cpus, NR_CPUS);
4262 unsigned long count;
4263 struct location *loc;
4266 static void free_loc_track(struct loc_track *t)
4269 free_pages((unsigned long)t->loc,
4270 get_order(sizeof(struct location) * t->max));
4273 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4278 order = get_order(sizeof(struct location) * max);
4280 l = (void *)__get_free_pages(flags, order);
4285 memcpy(l, t->loc, sizeof(struct location) * t->count);
4293 static int add_location(struct loc_track *t, struct kmem_cache *s,
4294 const struct track *track)
4296 long start, end, pos;
4298 unsigned long caddr;
4299 unsigned long age = jiffies - track->when;
4305 pos = start + (end - start + 1) / 2;
4308 * There is nothing at "end". If we end up there
4309 * we need to add something to before end.
4314 caddr = t->loc[pos].addr;
4315 if (track->addr == caddr) {
4321 if (age < l->min_time)
4323 if (age > l->max_time)
4326 if (track->pid < l->min_pid)
4327 l->min_pid = track->pid;
4328 if (track->pid > l->max_pid)
4329 l->max_pid = track->pid;
4331 cpumask_set_cpu(track->cpu,
4332 to_cpumask(l->cpus));
4334 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4338 if (track->addr < caddr)
4345 * Not found. Insert new tracking element.
4347 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4353 (t->count - pos) * sizeof(struct location));
4356 l->addr = track->addr;
4360 l->min_pid = track->pid;
4361 l->max_pid = track->pid;
4362 cpumask_clear(to_cpumask(l->cpus));
4363 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4364 nodes_clear(l->nodes);
4365 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4369 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4370 struct page *page, enum track_item alloc,
4373 void *addr = page_address(page);
4376 bitmap_zero(map, page->objects);
4377 get_map(s, page, map);
4379 for_each_object(p, s, addr, page->objects)
4380 if (!test_bit(slab_index(p, s, addr), map))
4381 add_location(t, s, get_track(s, p, alloc));
4384 static int list_locations(struct kmem_cache *s, char *buf,
4385 enum track_item alloc)
4389 struct loc_track t = { 0, 0, NULL };
4391 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4392 sizeof(unsigned long), GFP_KERNEL);
4393 struct kmem_cache_node *n;
4395 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4398 return sprintf(buf, "Out of memory\n");
4400 /* Push back cpu slabs */
4403 for_each_kmem_cache_node(s, node, n) {
4404 unsigned long flags;
4407 if (!atomic_long_read(&n->nr_slabs))
4410 raw_spin_lock_irqsave(&n->list_lock, flags);
4411 list_for_each_entry(page, &n->partial, lru)
4412 process_slab(&t, s, page, alloc, map);
4413 list_for_each_entry(page, &n->full, lru)
4414 process_slab(&t, s, page, alloc, map);
4415 raw_spin_unlock_irqrestore(&n->list_lock, flags);
4418 for (i = 0; i < t.count; i++) {
4419 struct location *l = &t.loc[i];
4421 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4423 len += sprintf(buf + len, "%7ld ", l->count);
4426 len += sprintf(buf + len, "%pS", (void *)l->addr);
4428 len += sprintf(buf + len, "<not-available>");
4430 if (l->sum_time != l->min_time) {
4431 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4433 (long)div_u64(l->sum_time, l->count),
4436 len += sprintf(buf + len, " age=%ld",
4439 if (l->min_pid != l->max_pid)
4440 len += sprintf(buf + len, " pid=%ld-%ld",
4441 l->min_pid, l->max_pid);
4443 len += sprintf(buf + len, " pid=%ld",
4446 if (num_online_cpus() > 1 &&
4447 !cpumask_empty(to_cpumask(l->cpus)) &&
4448 len < PAGE_SIZE - 60)
4449 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4451 cpumask_pr_args(to_cpumask(l->cpus)));
4453 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4454 len < PAGE_SIZE - 60)
4455 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4457 nodemask_pr_args(&l->nodes));
4459 len += sprintf(buf + len, "\n");
4465 len += sprintf(buf, "No data\n");
4470 #ifdef SLUB_RESILIENCY_TEST
4471 static void __init resiliency_test(void)
4475 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4477 pr_err("SLUB resiliency testing\n");
4478 pr_err("-----------------------\n");
4479 pr_err("A. Corruption after allocation\n");
4481 p = kzalloc(16, GFP_KERNEL);
4483 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4486 validate_slab_cache(kmalloc_caches[4]);
4488 /* Hmmm... The next two are dangerous */
4489 p = kzalloc(32, GFP_KERNEL);
4490 p[32 + sizeof(void *)] = 0x34;
4491 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4493 pr_err("If allocated object is overwritten then not detectable\n\n");
4495 validate_slab_cache(kmalloc_caches[5]);
4496 p = kzalloc(64, GFP_KERNEL);
4497 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4499 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4501 pr_err("If allocated object is overwritten then not detectable\n\n");
4502 validate_slab_cache(kmalloc_caches[6]);
4504 pr_err("\nB. Corruption after free\n");
4505 p = kzalloc(128, GFP_KERNEL);
4508 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4509 validate_slab_cache(kmalloc_caches[7]);
4511 p = kzalloc(256, GFP_KERNEL);
4514 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4515 validate_slab_cache(kmalloc_caches[8]);
4517 p = kzalloc(512, GFP_KERNEL);
4520 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4521 validate_slab_cache(kmalloc_caches[9]);
4525 static void resiliency_test(void) {};
4530 enum slab_stat_type {
4531 SL_ALL, /* All slabs */
4532 SL_PARTIAL, /* Only partially allocated slabs */
4533 SL_CPU, /* Only slabs used for cpu caches */
4534 SL_OBJECTS, /* Determine allocated objects not slabs */
4535 SL_TOTAL /* Determine object capacity not slabs */
4538 #define SO_ALL (1 << SL_ALL)
4539 #define SO_PARTIAL (1 << SL_PARTIAL)
4540 #define SO_CPU (1 << SL_CPU)
4541 #define SO_OBJECTS (1 << SL_OBJECTS)
4542 #define SO_TOTAL (1 << SL_TOTAL)
4544 static ssize_t show_slab_objects(struct kmem_cache *s,
4545 char *buf, unsigned long flags)
4547 unsigned long total = 0;
4550 unsigned long *nodes;
4552 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4556 if (flags & SO_CPU) {
4559 for_each_possible_cpu(cpu) {
4560 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4565 page = READ_ONCE(c->page);
4569 node = page_to_nid(page);
4570 if (flags & SO_TOTAL)
4572 else if (flags & SO_OBJECTS)
4580 page = READ_ONCE(c->partial);
4582 node = page_to_nid(page);
4583 if (flags & SO_TOTAL)
4585 else if (flags & SO_OBJECTS)
4596 #ifdef CONFIG_SLUB_DEBUG
4597 if (flags & SO_ALL) {
4598 struct kmem_cache_node *n;
4600 for_each_kmem_cache_node(s, node, n) {
4602 if (flags & SO_TOTAL)
4603 x = atomic_long_read(&n->total_objects);
4604 else if (flags & SO_OBJECTS)
4605 x = atomic_long_read(&n->total_objects) -
4606 count_partial(n, count_free);
4608 x = atomic_long_read(&n->nr_slabs);
4615 if (flags & SO_PARTIAL) {
4616 struct kmem_cache_node *n;
4618 for_each_kmem_cache_node(s, node, n) {
4619 if (flags & SO_TOTAL)
4620 x = count_partial(n, count_total);
4621 else if (flags & SO_OBJECTS)
4622 x = count_partial(n, count_inuse);
4629 x = sprintf(buf, "%lu", total);
4631 for (node = 0; node < nr_node_ids; node++)
4633 x += sprintf(buf + x, " N%d=%lu",
4638 return x + sprintf(buf + x, "\n");
4641 #ifdef CONFIG_SLUB_DEBUG
4642 static int any_slab_objects(struct kmem_cache *s)
4645 struct kmem_cache_node *n;
4647 for_each_kmem_cache_node(s, node, n)
4648 if (atomic_long_read(&n->total_objects))
4655 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4656 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4658 struct slab_attribute {
4659 struct attribute attr;
4660 ssize_t (*show)(struct kmem_cache *s, char *buf);
4661 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4664 #define SLAB_ATTR_RO(_name) \
4665 static struct slab_attribute _name##_attr = \
4666 __ATTR(_name, 0400, _name##_show, NULL)
4668 #define SLAB_ATTR(_name) \
4669 static struct slab_attribute _name##_attr = \
4670 __ATTR(_name, 0600, _name##_show, _name##_store)
4672 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4674 return sprintf(buf, "%d\n", s->size);
4676 SLAB_ATTR_RO(slab_size);
4678 static ssize_t align_show(struct kmem_cache *s, char *buf)
4680 return sprintf(buf, "%d\n", s->align);
4682 SLAB_ATTR_RO(align);
4684 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4686 return sprintf(buf, "%d\n", s->object_size);
4688 SLAB_ATTR_RO(object_size);
4690 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4692 return sprintf(buf, "%d\n", oo_objects(s->oo));
4694 SLAB_ATTR_RO(objs_per_slab);
4696 static ssize_t order_store(struct kmem_cache *s,
4697 const char *buf, size_t length)
4699 unsigned long order;
4702 err = kstrtoul(buf, 10, &order);
4706 if (order > slub_max_order || order < slub_min_order)
4709 calculate_sizes(s, order);
4713 static ssize_t order_show(struct kmem_cache *s, char *buf)
4715 return sprintf(buf, "%d\n", oo_order(s->oo));
4719 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4721 return sprintf(buf, "%lu\n", s->min_partial);
4724 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4730 err = kstrtoul(buf, 10, &min);
4734 set_min_partial(s, min);
4737 SLAB_ATTR(min_partial);
4739 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4741 return sprintf(buf, "%u\n", s->cpu_partial);
4744 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4747 unsigned long objects;
4750 err = kstrtoul(buf, 10, &objects);
4753 if (objects && !kmem_cache_has_cpu_partial(s))
4756 s->cpu_partial = objects;
4760 SLAB_ATTR(cpu_partial);
4762 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4766 return sprintf(buf, "%pS\n", s->ctor);
4770 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4772 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4774 SLAB_ATTR_RO(aliases);
4776 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4778 return show_slab_objects(s, buf, SO_PARTIAL);
4780 SLAB_ATTR_RO(partial);
4782 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4784 return show_slab_objects(s, buf, SO_CPU);
4786 SLAB_ATTR_RO(cpu_slabs);
4788 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4790 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4792 SLAB_ATTR_RO(objects);
4794 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4796 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4798 SLAB_ATTR_RO(objects_partial);
4800 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4807 for_each_online_cpu(cpu) {
4808 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4811 pages += page->pages;
4812 objects += page->pobjects;
4816 len = sprintf(buf, "%d(%d)", objects, pages);
4819 for_each_online_cpu(cpu) {
4820 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4822 if (page && len < PAGE_SIZE - 20)
4823 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4824 page->pobjects, page->pages);
4827 return len + sprintf(buf + len, "\n");
4829 SLAB_ATTR_RO(slabs_cpu_partial);
4831 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4833 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4836 static ssize_t reclaim_account_store(struct kmem_cache *s,
4837 const char *buf, size_t length)
4839 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4841 s->flags |= SLAB_RECLAIM_ACCOUNT;
4844 SLAB_ATTR(reclaim_account);
4846 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4848 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4850 SLAB_ATTR_RO(hwcache_align);
4852 #ifdef CONFIG_ZONE_DMA
4853 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4855 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4857 SLAB_ATTR_RO(cache_dma);
4860 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4862 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4864 SLAB_ATTR_RO(destroy_by_rcu);
4866 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4868 return sprintf(buf, "%d\n", s->reserved);
4870 SLAB_ATTR_RO(reserved);
4872 #ifdef CONFIG_SLUB_DEBUG
4873 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4875 return show_slab_objects(s, buf, SO_ALL);
4877 SLAB_ATTR_RO(slabs);
4879 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4881 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4883 SLAB_ATTR_RO(total_objects);
4885 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4887 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4890 static ssize_t sanity_checks_store(struct kmem_cache *s,
4891 const char *buf, size_t length)
4893 s->flags &= ~SLAB_DEBUG_FREE;
4894 if (buf[0] == '1') {
4895 s->flags &= ~__CMPXCHG_DOUBLE;
4896 s->flags |= SLAB_DEBUG_FREE;
4900 SLAB_ATTR(sanity_checks);
4902 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4904 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4907 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4911 * Tracing a merged cache is going to give confusing results
4912 * as well as cause other issues like converting a mergeable
4913 * cache into an umergeable one.
4915 if (s->refcount > 1)
4918 s->flags &= ~SLAB_TRACE;
4919 if (buf[0] == '1') {
4920 s->flags &= ~__CMPXCHG_DOUBLE;
4921 s->flags |= SLAB_TRACE;
4927 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4929 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4932 static ssize_t red_zone_store(struct kmem_cache *s,
4933 const char *buf, size_t length)
4935 if (any_slab_objects(s))
4938 s->flags &= ~SLAB_RED_ZONE;
4939 if (buf[0] == '1') {
4940 s->flags &= ~__CMPXCHG_DOUBLE;
4941 s->flags |= SLAB_RED_ZONE;
4943 calculate_sizes(s, -1);
4946 SLAB_ATTR(red_zone);
4948 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4950 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4953 static ssize_t poison_store(struct kmem_cache *s,
4954 const char *buf, size_t length)
4956 if (any_slab_objects(s))
4959 s->flags &= ~SLAB_POISON;
4960 if (buf[0] == '1') {
4961 s->flags &= ~__CMPXCHG_DOUBLE;
4962 s->flags |= SLAB_POISON;
4964 calculate_sizes(s, -1);
4969 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4971 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4974 static ssize_t store_user_store(struct kmem_cache *s,
4975 const char *buf, size_t length)
4977 if (any_slab_objects(s))
4980 s->flags &= ~SLAB_STORE_USER;
4981 if (buf[0] == '1') {
4982 s->flags &= ~__CMPXCHG_DOUBLE;
4983 s->flags |= SLAB_STORE_USER;
4985 calculate_sizes(s, -1);
4988 SLAB_ATTR(store_user);
4990 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4995 static ssize_t validate_store(struct kmem_cache *s,
4996 const char *buf, size_t length)
5000 if (buf[0] == '1') {
5001 ret = validate_slab_cache(s);
5007 SLAB_ATTR(validate);
5009 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5011 if (!(s->flags & SLAB_STORE_USER))
5013 return list_locations(s, buf, TRACK_ALLOC);
5015 SLAB_ATTR_RO(alloc_calls);
5017 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5019 if (!(s->flags & SLAB_STORE_USER))
5021 return list_locations(s, buf, TRACK_FREE);
5023 SLAB_ATTR_RO(free_calls);
5024 #endif /* CONFIG_SLUB_DEBUG */
5026 #ifdef CONFIG_FAILSLAB
5027 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5029 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5032 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5035 if (s->refcount > 1)
5038 s->flags &= ~SLAB_FAILSLAB;
5040 s->flags |= SLAB_FAILSLAB;
5043 SLAB_ATTR(failslab);
5046 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5051 static ssize_t shrink_store(struct kmem_cache *s,
5052 const char *buf, size_t length)
5055 kmem_cache_shrink(s);
5063 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5065 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5068 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5069 const char *buf, size_t length)
5071 unsigned long ratio;
5074 err = kstrtoul(buf, 10, &ratio);
5079 s->remote_node_defrag_ratio = ratio * 10;
5083 SLAB_ATTR(remote_node_defrag_ratio);
5086 #ifdef CONFIG_SLUB_STATS
5087 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5089 unsigned long sum = 0;
5092 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5097 for_each_online_cpu(cpu) {
5098 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5104 len = sprintf(buf, "%lu", sum);
5107 for_each_online_cpu(cpu) {
5108 if (data[cpu] && len < PAGE_SIZE - 20)
5109 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5113 return len + sprintf(buf + len, "\n");
5116 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5120 for_each_online_cpu(cpu)
5121 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5124 #define STAT_ATTR(si, text) \
5125 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5127 return show_stat(s, buf, si); \
5129 static ssize_t text##_store(struct kmem_cache *s, \
5130 const char *buf, size_t length) \
5132 if (buf[0] != '0') \
5134 clear_stat(s, si); \
5139 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5140 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5141 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5142 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5143 STAT_ATTR(FREE_FROZEN, free_frozen);
5144 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5145 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5146 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5147 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5148 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5149 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5150 STAT_ATTR(FREE_SLAB, free_slab);
5151 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5152 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5153 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5154 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5155 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5156 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5157 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5158 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5159 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5160 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5161 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5162 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5163 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5164 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5167 static struct attribute *slab_attrs[] = {
5168 &slab_size_attr.attr,
5169 &object_size_attr.attr,
5170 &objs_per_slab_attr.attr,
5172 &min_partial_attr.attr,
5173 &cpu_partial_attr.attr,
5175 &objects_partial_attr.attr,
5177 &cpu_slabs_attr.attr,
5181 &hwcache_align_attr.attr,
5182 &reclaim_account_attr.attr,
5183 &destroy_by_rcu_attr.attr,
5185 &reserved_attr.attr,
5186 &slabs_cpu_partial_attr.attr,
5187 #ifdef CONFIG_SLUB_DEBUG
5188 &total_objects_attr.attr,
5190 &sanity_checks_attr.attr,
5192 &red_zone_attr.attr,
5194 &store_user_attr.attr,
5195 &validate_attr.attr,
5196 &alloc_calls_attr.attr,
5197 &free_calls_attr.attr,
5199 #ifdef CONFIG_ZONE_DMA
5200 &cache_dma_attr.attr,
5203 &remote_node_defrag_ratio_attr.attr,
5205 #ifdef CONFIG_SLUB_STATS
5206 &alloc_fastpath_attr.attr,
5207 &alloc_slowpath_attr.attr,
5208 &free_fastpath_attr.attr,
5209 &free_slowpath_attr.attr,
5210 &free_frozen_attr.attr,
5211 &free_add_partial_attr.attr,
5212 &free_remove_partial_attr.attr,
5213 &alloc_from_partial_attr.attr,
5214 &alloc_slab_attr.attr,
5215 &alloc_refill_attr.attr,
5216 &alloc_node_mismatch_attr.attr,
5217 &free_slab_attr.attr,
5218 &cpuslab_flush_attr.attr,
5219 &deactivate_full_attr.attr,
5220 &deactivate_empty_attr.attr,
5221 &deactivate_to_head_attr.attr,
5222 &deactivate_to_tail_attr.attr,
5223 &deactivate_remote_frees_attr.attr,
5224 &deactivate_bypass_attr.attr,
5225 &order_fallback_attr.attr,
5226 &cmpxchg_double_fail_attr.attr,
5227 &cmpxchg_double_cpu_fail_attr.attr,
5228 &cpu_partial_alloc_attr.attr,
5229 &cpu_partial_free_attr.attr,
5230 &cpu_partial_node_attr.attr,
5231 &cpu_partial_drain_attr.attr,
5233 #ifdef CONFIG_FAILSLAB
5234 &failslab_attr.attr,
5240 static struct attribute_group slab_attr_group = {
5241 .attrs = slab_attrs,
5244 static ssize_t slab_attr_show(struct kobject *kobj,
5245 struct attribute *attr,
5248 struct slab_attribute *attribute;
5249 struct kmem_cache *s;
5252 attribute = to_slab_attr(attr);
5255 if (!attribute->show)
5258 err = attribute->show(s, buf);
5263 static ssize_t slab_attr_store(struct kobject *kobj,
5264 struct attribute *attr,
5265 const char *buf, size_t len)
5267 struct slab_attribute *attribute;
5268 struct kmem_cache *s;
5271 attribute = to_slab_attr(attr);
5274 if (!attribute->store)
5277 err = attribute->store(s, buf, len);
5278 #ifdef CONFIG_MEMCG_KMEM
5279 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5280 struct kmem_cache *c;
5282 mutex_lock(&slab_mutex);
5283 if (s->max_attr_size < len)
5284 s->max_attr_size = len;
5287 * This is a best effort propagation, so this function's return
5288 * value will be determined by the parent cache only. This is
5289 * basically because not all attributes will have a well
5290 * defined semantics for rollbacks - most of the actions will
5291 * have permanent effects.
5293 * Returning the error value of any of the children that fail
5294 * is not 100 % defined, in the sense that users seeing the
5295 * error code won't be able to know anything about the state of
5298 * Only returning the error code for the parent cache at least
5299 * has well defined semantics. The cache being written to
5300 * directly either failed or succeeded, in which case we loop
5301 * through the descendants with best-effort propagation.
5303 for_each_memcg_cache(c, s)
5304 attribute->store(c, buf, len);
5305 mutex_unlock(&slab_mutex);
5311 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5313 #ifdef CONFIG_MEMCG_KMEM
5315 char *buffer = NULL;
5316 struct kmem_cache *root_cache;
5318 if (is_root_cache(s))
5321 root_cache = s->memcg_params.root_cache;
5324 * This mean this cache had no attribute written. Therefore, no point
5325 * in copying default values around
5327 if (!root_cache->max_attr_size)
5330 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5333 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5335 if (!attr || !attr->store || !attr->show)
5339 * It is really bad that we have to allocate here, so we will
5340 * do it only as a fallback. If we actually allocate, though,
5341 * we can just use the allocated buffer until the end.
5343 * Most of the slub attributes will tend to be very small in
5344 * size, but sysfs allows buffers up to a page, so they can
5345 * theoretically happen.
5349 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5352 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5353 if (WARN_ON(!buffer))
5358 attr->show(root_cache, buf);
5359 attr->store(s, buf, strlen(buf));
5363 free_page((unsigned long)buffer);
5367 static void kmem_cache_release(struct kobject *k)
5369 slab_kmem_cache_release(to_slab(k));
5372 static const struct sysfs_ops slab_sysfs_ops = {
5373 .show = slab_attr_show,
5374 .store = slab_attr_store,
5377 static struct kobj_type slab_ktype = {
5378 .sysfs_ops = &slab_sysfs_ops,
5379 .release = kmem_cache_release,
5382 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5384 struct kobj_type *ktype = get_ktype(kobj);
5386 if (ktype == &slab_ktype)
5391 static const struct kset_uevent_ops slab_uevent_ops = {
5392 .filter = uevent_filter,
5395 static struct kset *slab_kset;
5397 static inline struct kset *cache_kset(struct kmem_cache *s)
5399 #ifdef CONFIG_MEMCG_KMEM
5400 if (!is_root_cache(s))
5401 return s->memcg_params.root_cache->memcg_kset;
5406 #define ID_STR_LENGTH 64
5408 /* Create a unique string id for a slab cache:
5410 * Format :[flags-]size
5412 static char *create_unique_id(struct kmem_cache *s)
5414 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5421 * First flags affecting slabcache operations. We will only
5422 * get here for aliasable slabs so we do not need to support
5423 * too many flags. The flags here must cover all flags that
5424 * are matched during merging to guarantee that the id is
5427 if (s->flags & SLAB_CACHE_DMA)
5429 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5431 if (s->flags & SLAB_DEBUG_FREE)
5433 if (!(s->flags & SLAB_NOTRACK))
5437 p += sprintf(p, "%07d", s->size);
5439 BUG_ON(p > name + ID_STR_LENGTH - 1);
5443 static int sysfs_slab_add(struct kmem_cache *s)
5447 int unmergeable = slab_unmergeable(s);
5451 * Slabcache can never be merged so we can use the name proper.
5452 * This is typically the case for debug situations. In that
5453 * case we can catch duplicate names easily.
5455 sysfs_remove_link(&slab_kset->kobj, s->name);
5459 * Create a unique name for the slab as a target
5462 name = create_unique_id(s);
5465 s->kobj.kset = cache_kset(s);
5466 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5470 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5474 #ifdef CONFIG_MEMCG_KMEM
5475 if (is_root_cache(s)) {
5476 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5477 if (!s->memcg_kset) {
5484 kobject_uevent(&s->kobj, KOBJ_ADD);
5486 /* Setup first alias */
5487 sysfs_slab_alias(s, s->name);
5494 kobject_del(&s->kobj);
5498 void sysfs_slab_remove(struct kmem_cache *s)
5500 if (slab_state < FULL)
5502 * Sysfs has not been setup yet so no need to remove the
5507 #ifdef CONFIG_MEMCG_KMEM
5508 kset_unregister(s->memcg_kset);
5510 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5511 kobject_del(&s->kobj);
5512 kobject_put(&s->kobj);
5516 * Need to buffer aliases during bootup until sysfs becomes
5517 * available lest we lose that information.
5519 struct saved_alias {
5520 struct kmem_cache *s;
5522 struct saved_alias *next;
5525 static struct saved_alias *alias_list;
5527 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5529 struct saved_alias *al;
5531 if (slab_state == FULL) {
5533 * If we have a leftover link then remove it.
5535 sysfs_remove_link(&slab_kset->kobj, name);
5536 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5539 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5545 al->next = alias_list;
5550 static int __init slab_sysfs_init(void)
5552 struct kmem_cache *s;
5555 mutex_lock(&slab_mutex);
5557 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5559 mutex_unlock(&slab_mutex);
5560 pr_err("Cannot register slab subsystem.\n");
5566 list_for_each_entry(s, &slab_caches, list) {
5567 err = sysfs_slab_add(s);
5569 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5573 while (alias_list) {
5574 struct saved_alias *al = alias_list;
5576 alias_list = alias_list->next;
5577 err = sysfs_slab_alias(al->s, al->name);
5579 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5584 mutex_unlock(&slab_mutex);
5589 __initcall(slab_sysfs_init);
5590 #endif /* CONFIG_SYSFS */
5593 * The /proc/slabinfo ABI
5595 #ifdef CONFIG_SLABINFO
5596 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5598 unsigned long nr_slabs = 0;
5599 unsigned long nr_objs = 0;
5600 unsigned long nr_free = 0;
5602 struct kmem_cache_node *n;
5604 for_each_kmem_cache_node(s, node, n) {
5605 nr_slabs += node_nr_slabs(n);
5606 nr_objs += node_nr_objs(n);
5607 nr_free += count_partial(n, count_free);
5610 sinfo->active_objs = nr_objs - nr_free;
5611 sinfo->num_objs = nr_objs;
5612 sinfo->active_slabs = nr_slabs;
5613 sinfo->num_slabs = nr_slabs;
5614 sinfo->objects_per_slab = oo_objects(s->oo);
5615 sinfo->cache_order = oo_order(s->oo);
5618 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5622 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5623 size_t count, loff_t *ppos)
5627 #endif /* CONFIG_SLABINFO */