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 VM_BUG_ON_PAGE(PageTail(page), page);
342 bit_spin_lock(PG_locked, &page->flags);
345 static __always_inline void slab_unlock(struct page *page)
347 VM_BUG_ON_PAGE(PageTail(page), page);
348 __bit_spin_unlock(PG_locked, &page->flags);
351 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
354 tmp.counters = counters_new;
356 * page->counters can cover frozen/inuse/objects as well
357 * as page->_count. If we assign to ->counters directly
358 * we run the risk of losing updates to page->_count, so
359 * be careful and only assign to the fields we need.
361 page->frozen = tmp.frozen;
362 page->inuse = tmp.inuse;
363 page->objects = tmp.objects;
366 /* Interrupts must be disabled (for the fallback code to work right) */
367 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
368 void *freelist_old, unsigned long counters_old,
369 void *freelist_new, unsigned long counters_new,
372 VM_BUG_ON(!irqs_disabled());
373 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
374 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
375 if (s->flags & __CMPXCHG_DOUBLE) {
376 if (cmpxchg_double(&page->freelist, &page->counters,
377 freelist_old, counters_old,
378 freelist_new, counters_new))
384 if (page->freelist == freelist_old &&
385 page->counters == counters_old) {
386 page->freelist = freelist_new;
387 set_page_slub_counters(page, counters_new);
395 stat(s, CMPXCHG_DOUBLE_FAIL);
397 #ifdef SLUB_DEBUG_CMPXCHG
398 pr_info("%s %s: cmpxchg double redo ", n, s->name);
404 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
405 void *freelist_old, unsigned long counters_old,
406 void *freelist_new, unsigned long counters_new,
409 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
410 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
411 if (s->flags & __CMPXCHG_DOUBLE) {
412 if (cmpxchg_double(&page->freelist, &page->counters,
413 freelist_old, counters_old,
414 freelist_new, counters_new))
421 local_irq_save(flags);
423 if (page->freelist == freelist_old &&
424 page->counters == counters_old) {
425 page->freelist = freelist_new;
426 set_page_slub_counters(page, counters_new);
428 local_irq_restore(flags);
432 local_irq_restore(flags);
436 stat(s, CMPXCHG_DOUBLE_FAIL);
438 #ifdef SLUB_DEBUG_CMPXCHG
439 pr_info("%s %s: cmpxchg double redo ", n, s->name);
445 #ifdef CONFIG_SLUB_DEBUG
447 * Determine a map of object in use on a page.
449 * Node listlock must be held to guarantee that the page does
450 * not vanish from under us.
452 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
455 void *addr = page_address(page);
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(slab_index(p, s, addr), map);
464 #ifdef CONFIG_SLUB_DEBUG_ON
465 static int slub_debug = DEBUG_DEFAULT_FLAGS;
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 static noinline struct kmem_cache_node *free_debug_processing(
1069 struct kmem_cache *s, struct page *page, void *object,
1070 unsigned long addr, unsigned long *flags)
1072 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1074 spin_lock_irqsave(&n->list_lock, *flags);
1077 if (!check_slab(s, page))
1080 if (!check_valid_pointer(s, page, object)) {
1081 slab_err(s, page, "Invalid object pointer 0x%p", object);
1085 if (on_freelist(s, page, object)) {
1086 object_err(s, page, object, "Object already free");
1090 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1093 if (unlikely(s != page->slab_cache)) {
1094 if (!PageSlab(page)) {
1095 slab_err(s, page, "Attempt to free object(0x%p) "
1096 "outside of slab", object);
1097 } else if (!page->slab_cache) {
1098 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1102 object_err(s, page, object,
1103 "page slab pointer corrupt.");
1107 if (s->flags & SLAB_STORE_USER)
1108 set_track(s, object, TRACK_FREE, addr);
1109 trace(s, page, object, 0);
1110 init_object(s, object, SLUB_RED_INACTIVE);
1114 * Keep node_lock to preserve integrity
1115 * until the object is actually freed
1121 spin_unlock_irqrestore(&n->list_lock, *flags);
1122 slab_fix(s, "Object at 0x%p not freed", object);
1126 static int __init setup_slub_debug(char *str)
1128 slub_debug = DEBUG_DEFAULT_FLAGS;
1129 if (*str++ != '=' || !*str)
1131 * No options specified. Switch on full debugging.
1137 * No options but restriction on slabs. This means full
1138 * debugging for slabs matching a pattern.
1145 * Switch off all debugging measures.
1150 * Determine which debug features should be switched on
1152 for (; *str && *str != ','; str++) {
1153 switch (tolower(*str)) {
1155 slub_debug |= SLAB_DEBUG_FREE;
1158 slub_debug |= SLAB_RED_ZONE;
1161 slub_debug |= SLAB_POISON;
1164 slub_debug |= SLAB_STORE_USER;
1167 slub_debug |= SLAB_TRACE;
1170 slub_debug |= SLAB_FAILSLAB;
1174 * Avoid enabling debugging on caches if its minimum
1175 * order would increase as a result.
1177 disable_higher_order_debug = 1;
1180 pr_err("slub_debug option '%c' unknown. skipped\n",
1187 slub_debug_slabs = str + 1;
1192 __setup("slub_debug", setup_slub_debug);
1194 unsigned long kmem_cache_flags(unsigned long object_size,
1195 unsigned long flags, const char *name,
1196 void (*ctor)(void *))
1199 * Enable debugging if selected on the kernel commandline.
1201 if (slub_debug && (!slub_debug_slabs || (name &&
1202 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1203 flags |= slub_debug;
1208 static inline void setup_object_debug(struct kmem_cache *s,
1209 struct page *page, void *object) {}
1211 static inline int alloc_debug_processing(struct kmem_cache *s,
1212 struct page *page, void *object, unsigned long addr) { return 0; }
1214 static inline struct kmem_cache_node *free_debug_processing(
1215 struct kmem_cache *s, struct page *page, void *object,
1216 unsigned long addr, unsigned long *flags) { return NULL; }
1218 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1220 static inline int check_object(struct kmem_cache *s, struct page *page,
1221 void *object, u8 val) { return 1; }
1222 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1223 struct page *page) {}
1224 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1225 struct page *page) {}
1226 unsigned long kmem_cache_flags(unsigned long object_size,
1227 unsigned long flags, const char *name,
1228 void (*ctor)(void *))
1232 #define slub_debug 0
1234 #define disable_higher_order_debug 0
1236 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1238 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1240 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1242 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1245 #endif /* CONFIG_SLUB_DEBUG */
1248 * Hooks for other subsystems that check memory allocations. In a typical
1249 * production configuration these hooks all should produce no code at all.
1251 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1253 kmemleak_alloc(ptr, size, 1, flags);
1254 kasan_kmalloc_large(ptr, size);
1257 static inline void kfree_hook(const void *x)
1260 kasan_kfree_large(x);
1263 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1266 flags &= gfp_allowed_mask;
1267 lockdep_trace_alloc(flags);
1268 might_sleep_if(flags & __GFP_WAIT);
1270 if (should_failslab(s->object_size, flags, s->flags))
1273 return memcg_kmem_get_cache(s, flags);
1276 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1277 gfp_t flags, void *object)
1279 flags &= gfp_allowed_mask;
1280 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1281 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1282 memcg_kmem_put_cache(s);
1283 kasan_slab_alloc(s, object);
1286 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1288 kmemleak_free_recursive(x, s->flags);
1291 * Trouble is that we may no longer disable interrupts in the fast path
1292 * So in order to make the debug calls that expect irqs to be
1293 * disabled we need to disable interrupts temporarily.
1295 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1297 unsigned long flags;
1299 local_irq_save(flags);
1300 kmemcheck_slab_free(s, x, s->object_size);
1301 debug_check_no_locks_freed(x, s->object_size);
1302 local_irq_restore(flags);
1305 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1306 debug_check_no_obj_freed(x, s->object_size);
1308 kasan_slab_free(s, x);
1312 * Slab allocation and freeing
1314 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1315 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1318 int order = oo_order(oo);
1320 flags |= __GFP_NOTRACK;
1322 if (memcg_charge_slab(s, flags, order))
1325 if (node == NUMA_NO_NODE)
1326 page = alloc_pages(flags, order);
1328 page = alloc_pages_exact_node(node, flags, order);
1331 memcg_uncharge_slab(s, order);
1336 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1339 struct kmem_cache_order_objects oo = s->oo;
1342 flags &= gfp_allowed_mask;
1344 if (flags & __GFP_WAIT)
1347 flags |= s->allocflags;
1350 * Let the initial higher-order allocation fail under memory pressure
1351 * so we fall-back to the minimum order allocation.
1353 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1355 page = alloc_slab_page(s, alloc_gfp, node, oo);
1356 if (unlikely(!page)) {
1360 * Allocation may have failed due to fragmentation.
1361 * Try a lower order alloc if possible
1363 page = alloc_slab_page(s, alloc_gfp, node, oo);
1366 stat(s, ORDER_FALLBACK);
1369 if (kmemcheck_enabled && page
1370 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1371 int pages = 1 << oo_order(oo);
1373 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1376 * Objects from caches that have a constructor don't get
1377 * cleared when they're allocated, so we need to do it here.
1380 kmemcheck_mark_uninitialized_pages(page, pages);
1382 kmemcheck_mark_unallocated_pages(page, pages);
1385 if (flags & __GFP_WAIT)
1386 local_irq_disable();
1390 page->objects = oo_objects(oo);
1391 mod_zone_page_state(page_zone(page),
1392 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1393 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1399 static void setup_object(struct kmem_cache *s, struct page *page,
1402 setup_object_debug(s, page, object);
1403 if (unlikely(s->ctor)) {
1404 kasan_unpoison_object_data(s, object);
1406 kasan_poison_object_data(s, object);
1410 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1418 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1419 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1423 page = allocate_slab(s,
1424 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1428 order = compound_order(page);
1429 inc_slabs_node(s, page_to_nid(page), page->objects);
1430 page->slab_cache = s;
1431 __SetPageSlab(page);
1432 if (page->pfmemalloc)
1433 SetPageSlabPfmemalloc(page);
1435 start = page_address(page);
1437 if (unlikely(s->flags & SLAB_POISON))
1438 memset(start, POISON_INUSE, PAGE_SIZE << order);
1440 kasan_poison_slab(page);
1442 for_each_object_idx(p, idx, s, start, page->objects) {
1443 setup_object(s, page, p);
1444 if (likely(idx < page->objects))
1445 set_freepointer(s, p, p + s->size);
1447 set_freepointer(s, p, NULL);
1450 page->freelist = start;
1451 page->inuse = page->objects;
1457 static void __free_slab(struct kmem_cache *s, struct page *page)
1459 int order = compound_order(page);
1460 int pages = 1 << order;
1462 if (kmem_cache_debug(s)) {
1465 slab_pad_check(s, page);
1466 for_each_object(p, s, page_address(page),
1468 check_object(s, page, p, SLUB_RED_INACTIVE);
1471 kmemcheck_free_shadow(page, compound_order(page));
1473 mod_zone_page_state(page_zone(page),
1474 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1475 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1478 __ClearPageSlabPfmemalloc(page);
1479 __ClearPageSlab(page);
1481 page_mapcount_reset(page);
1482 if (current->reclaim_state)
1483 current->reclaim_state->reclaimed_slab += pages;
1484 __free_pages(page, order);
1485 memcg_uncharge_slab(s, order);
1488 #define need_reserve_slab_rcu \
1489 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1491 static void rcu_free_slab(struct rcu_head *h)
1495 if (need_reserve_slab_rcu)
1496 page = virt_to_head_page(h);
1498 page = container_of((struct list_head *)h, struct page, lru);
1500 __free_slab(page->slab_cache, page);
1503 static void free_slab(struct kmem_cache *s, struct page *page)
1505 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1506 struct rcu_head *head;
1508 if (need_reserve_slab_rcu) {
1509 int order = compound_order(page);
1510 int offset = (PAGE_SIZE << order) - s->reserved;
1512 VM_BUG_ON(s->reserved != sizeof(*head));
1513 head = page_address(page) + offset;
1516 * RCU free overloads the RCU head over the LRU
1518 head = (void *)&page->lru;
1521 call_rcu(head, rcu_free_slab);
1523 __free_slab(s, page);
1526 static void discard_slab(struct kmem_cache *s, struct page *page)
1528 dec_slabs_node(s, page_to_nid(page), page->objects);
1533 * Management of partially allocated slabs.
1536 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1539 if (tail == DEACTIVATE_TO_TAIL)
1540 list_add_tail(&page->lru, &n->partial);
1542 list_add(&page->lru, &n->partial);
1545 static inline void add_partial(struct kmem_cache_node *n,
1546 struct page *page, int tail)
1548 lockdep_assert_held(&n->list_lock);
1549 __add_partial(n, page, tail);
1553 __remove_partial(struct kmem_cache_node *n, struct page *page)
1555 list_del(&page->lru);
1559 static inline void remove_partial(struct kmem_cache_node *n,
1562 lockdep_assert_held(&n->list_lock);
1563 __remove_partial(n, page);
1567 * Remove slab from the partial list, freeze it and
1568 * return the pointer to the freelist.
1570 * Returns a list of objects or NULL if it fails.
1572 static inline void *acquire_slab(struct kmem_cache *s,
1573 struct kmem_cache_node *n, struct page *page,
1574 int mode, int *objects)
1577 unsigned long counters;
1580 lockdep_assert_held(&n->list_lock);
1583 * Zap the freelist and set the frozen bit.
1584 * The old freelist is the list of objects for the
1585 * per cpu allocation list.
1587 freelist = page->freelist;
1588 counters = page->counters;
1589 new.counters = counters;
1590 *objects = new.objects - new.inuse;
1592 new.inuse = page->objects;
1593 new.freelist = NULL;
1595 new.freelist = freelist;
1598 VM_BUG_ON(new.frozen);
1601 if (!__cmpxchg_double_slab(s, page,
1603 new.freelist, new.counters,
1607 remove_partial(n, page);
1612 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1613 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1616 * Try to allocate a partial slab from a specific node.
1618 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1619 struct kmem_cache_cpu *c, gfp_t flags)
1621 struct page *page, *page2;
1622 void *object = NULL;
1627 * Racy check. If we mistakenly see no partial slabs then we
1628 * just allocate an empty slab. If we mistakenly try to get a
1629 * partial slab and there is none available then get_partials()
1632 if (!n || !n->nr_partial)
1635 spin_lock(&n->list_lock);
1636 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1639 if (!pfmemalloc_match(page, flags))
1642 t = acquire_slab(s, n, page, object == NULL, &objects);
1646 available += objects;
1649 stat(s, ALLOC_FROM_PARTIAL);
1652 put_cpu_partial(s, page, 0);
1653 stat(s, CPU_PARTIAL_NODE);
1655 if (!kmem_cache_has_cpu_partial(s)
1656 || available > s->cpu_partial / 2)
1660 spin_unlock(&n->list_lock);
1665 * Get a page from somewhere. Search in increasing NUMA distances.
1667 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1668 struct kmem_cache_cpu *c)
1671 struct zonelist *zonelist;
1674 enum zone_type high_zoneidx = gfp_zone(flags);
1676 unsigned int cpuset_mems_cookie;
1679 * The defrag ratio allows a configuration of the tradeoffs between
1680 * inter node defragmentation and node local allocations. A lower
1681 * defrag_ratio increases the tendency to do local allocations
1682 * instead of attempting to obtain partial slabs from other nodes.
1684 * If the defrag_ratio is set to 0 then kmalloc() always
1685 * returns node local objects. If the ratio is higher then kmalloc()
1686 * may return off node objects because partial slabs are obtained
1687 * from other nodes and filled up.
1689 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1690 * defrag_ratio = 1000) then every (well almost) allocation will
1691 * first attempt to defrag slab caches on other nodes. This means
1692 * scanning over all nodes to look for partial slabs which may be
1693 * expensive if we do it every time we are trying to find a slab
1694 * with available objects.
1696 if (!s->remote_node_defrag_ratio ||
1697 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1701 cpuset_mems_cookie = read_mems_allowed_begin();
1702 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1703 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1704 struct kmem_cache_node *n;
1706 n = get_node(s, zone_to_nid(zone));
1708 if (n && cpuset_zone_allowed(zone, flags) &&
1709 n->nr_partial > s->min_partial) {
1710 object = get_partial_node(s, n, c, flags);
1713 * Don't check read_mems_allowed_retry()
1714 * here - if mems_allowed was updated in
1715 * parallel, that was a harmless race
1716 * between allocation and the cpuset
1723 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1729 * Get a partial page, lock it and return it.
1731 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1732 struct kmem_cache_cpu *c)
1735 int searchnode = node;
1737 if (node == NUMA_NO_NODE)
1738 searchnode = numa_mem_id();
1739 else if (!node_present_pages(node))
1740 searchnode = node_to_mem_node(node);
1742 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1743 if (object || node != NUMA_NO_NODE)
1746 return get_any_partial(s, flags, c);
1749 #ifdef CONFIG_PREEMPT
1751 * Calculate the next globally unique transaction for disambiguiation
1752 * during cmpxchg. The transactions start with the cpu number and are then
1753 * incremented by CONFIG_NR_CPUS.
1755 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1758 * No preemption supported therefore also no need to check for
1764 static inline unsigned long next_tid(unsigned long tid)
1766 return tid + TID_STEP;
1769 static inline unsigned int tid_to_cpu(unsigned long tid)
1771 return tid % TID_STEP;
1774 static inline unsigned long tid_to_event(unsigned long tid)
1776 return tid / TID_STEP;
1779 static inline unsigned int init_tid(int cpu)
1784 static inline void note_cmpxchg_failure(const char *n,
1785 const struct kmem_cache *s, unsigned long tid)
1787 #ifdef SLUB_DEBUG_CMPXCHG
1788 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1790 pr_info("%s %s: cmpxchg redo ", n, s->name);
1792 #ifdef CONFIG_PREEMPT
1793 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1794 pr_warn("due to cpu change %d -> %d\n",
1795 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1798 if (tid_to_event(tid) != tid_to_event(actual_tid))
1799 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1800 tid_to_event(tid), tid_to_event(actual_tid));
1802 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1803 actual_tid, tid, next_tid(tid));
1805 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1808 static void init_kmem_cache_cpus(struct kmem_cache *s)
1812 for_each_possible_cpu(cpu)
1813 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1817 * Remove the cpu slab
1819 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1822 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1823 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1825 enum slab_modes l = M_NONE, m = M_NONE;
1827 int tail = DEACTIVATE_TO_HEAD;
1831 if (page->freelist) {
1832 stat(s, DEACTIVATE_REMOTE_FREES);
1833 tail = DEACTIVATE_TO_TAIL;
1837 * Stage one: Free all available per cpu objects back
1838 * to the page freelist while it is still frozen. Leave the
1841 * There is no need to take the list->lock because the page
1844 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1846 unsigned long counters;
1849 prior = page->freelist;
1850 counters = page->counters;
1851 set_freepointer(s, freelist, prior);
1852 new.counters = counters;
1854 VM_BUG_ON(!new.frozen);
1856 } while (!__cmpxchg_double_slab(s, page,
1858 freelist, new.counters,
1859 "drain percpu freelist"));
1861 freelist = nextfree;
1865 * Stage two: Ensure that the page is unfrozen while the
1866 * list presence reflects the actual number of objects
1869 * We setup the list membership and then perform a cmpxchg
1870 * with the count. If there is a mismatch then the page
1871 * is not unfrozen but the page is on the wrong list.
1873 * Then we restart the process which may have to remove
1874 * the page from the list that we just put it on again
1875 * because the number of objects in the slab may have
1880 old.freelist = page->freelist;
1881 old.counters = page->counters;
1882 VM_BUG_ON(!old.frozen);
1884 /* Determine target state of the slab */
1885 new.counters = old.counters;
1888 set_freepointer(s, freelist, old.freelist);
1889 new.freelist = freelist;
1891 new.freelist = old.freelist;
1895 if (!new.inuse && n->nr_partial >= s->min_partial)
1897 else if (new.freelist) {
1902 * Taking the spinlock removes the possiblity
1903 * that acquire_slab() will see a slab page that
1906 spin_lock(&n->list_lock);
1910 if (kmem_cache_debug(s) && !lock) {
1913 * This also ensures that the scanning of full
1914 * slabs from diagnostic functions will not see
1917 spin_lock(&n->list_lock);
1925 remove_partial(n, page);
1927 else if (l == M_FULL)
1929 remove_full(s, n, page);
1931 if (m == M_PARTIAL) {
1933 add_partial(n, page, tail);
1936 } else if (m == M_FULL) {
1938 stat(s, DEACTIVATE_FULL);
1939 add_full(s, n, page);
1945 if (!__cmpxchg_double_slab(s, page,
1946 old.freelist, old.counters,
1947 new.freelist, new.counters,
1952 spin_unlock(&n->list_lock);
1955 stat(s, DEACTIVATE_EMPTY);
1956 discard_slab(s, page);
1962 * Unfreeze all the cpu partial slabs.
1964 * This function must be called with interrupts disabled
1965 * for the cpu using c (or some other guarantee must be there
1966 * to guarantee no concurrent accesses).
1968 static void unfreeze_partials(struct kmem_cache *s,
1969 struct kmem_cache_cpu *c)
1971 #ifdef CONFIG_SLUB_CPU_PARTIAL
1972 struct kmem_cache_node *n = NULL, *n2 = NULL;
1973 struct page *page, *discard_page = NULL;
1975 while ((page = c->partial)) {
1979 c->partial = page->next;
1981 n2 = get_node(s, page_to_nid(page));
1984 spin_unlock(&n->list_lock);
1987 spin_lock(&n->list_lock);
1992 old.freelist = page->freelist;
1993 old.counters = page->counters;
1994 VM_BUG_ON(!old.frozen);
1996 new.counters = old.counters;
1997 new.freelist = old.freelist;
2001 } while (!__cmpxchg_double_slab(s, page,
2002 old.freelist, old.counters,
2003 new.freelist, new.counters,
2004 "unfreezing slab"));
2006 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2007 page->next = discard_page;
2008 discard_page = page;
2010 add_partial(n, page, DEACTIVATE_TO_TAIL);
2011 stat(s, FREE_ADD_PARTIAL);
2016 spin_unlock(&n->list_lock);
2018 while (discard_page) {
2019 page = discard_page;
2020 discard_page = discard_page->next;
2022 stat(s, DEACTIVATE_EMPTY);
2023 discard_slab(s, page);
2030 * Put a page that was just frozen (in __slab_free) into a partial page
2031 * slot if available. This is done without interrupts disabled and without
2032 * preemption disabled. The cmpxchg is racy and may put the partial page
2033 * onto a random cpus partial slot.
2035 * If we did not find a slot then simply move all the partials to the
2036 * per node partial list.
2038 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2040 #ifdef CONFIG_SLUB_CPU_PARTIAL
2041 struct page *oldpage;
2049 oldpage = this_cpu_read(s->cpu_slab->partial);
2052 pobjects = oldpage->pobjects;
2053 pages = oldpage->pages;
2054 if (drain && pobjects > s->cpu_partial) {
2055 unsigned long flags;
2057 * partial array is full. Move the existing
2058 * set to the per node partial list.
2060 local_irq_save(flags);
2061 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2062 local_irq_restore(flags);
2066 stat(s, CPU_PARTIAL_DRAIN);
2071 pobjects += page->objects - page->inuse;
2073 page->pages = pages;
2074 page->pobjects = pobjects;
2075 page->next = oldpage;
2077 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2079 if (unlikely(!s->cpu_partial)) {
2080 unsigned long flags;
2082 local_irq_save(flags);
2083 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2084 local_irq_restore(flags);
2090 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2092 stat(s, CPUSLAB_FLUSH);
2093 deactivate_slab(s, c->page, c->freelist);
2095 c->tid = next_tid(c->tid);
2103 * Called from IPI handler with interrupts disabled.
2105 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2107 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2113 unfreeze_partials(s, c);
2117 static void flush_cpu_slab(void *d)
2119 struct kmem_cache *s = d;
2121 __flush_cpu_slab(s, smp_processor_id());
2124 static bool has_cpu_slab(int cpu, void *info)
2126 struct kmem_cache *s = info;
2127 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2129 return c->page || c->partial;
2132 static void flush_all(struct kmem_cache *s)
2134 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2138 * Check if the objects in a per cpu structure fit numa
2139 * locality expectations.
2141 static inline int node_match(struct page *page, int node)
2144 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2150 #ifdef CONFIG_SLUB_DEBUG
2151 static int count_free(struct page *page)
2153 return page->objects - page->inuse;
2156 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2158 return atomic_long_read(&n->total_objects);
2160 #endif /* CONFIG_SLUB_DEBUG */
2162 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2163 static unsigned long count_partial(struct kmem_cache_node *n,
2164 int (*get_count)(struct page *))
2166 unsigned long flags;
2167 unsigned long x = 0;
2170 spin_lock_irqsave(&n->list_lock, flags);
2171 list_for_each_entry(page, &n->partial, lru)
2172 x += get_count(page);
2173 spin_unlock_irqrestore(&n->list_lock, flags);
2176 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2178 static noinline void
2179 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2181 #ifdef CONFIG_SLUB_DEBUG
2182 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2183 DEFAULT_RATELIMIT_BURST);
2185 struct kmem_cache_node *n;
2187 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2190 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2192 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2193 s->name, s->object_size, s->size, oo_order(s->oo),
2196 if (oo_order(s->min) > get_order(s->object_size))
2197 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2200 for_each_kmem_cache_node(s, node, n) {
2201 unsigned long nr_slabs;
2202 unsigned long nr_objs;
2203 unsigned long nr_free;
2205 nr_free = count_partial(n, count_free);
2206 nr_slabs = node_nr_slabs(n);
2207 nr_objs = node_nr_objs(n);
2209 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2210 node, nr_slabs, nr_objs, nr_free);
2215 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2216 int node, struct kmem_cache_cpu **pc)
2219 struct kmem_cache_cpu *c = *pc;
2222 freelist = get_partial(s, flags, node, c);
2227 page = new_slab(s, flags, node);
2229 c = raw_cpu_ptr(s->cpu_slab);
2234 * No other reference to the page yet so we can
2235 * muck around with it freely without cmpxchg
2237 freelist = page->freelist;
2238 page->freelist = NULL;
2240 stat(s, ALLOC_SLAB);
2249 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2251 if (unlikely(PageSlabPfmemalloc(page)))
2252 return gfp_pfmemalloc_allowed(gfpflags);
2258 * Check the page->freelist of a page and either transfer the freelist to the
2259 * per cpu freelist or deactivate the page.
2261 * The page is still frozen if the return value is not NULL.
2263 * If this function returns NULL then the page has been unfrozen.
2265 * This function must be called with interrupt disabled.
2267 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2270 unsigned long counters;
2274 freelist = page->freelist;
2275 counters = page->counters;
2277 new.counters = counters;
2278 VM_BUG_ON(!new.frozen);
2280 new.inuse = page->objects;
2281 new.frozen = freelist != NULL;
2283 } while (!__cmpxchg_double_slab(s, page,
2292 * Slow path. The lockless freelist is empty or we need to perform
2295 * Processing is still very fast if new objects have been freed to the
2296 * regular freelist. In that case we simply take over the regular freelist
2297 * as the lockless freelist and zap the regular freelist.
2299 * If that is not working then we fall back to the partial lists. We take the
2300 * first element of the freelist as the object to allocate now and move the
2301 * rest of the freelist to the lockless freelist.
2303 * And if we were unable to get a new slab from the partial slab lists then
2304 * we need to allocate a new slab. This is the slowest path since it involves
2305 * a call to the page allocator and the setup of a new slab.
2307 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2308 unsigned long addr, struct kmem_cache_cpu *c)
2312 unsigned long flags;
2314 local_irq_save(flags);
2315 #ifdef CONFIG_PREEMPT
2317 * We may have been preempted and rescheduled on a different
2318 * cpu before disabling interrupts. Need to reload cpu area
2321 c = this_cpu_ptr(s->cpu_slab);
2329 if (unlikely(!node_match(page, node))) {
2330 int searchnode = node;
2332 if (node != NUMA_NO_NODE && !node_present_pages(node))
2333 searchnode = node_to_mem_node(node);
2335 if (unlikely(!node_match(page, searchnode))) {
2336 stat(s, ALLOC_NODE_MISMATCH);
2337 deactivate_slab(s, page, c->freelist);
2345 * By rights, we should be searching for a slab page that was
2346 * PFMEMALLOC but right now, we are losing the pfmemalloc
2347 * information when the page leaves the per-cpu allocator
2349 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2350 deactivate_slab(s, page, c->freelist);
2356 /* must check again c->freelist in case of cpu migration or IRQ */
2357 freelist = c->freelist;
2361 freelist = get_freelist(s, page);
2365 stat(s, DEACTIVATE_BYPASS);
2369 stat(s, ALLOC_REFILL);
2373 * freelist is pointing to the list of objects to be used.
2374 * page is pointing to the page from which the objects are obtained.
2375 * That page must be frozen for per cpu allocations to work.
2377 VM_BUG_ON(!c->page->frozen);
2378 c->freelist = get_freepointer(s, freelist);
2379 c->tid = next_tid(c->tid);
2380 local_irq_restore(flags);
2386 page = c->page = c->partial;
2387 c->partial = page->next;
2388 stat(s, CPU_PARTIAL_ALLOC);
2393 freelist = new_slab_objects(s, gfpflags, node, &c);
2395 if (unlikely(!freelist)) {
2396 slab_out_of_memory(s, gfpflags, node);
2397 local_irq_restore(flags);
2402 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2405 /* Only entered in the debug case */
2406 if (kmem_cache_debug(s) &&
2407 !alloc_debug_processing(s, page, freelist, addr))
2408 goto new_slab; /* Slab failed checks. Next slab needed */
2410 deactivate_slab(s, page, get_freepointer(s, freelist));
2413 local_irq_restore(flags);
2418 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2419 * have the fastpath folded into their functions. So no function call
2420 * overhead for requests that can be satisfied on the fastpath.
2422 * The fastpath works by first checking if the lockless freelist can be used.
2423 * If not then __slab_alloc is called for slow processing.
2425 * Otherwise we can simply pick the next object from the lockless free list.
2427 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2428 gfp_t gfpflags, int node, unsigned long addr)
2431 struct kmem_cache_cpu *c;
2435 s = slab_pre_alloc_hook(s, gfpflags);
2440 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2441 * enabled. We may switch back and forth between cpus while
2442 * reading from one cpu area. That does not matter as long
2443 * as we end up on the original cpu again when doing the cmpxchg.
2445 * We should guarantee that tid and kmem_cache are retrieved on
2446 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2447 * to check if it is matched or not.
2450 tid = this_cpu_read(s->cpu_slab->tid);
2451 c = raw_cpu_ptr(s->cpu_slab);
2452 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2453 unlikely(tid != READ_ONCE(c->tid)));
2456 * Irqless object alloc/free algorithm used here depends on sequence
2457 * of fetching cpu_slab's data. tid should be fetched before anything
2458 * on c to guarantee that object and page associated with previous tid
2459 * won't be used with current tid. If we fetch tid first, object and
2460 * page could be one associated with next tid and our alloc/free
2461 * request will be failed. In this case, we will retry. So, no problem.
2466 * The transaction ids are globally unique per cpu and per operation on
2467 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2468 * occurs on the right processor and that there was no operation on the
2469 * linked list in between.
2472 object = c->freelist;
2474 if (unlikely(!object || !node_match(page, node))) {
2475 object = __slab_alloc(s, gfpflags, node, addr, c);
2476 stat(s, ALLOC_SLOWPATH);
2478 void *next_object = get_freepointer_safe(s, object);
2481 * The cmpxchg will only match if there was no additional
2482 * operation and if we are on the right processor.
2484 * The cmpxchg does the following atomically (without lock
2486 * 1. Relocate first pointer to the current per cpu area.
2487 * 2. Verify that tid and freelist have not been changed
2488 * 3. If they were not changed replace tid and freelist
2490 * Since this is without lock semantics the protection is only
2491 * against code executing on this cpu *not* from access by
2494 if (unlikely(!this_cpu_cmpxchg_double(
2495 s->cpu_slab->freelist, s->cpu_slab->tid,
2497 next_object, next_tid(tid)))) {
2499 note_cmpxchg_failure("slab_alloc", s, tid);
2502 prefetch_freepointer(s, next_object);
2503 stat(s, ALLOC_FASTPATH);
2506 if (unlikely(gfpflags & __GFP_ZERO) && object)
2507 memset(object, 0, s->object_size);
2509 slab_post_alloc_hook(s, gfpflags, object);
2514 static __always_inline void *slab_alloc(struct kmem_cache *s,
2515 gfp_t gfpflags, unsigned long addr)
2517 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2520 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2522 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2524 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2529 EXPORT_SYMBOL(kmem_cache_alloc);
2531 #ifdef CONFIG_TRACING
2532 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2534 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2535 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2536 kasan_kmalloc(s, ret, size);
2539 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2543 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2545 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2547 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2548 s->object_size, s->size, gfpflags, node);
2552 EXPORT_SYMBOL(kmem_cache_alloc_node);
2554 #ifdef CONFIG_TRACING
2555 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2557 int node, size_t size)
2559 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2561 trace_kmalloc_node(_RET_IP_, ret,
2562 size, s->size, gfpflags, node);
2564 kasan_kmalloc(s, ret, size);
2567 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2572 * Slow path handling. This may still be called frequently since objects
2573 * have a longer lifetime than the cpu slabs in most processing loads.
2575 * So we still attempt to reduce cache line usage. Just take the slab
2576 * lock and free the item. If there is no additional partial page
2577 * handling required then we can return immediately.
2579 static void __slab_free(struct kmem_cache *s, struct page *page,
2580 void *x, unsigned long addr)
2583 void **object = (void *)x;
2586 unsigned long counters;
2587 struct kmem_cache_node *n = NULL;
2588 unsigned long uninitialized_var(flags);
2590 stat(s, FREE_SLOWPATH);
2592 if (kmem_cache_debug(s) &&
2593 !(n = free_debug_processing(s, page, x, addr, &flags)))
2598 spin_unlock_irqrestore(&n->list_lock, flags);
2601 prior = page->freelist;
2602 counters = page->counters;
2603 set_freepointer(s, object, prior);
2604 new.counters = counters;
2605 was_frozen = new.frozen;
2607 if ((!new.inuse || !prior) && !was_frozen) {
2609 if (kmem_cache_has_cpu_partial(s) && !prior) {
2612 * Slab was on no list before and will be
2614 * We can defer the list move and instead
2619 } else { /* Needs to be taken off a list */
2621 n = get_node(s, page_to_nid(page));
2623 * Speculatively acquire the list_lock.
2624 * If the cmpxchg does not succeed then we may
2625 * drop the list_lock without any processing.
2627 * Otherwise the list_lock will synchronize with
2628 * other processors updating the list of slabs.
2630 spin_lock_irqsave(&n->list_lock, flags);
2635 } while (!cmpxchg_double_slab(s, page,
2637 object, new.counters,
2643 * If we just froze the page then put it onto the
2644 * per cpu partial list.
2646 if (new.frozen && !was_frozen) {
2647 put_cpu_partial(s, page, 1);
2648 stat(s, CPU_PARTIAL_FREE);
2651 * The list lock was not taken therefore no list
2652 * activity can be necessary.
2655 stat(s, FREE_FROZEN);
2659 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2663 * Objects left in the slab. If it was not on the partial list before
2666 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2667 if (kmem_cache_debug(s))
2668 remove_full(s, n, page);
2669 add_partial(n, page, DEACTIVATE_TO_TAIL);
2670 stat(s, FREE_ADD_PARTIAL);
2672 spin_unlock_irqrestore(&n->list_lock, flags);
2678 * Slab on the partial list.
2680 remove_partial(n, page);
2681 stat(s, FREE_REMOVE_PARTIAL);
2683 /* Slab must be on the full list */
2684 remove_full(s, n, page);
2687 spin_unlock_irqrestore(&n->list_lock, flags);
2689 discard_slab(s, page);
2693 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2694 * can perform fastpath freeing without additional function calls.
2696 * The fastpath is only possible if we are freeing to the current cpu slab
2697 * of this processor. This typically the case if we have just allocated
2700 * If fastpath is not possible then fall back to __slab_free where we deal
2701 * with all sorts of special processing.
2703 static __always_inline void slab_free(struct kmem_cache *s,
2704 struct page *page, void *x, unsigned long addr)
2706 void **object = (void *)x;
2707 struct kmem_cache_cpu *c;
2710 slab_free_hook(s, x);
2714 * Determine the currently cpus per cpu slab.
2715 * The cpu may change afterward. However that does not matter since
2716 * data is retrieved via this pointer. If we are on the same cpu
2717 * during the cmpxchg then the free will succedd.
2720 tid = this_cpu_read(s->cpu_slab->tid);
2721 c = raw_cpu_ptr(s->cpu_slab);
2722 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2723 unlikely(tid != READ_ONCE(c->tid)));
2725 /* Same with comment on barrier() in slab_alloc_node() */
2728 if (likely(page == c->page)) {
2729 set_freepointer(s, object, c->freelist);
2731 if (unlikely(!this_cpu_cmpxchg_double(
2732 s->cpu_slab->freelist, s->cpu_slab->tid,
2734 object, next_tid(tid)))) {
2736 note_cmpxchg_failure("slab_free", s, tid);
2739 stat(s, FREE_FASTPATH);
2741 __slab_free(s, page, x, addr);
2745 void kmem_cache_free(struct kmem_cache *s, void *x)
2747 s = cache_from_obj(s, x);
2750 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2751 trace_kmem_cache_free(_RET_IP_, x);
2753 EXPORT_SYMBOL(kmem_cache_free);
2755 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2757 __kmem_cache_free_bulk(s, size, p);
2759 EXPORT_SYMBOL(kmem_cache_free_bulk);
2761 bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2764 return kmem_cache_alloc_bulk(s, flags, size, p);
2766 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2770 * Object placement in a slab is made very easy because we always start at
2771 * offset 0. If we tune the size of the object to the alignment then we can
2772 * get the required alignment by putting one properly sized object after
2775 * Notice that the allocation order determines the sizes of the per cpu
2776 * caches. Each processor has always one slab available for allocations.
2777 * Increasing the allocation order reduces the number of times that slabs
2778 * must be moved on and off the partial lists and is therefore a factor in
2783 * Mininum / Maximum order of slab pages. This influences locking overhead
2784 * and slab fragmentation. A higher order reduces the number of partial slabs
2785 * and increases the number of allocations possible without having to
2786 * take the list_lock.
2788 static int slub_min_order;
2789 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2790 static int slub_min_objects;
2793 * Calculate the order of allocation given an slab object size.
2795 * The order of allocation has significant impact on performance and other
2796 * system components. Generally order 0 allocations should be preferred since
2797 * order 0 does not cause fragmentation in the page allocator. Larger objects
2798 * be problematic to put into order 0 slabs because there may be too much
2799 * unused space left. We go to a higher order if more than 1/16th of the slab
2802 * In order to reach satisfactory performance we must ensure that a minimum
2803 * number of objects is in one slab. Otherwise we may generate too much
2804 * activity on the partial lists which requires taking the list_lock. This is
2805 * less a concern for large slabs though which are rarely used.
2807 * slub_max_order specifies the order where we begin to stop considering the
2808 * number of objects in a slab as critical. If we reach slub_max_order then
2809 * we try to keep the page order as low as possible. So we accept more waste
2810 * of space in favor of a small page order.
2812 * Higher order allocations also allow the placement of more objects in a
2813 * slab and thereby reduce object handling overhead. If the user has
2814 * requested a higher mininum order then we start with that one instead of
2815 * the smallest order which will fit the object.
2817 static inline int slab_order(int size, int min_objects,
2818 int max_order, int fract_leftover, int reserved)
2822 int min_order = slub_min_order;
2824 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2825 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2827 for (order = max(min_order,
2828 fls(min_objects * size - 1) - PAGE_SHIFT);
2829 order <= max_order; order++) {
2831 unsigned long slab_size = PAGE_SIZE << order;
2833 if (slab_size < min_objects * size + reserved)
2836 rem = (slab_size - reserved) % size;
2838 if (rem <= slab_size / fract_leftover)
2846 static inline int calculate_order(int size, int reserved)
2854 * Attempt to find best configuration for a slab. This
2855 * works by first attempting to generate a layout with
2856 * the best configuration and backing off gradually.
2858 * First we reduce the acceptable waste in a slab. Then
2859 * we reduce the minimum objects required in a slab.
2861 min_objects = slub_min_objects;
2863 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2864 max_objects = order_objects(slub_max_order, size, reserved);
2865 min_objects = min(min_objects, max_objects);
2867 while (min_objects > 1) {
2869 while (fraction >= 4) {
2870 order = slab_order(size, min_objects,
2871 slub_max_order, fraction, reserved);
2872 if (order <= slub_max_order)
2880 * We were unable to place multiple objects in a slab. Now
2881 * lets see if we can place a single object there.
2883 order = slab_order(size, 1, slub_max_order, 1, reserved);
2884 if (order <= slub_max_order)
2888 * Doh this slab cannot be placed using slub_max_order.
2890 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2891 if (order < MAX_ORDER)
2897 init_kmem_cache_node(struct kmem_cache_node *n)
2900 spin_lock_init(&n->list_lock);
2901 INIT_LIST_HEAD(&n->partial);
2902 #ifdef CONFIG_SLUB_DEBUG
2903 atomic_long_set(&n->nr_slabs, 0);
2904 atomic_long_set(&n->total_objects, 0);
2905 INIT_LIST_HEAD(&n->full);
2909 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2911 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2912 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2915 * Must align to double word boundary for the double cmpxchg
2916 * instructions to work; see __pcpu_double_call_return_bool().
2918 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2919 2 * sizeof(void *));
2924 init_kmem_cache_cpus(s);
2929 static struct kmem_cache *kmem_cache_node;
2932 * No kmalloc_node yet so do it by hand. We know that this is the first
2933 * slab on the node for this slabcache. There are no concurrent accesses
2936 * Note that this function only works on the kmem_cache_node
2937 * when allocating for the kmem_cache_node. This is used for bootstrapping
2938 * memory on a fresh node that has no slab structures yet.
2940 static void early_kmem_cache_node_alloc(int node)
2943 struct kmem_cache_node *n;
2945 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2947 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2950 if (page_to_nid(page) != node) {
2951 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2952 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2957 page->freelist = get_freepointer(kmem_cache_node, n);
2960 kmem_cache_node->node[node] = n;
2961 #ifdef CONFIG_SLUB_DEBUG
2962 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2963 init_tracking(kmem_cache_node, n);
2965 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
2966 init_kmem_cache_node(n);
2967 inc_slabs_node(kmem_cache_node, node, page->objects);
2970 * No locks need to be taken here as it has just been
2971 * initialized and there is no concurrent access.
2973 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2976 static void free_kmem_cache_nodes(struct kmem_cache *s)
2979 struct kmem_cache_node *n;
2981 for_each_kmem_cache_node(s, node, n) {
2982 kmem_cache_free(kmem_cache_node, n);
2983 s->node[node] = NULL;
2987 static int init_kmem_cache_nodes(struct kmem_cache *s)
2991 for_each_node_state(node, N_NORMAL_MEMORY) {
2992 struct kmem_cache_node *n;
2994 if (slab_state == DOWN) {
2995 early_kmem_cache_node_alloc(node);
2998 n = kmem_cache_alloc_node(kmem_cache_node,
3002 free_kmem_cache_nodes(s);
3007 init_kmem_cache_node(n);
3012 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3014 if (min < MIN_PARTIAL)
3016 else if (min > MAX_PARTIAL)
3018 s->min_partial = min;
3022 * calculate_sizes() determines the order and the distribution of data within
3025 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3027 unsigned long flags = s->flags;
3028 unsigned long size = s->object_size;
3032 * Round up object size to the next word boundary. We can only
3033 * place the free pointer at word boundaries and this determines
3034 * the possible location of the free pointer.
3036 size = ALIGN(size, sizeof(void *));
3038 #ifdef CONFIG_SLUB_DEBUG
3040 * Determine if we can poison the object itself. If the user of
3041 * the slab may touch the object after free or before allocation
3042 * then we should never poison the object itself.
3044 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3046 s->flags |= __OBJECT_POISON;
3048 s->flags &= ~__OBJECT_POISON;
3052 * If we are Redzoning then check if there is some space between the
3053 * end of the object and the free pointer. If not then add an
3054 * additional word to have some bytes to store Redzone information.
3056 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3057 size += sizeof(void *);
3061 * With that we have determined the number of bytes in actual use
3062 * by the object. This is the potential offset to the free pointer.
3066 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3069 * Relocate free pointer after the object if it is not
3070 * permitted to overwrite the first word of the object on
3073 * This is the case if we do RCU, have a constructor or
3074 * destructor or are poisoning the objects.
3077 size += sizeof(void *);
3080 #ifdef CONFIG_SLUB_DEBUG
3081 if (flags & SLAB_STORE_USER)
3083 * Need to store information about allocs and frees after
3086 size += 2 * sizeof(struct track);
3088 if (flags & SLAB_RED_ZONE)
3090 * Add some empty padding so that we can catch
3091 * overwrites from earlier objects rather than let
3092 * tracking information or the free pointer be
3093 * corrupted if a user writes before the start
3096 size += sizeof(void *);
3100 * SLUB stores one object immediately after another beginning from
3101 * offset 0. In order to align the objects we have to simply size
3102 * each object to conform to the alignment.
3104 size = ALIGN(size, s->align);
3106 if (forced_order >= 0)
3107 order = forced_order;
3109 order = calculate_order(size, s->reserved);
3116 s->allocflags |= __GFP_COMP;
3118 if (s->flags & SLAB_CACHE_DMA)
3119 s->allocflags |= GFP_DMA;
3121 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3122 s->allocflags |= __GFP_RECLAIMABLE;
3125 * Determine the number of objects per slab
3127 s->oo = oo_make(order, size, s->reserved);
3128 s->min = oo_make(get_order(size), size, s->reserved);
3129 if (oo_objects(s->oo) > oo_objects(s->max))
3132 return !!oo_objects(s->oo);
3135 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3137 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3140 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3141 s->reserved = sizeof(struct rcu_head);
3143 if (!calculate_sizes(s, -1))
3145 if (disable_higher_order_debug) {
3147 * Disable debugging flags that store metadata if the min slab
3150 if (get_order(s->size) > get_order(s->object_size)) {
3151 s->flags &= ~DEBUG_METADATA_FLAGS;
3153 if (!calculate_sizes(s, -1))
3158 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3159 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3160 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3161 /* Enable fast mode */
3162 s->flags |= __CMPXCHG_DOUBLE;
3166 * The larger the object size is, the more pages we want on the partial
3167 * list to avoid pounding the page allocator excessively.
3169 set_min_partial(s, ilog2(s->size) / 2);
3172 * cpu_partial determined the maximum number of objects kept in the
3173 * per cpu partial lists of a processor.
3175 * Per cpu partial lists mainly contain slabs that just have one
3176 * object freed. If they are used for allocation then they can be
3177 * filled up again with minimal effort. The slab will never hit the
3178 * per node partial lists and therefore no locking will be required.
3180 * This setting also determines
3182 * A) The number of objects from per cpu partial slabs dumped to the
3183 * per node list when we reach the limit.
3184 * B) The number of objects in cpu partial slabs to extract from the
3185 * per node list when we run out of per cpu objects. We only fetch
3186 * 50% to keep some capacity around for frees.
3188 if (!kmem_cache_has_cpu_partial(s))
3190 else if (s->size >= PAGE_SIZE)
3192 else if (s->size >= 1024)
3194 else if (s->size >= 256)
3195 s->cpu_partial = 13;
3197 s->cpu_partial = 30;
3200 s->remote_node_defrag_ratio = 1000;
3202 if (!init_kmem_cache_nodes(s))
3205 if (alloc_kmem_cache_cpus(s))
3208 free_kmem_cache_nodes(s);
3210 if (flags & SLAB_PANIC)
3211 panic("Cannot create slab %s size=%lu realsize=%u "
3212 "order=%u offset=%u flags=%lx\n",
3213 s->name, (unsigned long)s->size, s->size,
3214 oo_order(s->oo), s->offset, flags);
3218 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3221 #ifdef CONFIG_SLUB_DEBUG
3222 void *addr = page_address(page);
3224 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3225 sizeof(long), GFP_ATOMIC);
3228 slab_err(s, page, text, s->name);
3231 get_map(s, page, map);
3232 for_each_object(p, s, addr, page->objects) {
3234 if (!test_bit(slab_index(p, s, addr), map)) {
3235 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3236 print_tracking(s, p);
3245 * Attempt to free all partial slabs on a node.
3246 * This is called from kmem_cache_close(). We must be the last thread
3247 * using the cache and therefore we do not need to lock anymore.
3249 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3251 struct page *page, *h;
3253 list_for_each_entry_safe(page, h, &n->partial, lru) {
3255 __remove_partial(n, page);
3256 discard_slab(s, page);
3258 list_slab_objects(s, page,
3259 "Objects remaining in %s on kmem_cache_close()");
3265 * Release all resources used by a slab cache.
3267 static inline int kmem_cache_close(struct kmem_cache *s)
3270 struct kmem_cache_node *n;
3273 /* Attempt to free all objects */
3274 for_each_kmem_cache_node(s, node, n) {
3276 if (n->nr_partial || slabs_node(s, node))
3279 free_percpu(s->cpu_slab);
3280 free_kmem_cache_nodes(s);
3284 int __kmem_cache_shutdown(struct kmem_cache *s)
3286 return kmem_cache_close(s);
3289 /********************************************************************
3291 *******************************************************************/
3293 static int __init setup_slub_min_order(char *str)
3295 get_option(&str, &slub_min_order);
3300 __setup("slub_min_order=", setup_slub_min_order);
3302 static int __init setup_slub_max_order(char *str)
3304 get_option(&str, &slub_max_order);
3305 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3310 __setup("slub_max_order=", setup_slub_max_order);
3312 static int __init setup_slub_min_objects(char *str)
3314 get_option(&str, &slub_min_objects);
3319 __setup("slub_min_objects=", setup_slub_min_objects);
3321 void *__kmalloc(size_t size, gfp_t flags)
3323 struct kmem_cache *s;
3326 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3327 return kmalloc_large(size, flags);
3329 s = kmalloc_slab(size, flags);
3331 if (unlikely(ZERO_OR_NULL_PTR(s)))
3334 ret = slab_alloc(s, flags, _RET_IP_);
3336 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3338 kasan_kmalloc(s, ret, size);
3342 EXPORT_SYMBOL(__kmalloc);
3345 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3350 flags |= __GFP_COMP | __GFP_NOTRACK;
3351 page = alloc_kmem_pages_node(node, flags, get_order(size));
3353 ptr = page_address(page);
3355 kmalloc_large_node_hook(ptr, size, flags);
3359 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3361 struct kmem_cache *s;
3364 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3365 ret = kmalloc_large_node(size, flags, node);
3367 trace_kmalloc_node(_RET_IP_, ret,
3368 size, PAGE_SIZE << get_order(size),
3374 s = kmalloc_slab(size, flags);
3376 if (unlikely(ZERO_OR_NULL_PTR(s)))
3379 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3381 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3383 kasan_kmalloc(s, ret, size);
3387 EXPORT_SYMBOL(__kmalloc_node);
3390 static size_t __ksize(const void *object)
3394 if (unlikely(object == ZERO_SIZE_PTR))
3397 page = virt_to_head_page(object);
3399 if (unlikely(!PageSlab(page))) {
3400 WARN_ON(!PageCompound(page));
3401 return PAGE_SIZE << compound_order(page);
3404 return slab_ksize(page->slab_cache);
3407 size_t ksize(const void *object)
3409 size_t size = __ksize(object);
3410 /* We assume that ksize callers could use whole allocated area,
3411 so we need unpoison this area. */
3412 kasan_krealloc(object, size);
3415 EXPORT_SYMBOL(ksize);
3417 void kfree(const void *x)
3420 void *object = (void *)x;
3422 trace_kfree(_RET_IP_, x);
3424 if (unlikely(ZERO_OR_NULL_PTR(x)))
3427 page = virt_to_head_page(x);
3428 if (unlikely(!PageSlab(page))) {
3429 BUG_ON(!PageCompound(page));
3431 __free_kmem_pages(page, compound_order(page));
3434 slab_free(page->slab_cache, page, object, _RET_IP_);
3436 EXPORT_SYMBOL(kfree);
3438 #define SHRINK_PROMOTE_MAX 32
3441 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3442 * up most to the head of the partial lists. New allocations will then
3443 * fill those up and thus they can be removed from the partial lists.
3445 * The slabs with the least items are placed last. This results in them
3446 * being allocated from last increasing the chance that the last objects
3447 * are freed in them.
3449 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3453 struct kmem_cache_node *n;
3456 struct list_head discard;
3457 struct list_head promote[SHRINK_PROMOTE_MAX];
3458 unsigned long flags;
3463 * Disable empty slabs caching. Used to avoid pinning offline
3464 * memory cgroups by kmem pages that can be freed.
3470 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3471 * so we have to make sure the change is visible.
3473 kick_all_cpus_sync();
3477 for_each_kmem_cache_node(s, node, n) {
3478 INIT_LIST_HEAD(&discard);
3479 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3480 INIT_LIST_HEAD(promote + i);
3482 spin_lock_irqsave(&n->list_lock, flags);
3485 * Build lists of slabs to discard or promote.
3487 * Note that concurrent frees may occur while we hold the
3488 * list_lock. page->inuse here is the upper limit.
3490 list_for_each_entry_safe(page, t, &n->partial, lru) {
3491 int free = page->objects - page->inuse;
3493 /* Do not reread page->inuse */
3496 /* We do not keep full slabs on the list */
3499 if (free == page->objects) {
3500 list_move(&page->lru, &discard);
3502 } else if (free <= SHRINK_PROMOTE_MAX)
3503 list_move(&page->lru, promote + free - 1);
3507 * Promote the slabs filled up most to the head of the
3510 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3511 list_splice(promote + i, &n->partial);
3513 spin_unlock_irqrestore(&n->list_lock, flags);
3515 /* Release empty slabs */
3516 list_for_each_entry_safe(page, t, &discard, lru)
3517 discard_slab(s, page);
3519 if (slabs_node(s, node))
3526 static int slab_mem_going_offline_callback(void *arg)
3528 struct kmem_cache *s;
3530 mutex_lock(&slab_mutex);
3531 list_for_each_entry(s, &slab_caches, list)
3532 __kmem_cache_shrink(s, false);
3533 mutex_unlock(&slab_mutex);
3538 static void slab_mem_offline_callback(void *arg)
3540 struct kmem_cache_node *n;
3541 struct kmem_cache *s;
3542 struct memory_notify *marg = arg;
3545 offline_node = marg->status_change_nid_normal;
3548 * If the node still has available memory. we need kmem_cache_node
3551 if (offline_node < 0)
3554 mutex_lock(&slab_mutex);
3555 list_for_each_entry(s, &slab_caches, list) {
3556 n = get_node(s, offline_node);
3559 * if n->nr_slabs > 0, slabs still exist on the node
3560 * that is going down. We were unable to free them,
3561 * and offline_pages() function shouldn't call this
3562 * callback. So, we must fail.
3564 BUG_ON(slabs_node(s, offline_node));
3566 s->node[offline_node] = NULL;
3567 kmem_cache_free(kmem_cache_node, n);
3570 mutex_unlock(&slab_mutex);
3573 static int slab_mem_going_online_callback(void *arg)
3575 struct kmem_cache_node *n;
3576 struct kmem_cache *s;
3577 struct memory_notify *marg = arg;
3578 int nid = marg->status_change_nid_normal;
3582 * If the node's memory is already available, then kmem_cache_node is
3583 * already created. Nothing to do.
3589 * We are bringing a node online. No memory is available yet. We must
3590 * allocate a kmem_cache_node structure in order to bring the node
3593 mutex_lock(&slab_mutex);
3594 list_for_each_entry(s, &slab_caches, list) {
3596 * XXX: kmem_cache_alloc_node will fallback to other nodes
3597 * since memory is not yet available from the node that
3600 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3605 init_kmem_cache_node(n);
3609 mutex_unlock(&slab_mutex);
3613 static int slab_memory_callback(struct notifier_block *self,
3614 unsigned long action, void *arg)
3619 case MEM_GOING_ONLINE:
3620 ret = slab_mem_going_online_callback(arg);
3622 case MEM_GOING_OFFLINE:
3623 ret = slab_mem_going_offline_callback(arg);
3626 case MEM_CANCEL_ONLINE:
3627 slab_mem_offline_callback(arg);
3630 case MEM_CANCEL_OFFLINE:
3634 ret = notifier_from_errno(ret);
3640 static struct notifier_block slab_memory_callback_nb = {
3641 .notifier_call = slab_memory_callback,
3642 .priority = SLAB_CALLBACK_PRI,
3645 /********************************************************************
3646 * Basic setup of slabs
3647 *******************************************************************/
3650 * Used for early kmem_cache structures that were allocated using
3651 * the page allocator. Allocate them properly then fix up the pointers
3652 * that may be pointing to the wrong kmem_cache structure.
3655 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3658 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3659 struct kmem_cache_node *n;
3661 memcpy(s, static_cache, kmem_cache->object_size);
3664 * This runs very early, and only the boot processor is supposed to be
3665 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3668 __flush_cpu_slab(s, smp_processor_id());
3669 for_each_kmem_cache_node(s, node, n) {
3672 list_for_each_entry(p, &n->partial, lru)
3675 #ifdef CONFIG_SLUB_DEBUG
3676 list_for_each_entry(p, &n->full, lru)
3680 slab_init_memcg_params(s);
3681 list_add(&s->list, &slab_caches);
3685 void __init kmem_cache_init(void)
3687 static __initdata struct kmem_cache boot_kmem_cache,
3688 boot_kmem_cache_node;
3690 if (debug_guardpage_minorder())
3693 kmem_cache_node = &boot_kmem_cache_node;
3694 kmem_cache = &boot_kmem_cache;
3696 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3697 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3699 register_hotmemory_notifier(&slab_memory_callback_nb);
3701 /* Able to allocate the per node structures */
3702 slab_state = PARTIAL;
3704 create_boot_cache(kmem_cache, "kmem_cache",
3705 offsetof(struct kmem_cache, node) +
3706 nr_node_ids * sizeof(struct kmem_cache_node *),
3707 SLAB_HWCACHE_ALIGN);
3709 kmem_cache = bootstrap(&boot_kmem_cache);
3712 * Allocate kmem_cache_node properly from the kmem_cache slab.
3713 * kmem_cache_node is separately allocated so no need to
3714 * update any list pointers.
3716 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3718 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3719 create_kmalloc_caches(0);
3722 register_cpu_notifier(&slab_notifier);
3725 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3727 slub_min_order, slub_max_order, slub_min_objects,
3728 nr_cpu_ids, nr_node_ids);
3731 void __init kmem_cache_init_late(void)
3736 __kmem_cache_alias(const char *name, size_t size, size_t align,
3737 unsigned long flags, void (*ctor)(void *))
3739 struct kmem_cache *s, *c;
3741 s = find_mergeable(size, align, flags, name, ctor);
3746 * Adjust the object sizes so that we clear
3747 * the complete object on kzalloc.
3749 s->object_size = max(s->object_size, (int)size);
3750 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3752 for_each_memcg_cache(c, s) {
3753 c->object_size = s->object_size;
3754 c->inuse = max_t(int, c->inuse,
3755 ALIGN(size, sizeof(void *)));
3758 if (sysfs_slab_alias(s, name)) {
3767 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3771 err = kmem_cache_open(s, flags);
3775 /* Mutex is not taken during early boot */
3776 if (slab_state <= UP)
3779 memcg_propagate_slab_attrs(s);
3780 err = sysfs_slab_add(s);
3782 kmem_cache_close(s);
3789 * Use the cpu notifier to insure that the cpu slabs are flushed when
3792 static int slab_cpuup_callback(struct notifier_block *nfb,
3793 unsigned long action, void *hcpu)
3795 long cpu = (long)hcpu;
3796 struct kmem_cache *s;
3797 unsigned long flags;
3800 case CPU_UP_CANCELED:
3801 case CPU_UP_CANCELED_FROZEN:
3803 case CPU_DEAD_FROZEN:
3804 mutex_lock(&slab_mutex);
3805 list_for_each_entry(s, &slab_caches, list) {
3806 local_irq_save(flags);
3807 __flush_cpu_slab(s, cpu);
3808 local_irq_restore(flags);
3810 mutex_unlock(&slab_mutex);
3818 static struct notifier_block slab_notifier = {
3819 .notifier_call = slab_cpuup_callback
3824 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3826 struct kmem_cache *s;
3829 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3830 return kmalloc_large(size, gfpflags);
3832 s = kmalloc_slab(size, gfpflags);
3834 if (unlikely(ZERO_OR_NULL_PTR(s)))
3837 ret = slab_alloc(s, gfpflags, caller);
3839 /* Honor the call site pointer we received. */
3840 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3846 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3847 int node, unsigned long caller)
3849 struct kmem_cache *s;
3852 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3853 ret = kmalloc_large_node(size, gfpflags, node);
3855 trace_kmalloc_node(caller, ret,
3856 size, PAGE_SIZE << get_order(size),
3862 s = kmalloc_slab(size, gfpflags);
3864 if (unlikely(ZERO_OR_NULL_PTR(s)))
3867 ret = slab_alloc_node(s, gfpflags, node, caller);
3869 /* Honor the call site pointer we received. */
3870 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3877 static int count_inuse(struct page *page)
3882 static int count_total(struct page *page)
3884 return page->objects;
3888 #ifdef CONFIG_SLUB_DEBUG
3889 static int validate_slab(struct kmem_cache *s, struct page *page,
3893 void *addr = page_address(page);
3895 if (!check_slab(s, page) ||
3896 !on_freelist(s, page, NULL))
3899 /* Now we know that a valid freelist exists */
3900 bitmap_zero(map, page->objects);
3902 get_map(s, page, map);
3903 for_each_object(p, s, addr, page->objects) {
3904 if (test_bit(slab_index(p, s, addr), map))
3905 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3909 for_each_object(p, s, addr, page->objects)
3910 if (!test_bit(slab_index(p, s, addr), map))
3911 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3916 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3920 validate_slab(s, page, map);
3924 static int validate_slab_node(struct kmem_cache *s,
3925 struct kmem_cache_node *n, unsigned long *map)
3927 unsigned long count = 0;
3929 unsigned long flags;
3931 spin_lock_irqsave(&n->list_lock, flags);
3933 list_for_each_entry(page, &n->partial, lru) {
3934 validate_slab_slab(s, page, map);
3937 if (count != n->nr_partial)
3938 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3939 s->name, count, n->nr_partial);
3941 if (!(s->flags & SLAB_STORE_USER))
3944 list_for_each_entry(page, &n->full, lru) {
3945 validate_slab_slab(s, page, map);
3948 if (count != atomic_long_read(&n->nr_slabs))
3949 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3950 s->name, count, atomic_long_read(&n->nr_slabs));
3953 spin_unlock_irqrestore(&n->list_lock, flags);
3957 static long validate_slab_cache(struct kmem_cache *s)
3960 unsigned long count = 0;
3961 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3962 sizeof(unsigned long), GFP_KERNEL);
3963 struct kmem_cache_node *n;
3969 for_each_kmem_cache_node(s, node, n)
3970 count += validate_slab_node(s, n, map);
3975 * Generate lists of code addresses where slabcache objects are allocated
3980 unsigned long count;
3987 DECLARE_BITMAP(cpus, NR_CPUS);
3993 unsigned long count;
3994 struct location *loc;
3997 static void free_loc_track(struct loc_track *t)
4000 free_pages((unsigned long)t->loc,
4001 get_order(sizeof(struct location) * t->max));
4004 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4009 order = get_order(sizeof(struct location) * max);
4011 l = (void *)__get_free_pages(flags, order);
4016 memcpy(l, t->loc, sizeof(struct location) * t->count);
4024 static int add_location(struct loc_track *t, struct kmem_cache *s,
4025 const struct track *track)
4027 long start, end, pos;
4029 unsigned long caddr;
4030 unsigned long age = jiffies - track->when;
4036 pos = start + (end - start + 1) / 2;
4039 * There is nothing at "end". If we end up there
4040 * we need to add something to before end.
4045 caddr = t->loc[pos].addr;
4046 if (track->addr == caddr) {
4052 if (age < l->min_time)
4054 if (age > l->max_time)
4057 if (track->pid < l->min_pid)
4058 l->min_pid = track->pid;
4059 if (track->pid > l->max_pid)
4060 l->max_pid = track->pid;
4062 cpumask_set_cpu(track->cpu,
4063 to_cpumask(l->cpus));
4065 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4069 if (track->addr < caddr)
4076 * Not found. Insert new tracking element.
4078 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4084 (t->count - pos) * sizeof(struct location));
4087 l->addr = track->addr;
4091 l->min_pid = track->pid;
4092 l->max_pid = track->pid;
4093 cpumask_clear(to_cpumask(l->cpus));
4094 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4095 nodes_clear(l->nodes);
4096 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4100 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4101 struct page *page, enum track_item alloc,
4104 void *addr = page_address(page);
4107 bitmap_zero(map, page->objects);
4108 get_map(s, page, map);
4110 for_each_object(p, s, addr, page->objects)
4111 if (!test_bit(slab_index(p, s, addr), map))
4112 add_location(t, s, get_track(s, p, alloc));
4115 static int list_locations(struct kmem_cache *s, char *buf,
4116 enum track_item alloc)
4120 struct loc_track t = { 0, 0, NULL };
4122 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4123 sizeof(unsigned long), GFP_KERNEL);
4124 struct kmem_cache_node *n;
4126 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4129 return sprintf(buf, "Out of memory\n");
4131 /* Push back cpu slabs */
4134 for_each_kmem_cache_node(s, node, n) {
4135 unsigned long flags;
4138 if (!atomic_long_read(&n->nr_slabs))
4141 spin_lock_irqsave(&n->list_lock, flags);
4142 list_for_each_entry(page, &n->partial, lru)
4143 process_slab(&t, s, page, alloc, map);
4144 list_for_each_entry(page, &n->full, lru)
4145 process_slab(&t, s, page, alloc, map);
4146 spin_unlock_irqrestore(&n->list_lock, flags);
4149 for (i = 0; i < t.count; i++) {
4150 struct location *l = &t.loc[i];
4152 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4154 len += sprintf(buf + len, "%7ld ", l->count);
4157 len += sprintf(buf + len, "%pS", (void *)l->addr);
4159 len += sprintf(buf + len, "<not-available>");
4161 if (l->sum_time != l->min_time) {
4162 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4164 (long)div_u64(l->sum_time, l->count),
4167 len += sprintf(buf + len, " age=%ld",
4170 if (l->min_pid != l->max_pid)
4171 len += sprintf(buf + len, " pid=%ld-%ld",
4172 l->min_pid, l->max_pid);
4174 len += sprintf(buf + len, " pid=%ld",
4177 if (num_online_cpus() > 1 &&
4178 !cpumask_empty(to_cpumask(l->cpus)) &&
4179 len < PAGE_SIZE - 60)
4180 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4182 cpumask_pr_args(to_cpumask(l->cpus)));
4184 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4185 len < PAGE_SIZE - 60)
4186 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4188 nodemask_pr_args(&l->nodes));
4190 len += sprintf(buf + len, "\n");
4196 len += sprintf(buf, "No data\n");
4201 #ifdef SLUB_RESILIENCY_TEST
4202 static void __init resiliency_test(void)
4206 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4208 pr_err("SLUB resiliency testing\n");
4209 pr_err("-----------------------\n");
4210 pr_err("A. Corruption after allocation\n");
4212 p = kzalloc(16, GFP_KERNEL);
4214 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4217 validate_slab_cache(kmalloc_caches[4]);
4219 /* Hmmm... The next two are dangerous */
4220 p = kzalloc(32, GFP_KERNEL);
4221 p[32 + sizeof(void *)] = 0x34;
4222 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4224 pr_err("If allocated object is overwritten then not detectable\n\n");
4226 validate_slab_cache(kmalloc_caches[5]);
4227 p = kzalloc(64, GFP_KERNEL);
4228 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4230 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4232 pr_err("If allocated object is overwritten then not detectable\n\n");
4233 validate_slab_cache(kmalloc_caches[6]);
4235 pr_err("\nB. Corruption after free\n");
4236 p = kzalloc(128, GFP_KERNEL);
4239 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4240 validate_slab_cache(kmalloc_caches[7]);
4242 p = kzalloc(256, GFP_KERNEL);
4245 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4246 validate_slab_cache(kmalloc_caches[8]);
4248 p = kzalloc(512, GFP_KERNEL);
4251 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4252 validate_slab_cache(kmalloc_caches[9]);
4256 static void resiliency_test(void) {};
4261 enum slab_stat_type {
4262 SL_ALL, /* All slabs */
4263 SL_PARTIAL, /* Only partially allocated slabs */
4264 SL_CPU, /* Only slabs used for cpu caches */
4265 SL_OBJECTS, /* Determine allocated objects not slabs */
4266 SL_TOTAL /* Determine object capacity not slabs */
4269 #define SO_ALL (1 << SL_ALL)
4270 #define SO_PARTIAL (1 << SL_PARTIAL)
4271 #define SO_CPU (1 << SL_CPU)
4272 #define SO_OBJECTS (1 << SL_OBJECTS)
4273 #define SO_TOTAL (1 << SL_TOTAL)
4275 static ssize_t show_slab_objects(struct kmem_cache *s,
4276 char *buf, unsigned long flags)
4278 unsigned long total = 0;
4281 unsigned long *nodes;
4283 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4287 if (flags & SO_CPU) {
4290 for_each_possible_cpu(cpu) {
4291 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4296 page = READ_ONCE(c->page);
4300 node = page_to_nid(page);
4301 if (flags & SO_TOTAL)
4303 else if (flags & SO_OBJECTS)
4311 page = READ_ONCE(c->partial);
4313 node = page_to_nid(page);
4314 if (flags & SO_TOTAL)
4316 else if (flags & SO_OBJECTS)
4327 #ifdef CONFIG_SLUB_DEBUG
4328 if (flags & SO_ALL) {
4329 struct kmem_cache_node *n;
4331 for_each_kmem_cache_node(s, node, n) {
4333 if (flags & SO_TOTAL)
4334 x = atomic_long_read(&n->total_objects);
4335 else if (flags & SO_OBJECTS)
4336 x = atomic_long_read(&n->total_objects) -
4337 count_partial(n, count_free);
4339 x = atomic_long_read(&n->nr_slabs);
4346 if (flags & SO_PARTIAL) {
4347 struct kmem_cache_node *n;
4349 for_each_kmem_cache_node(s, node, n) {
4350 if (flags & SO_TOTAL)
4351 x = count_partial(n, count_total);
4352 else if (flags & SO_OBJECTS)
4353 x = count_partial(n, count_inuse);
4360 x = sprintf(buf, "%lu", total);
4362 for (node = 0; node < nr_node_ids; node++)
4364 x += sprintf(buf + x, " N%d=%lu",
4369 return x + sprintf(buf + x, "\n");
4372 #ifdef CONFIG_SLUB_DEBUG
4373 static int any_slab_objects(struct kmem_cache *s)
4376 struct kmem_cache_node *n;
4378 for_each_kmem_cache_node(s, node, n)
4379 if (atomic_long_read(&n->total_objects))
4386 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4387 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4389 struct slab_attribute {
4390 struct attribute attr;
4391 ssize_t (*show)(struct kmem_cache *s, char *buf);
4392 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4395 #define SLAB_ATTR_RO(_name) \
4396 static struct slab_attribute _name##_attr = \
4397 __ATTR(_name, 0400, _name##_show, NULL)
4399 #define SLAB_ATTR(_name) \
4400 static struct slab_attribute _name##_attr = \
4401 __ATTR(_name, 0600, _name##_show, _name##_store)
4403 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4405 return sprintf(buf, "%d\n", s->size);
4407 SLAB_ATTR_RO(slab_size);
4409 static ssize_t align_show(struct kmem_cache *s, char *buf)
4411 return sprintf(buf, "%d\n", s->align);
4413 SLAB_ATTR_RO(align);
4415 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4417 return sprintf(buf, "%d\n", s->object_size);
4419 SLAB_ATTR_RO(object_size);
4421 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4423 return sprintf(buf, "%d\n", oo_objects(s->oo));
4425 SLAB_ATTR_RO(objs_per_slab);
4427 static ssize_t order_store(struct kmem_cache *s,
4428 const char *buf, size_t length)
4430 unsigned long order;
4433 err = kstrtoul(buf, 10, &order);
4437 if (order > slub_max_order || order < slub_min_order)
4440 calculate_sizes(s, order);
4444 static ssize_t order_show(struct kmem_cache *s, char *buf)
4446 return sprintf(buf, "%d\n", oo_order(s->oo));
4450 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4452 return sprintf(buf, "%lu\n", s->min_partial);
4455 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4461 err = kstrtoul(buf, 10, &min);
4465 set_min_partial(s, min);
4468 SLAB_ATTR(min_partial);
4470 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4472 return sprintf(buf, "%u\n", s->cpu_partial);
4475 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4478 unsigned long objects;
4481 err = kstrtoul(buf, 10, &objects);
4484 if (objects && !kmem_cache_has_cpu_partial(s))
4487 s->cpu_partial = objects;
4491 SLAB_ATTR(cpu_partial);
4493 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4497 return sprintf(buf, "%pS\n", s->ctor);
4501 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4503 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4505 SLAB_ATTR_RO(aliases);
4507 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4509 return show_slab_objects(s, buf, SO_PARTIAL);
4511 SLAB_ATTR_RO(partial);
4513 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4515 return show_slab_objects(s, buf, SO_CPU);
4517 SLAB_ATTR_RO(cpu_slabs);
4519 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4521 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4523 SLAB_ATTR_RO(objects);
4525 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4527 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4529 SLAB_ATTR_RO(objects_partial);
4531 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4538 for_each_online_cpu(cpu) {
4539 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4542 pages += page->pages;
4543 objects += page->pobjects;
4547 len = sprintf(buf, "%d(%d)", objects, pages);
4550 for_each_online_cpu(cpu) {
4551 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4553 if (page && len < PAGE_SIZE - 20)
4554 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4555 page->pobjects, page->pages);
4558 return len + sprintf(buf + len, "\n");
4560 SLAB_ATTR_RO(slabs_cpu_partial);
4562 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4564 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4567 static ssize_t reclaim_account_store(struct kmem_cache *s,
4568 const char *buf, size_t length)
4570 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4572 s->flags |= SLAB_RECLAIM_ACCOUNT;
4575 SLAB_ATTR(reclaim_account);
4577 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4579 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4581 SLAB_ATTR_RO(hwcache_align);
4583 #ifdef CONFIG_ZONE_DMA
4584 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4586 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4588 SLAB_ATTR_RO(cache_dma);
4591 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4593 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4595 SLAB_ATTR_RO(destroy_by_rcu);
4597 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4599 return sprintf(buf, "%d\n", s->reserved);
4601 SLAB_ATTR_RO(reserved);
4603 #ifdef CONFIG_SLUB_DEBUG
4604 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4606 return show_slab_objects(s, buf, SO_ALL);
4608 SLAB_ATTR_RO(slabs);
4610 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4612 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4614 SLAB_ATTR_RO(total_objects);
4616 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4618 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4621 static ssize_t sanity_checks_store(struct kmem_cache *s,
4622 const char *buf, size_t length)
4624 s->flags &= ~SLAB_DEBUG_FREE;
4625 if (buf[0] == '1') {
4626 s->flags &= ~__CMPXCHG_DOUBLE;
4627 s->flags |= SLAB_DEBUG_FREE;
4631 SLAB_ATTR(sanity_checks);
4633 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4635 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4638 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4642 * Tracing a merged cache is going to give confusing results
4643 * as well as cause other issues like converting a mergeable
4644 * cache into an umergeable one.
4646 if (s->refcount > 1)
4649 s->flags &= ~SLAB_TRACE;
4650 if (buf[0] == '1') {
4651 s->flags &= ~__CMPXCHG_DOUBLE;
4652 s->flags |= SLAB_TRACE;
4658 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4660 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4663 static ssize_t red_zone_store(struct kmem_cache *s,
4664 const char *buf, size_t length)
4666 if (any_slab_objects(s))
4669 s->flags &= ~SLAB_RED_ZONE;
4670 if (buf[0] == '1') {
4671 s->flags &= ~__CMPXCHG_DOUBLE;
4672 s->flags |= SLAB_RED_ZONE;
4674 calculate_sizes(s, -1);
4677 SLAB_ATTR(red_zone);
4679 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4681 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4684 static ssize_t poison_store(struct kmem_cache *s,
4685 const char *buf, size_t length)
4687 if (any_slab_objects(s))
4690 s->flags &= ~SLAB_POISON;
4691 if (buf[0] == '1') {
4692 s->flags &= ~__CMPXCHG_DOUBLE;
4693 s->flags |= SLAB_POISON;
4695 calculate_sizes(s, -1);
4700 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4702 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4705 static ssize_t store_user_store(struct kmem_cache *s,
4706 const char *buf, size_t length)
4708 if (any_slab_objects(s))
4711 s->flags &= ~SLAB_STORE_USER;
4712 if (buf[0] == '1') {
4713 s->flags &= ~__CMPXCHG_DOUBLE;
4714 s->flags |= SLAB_STORE_USER;
4716 calculate_sizes(s, -1);
4719 SLAB_ATTR(store_user);
4721 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4726 static ssize_t validate_store(struct kmem_cache *s,
4727 const char *buf, size_t length)
4731 if (buf[0] == '1') {
4732 ret = validate_slab_cache(s);
4738 SLAB_ATTR(validate);
4740 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4742 if (!(s->flags & SLAB_STORE_USER))
4744 return list_locations(s, buf, TRACK_ALLOC);
4746 SLAB_ATTR_RO(alloc_calls);
4748 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4750 if (!(s->flags & SLAB_STORE_USER))
4752 return list_locations(s, buf, TRACK_FREE);
4754 SLAB_ATTR_RO(free_calls);
4755 #endif /* CONFIG_SLUB_DEBUG */
4757 #ifdef CONFIG_FAILSLAB
4758 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4760 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4763 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4766 if (s->refcount > 1)
4769 s->flags &= ~SLAB_FAILSLAB;
4771 s->flags |= SLAB_FAILSLAB;
4774 SLAB_ATTR(failslab);
4777 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4782 static ssize_t shrink_store(struct kmem_cache *s,
4783 const char *buf, size_t length)
4786 kmem_cache_shrink(s);
4794 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4796 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4799 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4800 const char *buf, size_t length)
4802 unsigned long ratio;
4805 err = kstrtoul(buf, 10, &ratio);
4810 s->remote_node_defrag_ratio = ratio * 10;
4814 SLAB_ATTR(remote_node_defrag_ratio);
4817 #ifdef CONFIG_SLUB_STATS
4818 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4820 unsigned long sum = 0;
4823 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4828 for_each_online_cpu(cpu) {
4829 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4835 len = sprintf(buf, "%lu", sum);
4838 for_each_online_cpu(cpu) {
4839 if (data[cpu] && len < PAGE_SIZE - 20)
4840 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4844 return len + sprintf(buf + len, "\n");
4847 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4851 for_each_online_cpu(cpu)
4852 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4855 #define STAT_ATTR(si, text) \
4856 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4858 return show_stat(s, buf, si); \
4860 static ssize_t text##_store(struct kmem_cache *s, \
4861 const char *buf, size_t length) \
4863 if (buf[0] != '0') \
4865 clear_stat(s, si); \
4870 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4871 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4872 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4873 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4874 STAT_ATTR(FREE_FROZEN, free_frozen);
4875 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4876 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4877 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4878 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4879 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4880 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4881 STAT_ATTR(FREE_SLAB, free_slab);
4882 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4883 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4884 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4885 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4886 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4887 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4888 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4889 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4890 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4891 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4892 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4893 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4894 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4895 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4898 static struct attribute *slab_attrs[] = {
4899 &slab_size_attr.attr,
4900 &object_size_attr.attr,
4901 &objs_per_slab_attr.attr,
4903 &min_partial_attr.attr,
4904 &cpu_partial_attr.attr,
4906 &objects_partial_attr.attr,
4908 &cpu_slabs_attr.attr,
4912 &hwcache_align_attr.attr,
4913 &reclaim_account_attr.attr,
4914 &destroy_by_rcu_attr.attr,
4916 &reserved_attr.attr,
4917 &slabs_cpu_partial_attr.attr,
4918 #ifdef CONFIG_SLUB_DEBUG
4919 &total_objects_attr.attr,
4921 &sanity_checks_attr.attr,
4923 &red_zone_attr.attr,
4925 &store_user_attr.attr,
4926 &validate_attr.attr,
4927 &alloc_calls_attr.attr,
4928 &free_calls_attr.attr,
4930 #ifdef CONFIG_ZONE_DMA
4931 &cache_dma_attr.attr,
4934 &remote_node_defrag_ratio_attr.attr,
4936 #ifdef CONFIG_SLUB_STATS
4937 &alloc_fastpath_attr.attr,
4938 &alloc_slowpath_attr.attr,
4939 &free_fastpath_attr.attr,
4940 &free_slowpath_attr.attr,
4941 &free_frozen_attr.attr,
4942 &free_add_partial_attr.attr,
4943 &free_remove_partial_attr.attr,
4944 &alloc_from_partial_attr.attr,
4945 &alloc_slab_attr.attr,
4946 &alloc_refill_attr.attr,
4947 &alloc_node_mismatch_attr.attr,
4948 &free_slab_attr.attr,
4949 &cpuslab_flush_attr.attr,
4950 &deactivate_full_attr.attr,
4951 &deactivate_empty_attr.attr,
4952 &deactivate_to_head_attr.attr,
4953 &deactivate_to_tail_attr.attr,
4954 &deactivate_remote_frees_attr.attr,
4955 &deactivate_bypass_attr.attr,
4956 &order_fallback_attr.attr,
4957 &cmpxchg_double_fail_attr.attr,
4958 &cmpxchg_double_cpu_fail_attr.attr,
4959 &cpu_partial_alloc_attr.attr,
4960 &cpu_partial_free_attr.attr,
4961 &cpu_partial_node_attr.attr,
4962 &cpu_partial_drain_attr.attr,
4964 #ifdef CONFIG_FAILSLAB
4965 &failslab_attr.attr,
4971 static struct attribute_group slab_attr_group = {
4972 .attrs = slab_attrs,
4975 static ssize_t slab_attr_show(struct kobject *kobj,
4976 struct attribute *attr,
4979 struct slab_attribute *attribute;
4980 struct kmem_cache *s;
4983 attribute = to_slab_attr(attr);
4986 if (!attribute->show)
4989 err = attribute->show(s, buf);
4994 static ssize_t slab_attr_store(struct kobject *kobj,
4995 struct attribute *attr,
4996 const char *buf, size_t len)
4998 struct slab_attribute *attribute;
4999 struct kmem_cache *s;
5002 attribute = to_slab_attr(attr);
5005 if (!attribute->store)
5008 err = attribute->store(s, buf, len);
5009 #ifdef CONFIG_MEMCG_KMEM
5010 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5011 struct kmem_cache *c;
5013 mutex_lock(&slab_mutex);
5014 if (s->max_attr_size < len)
5015 s->max_attr_size = len;
5018 * This is a best effort propagation, so this function's return
5019 * value will be determined by the parent cache only. This is
5020 * basically because not all attributes will have a well
5021 * defined semantics for rollbacks - most of the actions will
5022 * have permanent effects.
5024 * Returning the error value of any of the children that fail
5025 * is not 100 % defined, in the sense that users seeing the
5026 * error code won't be able to know anything about the state of
5029 * Only returning the error code for the parent cache at least
5030 * has well defined semantics. The cache being written to
5031 * directly either failed or succeeded, in which case we loop
5032 * through the descendants with best-effort propagation.
5034 for_each_memcg_cache(c, s)
5035 attribute->store(c, buf, len);
5036 mutex_unlock(&slab_mutex);
5042 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5044 #ifdef CONFIG_MEMCG_KMEM
5046 char *buffer = NULL;
5047 struct kmem_cache *root_cache;
5049 if (is_root_cache(s))
5052 root_cache = s->memcg_params.root_cache;
5055 * This mean this cache had no attribute written. Therefore, no point
5056 * in copying default values around
5058 if (!root_cache->max_attr_size)
5061 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5064 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5066 if (!attr || !attr->store || !attr->show)
5070 * It is really bad that we have to allocate here, so we will
5071 * do it only as a fallback. If we actually allocate, though,
5072 * we can just use the allocated buffer until the end.
5074 * Most of the slub attributes will tend to be very small in
5075 * size, but sysfs allows buffers up to a page, so they can
5076 * theoretically happen.
5080 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5083 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5084 if (WARN_ON(!buffer))
5089 attr->show(root_cache, buf);
5090 attr->store(s, buf, strlen(buf));
5094 free_page((unsigned long)buffer);
5098 static void kmem_cache_release(struct kobject *k)
5100 slab_kmem_cache_release(to_slab(k));
5103 static const struct sysfs_ops slab_sysfs_ops = {
5104 .show = slab_attr_show,
5105 .store = slab_attr_store,
5108 static struct kobj_type slab_ktype = {
5109 .sysfs_ops = &slab_sysfs_ops,
5110 .release = kmem_cache_release,
5113 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5115 struct kobj_type *ktype = get_ktype(kobj);
5117 if (ktype == &slab_ktype)
5122 static const struct kset_uevent_ops slab_uevent_ops = {
5123 .filter = uevent_filter,
5126 static struct kset *slab_kset;
5128 static inline struct kset *cache_kset(struct kmem_cache *s)
5130 #ifdef CONFIG_MEMCG_KMEM
5131 if (!is_root_cache(s))
5132 return s->memcg_params.root_cache->memcg_kset;
5137 #define ID_STR_LENGTH 64
5139 /* Create a unique string id for a slab cache:
5141 * Format :[flags-]size
5143 static char *create_unique_id(struct kmem_cache *s)
5145 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5152 * First flags affecting slabcache operations. We will only
5153 * get here for aliasable slabs so we do not need to support
5154 * too many flags. The flags here must cover all flags that
5155 * are matched during merging to guarantee that the id is
5158 if (s->flags & SLAB_CACHE_DMA)
5160 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5162 if (s->flags & SLAB_DEBUG_FREE)
5164 if (!(s->flags & SLAB_NOTRACK))
5168 p += sprintf(p, "%07d", s->size);
5170 BUG_ON(p > name + ID_STR_LENGTH - 1);
5174 static int sysfs_slab_add(struct kmem_cache *s)
5178 int unmergeable = slab_unmergeable(s);
5182 * Slabcache can never be merged so we can use the name proper.
5183 * This is typically the case for debug situations. In that
5184 * case we can catch duplicate names easily.
5186 sysfs_remove_link(&slab_kset->kobj, s->name);
5190 * Create a unique name for the slab as a target
5193 name = create_unique_id(s);
5196 s->kobj.kset = cache_kset(s);
5197 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5201 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5205 #ifdef CONFIG_MEMCG_KMEM
5206 if (is_root_cache(s)) {
5207 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5208 if (!s->memcg_kset) {
5215 kobject_uevent(&s->kobj, KOBJ_ADD);
5217 /* Setup first alias */
5218 sysfs_slab_alias(s, s->name);
5225 kobject_del(&s->kobj);
5227 kobject_put(&s->kobj);
5231 void sysfs_slab_remove(struct kmem_cache *s)
5233 if (slab_state < FULL)
5235 * Sysfs has not been setup yet so no need to remove the
5240 #ifdef CONFIG_MEMCG_KMEM
5241 kset_unregister(s->memcg_kset);
5243 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5244 kobject_del(&s->kobj);
5245 kobject_put(&s->kobj);
5249 * Need to buffer aliases during bootup until sysfs becomes
5250 * available lest we lose that information.
5252 struct saved_alias {
5253 struct kmem_cache *s;
5255 struct saved_alias *next;
5258 static struct saved_alias *alias_list;
5260 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5262 struct saved_alias *al;
5264 if (slab_state == FULL) {
5266 * If we have a leftover link then remove it.
5268 sysfs_remove_link(&slab_kset->kobj, name);
5269 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5272 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5278 al->next = alias_list;
5283 static int __init slab_sysfs_init(void)
5285 struct kmem_cache *s;
5288 mutex_lock(&slab_mutex);
5290 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5292 mutex_unlock(&slab_mutex);
5293 pr_err("Cannot register slab subsystem.\n");
5299 list_for_each_entry(s, &slab_caches, list) {
5300 err = sysfs_slab_add(s);
5302 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5306 while (alias_list) {
5307 struct saved_alias *al = alias_list;
5309 alias_list = alias_list->next;
5310 err = sysfs_slab_alias(al->s, al->name);
5312 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5317 mutex_unlock(&slab_mutex);
5322 __initcall(slab_sysfs_init);
5323 #endif /* CONFIG_SYSFS */
5326 * The /proc/slabinfo ABI
5328 #ifdef CONFIG_SLABINFO
5329 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5331 unsigned long nr_slabs = 0;
5332 unsigned long nr_objs = 0;
5333 unsigned long nr_free = 0;
5335 struct kmem_cache_node *n;
5337 for_each_kmem_cache_node(s, node, n) {
5338 nr_slabs += node_nr_slabs(n);
5339 nr_objs += node_nr_objs(n);
5340 nr_free += count_partial(n, count_free);
5343 sinfo->active_objs = nr_objs - nr_free;
5344 sinfo->num_objs = nr_objs;
5345 sinfo->active_slabs = nr_slabs;
5346 sinfo->num_slabs = nr_slabs;
5347 sinfo->objects_per_slab = oo_objects(s->oo);
5348 sinfo->cache_order = oo_order(s->oo);
5351 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5355 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5356 size_t count, loff_t *ppos)
5360 #endif /* CONFIG_SLABINFO */