2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * Debugging flags that require metadata to be stored in the slab. These get
168 * disabled when slub_debug=O is used and a cache's min order increases with
171 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
174 #define OO_MASK ((1 << OO_SHIFT) - 1)
175 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
177 /* Internal SLUB flags */
178 #define __OBJECT_POISON 0x80000000UL /* Poison object */
179 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
182 static struct notifier_block slab_notifier;
186 * Tracking user of a slab.
188 #define TRACK_ADDRS_COUNT 16
190 unsigned long addr; /* Called from address */
191 #ifdef CONFIG_STACKTRACE
192 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
194 int cpu; /* Was running on cpu */
195 int pid; /* Pid context */
196 unsigned long when; /* When did the operation occur */
199 enum track_item { TRACK_ALLOC, TRACK_FREE };
202 static int sysfs_slab_add(struct kmem_cache *);
203 static int sysfs_slab_alias(struct kmem_cache *, const char *);
204 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
206 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
207 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
209 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
212 static inline void stat(const struct kmem_cache *s, enum stat_item si)
214 #ifdef CONFIG_SLUB_STATS
216 * The rmw is racy on a preemptible kernel but this is acceptable, so
217 * avoid this_cpu_add()'s irq-disable overhead.
219 raw_cpu_inc(s->cpu_slab->stat[si]);
223 /********************************************************************
224 * Core slab cache functions
225 *******************************************************************/
227 /* Verify that a pointer has an address that is valid within a slab page */
228 static inline int check_valid_pointer(struct kmem_cache *s,
229 struct page *page, const void *object)
236 base = page_address(page);
237 if (object < base || object >= base + page->objects * s->size ||
238 (object - base) % s->size) {
245 static inline void *get_freepointer(struct kmem_cache *s, void *object)
247 return *(void **)(object + s->offset);
250 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
252 prefetch(object + s->offset);
255 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
259 #ifdef CONFIG_DEBUG_PAGEALLOC
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
262 p = get_freepointer(s, object);
267 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
269 *(void **)(object + s->offset) = fp;
272 /* Loop over all objects in a slab */
273 #define for_each_object(__p, __s, __addr, __objects) \
274 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
277 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
278 for (__p = (__addr), __idx = 1; __idx <= __objects;\
279 __p += (__s)->size, __idx++)
281 /* Determine object index from a given position */
282 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
284 return (p - addr) / s->size;
287 static inline size_t slab_ksize(const struct kmem_cache *s)
289 #ifdef CONFIG_SLUB_DEBUG
291 * Debugging requires use of the padding between object
292 * and whatever may come after it.
294 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
295 return s->object_size;
299 * If we have the need to store the freelist pointer
300 * back there or track user information then we can
301 * only use the space before that information.
303 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
306 * Else we can use all the padding etc for the allocation
311 static inline int order_objects(int order, unsigned long size, int reserved)
313 return ((PAGE_SIZE << order) - reserved) / size;
316 static inline struct kmem_cache_order_objects oo_make(int order,
317 unsigned long size, int reserved)
319 struct kmem_cache_order_objects x = {
320 (order << OO_SHIFT) + order_objects(order, size, reserved)
326 static inline int oo_order(struct kmem_cache_order_objects x)
328 return x.x >> OO_SHIFT;
331 static inline int oo_objects(struct kmem_cache_order_objects x)
333 return x.x & OO_MASK;
337 * Per slab locking using the pagelock
339 static __always_inline void slab_lock(struct page *page)
341 bit_spin_lock(PG_locked, &page->flags);
344 static __always_inline void slab_unlock(struct page *page)
346 __bit_spin_unlock(PG_locked, &page->flags);
349 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
352 tmp.counters = counters_new;
354 * page->counters can cover frozen/inuse/objects as well
355 * as page->_count. If we assign to ->counters directly
356 * we run the risk of losing updates to page->_count, so
357 * be careful and only assign to the fields we need.
359 page->frozen = tmp.frozen;
360 page->inuse = tmp.inuse;
361 page->objects = tmp.objects;
364 /* Interrupts must be disabled (for the fallback code to work right) */
365 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
370 VM_BUG_ON(!irqs_disabled());
371 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 set_page_slub_counters(page, counters_new);
393 stat(s, CMPXCHG_DOUBLE_FAIL);
395 #ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
402 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
407 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
419 local_irq_save(flags);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 set_page_slub_counters(page, counters_new);
426 local_irq_restore(flags);
430 local_irq_restore(flags);
434 stat(s, CMPXCHG_DOUBLE_FAIL);
436 #ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
443 #ifdef CONFIG_SLUB_DEBUG
445 * Determine a map of object in use on a page.
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
450 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 void *addr = page_address(page);
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
462 #ifdef CONFIG_SLUB_DEBUG_ON
463 static int slub_debug = DEBUG_DEFAULT_FLAGS;
465 static int slub_debug;
468 static char *slub_debug_slabs;
469 static int disable_higher_order_debug;
472 * slub is about to manipulate internal object metadata. This memory lies
473 * outside the range of the allocated object, so accessing it would normally
474 * be reported by kasan as a bounds error. metadata_access_enable() is used
475 * to tell kasan that these accesses are OK.
477 static inline void metadata_access_enable(void)
479 kasan_disable_current();
482 static inline void metadata_access_disable(void)
484 kasan_enable_current();
490 static void print_section(char *text, u8 *addr, unsigned int length)
492 metadata_access_enable();
493 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
495 metadata_access_disable();
498 static struct track *get_track(struct kmem_cache *s, void *object,
499 enum track_item alloc)
504 p = object + s->offset + sizeof(void *);
506 p = object + s->inuse;
511 static void set_track(struct kmem_cache *s, void *object,
512 enum track_item alloc, unsigned long addr)
514 struct track *p = get_track(s, object, alloc);
517 #ifdef CONFIG_STACKTRACE
518 struct stack_trace trace;
521 trace.nr_entries = 0;
522 trace.max_entries = TRACK_ADDRS_COUNT;
523 trace.entries = p->addrs;
525 metadata_access_enable();
526 save_stack_trace(&trace);
527 metadata_access_disable();
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
542 memset(p, 0, sizeof(struct track));
545 static void init_tracking(struct kmem_cache *s, void *object)
547 if (!(s->flags & SLAB_STORE_USER))
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
554 static void print_track(const char *s, struct track *t)
559 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561 #ifdef CONFIG_STACKTRACE
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
566 pr_err("\t%pS\n", (void *)t->addrs[i]);
573 static void print_tracking(struct kmem_cache *s, void *object)
575 if (!(s->flags & SLAB_STORE_USER))
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
582 static void print_page_info(struct page *page)
584 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
589 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
591 struct va_format vaf;
597 pr_err("=============================================================================\n");
598 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
599 pr_err("-----------------------------------------------------------------------------\n\n");
601 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
605 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
607 struct va_format vaf;
613 pr_err("FIX %s: %pV\n", s->name, &vaf);
617 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
619 unsigned int off; /* Offset of last byte */
620 u8 *addr = page_address(page);
622 print_tracking(s, p);
624 print_page_info(page);
626 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
627 p, p - addr, get_freepointer(s, p));
630 print_section("Bytes b4 ", p - 16, 16);
632 print_section("Object ", p, min_t(unsigned long, s->object_size,
634 if (s->flags & SLAB_RED_ZONE)
635 print_section("Redzone ", p + s->object_size,
636 s->inuse - s->object_size);
639 off = s->offset + sizeof(void *);
643 if (s->flags & SLAB_STORE_USER)
644 off += 2 * sizeof(struct track);
647 /* Beginning of the filler is the free pointer */
648 print_section("Padding ", p + off, s->size - off);
653 void object_err(struct kmem_cache *s, struct page *page,
654 u8 *object, char *reason)
656 slab_bug(s, "%s", reason);
657 print_trailer(s, page, object);
660 static void slab_err(struct kmem_cache *s, struct page *page,
661 const char *fmt, ...)
667 vsnprintf(buf, sizeof(buf), fmt, args);
669 slab_bug(s, "%s", buf);
670 print_page_info(page);
674 static void init_object(struct kmem_cache *s, void *object, u8 val)
678 if (s->flags & __OBJECT_POISON) {
679 memset(p, POISON_FREE, s->object_size - 1);
680 p[s->object_size - 1] = POISON_END;
683 if (s->flags & SLAB_RED_ZONE)
684 memset(p + s->object_size, val, s->inuse - s->object_size);
687 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
688 void *from, void *to)
690 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
691 memset(from, data, to - from);
694 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
695 u8 *object, char *what,
696 u8 *start, unsigned int value, unsigned int bytes)
701 metadata_access_enable();
702 fault = memchr_inv(start, value, bytes);
703 metadata_access_disable();
708 while (end > fault && end[-1] == value)
711 slab_bug(s, "%s overwritten", what);
712 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
713 fault, end - 1, fault[0], value);
714 print_trailer(s, page, object);
716 restore_bytes(s, what, value, fault, end);
724 * Bytes of the object to be managed.
725 * If the freepointer may overlay the object then the free
726 * pointer is the first word of the object.
728 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
731 * object + s->object_size
732 * Padding to reach word boundary. This is also used for Redzoning.
733 * Padding is extended by another word if Redzoning is enabled and
734 * object_size == inuse.
736 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
737 * 0xcc (RED_ACTIVE) for objects in use.
740 * Meta data starts here.
742 * A. Free pointer (if we cannot overwrite object on free)
743 * B. Tracking data for SLAB_STORE_USER
744 * C. Padding to reach required alignment boundary or at mininum
745 * one word if debugging is on to be able to detect writes
746 * before the word boundary.
748 * Padding is done using 0x5a (POISON_INUSE)
751 * Nothing is used beyond s->size.
753 * If slabcaches are merged then the object_size and inuse boundaries are mostly
754 * ignored. And therefore no slab options that rely on these boundaries
755 * may be used with merged slabcaches.
758 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
760 unsigned long off = s->inuse; /* The end of info */
763 /* Freepointer is placed after the object. */
764 off += sizeof(void *);
766 if (s->flags & SLAB_STORE_USER)
767 /* We also have user information there */
768 off += 2 * sizeof(struct track);
773 return check_bytes_and_report(s, page, p, "Object padding",
774 p + off, POISON_INUSE, s->size - off);
777 /* Check the pad bytes at the end of a slab page */
778 static int slab_pad_check(struct kmem_cache *s, struct page *page)
786 if (!(s->flags & SLAB_POISON))
789 start = page_address(page);
790 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
791 end = start + length;
792 remainder = length % s->size;
796 metadata_access_enable();
797 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
798 metadata_access_disable();
801 while (end > fault && end[-1] == POISON_INUSE)
804 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
805 print_section("Padding ", end - remainder, remainder);
807 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
811 static int check_object(struct kmem_cache *s, struct page *page,
812 void *object, u8 val)
815 u8 *endobject = object + s->object_size;
817 if (s->flags & SLAB_RED_ZONE) {
818 if (!check_bytes_and_report(s, page, object, "Redzone",
819 endobject, val, s->inuse - s->object_size))
822 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
823 check_bytes_and_report(s, page, p, "Alignment padding",
824 endobject, POISON_INUSE,
825 s->inuse - s->object_size);
829 if (s->flags & SLAB_POISON) {
830 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
831 (!check_bytes_and_report(s, page, p, "Poison", p,
832 POISON_FREE, s->object_size - 1) ||
833 !check_bytes_and_report(s, page, p, "Poison",
834 p + s->object_size - 1, POISON_END, 1)))
837 * check_pad_bytes cleans up on its own.
839 check_pad_bytes(s, page, p);
842 if (!s->offset && val == SLUB_RED_ACTIVE)
844 * Object and freepointer overlap. Cannot check
845 * freepointer while object is allocated.
849 /* Check free pointer validity */
850 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
851 object_err(s, page, p, "Freepointer corrupt");
853 * No choice but to zap it and thus lose the remainder
854 * of the free objects in this slab. May cause
855 * another error because the object count is now wrong.
857 set_freepointer(s, p, NULL);
863 static int check_slab(struct kmem_cache *s, struct page *page)
867 VM_BUG_ON(!irqs_disabled());
869 if (!PageSlab(page)) {
870 slab_err(s, page, "Not a valid slab page");
874 maxobj = order_objects(compound_order(page), s->size, s->reserved);
875 if (page->objects > maxobj) {
876 slab_err(s, page, "objects %u > max %u",
877 page->objects, maxobj);
880 if (page->inuse > page->objects) {
881 slab_err(s, page, "inuse %u > max %u",
882 page->inuse, page->objects);
885 /* Slab_pad_check fixes things up after itself */
886 slab_pad_check(s, page);
891 * Determine if a certain object on a page is on the freelist. Must hold the
892 * slab lock to guarantee that the chains are in a consistent state.
894 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
902 while (fp && nr <= page->objects) {
905 if (!check_valid_pointer(s, page, fp)) {
907 object_err(s, page, object,
908 "Freechain corrupt");
909 set_freepointer(s, object, NULL);
911 slab_err(s, page, "Freepointer corrupt");
912 page->freelist = NULL;
913 page->inuse = page->objects;
914 slab_fix(s, "Freelist cleared");
920 fp = get_freepointer(s, object);
924 max_objects = order_objects(compound_order(page), s->size, s->reserved);
925 if (max_objects > MAX_OBJS_PER_PAGE)
926 max_objects = MAX_OBJS_PER_PAGE;
928 if (page->objects != max_objects) {
929 slab_err(s, page, "Wrong number of objects. Found %d but "
930 "should be %d", page->objects, max_objects);
931 page->objects = max_objects;
932 slab_fix(s, "Number of objects adjusted.");
934 if (page->inuse != page->objects - nr) {
935 slab_err(s, page, "Wrong object count. Counter is %d but "
936 "counted were %d", page->inuse, page->objects - nr);
937 page->inuse = page->objects - nr;
938 slab_fix(s, "Object count adjusted.");
940 return search == NULL;
943 static void trace(struct kmem_cache *s, struct page *page, void *object,
946 if (s->flags & SLAB_TRACE) {
947 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
949 alloc ? "alloc" : "free",
954 print_section("Object ", (void *)object,
962 * Tracking of fully allocated slabs for debugging purposes.
964 static void add_full(struct kmem_cache *s,
965 struct kmem_cache_node *n, struct page *page)
967 if (!(s->flags & SLAB_STORE_USER))
970 lockdep_assert_held(&n->list_lock);
971 list_add(&page->lru, &n->full);
974 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
976 if (!(s->flags & SLAB_STORE_USER))
979 lockdep_assert_held(&n->list_lock);
980 list_del(&page->lru);
983 /* Tracking of the number of slabs for debugging purposes */
984 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
986 struct kmem_cache_node *n = get_node(s, node);
988 return atomic_long_read(&n->nr_slabs);
991 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
993 return atomic_long_read(&n->nr_slabs);
996 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
998 struct kmem_cache_node *n = get_node(s, node);
1001 * May be called early in order to allocate a slab for the
1002 * kmem_cache_node structure. Solve the chicken-egg
1003 * dilemma by deferring the increment of the count during
1004 * bootstrap (see early_kmem_cache_node_alloc).
1007 atomic_long_inc(&n->nr_slabs);
1008 atomic_long_add(objects, &n->total_objects);
1011 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1013 struct kmem_cache_node *n = get_node(s, node);
1015 atomic_long_dec(&n->nr_slabs);
1016 atomic_long_sub(objects, &n->total_objects);
1019 /* Object debug checks for alloc/free paths */
1020 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1023 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1026 init_object(s, object, SLUB_RED_INACTIVE);
1027 init_tracking(s, object);
1030 static noinline int alloc_debug_processing(struct kmem_cache *s,
1032 void *object, unsigned long addr)
1034 if (!check_slab(s, page))
1037 if (!check_valid_pointer(s, page, object)) {
1038 object_err(s, page, object, "Freelist Pointer check fails");
1042 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1045 /* Success perform special debug activities for allocs */
1046 if (s->flags & SLAB_STORE_USER)
1047 set_track(s, object, TRACK_ALLOC, addr);
1048 trace(s, page, object, 1);
1049 init_object(s, object, SLUB_RED_ACTIVE);
1053 if (PageSlab(page)) {
1055 * If this is a slab page then lets do the best we can
1056 * to avoid issues in the future. Marking all objects
1057 * as used avoids touching the remaining objects.
1059 slab_fix(s, "Marking all objects used");
1060 page->inuse = page->objects;
1061 page->freelist = NULL;
1066 static noinline struct kmem_cache_node *free_debug_processing(
1067 struct kmem_cache *s, struct page *page, void *object,
1068 unsigned long addr, unsigned long *flags)
1070 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1072 raw_spin_lock_irqsave(&n->list_lock, *flags);
1075 if (!check_slab(s, page))
1078 if (!check_valid_pointer(s, page, object)) {
1079 slab_err(s, page, "Invalid object pointer 0x%p", object);
1083 if (on_freelist(s, page, object)) {
1084 object_err(s, page, object, "Object already free");
1088 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1091 if (unlikely(s != page->slab_cache)) {
1092 if (!PageSlab(page)) {
1093 slab_err(s, page, "Attempt to free object(0x%p) "
1094 "outside of slab", object);
1095 } else if (!page->slab_cache) {
1096 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1100 object_err(s, page, object,
1101 "page slab pointer corrupt.");
1105 if (s->flags & SLAB_STORE_USER)
1106 set_track(s, object, TRACK_FREE, addr);
1107 trace(s, page, object, 0);
1108 init_object(s, object, SLUB_RED_INACTIVE);
1112 * Keep node_lock to preserve integrity
1113 * until the object is actually freed
1119 raw_spin_unlock_irqrestore(&n->list_lock, *flags);
1120 slab_fix(s, "Object at 0x%p not freed", object);
1124 static int __init setup_slub_debug(char *str)
1126 slub_debug = DEBUG_DEFAULT_FLAGS;
1127 if (*str++ != '=' || !*str)
1129 * No options specified. Switch on full debugging.
1135 * No options but restriction on slabs. This means full
1136 * debugging for slabs matching a pattern.
1143 * Switch off all debugging measures.
1148 * Determine which debug features should be switched on
1150 for (; *str && *str != ','; str++) {
1151 switch (tolower(*str)) {
1153 slub_debug |= SLAB_DEBUG_FREE;
1156 slub_debug |= SLAB_RED_ZONE;
1159 slub_debug |= SLAB_POISON;
1162 slub_debug |= SLAB_STORE_USER;
1165 slub_debug |= SLAB_TRACE;
1168 slub_debug |= SLAB_FAILSLAB;
1172 * Avoid enabling debugging on caches if its minimum
1173 * order would increase as a result.
1175 disable_higher_order_debug = 1;
1178 pr_err("slub_debug option '%c' unknown. skipped\n",
1185 slub_debug_slabs = str + 1;
1190 __setup("slub_debug", setup_slub_debug);
1192 unsigned long kmem_cache_flags(unsigned long object_size,
1193 unsigned long flags, const char *name,
1194 void (*ctor)(void *))
1197 * Enable debugging if selected on the kernel commandline.
1199 if (slub_debug && (!slub_debug_slabs || (name &&
1200 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1201 flags |= slub_debug;
1206 static inline void setup_object_debug(struct kmem_cache *s,
1207 struct page *page, void *object) {}
1209 static inline int alloc_debug_processing(struct kmem_cache *s,
1210 struct page *page, void *object, unsigned long addr) { return 0; }
1212 static inline struct kmem_cache_node *free_debug_processing(
1213 struct kmem_cache *s, struct page *page, void *object,
1214 unsigned long addr, unsigned long *flags) { return NULL; }
1216 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1218 static inline int check_object(struct kmem_cache *s, struct page *page,
1219 void *object, u8 val) { return 1; }
1220 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1221 struct page *page) {}
1222 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1223 struct page *page) {}
1224 unsigned long kmem_cache_flags(unsigned long object_size,
1225 unsigned long flags, const char *name,
1226 void (*ctor)(void *))
1230 #define slub_debug 0
1232 #define disable_higher_order_debug 0
1234 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1236 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1238 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1240 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1243 #endif /* CONFIG_SLUB_DEBUG */
1245 struct slub_free_list {
1246 raw_spinlock_t lock;
1247 struct list_head list;
1249 static DEFINE_PER_CPU(struct slub_free_list, slub_free_list);
1252 * Hooks for other subsystems that check memory allocations. In a typical
1253 * production configuration these hooks all should produce no code at all.
1255 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1257 kmemleak_alloc(ptr, size, 1, flags);
1258 kasan_kmalloc_large(ptr, size);
1261 static inline void kfree_hook(const void *x)
1264 kasan_kfree_large(x);
1267 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1270 flags &= gfp_allowed_mask;
1271 lockdep_trace_alloc(flags);
1272 might_sleep_if(flags & __GFP_WAIT);
1274 if (should_failslab(s->object_size, flags, s->flags))
1277 return memcg_kmem_get_cache(s, flags);
1280 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1281 gfp_t flags, void *object)
1283 flags &= gfp_allowed_mask;
1284 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1285 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1286 memcg_kmem_put_cache(s);
1287 kasan_slab_alloc(s, object);
1290 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1292 kmemleak_free_recursive(x, s->flags);
1295 * Trouble is that we may no longer disable interrupts in the fast path
1296 * So in order to make the debug calls that expect irqs to be
1297 * disabled we need to disable interrupts temporarily.
1299 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1301 unsigned long flags;
1303 local_irq_save(flags);
1304 kmemcheck_slab_free(s, x, s->object_size);
1305 debug_check_no_locks_freed(x, s->object_size);
1306 local_irq_restore(flags);
1309 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1310 debug_check_no_obj_freed(x, s->object_size);
1312 kasan_slab_free(s, x);
1315 static void setup_object(struct kmem_cache *s, struct page *page,
1318 setup_object_debug(s, page, object);
1319 if (unlikely(s->ctor)) {
1320 kasan_unpoison_object_data(s, object);
1322 kasan_poison_object_data(s, object);
1327 * Slab allocation and freeing
1329 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1330 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1333 int order = oo_order(oo);
1335 flags |= __GFP_NOTRACK;
1337 if (memcg_charge_slab(s, flags, order))
1340 if (node == NUMA_NO_NODE)
1341 page = alloc_pages(flags, order);
1343 page = alloc_pages_exact_node(node, flags, order);
1346 memcg_uncharge_slab(s, order);
1351 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1354 struct kmem_cache_order_objects oo = s->oo;
1360 flags &= gfp_allowed_mask;
1362 enableirqs = (flags & __GFP_WAIT) != 0;
1363 #ifdef CONFIG_PREEMPT_RT_FULL
1364 enableirqs |= system_state == SYSTEM_RUNNING;
1369 flags |= s->allocflags;
1372 * Let the initial higher-order allocation fail under memory pressure
1373 * so we fall-back to the minimum order allocation.
1375 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1377 page = alloc_slab_page(s, alloc_gfp, node, oo);
1378 if (unlikely(!page)) {
1382 * Allocation may have failed due to fragmentation.
1383 * Try a lower order alloc if possible
1385 page = alloc_slab_page(s, alloc_gfp, node, oo);
1386 if (unlikely(!page))
1388 stat(s, ORDER_FALLBACK);
1391 if (kmemcheck_enabled &&
1392 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1393 int pages = 1 << oo_order(oo);
1395 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1398 * Objects from caches that have a constructor don't get
1399 * cleared when they're allocated, so we need to do it here.
1402 kmemcheck_mark_uninitialized_pages(page, pages);
1404 kmemcheck_mark_unallocated_pages(page, pages);
1407 page->objects = oo_objects(oo);
1409 order = compound_order(page);
1410 page->slab_cache = s;
1411 __SetPageSlab(page);
1412 if (page_is_pfmemalloc(page))
1413 SetPageSlabPfmemalloc(page);
1415 start = page_address(page);
1417 if (unlikely(s->flags & SLAB_POISON))
1418 memset(start, POISON_INUSE, PAGE_SIZE << order);
1420 kasan_poison_slab(page);
1422 for_each_object_idx(p, idx, s, start, page->objects) {
1423 setup_object(s, page, p);
1424 if (likely(idx < page->objects))
1425 set_freepointer(s, p, p + s->size);
1427 set_freepointer(s, p, NULL);
1430 page->freelist = start;
1431 page->inuse = page->objects;
1436 local_irq_disable();
1440 mod_zone_page_state(page_zone(page),
1441 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1442 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1445 inc_slabs_node(s, page_to_nid(page), page->objects);
1450 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1452 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1453 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1457 return allocate_slab(s,
1458 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1461 static void __free_slab(struct kmem_cache *s, struct page *page)
1463 int order = compound_order(page);
1464 int pages = 1 << order;
1466 if (kmem_cache_debug(s)) {
1469 slab_pad_check(s, page);
1470 for_each_object(p, s, page_address(page),
1472 check_object(s, page, p, SLUB_RED_INACTIVE);
1475 kmemcheck_free_shadow(page, compound_order(page));
1477 mod_zone_page_state(page_zone(page),
1478 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1479 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1482 __ClearPageSlabPfmemalloc(page);
1483 __ClearPageSlab(page);
1485 page_mapcount_reset(page);
1486 if (current->reclaim_state)
1487 current->reclaim_state->reclaimed_slab += pages;
1488 __free_pages(page, order);
1489 memcg_uncharge_slab(s, order);
1492 static void free_delayed(struct list_head *h)
1494 while(!list_empty(h)) {
1495 struct page *page = list_first_entry(h, struct page, lru);
1497 list_del(&page->lru);
1498 __free_slab(page->slab_cache, page);
1502 #define need_reserve_slab_rcu \
1503 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1505 static void rcu_free_slab(struct rcu_head *h)
1509 if (need_reserve_slab_rcu)
1510 page = virt_to_head_page(h);
1512 page = container_of((struct list_head *)h, struct page, lru);
1514 __free_slab(page->slab_cache, page);
1517 static void free_slab(struct kmem_cache *s, struct page *page)
1519 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1520 struct rcu_head *head;
1522 if (need_reserve_slab_rcu) {
1523 int order = compound_order(page);
1524 int offset = (PAGE_SIZE << order) - s->reserved;
1526 VM_BUG_ON(s->reserved != sizeof(*head));
1527 head = page_address(page) + offset;
1530 * RCU free overloads the RCU head over the LRU
1532 head = (void *)&page->lru;
1535 call_rcu(head, rcu_free_slab);
1536 } else if (irqs_disabled()) {
1537 struct slub_free_list *f = this_cpu_ptr(&slub_free_list);
1539 raw_spin_lock(&f->lock);
1540 list_add(&page->lru, &f->list);
1541 raw_spin_unlock(&f->lock);
1543 __free_slab(s, page);
1546 static void discard_slab(struct kmem_cache *s, struct page *page)
1548 dec_slabs_node(s, page_to_nid(page), page->objects);
1553 * Management of partially allocated slabs.
1556 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1559 if (tail == DEACTIVATE_TO_TAIL)
1560 list_add_tail(&page->lru, &n->partial);
1562 list_add(&page->lru, &n->partial);
1565 static inline void add_partial(struct kmem_cache_node *n,
1566 struct page *page, int tail)
1568 lockdep_assert_held(&n->list_lock);
1569 __add_partial(n, page, tail);
1573 __remove_partial(struct kmem_cache_node *n, struct page *page)
1575 list_del(&page->lru);
1579 static inline void remove_partial(struct kmem_cache_node *n,
1582 lockdep_assert_held(&n->list_lock);
1583 __remove_partial(n, page);
1587 * Remove slab from the partial list, freeze it and
1588 * return the pointer to the freelist.
1590 * Returns a list of objects or NULL if it fails.
1592 static inline void *acquire_slab(struct kmem_cache *s,
1593 struct kmem_cache_node *n, struct page *page,
1594 int mode, int *objects)
1597 unsigned long counters;
1600 lockdep_assert_held(&n->list_lock);
1603 * Zap the freelist and set the frozen bit.
1604 * The old freelist is the list of objects for the
1605 * per cpu allocation list.
1607 freelist = page->freelist;
1608 counters = page->counters;
1609 new.counters = counters;
1610 *objects = new.objects - new.inuse;
1612 new.inuse = page->objects;
1613 new.freelist = NULL;
1615 new.freelist = freelist;
1618 VM_BUG_ON(new.frozen);
1621 if (!__cmpxchg_double_slab(s, page,
1623 new.freelist, new.counters,
1627 remove_partial(n, page);
1632 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1633 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1636 * Try to allocate a partial slab from a specific node.
1638 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1639 struct kmem_cache_cpu *c, gfp_t flags)
1641 struct page *page, *page2;
1642 void *object = NULL;
1647 * Racy check. If we mistakenly see no partial slabs then we
1648 * just allocate an empty slab. If we mistakenly try to get a
1649 * partial slab and there is none available then get_partials()
1652 if (!n || !n->nr_partial)
1655 raw_spin_lock(&n->list_lock);
1656 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1659 if (!pfmemalloc_match(page, flags))
1662 t = acquire_slab(s, n, page, object == NULL, &objects);
1666 available += objects;
1669 stat(s, ALLOC_FROM_PARTIAL);
1672 put_cpu_partial(s, page, 0);
1673 stat(s, CPU_PARTIAL_NODE);
1675 if (!kmem_cache_has_cpu_partial(s)
1676 || available > s->cpu_partial / 2)
1680 raw_spin_unlock(&n->list_lock);
1685 * Get a page from somewhere. Search in increasing NUMA distances.
1687 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1688 struct kmem_cache_cpu *c)
1691 struct zonelist *zonelist;
1694 enum zone_type high_zoneidx = gfp_zone(flags);
1696 unsigned int cpuset_mems_cookie;
1699 * The defrag ratio allows a configuration of the tradeoffs between
1700 * inter node defragmentation and node local allocations. A lower
1701 * defrag_ratio increases the tendency to do local allocations
1702 * instead of attempting to obtain partial slabs from other nodes.
1704 * If the defrag_ratio is set to 0 then kmalloc() always
1705 * returns node local objects. If the ratio is higher then kmalloc()
1706 * may return off node objects because partial slabs are obtained
1707 * from other nodes and filled up.
1709 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1710 * defrag_ratio = 1000) then every (well almost) allocation will
1711 * first attempt to defrag slab caches on other nodes. This means
1712 * scanning over all nodes to look for partial slabs which may be
1713 * expensive if we do it every time we are trying to find a slab
1714 * with available objects.
1716 if (!s->remote_node_defrag_ratio ||
1717 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1721 cpuset_mems_cookie = read_mems_allowed_begin();
1722 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1723 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1724 struct kmem_cache_node *n;
1726 n = get_node(s, zone_to_nid(zone));
1728 if (n && cpuset_zone_allowed(zone, flags) &&
1729 n->nr_partial > s->min_partial) {
1730 object = get_partial_node(s, n, c, flags);
1733 * Don't check read_mems_allowed_retry()
1734 * here - if mems_allowed was updated in
1735 * parallel, that was a harmless race
1736 * between allocation and the cpuset
1743 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1749 * Get a partial page, lock it and return it.
1751 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1752 struct kmem_cache_cpu *c)
1755 int searchnode = node;
1757 if (node == NUMA_NO_NODE)
1758 searchnode = numa_mem_id();
1759 else if (!node_present_pages(node))
1760 searchnode = node_to_mem_node(node);
1762 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1763 if (object || node != NUMA_NO_NODE)
1766 return get_any_partial(s, flags, c);
1769 #ifdef CONFIG_PREEMPT
1771 * Calculate the next globally unique transaction for disambiguiation
1772 * during cmpxchg. The transactions start with the cpu number and are then
1773 * incremented by CONFIG_NR_CPUS.
1775 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1778 * No preemption supported therefore also no need to check for
1784 static inline unsigned long next_tid(unsigned long tid)
1786 return tid + TID_STEP;
1789 static inline unsigned int tid_to_cpu(unsigned long tid)
1791 return tid % TID_STEP;
1794 static inline unsigned long tid_to_event(unsigned long tid)
1796 return tid / TID_STEP;
1799 static inline unsigned int init_tid(int cpu)
1804 static inline void note_cmpxchg_failure(const char *n,
1805 const struct kmem_cache *s, unsigned long tid)
1807 #ifdef SLUB_DEBUG_CMPXCHG
1808 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1810 pr_info("%s %s: cmpxchg redo ", n, s->name);
1812 #ifdef CONFIG_PREEMPT
1813 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1814 pr_warn("due to cpu change %d -> %d\n",
1815 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1818 if (tid_to_event(tid) != tid_to_event(actual_tid))
1819 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1820 tid_to_event(tid), tid_to_event(actual_tid));
1822 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1823 actual_tid, tid, next_tid(tid));
1825 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1828 static void init_kmem_cache_cpus(struct kmem_cache *s)
1832 for_each_possible_cpu(cpu)
1833 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1837 * Remove the cpu slab
1839 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1842 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1843 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1845 enum slab_modes l = M_NONE, m = M_NONE;
1847 int tail = DEACTIVATE_TO_HEAD;
1851 if (page->freelist) {
1852 stat(s, DEACTIVATE_REMOTE_FREES);
1853 tail = DEACTIVATE_TO_TAIL;
1857 * Stage one: Free all available per cpu objects back
1858 * to the page freelist while it is still frozen. Leave the
1861 * There is no need to take the list->lock because the page
1864 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1866 unsigned long counters;
1869 prior = page->freelist;
1870 counters = page->counters;
1871 set_freepointer(s, freelist, prior);
1872 new.counters = counters;
1874 VM_BUG_ON(!new.frozen);
1876 } while (!__cmpxchg_double_slab(s, page,
1878 freelist, new.counters,
1879 "drain percpu freelist"));
1881 freelist = nextfree;
1885 * Stage two: Ensure that the page is unfrozen while the
1886 * list presence reflects the actual number of objects
1889 * We setup the list membership and then perform a cmpxchg
1890 * with the count. If there is a mismatch then the page
1891 * is not unfrozen but the page is on the wrong list.
1893 * Then we restart the process which may have to remove
1894 * the page from the list that we just put it on again
1895 * because the number of objects in the slab may have
1900 old.freelist = page->freelist;
1901 old.counters = page->counters;
1902 VM_BUG_ON(!old.frozen);
1904 /* Determine target state of the slab */
1905 new.counters = old.counters;
1908 set_freepointer(s, freelist, old.freelist);
1909 new.freelist = freelist;
1911 new.freelist = old.freelist;
1915 if (!new.inuse && n->nr_partial >= s->min_partial)
1917 else if (new.freelist) {
1922 * Taking the spinlock removes the possiblity
1923 * that acquire_slab() will see a slab page that
1926 raw_spin_lock(&n->list_lock);
1930 if (kmem_cache_debug(s) && !lock) {
1933 * This also ensures that the scanning of full
1934 * slabs from diagnostic functions will not see
1937 raw_spin_lock(&n->list_lock);
1945 remove_partial(n, page);
1947 else if (l == M_FULL)
1949 remove_full(s, n, page);
1951 if (m == M_PARTIAL) {
1953 add_partial(n, page, tail);
1956 } else if (m == M_FULL) {
1958 stat(s, DEACTIVATE_FULL);
1959 add_full(s, n, page);
1965 if (!__cmpxchg_double_slab(s, page,
1966 old.freelist, old.counters,
1967 new.freelist, new.counters,
1972 raw_spin_unlock(&n->list_lock);
1975 stat(s, DEACTIVATE_EMPTY);
1976 discard_slab(s, page);
1982 * Unfreeze all the cpu partial slabs.
1984 * This function must be called with interrupts disabled
1985 * for the cpu using c (or some other guarantee must be there
1986 * to guarantee no concurrent accesses).
1988 static void unfreeze_partials(struct kmem_cache *s,
1989 struct kmem_cache_cpu *c)
1991 #ifdef CONFIG_SLUB_CPU_PARTIAL
1992 struct kmem_cache_node *n = NULL, *n2 = NULL;
1993 struct page *page, *discard_page = NULL;
1995 while ((page = c->partial)) {
1999 c->partial = page->next;
2001 n2 = get_node(s, page_to_nid(page));
2004 raw_spin_unlock(&n->list_lock);
2007 raw_spin_lock(&n->list_lock);
2012 old.freelist = page->freelist;
2013 old.counters = page->counters;
2014 VM_BUG_ON(!old.frozen);
2016 new.counters = old.counters;
2017 new.freelist = old.freelist;
2021 } while (!__cmpxchg_double_slab(s, page,
2022 old.freelist, old.counters,
2023 new.freelist, new.counters,
2024 "unfreezing slab"));
2026 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2027 page->next = discard_page;
2028 discard_page = page;
2030 add_partial(n, page, DEACTIVATE_TO_TAIL);
2031 stat(s, FREE_ADD_PARTIAL);
2036 raw_spin_unlock(&n->list_lock);
2038 while (discard_page) {
2039 page = discard_page;
2040 discard_page = discard_page->next;
2042 stat(s, DEACTIVATE_EMPTY);
2043 discard_slab(s, page);
2050 * Put a page that was just frozen (in __slab_free) into a partial page
2051 * slot if available. This is done without interrupts disabled and without
2052 * preemption disabled. The cmpxchg is racy and may put the partial page
2053 * onto a random cpus partial slot.
2055 * If we did not find a slot then simply move all the partials to the
2056 * per node partial list.
2058 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2060 #ifdef CONFIG_SLUB_CPU_PARTIAL
2061 struct page *oldpage;
2069 oldpage = this_cpu_read(s->cpu_slab->partial);
2072 pobjects = oldpage->pobjects;
2073 pages = oldpage->pages;
2074 if (drain && pobjects > s->cpu_partial) {
2075 struct slub_free_list *f;
2076 unsigned long flags;
2079 * partial array is full. Move the existing
2080 * set to the per node partial list.
2082 local_irq_save(flags);
2083 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2084 f = this_cpu_ptr(&slub_free_list);
2085 raw_spin_lock(&f->lock);
2086 list_splice_init(&f->list, &tofree);
2087 raw_spin_unlock(&f->lock);
2088 local_irq_restore(flags);
2089 free_delayed(&tofree);
2093 stat(s, CPU_PARTIAL_DRAIN);
2098 pobjects += page->objects - page->inuse;
2100 page->pages = pages;
2101 page->pobjects = pobjects;
2102 page->next = oldpage;
2104 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2106 if (unlikely(!s->cpu_partial)) {
2107 unsigned long flags;
2109 local_irq_save(flags);
2110 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2111 local_irq_restore(flags);
2117 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2119 stat(s, CPUSLAB_FLUSH);
2120 deactivate_slab(s, c->page, c->freelist);
2122 c->tid = next_tid(c->tid);
2130 * Called from IPI handler with interrupts disabled.
2132 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2134 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2140 unfreeze_partials(s, c);
2144 static void flush_cpu_slab(void *d)
2146 struct kmem_cache *s = d;
2148 __flush_cpu_slab(s, smp_processor_id());
2151 static bool has_cpu_slab(int cpu, void *info)
2153 struct kmem_cache *s = info;
2154 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2156 return c->page || c->partial;
2159 static void flush_all(struct kmem_cache *s)
2164 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2165 for_each_online_cpu(cpu) {
2166 struct slub_free_list *f;
2168 if (!has_cpu_slab(cpu, s))
2171 f = &per_cpu(slub_free_list, cpu);
2172 raw_spin_lock_irq(&f->lock);
2173 list_splice_init(&f->list, &tofree);
2174 raw_spin_unlock_irq(&f->lock);
2175 free_delayed(&tofree);
2180 * Check if the objects in a per cpu structure fit numa
2181 * locality expectations.
2183 static inline int node_match(struct page *page, int node)
2186 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2192 #ifdef CONFIG_SLUB_DEBUG
2193 static int count_free(struct page *page)
2195 return page->objects - page->inuse;
2198 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2200 return atomic_long_read(&n->total_objects);
2202 #endif /* CONFIG_SLUB_DEBUG */
2204 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2205 static unsigned long count_partial(struct kmem_cache_node *n,
2206 int (*get_count)(struct page *))
2208 unsigned long flags;
2209 unsigned long x = 0;
2212 raw_spin_lock_irqsave(&n->list_lock, flags);
2213 list_for_each_entry(page, &n->partial, lru)
2214 x += get_count(page);
2215 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2218 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2220 static noinline void
2221 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2223 #ifdef CONFIG_SLUB_DEBUG
2224 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2225 DEFAULT_RATELIMIT_BURST);
2227 struct kmem_cache_node *n;
2229 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2232 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2234 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2235 s->name, s->object_size, s->size, oo_order(s->oo),
2238 if (oo_order(s->min) > get_order(s->object_size))
2239 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2242 for_each_kmem_cache_node(s, node, n) {
2243 unsigned long nr_slabs;
2244 unsigned long nr_objs;
2245 unsigned long nr_free;
2247 nr_free = count_partial(n, count_free);
2248 nr_slabs = node_nr_slabs(n);
2249 nr_objs = node_nr_objs(n);
2251 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2252 node, nr_slabs, nr_objs, nr_free);
2257 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2258 int node, struct kmem_cache_cpu **pc)
2261 struct kmem_cache_cpu *c = *pc;
2264 freelist = get_partial(s, flags, node, c);
2269 page = new_slab(s, flags, node);
2271 c = raw_cpu_ptr(s->cpu_slab);
2276 * No other reference to the page yet so we can
2277 * muck around with it freely without cmpxchg
2279 freelist = page->freelist;
2280 page->freelist = NULL;
2282 stat(s, ALLOC_SLAB);
2291 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2293 if (unlikely(PageSlabPfmemalloc(page)))
2294 return gfp_pfmemalloc_allowed(gfpflags);
2300 * Check the page->freelist of a page and either transfer the freelist to the
2301 * per cpu freelist or deactivate the page.
2303 * The page is still frozen if the return value is not NULL.
2305 * If this function returns NULL then the page has been unfrozen.
2307 * This function must be called with interrupt disabled.
2309 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2312 unsigned long counters;
2316 freelist = page->freelist;
2317 counters = page->counters;
2319 new.counters = counters;
2320 VM_BUG_ON(!new.frozen);
2322 new.inuse = page->objects;
2323 new.frozen = freelist != NULL;
2325 } while (!__cmpxchg_double_slab(s, page,
2334 * Slow path. The lockless freelist is empty or we need to perform
2337 * Processing is still very fast if new objects have been freed to the
2338 * regular freelist. In that case we simply take over the regular freelist
2339 * as the lockless freelist and zap the regular freelist.
2341 * If that is not working then we fall back to the partial lists. We take the
2342 * first element of the freelist as the object to allocate now and move the
2343 * rest of the freelist to the lockless freelist.
2345 * And if we were unable to get a new slab from the partial slab lists then
2346 * we need to allocate a new slab. This is the slowest path since it involves
2347 * a call to the page allocator and the setup of a new slab.
2349 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2350 unsigned long addr, struct kmem_cache_cpu *c)
2352 struct slub_free_list *f;
2355 unsigned long flags;
2358 local_irq_save(flags);
2359 #ifdef CONFIG_PREEMPT
2361 * We may have been preempted and rescheduled on a different
2362 * cpu before disabling interrupts. Need to reload cpu area
2365 c = this_cpu_ptr(s->cpu_slab);
2373 if (unlikely(!node_match(page, node))) {
2374 int searchnode = node;
2376 if (node != NUMA_NO_NODE && !node_present_pages(node))
2377 searchnode = node_to_mem_node(node);
2379 if (unlikely(!node_match(page, searchnode))) {
2380 stat(s, ALLOC_NODE_MISMATCH);
2381 deactivate_slab(s, page, c->freelist);
2389 * By rights, we should be searching for a slab page that was
2390 * PFMEMALLOC but right now, we are losing the pfmemalloc
2391 * information when the page leaves the per-cpu allocator
2393 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2394 deactivate_slab(s, page, c->freelist);
2400 /* must check again c->freelist in case of cpu migration or IRQ */
2401 freelist = c->freelist;
2405 freelist = get_freelist(s, page);
2409 stat(s, DEACTIVATE_BYPASS);
2413 stat(s, ALLOC_REFILL);
2417 * freelist is pointing to the list of objects to be used.
2418 * page is pointing to the page from which the objects are obtained.
2419 * That page must be frozen for per cpu allocations to work.
2421 VM_BUG_ON(!c->page->frozen);
2422 c->freelist = get_freepointer(s, freelist);
2423 c->tid = next_tid(c->tid);
2425 f = this_cpu_ptr(&slub_free_list);
2426 raw_spin_lock(&f->lock);
2427 list_splice_init(&f->list, &tofree);
2428 raw_spin_unlock(&f->lock);
2429 local_irq_restore(flags);
2430 free_delayed(&tofree);
2436 page = c->page = c->partial;
2437 c->partial = page->next;
2438 stat(s, CPU_PARTIAL_ALLOC);
2443 freelist = new_slab_objects(s, gfpflags, node, &c);
2445 if (unlikely(!freelist)) {
2446 slab_out_of_memory(s, gfpflags, node);
2451 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2454 /* Only entered in the debug case */
2455 if (kmem_cache_debug(s) &&
2456 !alloc_debug_processing(s, page, freelist, addr))
2457 goto new_slab; /* Slab failed checks. Next slab needed */
2459 deactivate_slab(s, page, get_freepointer(s, freelist));
2466 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2467 * have the fastpath folded into their functions. So no function call
2468 * overhead for requests that can be satisfied on the fastpath.
2470 * The fastpath works by first checking if the lockless freelist can be used.
2471 * If not then __slab_alloc is called for slow processing.
2473 * Otherwise we can simply pick the next object from the lockless free list.
2475 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2476 gfp_t gfpflags, int node, unsigned long addr)
2479 struct kmem_cache_cpu *c;
2483 s = slab_pre_alloc_hook(s, gfpflags);
2488 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2489 * enabled. We may switch back and forth between cpus while
2490 * reading from one cpu area. That does not matter as long
2491 * as we end up on the original cpu again when doing the cmpxchg.
2493 * We should guarantee that tid and kmem_cache are retrieved on
2494 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2495 * to check if it is matched or not.
2498 tid = this_cpu_read(s->cpu_slab->tid);
2499 c = raw_cpu_ptr(s->cpu_slab);
2500 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2501 unlikely(tid != READ_ONCE(c->tid)));
2504 * Irqless object alloc/free algorithm used here depends on sequence
2505 * of fetching cpu_slab's data. tid should be fetched before anything
2506 * on c to guarantee that object and page associated with previous tid
2507 * won't be used with current tid. If we fetch tid first, object and
2508 * page could be one associated with next tid and our alloc/free
2509 * request will be failed. In this case, we will retry. So, no problem.
2514 * The transaction ids are globally unique per cpu and per operation on
2515 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2516 * occurs on the right processor and that there was no operation on the
2517 * linked list in between.
2520 object = c->freelist;
2522 if (unlikely(!object || !node_match(page, node))) {
2523 object = __slab_alloc(s, gfpflags, node, addr, c);
2524 stat(s, ALLOC_SLOWPATH);
2526 void *next_object = get_freepointer_safe(s, object);
2529 * The cmpxchg will only match if there was no additional
2530 * operation and if we are on the right processor.
2532 * The cmpxchg does the following atomically (without lock
2534 * 1. Relocate first pointer to the current per cpu area.
2535 * 2. Verify that tid and freelist have not been changed
2536 * 3. If they were not changed replace tid and freelist
2538 * Since this is without lock semantics the protection is only
2539 * against code executing on this cpu *not* from access by
2542 if (unlikely(!this_cpu_cmpxchg_double(
2543 s->cpu_slab->freelist, s->cpu_slab->tid,
2545 next_object, next_tid(tid)))) {
2547 note_cmpxchg_failure("slab_alloc", s, tid);
2550 prefetch_freepointer(s, next_object);
2551 stat(s, ALLOC_FASTPATH);
2554 if (unlikely(gfpflags & __GFP_ZERO) && object)
2555 memset(object, 0, s->object_size);
2557 slab_post_alloc_hook(s, gfpflags, object);
2562 static __always_inline void *slab_alloc(struct kmem_cache *s,
2563 gfp_t gfpflags, unsigned long addr)
2565 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2568 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2570 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2572 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2577 EXPORT_SYMBOL(kmem_cache_alloc);
2579 #ifdef CONFIG_TRACING
2580 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2582 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2583 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2584 kasan_kmalloc(s, ret, size);
2587 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2591 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2593 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2595 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2596 s->object_size, s->size, gfpflags, node);
2600 EXPORT_SYMBOL(kmem_cache_alloc_node);
2602 #ifdef CONFIG_TRACING
2603 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2605 int node, size_t size)
2607 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2609 trace_kmalloc_node(_RET_IP_, ret,
2610 size, s->size, gfpflags, node);
2612 kasan_kmalloc(s, ret, size);
2615 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2620 * Slow path handling. This may still be called frequently since objects
2621 * have a longer lifetime than the cpu slabs in most processing loads.
2623 * So we still attempt to reduce cache line usage. Just take the slab
2624 * lock and free the item. If there is no additional partial page
2625 * handling required then we can return immediately.
2627 static void __slab_free(struct kmem_cache *s, struct page *page,
2628 void *x, unsigned long addr)
2631 void **object = (void *)x;
2634 unsigned long counters;
2635 struct kmem_cache_node *n = NULL;
2636 unsigned long uninitialized_var(flags);
2638 stat(s, FREE_SLOWPATH);
2640 if (kmem_cache_debug(s) &&
2641 !(n = free_debug_processing(s, page, x, addr, &flags)))
2646 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2649 prior = page->freelist;
2650 counters = page->counters;
2651 set_freepointer(s, object, prior);
2652 new.counters = counters;
2653 was_frozen = new.frozen;
2655 if ((!new.inuse || !prior) && !was_frozen) {
2657 if (kmem_cache_has_cpu_partial(s) && !prior) {
2660 * Slab was on no list before and will be
2662 * We can defer the list move and instead
2667 } else { /* Needs to be taken off a list */
2669 n = get_node(s, page_to_nid(page));
2671 * Speculatively acquire the list_lock.
2672 * If the cmpxchg does not succeed then we may
2673 * drop the list_lock without any processing.
2675 * Otherwise the list_lock will synchronize with
2676 * other processors updating the list of slabs.
2678 raw_spin_lock_irqsave(&n->list_lock, flags);
2683 } while (!cmpxchg_double_slab(s, page,
2685 object, new.counters,
2691 * If we just froze the page then put it onto the
2692 * per cpu partial list.
2694 if (new.frozen && !was_frozen) {
2695 put_cpu_partial(s, page, 1);
2696 stat(s, CPU_PARTIAL_FREE);
2699 * The list lock was not taken therefore no list
2700 * activity can be necessary.
2703 stat(s, FREE_FROZEN);
2707 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2711 * Objects left in the slab. If it was not on the partial list before
2714 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2715 if (kmem_cache_debug(s))
2716 remove_full(s, n, page);
2717 add_partial(n, page, DEACTIVATE_TO_TAIL);
2718 stat(s, FREE_ADD_PARTIAL);
2720 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2726 * Slab on the partial list.
2728 remove_partial(n, page);
2729 stat(s, FREE_REMOVE_PARTIAL);
2731 /* Slab must be on the full list */
2732 remove_full(s, n, page);
2735 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2737 discard_slab(s, page);
2741 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2742 * can perform fastpath freeing without additional function calls.
2744 * The fastpath is only possible if we are freeing to the current cpu slab
2745 * of this processor. This typically the case if we have just allocated
2748 * If fastpath is not possible then fall back to __slab_free where we deal
2749 * with all sorts of special processing.
2751 static __always_inline void slab_free(struct kmem_cache *s,
2752 struct page *page, void *x, unsigned long addr)
2754 void **object = (void *)x;
2755 struct kmem_cache_cpu *c;
2758 slab_free_hook(s, x);
2762 * Determine the currently cpus per cpu slab.
2763 * The cpu may change afterward. However that does not matter since
2764 * data is retrieved via this pointer. If we are on the same cpu
2765 * during the cmpxchg then the free will succedd.
2768 tid = this_cpu_read(s->cpu_slab->tid);
2769 c = raw_cpu_ptr(s->cpu_slab);
2770 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2771 unlikely(tid != READ_ONCE(c->tid)));
2773 /* Same with comment on barrier() in slab_alloc_node() */
2776 if (likely(page == c->page)) {
2777 set_freepointer(s, object, c->freelist);
2779 if (unlikely(!this_cpu_cmpxchg_double(
2780 s->cpu_slab->freelist, s->cpu_slab->tid,
2782 object, next_tid(tid)))) {
2784 note_cmpxchg_failure("slab_free", s, tid);
2787 stat(s, FREE_FASTPATH);
2789 __slab_free(s, page, x, addr);
2793 void kmem_cache_free(struct kmem_cache *s, void *x)
2795 s = cache_from_obj(s, x);
2798 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2799 trace_kmem_cache_free(_RET_IP_, x);
2801 EXPORT_SYMBOL(kmem_cache_free);
2804 * Object placement in a slab is made very easy because we always start at
2805 * offset 0. If we tune the size of the object to the alignment then we can
2806 * get the required alignment by putting one properly sized object after
2809 * Notice that the allocation order determines the sizes of the per cpu
2810 * caches. Each processor has always one slab available for allocations.
2811 * Increasing the allocation order reduces the number of times that slabs
2812 * must be moved on and off the partial lists and is therefore a factor in
2817 * Mininum / Maximum order of slab pages. This influences locking overhead
2818 * and slab fragmentation. A higher order reduces the number of partial slabs
2819 * and increases the number of allocations possible without having to
2820 * take the list_lock.
2822 static int slub_min_order;
2823 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2824 static int slub_min_objects;
2827 * Calculate the order of allocation given an slab object size.
2829 * The order of allocation has significant impact on performance and other
2830 * system components. Generally order 0 allocations should be preferred since
2831 * order 0 does not cause fragmentation in the page allocator. Larger objects
2832 * be problematic to put into order 0 slabs because there may be too much
2833 * unused space left. We go to a higher order if more than 1/16th of the slab
2836 * In order to reach satisfactory performance we must ensure that a minimum
2837 * number of objects is in one slab. Otherwise we may generate too much
2838 * activity on the partial lists which requires taking the list_lock. This is
2839 * less a concern for large slabs though which are rarely used.
2841 * slub_max_order specifies the order where we begin to stop considering the
2842 * number of objects in a slab as critical. If we reach slub_max_order then
2843 * we try to keep the page order as low as possible. So we accept more waste
2844 * of space in favor of a small page order.
2846 * Higher order allocations also allow the placement of more objects in a
2847 * slab and thereby reduce object handling overhead. If the user has
2848 * requested a higher mininum order then we start with that one instead of
2849 * the smallest order which will fit the object.
2851 static inline int slab_order(int size, int min_objects,
2852 int max_order, int fract_leftover, int reserved)
2856 int min_order = slub_min_order;
2858 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2859 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2861 for (order = max(min_order,
2862 fls(min_objects * size - 1) - PAGE_SHIFT);
2863 order <= max_order; order++) {
2865 unsigned long slab_size = PAGE_SIZE << order;
2867 if (slab_size < min_objects * size + reserved)
2870 rem = (slab_size - reserved) % size;
2872 if (rem <= slab_size / fract_leftover)
2880 static inline int calculate_order(int size, int reserved)
2888 * Attempt to find best configuration for a slab. This
2889 * works by first attempting to generate a layout with
2890 * the best configuration and backing off gradually.
2892 * First we reduce the acceptable waste in a slab. Then
2893 * we reduce the minimum objects required in a slab.
2895 min_objects = slub_min_objects;
2897 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2898 max_objects = order_objects(slub_max_order, size, reserved);
2899 min_objects = min(min_objects, max_objects);
2901 while (min_objects > 1) {
2903 while (fraction >= 4) {
2904 order = slab_order(size, min_objects,
2905 slub_max_order, fraction, reserved);
2906 if (order <= slub_max_order)
2914 * We were unable to place multiple objects in a slab. Now
2915 * lets see if we can place a single object there.
2917 order = slab_order(size, 1, slub_max_order, 1, reserved);
2918 if (order <= slub_max_order)
2922 * Doh this slab cannot be placed using slub_max_order.
2924 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2925 if (order < MAX_ORDER)
2931 init_kmem_cache_node(struct kmem_cache_node *n)
2934 raw_spin_lock_init(&n->list_lock);
2935 INIT_LIST_HEAD(&n->partial);
2936 #ifdef CONFIG_SLUB_DEBUG
2937 atomic_long_set(&n->nr_slabs, 0);
2938 atomic_long_set(&n->total_objects, 0);
2939 INIT_LIST_HEAD(&n->full);
2943 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2945 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2946 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2949 * Must align to double word boundary for the double cmpxchg
2950 * instructions to work; see __pcpu_double_call_return_bool().
2952 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2953 2 * sizeof(void *));
2958 init_kmem_cache_cpus(s);
2963 static struct kmem_cache *kmem_cache_node;
2966 * No kmalloc_node yet so do it by hand. We know that this is the first
2967 * slab on the node for this slabcache. There are no concurrent accesses
2970 * Note that this function only works on the kmem_cache_node
2971 * when allocating for the kmem_cache_node. This is used for bootstrapping
2972 * memory on a fresh node that has no slab structures yet.
2974 static void early_kmem_cache_node_alloc(int node)
2977 struct kmem_cache_node *n;
2979 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2981 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2984 if (page_to_nid(page) != node) {
2985 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2986 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2991 page->freelist = get_freepointer(kmem_cache_node, n);
2994 kmem_cache_node->node[node] = n;
2995 #ifdef CONFIG_SLUB_DEBUG
2996 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2997 init_tracking(kmem_cache_node, n);
2999 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3000 init_kmem_cache_node(n);
3001 inc_slabs_node(kmem_cache_node, node, page->objects);
3004 * No locks need to be taken here as it has just been
3005 * initialized and there is no concurrent access.
3007 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3010 static void free_kmem_cache_nodes(struct kmem_cache *s)
3013 struct kmem_cache_node *n;
3015 for_each_kmem_cache_node(s, node, n) {
3016 kmem_cache_free(kmem_cache_node, n);
3017 s->node[node] = NULL;
3021 static int init_kmem_cache_nodes(struct kmem_cache *s)
3025 for_each_node_state(node, N_NORMAL_MEMORY) {
3026 struct kmem_cache_node *n;
3028 if (slab_state == DOWN) {
3029 early_kmem_cache_node_alloc(node);
3032 n = kmem_cache_alloc_node(kmem_cache_node,
3036 free_kmem_cache_nodes(s);
3041 init_kmem_cache_node(n);
3046 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3048 if (min < MIN_PARTIAL)
3050 else if (min > MAX_PARTIAL)
3052 s->min_partial = min;
3056 * calculate_sizes() determines the order and the distribution of data within
3059 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3061 unsigned long flags = s->flags;
3062 unsigned long size = s->object_size;
3066 * Round up object size to the next word boundary. We can only
3067 * place the free pointer at word boundaries and this determines
3068 * the possible location of the free pointer.
3070 size = ALIGN(size, sizeof(void *));
3072 #ifdef CONFIG_SLUB_DEBUG
3074 * Determine if we can poison the object itself. If the user of
3075 * the slab may touch the object after free or before allocation
3076 * then we should never poison the object itself.
3078 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3080 s->flags |= __OBJECT_POISON;
3082 s->flags &= ~__OBJECT_POISON;
3086 * If we are Redzoning then check if there is some space between the
3087 * end of the object and the free pointer. If not then add an
3088 * additional word to have some bytes to store Redzone information.
3090 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3091 size += sizeof(void *);
3095 * With that we have determined the number of bytes in actual use
3096 * by the object. This is the potential offset to the free pointer.
3100 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3103 * Relocate free pointer after the object if it is not
3104 * permitted to overwrite the first word of the object on
3107 * This is the case if we do RCU, have a constructor or
3108 * destructor or are poisoning the objects.
3111 size += sizeof(void *);
3114 #ifdef CONFIG_SLUB_DEBUG
3115 if (flags & SLAB_STORE_USER)
3117 * Need to store information about allocs and frees after
3120 size += 2 * sizeof(struct track);
3122 if (flags & SLAB_RED_ZONE)
3124 * Add some empty padding so that we can catch
3125 * overwrites from earlier objects rather than let
3126 * tracking information or the free pointer be
3127 * corrupted if a user writes before the start
3130 size += sizeof(void *);
3134 * SLUB stores one object immediately after another beginning from
3135 * offset 0. In order to align the objects we have to simply size
3136 * each object to conform to the alignment.
3138 size = ALIGN(size, s->align);
3140 if (forced_order >= 0)
3141 order = forced_order;
3143 order = calculate_order(size, s->reserved);
3150 s->allocflags |= __GFP_COMP;
3152 if (s->flags & SLAB_CACHE_DMA)
3153 s->allocflags |= GFP_DMA;
3155 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3156 s->allocflags |= __GFP_RECLAIMABLE;
3159 * Determine the number of objects per slab
3161 s->oo = oo_make(order, size, s->reserved);
3162 s->min = oo_make(get_order(size), size, s->reserved);
3163 if (oo_objects(s->oo) > oo_objects(s->max))
3166 return !!oo_objects(s->oo);
3169 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3171 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3174 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3175 s->reserved = sizeof(struct rcu_head);
3177 if (!calculate_sizes(s, -1))
3179 if (disable_higher_order_debug) {
3181 * Disable debugging flags that store metadata if the min slab
3184 if (get_order(s->size) > get_order(s->object_size)) {
3185 s->flags &= ~DEBUG_METADATA_FLAGS;
3187 if (!calculate_sizes(s, -1))
3192 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3193 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3194 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3195 /* Enable fast mode */
3196 s->flags |= __CMPXCHG_DOUBLE;
3200 * The larger the object size is, the more pages we want on the partial
3201 * list to avoid pounding the page allocator excessively.
3203 set_min_partial(s, ilog2(s->size) / 2);
3206 * cpu_partial determined the maximum number of objects kept in the
3207 * per cpu partial lists of a processor.
3209 * Per cpu partial lists mainly contain slabs that just have one
3210 * object freed. If they are used for allocation then they can be
3211 * filled up again with minimal effort. The slab will never hit the
3212 * per node partial lists and therefore no locking will be required.
3214 * This setting also determines
3216 * A) The number of objects from per cpu partial slabs dumped to the
3217 * per node list when we reach the limit.
3218 * B) The number of objects in cpu partial slabs to extract from the
3219 * per node list when we run out of per cpu objects. We only fetch
3220 * 50% to keep some capacity around for frees.
3222 if (!kmem_cache_has_cpu_partial(s))
3224 else if (s->size >= PAGE_SIZE)
3226 else if (s->size >= 1024)
3228 else if (s->size >= 256)
3229 s->cpu_partial = 13;
3231 s->cpu_partial = 30;
3234 s->remote_node_defrag_ratio = 1000;
3236 if (!init_kmem_cache_nodes(s))
3239 if (alloc_kmem_cache_cpus(s))
3242 free_kmem_cache_nodes(s);
3244 if (flags & SLAB_PANIC)
3245 panic("Cannot create slab %s size=%lu realsize=%u "
3246 "order=%u offset=%u flags=%lx\n",
3247 s->name, (unsigned long)s->size, s->size,
3248 oo_order(s->oo), s->offset, flags);
3252 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3255 #ifdef CONFIG_SLUB_DEBUG
3256 void *addr = page_address(page);
3258 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3259 sizeof(long), GFP_ATOMIC);
3262 slab_err(s, page, text, s->name);
3265 get_map(s, page, map);
3266 for_each_object(p, s, addr, page->objects) {
3268 if (!test_bit(slab_index(p, s, addr), map)) {
3269 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3270 print_tracking(s, p);
3279 * Attempt to free all partial slabs on a node.
3280 * This is called from kmem_cache_close(). We must be the last thread
3281 * using the cache and therefore we do not need to lock anymore.
3283 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3285 struct page *page, *h;
3287 list_for_each_entry_safe(page, h, &n->partial, lru) {
3289 __remove_partial(n, page);
3290 discard_slab(s, page);
3292 list_slab_objects(s, page,
3293 "Objects remaining in %s on kmem_cache_close()");
3299 * Release all resources used by a slab cache.
3301 static inline int kmem_cache_close(struct kmem_cache *s)
3304 struct kmem_cache_node *n;
3307 /* Attempt to free all objects */
3308 for_each_kmem_cache_node(s, node, n) {
3310 if (n->nr_partial || slabs_node(s, node))
3313 free_percpu(s->cpu_slab);
3314 free_kmem_cache_nodes(s);
3318 int __kmem_cache_shutdown(struct kmem_cache *s)
3320 return kmem_cache_close(s);
3323 /********************************************************************
3325 *******************************************************************/
3327 static int __init setup_slub_min_order(char *str)
3329 get_option(&str, &slub_min_order);
3334 __setup("slub_min_order=", setup_slub_min_order);
3336 static int __init setup_slub_max_order(char *str)
3338 get_option(&str, &slub_max_order);
3339 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3344 __setup("slub_max_order=", setup_slub_max_order);
3346 static int __init setup_slub_min_objects(char *str)
3348 get_option(&str, &slub_min_objects);
3353 __setup("slub_min_objects=", setup_slub_min_objects);
3355 void *__kmalloc(size_t size, gfp_t flags)
3357 struct kmem_cache *s;
3360 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3361 return kmalloc_large(size, flags);
3363 s = kmalloc_slab(size, flags);
3365 if (unlikely(ZERO_OR_NULL_PTR(s)))
3368 ret = slab_alloc(s, flags, _RET_IP_);
3370 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3372 kasan_kmalloc(s, ret, size);
3376 EXPORT_SYMBOL(__kmalloc);
3379 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3384 flags |= __GFP_COMP | __GFP_NOTRACK;
3385 page = alloc_kmem_pages_node(node, flags, get_order(size));
3387 ptr = page_address(page);
3389 kmalloc_large_node_hook(ptr, size, flags);
3393 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3395 struct kmem_cache *s;
3398 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3399 ret = kmalloc_large_node(size, flags, node);
3401 trace_kmalloc_node(_RET_IP_, ret,
3402 size, PAGE_SIZE << get_order(size),
3408 s = kmalloc_slab(size, flags);
3410 if (unlikely(ZERO_OR_NULL_PTR(s)))
3413 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3415 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3417 kasan_kmalloc(s, ret, size);
3421 EXPORT_SYMBOL(__kmalloc_node);
3424 static size_t __ksize(const void *object)
3428 if (unlikely(object == ZERO_SIZE_PTR))
3431 page = virt_to_head_page(object);
3433 if (unlikely(!PageSlab(page))) {
3434 WARN_ON(!PageCompound(page));
3435 return PAGE_SIZE << compound_order(page);
3438 return slab_ksize(page->slab_cache);
3441 size_t ksize(const void *object)
3443 size_t size = __ksize(object);
3444 /* We assume that ksize callers could use whole allocated area,
3445 so we need unpoison this area. */
3446 kasan_krealloc(object, size);
3449 EXPORT_SYMBOL(ksize);
3451 void kfree(const void *x)
3454 void *object = (void *)x;
3456 trace_kfree(_RET_IP_, x);
3458 if (unlikely(ZERO_OR_NULL_PTR(x)))
3461 page = virt_to_head_page(x);
3462 if (unlikely(!PageSlab(page))) {
3463 BUG_ON(!PageCompound(page));
3465 __free_kmem_pages(page, compound_order(page));
3468 slab_free(page->slab_cache, page, object, _RET_IP_);
3470 EXPORT_SYMBOL(kfree);
3472 #define SHRINK_PROMOTE_MAX 32
3475 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3476 * up most to the head of the partial lists. New allocations will then
3477 * fill those up and thus they can be removed from the partial lists.
3479 * The slabs with the least items are placed last. This results in them
3480 * being allocated from last increasing the chance that the last objects
3481 * are freed in them.
3483 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3487 struct kmem_cache_node *n;
3490 struct list_head discard;
3491 struct list_head promote[SHRINK_PROMOTE_MAX];
3492 unsigned long flags;
3497 * Disable empty slabs caching. Used to avoid pinning offline
3498 * memory cgroups by kmem pages that can be freed.
3504 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3505 * so we have to make sure the change is visible.
3507 kick_all_cpus_sync();
3511 for_each_kmem_cache_node(s, node, n) {
3512 INIT_LIST_HEAD(&discard);
3513 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3514 INIT_LIST_HEAD(promote + i);
3516 raw_spin_lock_irqsave(&n->list_lock, flags);
3519 * Build lists of slabs to discard or promote.
3521 * Note that concurrent frees may occur while we hold the
3522 * list_lock. page->inuse here is the upper limit.
3524 list_for_each_entry_safe(page, t, &n->partial, lru) {
3525 int free = page->objects - page->inuse;
3527 /* Do not reread page->inuse */
3530 /* We do not keep full slabs on the list */
3533 if (free == page->objects) {
3534 list_move(&page->lru, &discard);
3536 } else if (free <= SHRINK_PROMOTE_MAX)
3537 list_move(&page->lru, promote + free - 1);
3541 * Promote the slabs filled up most to the head of the
3544 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3545 list_splice(promote + i, &n->partial);
3547 raw_spin_unlock_irqrestore(&n->list_lock, flags);
3549 /* Release empty slabs */
3550 list_for_each_entry_safe(page, t, &discard, lru)
3551 discard_slab(s, page);
3553 if (slabs_node(s, node))
3560 static int slab_mem_going_offline_callback(void *arg)
3562 struct kmem_cache *s;
3564 mutex_lock(&slab_mutex);
3565 list_for_each_entry(s, &slab_caches, list)
3566 __kmem_cache_shrink(s, false);
3567 mutex_unlock(&slab_mutex);
3572 static void slab_mem_offline_callback(void *arg)
3574 struct kmem_cache_node *n;
3575 struct kmem_cache *s;
3576 struct memory_notify *marg = arg;
3579 offline_node = marg->status_change_nid_normal;
3582 * If the node still has available memory. we need kmem_cache_node
3585 if (offline_node < 0)
3588 mutex_lock(&slab_mutex);
3589 list_for_each_entry(s, &slab_caches, list) {
3590 n = get_node(s, offline_node);
3593 * if n->nr_slabs > 0, slabs still exist on the node
3594 * that is going down. We were unable to free them,
3595 * and offline_pages() function shouldn't call this
3596 * callback. So, we must fail.
3598 BUG_ON(slabs_node(s, offline_node));
3600 s->node[offline_node] = NULL;
3601 kmem_cache_free(kmem_cache_node, n);
3604 mutex_unlock(&slab_mutex);
3607 static int slab_mem_going_online_callback(void *arg)
3609 struct kmem_cache_node *n;
3610 struct kmem_cache *s;
3611 struct memory_notify *marg = arg;
3612 int nid = marg->status_change_nid_normal;
3616 * If the node's memory is already available, then kmem_cache_node is
3617 * already created. Nothing to do.
3623 * We are bringing a node online. No memory is available yet. We must
3624 * allocate a kmem_cache_node structure in order to bring the node
3627 mutex_lock(&slab_mutex);
3628 list_for_each_entry(s, &slab_caches, list) {
3630 * XXX: kmem_cache_alloc_node will fallback to other nodes
3631 * since memory is not yet available from the node that
3634 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3639 init_kmem_cache_node(n);
3643 mutex_unlock(&slab_mutex);
3647 static int slab_memory_callback(struct notifier_block *self,
3648 unsigned long action, void *arg)
3653 case MEM_GOING_ONLINE:
3654 ret = slab_mem_going_online_callback(arg);
3656 case MEM_GOING_OFFLINE:
3657 ret = slab_mem_going_offline_callback(arg);
3660 case MEM_CANCEL_ONLINE:
3661 slab_mem_offline_callback(arg);
3664 case MEM_CANCEL_OFFLINE:
3668 ret = notifier_from_errno(ret);
3674 static struct notifier_block slab_memory_callback_nb = {
3675 .notifier_call = slab_memory_callback,
3676 .priority = SLAB_CALLBACK_PRI,
3679 /********************************************************************
3680 * Basic setup of slabs
3681 *******************************************************************/
3684 * Used for early kmem_cache structures that were allocated using
3685 * the page allocator. Allocate them properly then fix up the pointers
3686 * that may be pointing to the wrong kmem_cache structure.
3689 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3692 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3693 struct kmem_cache_node *n;
3695 memcpy(s, static_cache, kmem_cache->object_size);
3698 * This runs very early, and only the boot processor is supposed to be
3699 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3702 __flush_cpu_slab(s, smp_processor_id());
3703 for_each_kmem_cache_node(s, node, n) {
3706 list_for_each_entry(p, &n->partial, lru)
3709 #ifdef CONFIG_SLUB_DEBUG
3710 list_for_each_entry(p, &n->full, lru)
3714 slab_init_memcg_params(s);
3715 list_add(&s->list, &slab_caches);
3719 void __init kmem_cache_init(void)
3721 static __initdata struct kmem_cache boot_kmem_cache,
3722 boot_kmem_cache_node;
3725 for_each_possible_cpu(cpu) {
3726 raw_spin_lock_init(&per_cpu(slub_free_list, cpu).lock);
3727 INIT_LIST_HEAD(&per_cpu(slub_free_list, cpu).list);
3730 if (debug_guardpage_minorder())
3733 kmem_cache_node = &boot_kmem_cache_node;
3734 kmem_cache = &boot_kmem_cache;
3736 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3737 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3739 register_hotmemory_notifier(&slab_memory_callback_nb);
3741 /* Able to allocate the per node structures */
3742 slab_state = PARTIAL;
3744 create_boot_cache(kmem_cache, "kmem_cache",
3745 offsetof(struct kmem_cache, node) +
3746 nr_node_ids * sizeof(struct kmem_cache_node *),
3747 SLAB_HWCACHE_ALIGN);
3749 kmem_cache = bootstrap(&boot_kmem_cache);
3752 * Allocate kmem_cache_node properly from the kmem_cache slab.
3753 * kmem_cache_node is separately allocated so no need to
3754 * update any list pointers.
3756 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3758 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3759 create_kmalloc_caches(0);
3762 register_cpu_notifier(&slab_notifier);
3765 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3767 slub_min_order, slub_max_order, slub_min_objects,
3768 nr_cpu_ids, nr_node_ids);
3771 void __init kmem_cache_init_late(void)
3776 __kmem_cache_alias(const char *name, size_t size, size_t align,
3777 unsigned long flags, void (*ctor)(void *))
3779 struct kmem_cache *s, *c;
3781 s = find_mergeable(size, align, flags, name, ctor);
3786 * Adjust the object sizes so that we clear
3787 * the complete object on kzalloc.
3789 s->object_size = max(s->object_size, (int)size);
3790 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3792 for_each_memcg_cache(c, s) {
3793 c->object_size = s->object_size;
3794 c->inuse = max_t(int, c->inuse,
3795 ALIGN(size, sizeof(void *)));
3798 if (sysfs_slab_alias(s, name)) {
3807 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3811 err = kmem_cache_open(s, flags);
3815 /* Mutex is not taken during early boot */
3816 if (slab_state <= UP)
3819 memcg_propagate_slab_attrs(s);
3820 err = sysfs_slab_add(s);
3822 kmem_cache_close(s);
3829 * Use the cpu notifier to insure that the cpu slabs are flushed when
3832 static int slab_cpuup_callback(struct notifier_block *nfb,
3833 unsigned long action, void *hcpu)
3835 long cpu = (long)hcpu;
3836 struct kmem_cache *s;
3837 unsigned long flags;
3840 case CPU_UP_CANCELED:
3841 case CPU_UP_CANCELED_FROZEN:
3843 case CPU_DEAD_FROZEN:
3844 mutex_lock(&slab_mutex);
3845 list_for_each_entry(s, &slab_caches, list) {
3846 local_irq_save(flags);
3847 __flush_cpu_slab(s, cpu);
3848 local_irq_restore(flags);
3850 mutex_unlock(&slab_mutex);
3858 static struct notifier_block slab_notifier = {
3859 .notifier_call = slab_cpuup_callback
3864 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3866 struct kmem_cache *s;
3869 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3870 return kmalloc_large(size, gfpflags);
3872 s = kmalloc_slab(size, gfpflags);
3874 if (unlikely(ZERO_OR_NULL_PTR(s)))
3877 ret = slab_alloc(s, gfpflags, caller);
3879 /* Honor the call site pointer we received. */
3880 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3886 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3887 int node, unsigned long caller)
3889 struct kmem_cache *s;
3892 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3893 ret = kmalloc_large_node(size, gfpflags, node);
3895 trace_kmalloc_node(caller, ret,
3896 size, PAGE_SIZE << get_order(size),
3902 s = kmalloc_slab(size, gfpflags);
3904 if (unlikely(ZERO_OR_NULL_PTR(s)))
3907 ret = slab_alloc_node(s, gfpflags, node, caller);
3909 /* Honor the call site pointer we received. */
3910 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3917 static int count_inuse(struct page *page)
3922 static int count_total(struct page *page)
3924 return page->objects;
3928 #ifdef CONFIG_SLUB_DEBUG
3929 static int validate_slab(struct kmem_cache *s, struct page *page,
3933 void *addr = page_address(page);
3935 if (!check_slab(s, page) ||
3936 !on_freelist(s, page, NULL))
3939 /* Now we know that a valid freelist exists */
3940 bitmap_zero(map, page->objects);
3942 get_map(s, page, map);
3943 for_each_object(p, s, addr, page->objects) {
3944 if (test_bit(slab_index(p, s, addr), map))
3945 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3949 for_each_object(p, s, addr, page->objects)
3950 if (!test_bit(slab_index(p, s, addr), map))
3951 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3956 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3960 validate_slab(s, page, map);
3964 static int validate_slab_node(struct kmem_cache *s,
3965 struct kmem_cache_node *n, unsigned long *map)
3967 unsigned long count = 0;
3969 unsigned long flags;
3971 raw_spin_lock_irqsave(&n->list_lock, flags);
3973 list_for_each_entry(page, &n->partial, lru) {
3974 validate_slab_slab(s, page, map);
3977 if (count != n->nr_partial)
3978 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3979 s->name, count, n->nr_partial);
3981 if (!(s->flags & SLAB_STORE_USER))
3984 list_for_each_entry(page, &n->full, lru) {
3985 validate_slab_slab(s, page, map);
3988 if (count != atomic_long_read(&n->nr_slabs))
3989 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3990 s->name, count, atomic_long_read(&n->nr_slabs));
3993 raw_spin_unlock_irqrestore(&n->list_lock, flags);
3997 static long validate_slab_cache(struct kmem_cache *s)
4000 unsigned long count = 0;
4001 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4002 sizeof(unsigned long), GFP_KERNEL);
4003 struct kmem_cache_node *n;
4009 for_each_kmem_cache_node(s, node, n)
4010 count += validate_slab_node(s, n, map);
4015 * Generate lists of code addresses where slabcache objects are allocated
4020 unsigned long count;
4027 DECLARE_BITMAP(cpus, NR_CPUS);
4033 unsigned long count;
4034 struct location *loc;
4037 static void free_loc_track(struct loc_track *t)
4040 free_pages((unsigned long)t->loc,
4041 get_order(sizeof(struct location) * t->max));
4044 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4049 order = get_order(sizeof(struct location) * max);
4051 l = (void *)__get_free_pages(flags, order);
4056 memcpy(l, t->loc, sizeof(struct location) * t->count);
4064 static int add_location(struct loc_track *t, struct kmem_cache *s,
4065 const struct track *track)
4067 long start, end, pos;
4069 unsigned long caddr;
4070 unsigned long age = jiffies - track->when;
4076 pos = start + (end - start + 1) / 2;
4079 * There is nothing at "end". If we end up there
4080 * we need to add something to before end.
4085 caddr = t->loc[pos].addr;
4086 if (track->addr == caddr) {
4092 if (age < l->min_time)
4094 if (age > l->max_time)
4097 if (track->pid < l->min_pid)
4098 l->min_pid = track->pid;
4099 if (track->pid > l->max_pid)
4100 l->max_pid = track->pid;
4102 cpumask_set_cpu(track->cpu,
4103 to_cpumask(l->cpus));
4105 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4109 if (track->addr < caddr)
4116 * Not found. Insert new tracking element.
4118 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4124 (t->count - pos) * sizeof(struct location));
4127 l->addr = track->addr;
4131 l->min_pid = track->pid;
4132 l->max_pid = track->pid;
4133 cpumask_clear(to_cpumask(l->cpus));
4134 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4135 nodes_clear(l->nodes);
4136 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4140 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4141 struct page *page, enum track_item alloc,
4144 void *addr = page_address(page);
4147 bitmap_zero(map, page->objects);
4148 get_map(s, page, map);
4150 for_each_object(p, s, addr, page->objects)
4151 if (!test_bit(slab_index(p, s, addr), map))
4152 add_location(t, s, get_track(s, p, alloc));
4155 static int list_locations(struct kmem_cache *s, char *buf,
4156 enum track_item alloc)
4160 struct loc_track t = { 0, 0, NULL };
4162 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4163 sizeof(unsigned long), GFP_KERNEL);
4164 struct kmem_cache_node *n;
4166 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4169 return sprintf(buf, "Out of memory\n");
4171 /* Push back cpu slabs */
4174 for_each_kmem_cache_node(s, node, n) {
4175 unsigned long flags;
4178 if (!atomic_long_read(&n->nr_slabs))
4181 raw_spin_lock_irqsave(&n->list_lock, flags);
4182 list_for_each_entry(page, &n->partial, lru)
4183 process_slab(&t, s, page, alloc, map);
4184 list_for_each_entry(page, &n->full, lru)
4185 process_slab(&t, s, page, alloc, map);
4186 raw_spin_unlock_irqrestore(&n->list_lock, flags);
4189 for (i = 0; i < t.count; i++) {
4190 struct location *l = &t.loc[i];
4192 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4194 len += sprintf(buf + len, "%7ld ", l->count);
4197 len += sprintf(buf + len, "%pS", (void *)l->addr);
4199 len += sprintf(buf + len, "<not-available>");
4201 if (l->sum_time != l->min_time) {
4202 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4204 (long)div_u64(l->sum_time, l->count),
4207 len += sprintf(buf + len, " age=%ld",
4210 if (l->min_pid != l->max_pid)
4211 len += sprintf(buf + len, " pid=%ld-%ld",
4212 l->min_pid, l->max_pid);
4214 len += sprintf(buf + len, " pid=%ld",
4217 if (num_online_cpus() > 1 &&
4218 !cpumask_empty(to_cpumask(l->cpus)) &&
4219 len < PAGE_SIZE - 60)
4220 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4222 cpumask_pr_args(to_cpumask(l->cpus)));
4224 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4225 len < PAGE_SIZE - 60)
4226 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4228 nodemask_pr_args(&l->nodes));
4230 len += sprintf(buf + len, "\n");
4236 len += sprintf(buf, "No data\n");
4241 #ifdef SLUB_RESILIENCY_TEST
4242 static void __init resiliency_test(void)
4246 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4248 pr_err("SLUB resiliency testing\n");
4249 pr_err("-----------------------\n");
4250 pr_err("A. Corruption after allocation\n");
4252 p = kzalloc(16, GFP_KERNEL);
4254 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4257 validate_slab_cache(kmalloc_caches[4]);
4259 /* Hmmm... The next two are dangerous */
4260 p = kzalloc(32, GFP_KERNEL);
4261 p[32 + sizeof(void *)] = 0x34;
4262 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4264 pr_err("If allocated object is overwritten then not detectable\n\n");
4266 validate_slab_cache(kmalloc_caches[5]);
4267 p = kzalloc(64, GFP_KERNEL);
4268 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4270 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4272 pr_err("If allocated object is overwritten then not detectable\n\n");
4273 validate_slab_cache(kmalloc_caches[6]);
4275 pr_err("\nB. Corruption after free\n");
4276 p = kzalloc(128, GFP_KERNEL);
4279 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4280 validate_slab_cache(kmalloc_caches[7]);
4282 p = kzalloc(256, GFP_KERNEL);
4285 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4286 validate_slab_cache(kmalloc_caches[8]);
4288 p = kzalloc(512, GFP_KERNEL);
4291 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4292 validate_slab_cache(kmalloc_caches[9]);
4296 static void resiliency_test(void) {};
4301 enum slab_stat_type {
4302 SL_ALL, /* All slabs */
4303 SL_PARTIAL, /* Only partially allocated slabs */
4304 SL_CPU, /* Only slabs used for cpu caches */
4305 SL_OBJECTS, /* Determine allocated objects not slabs */
4306 SL_TOTAL /* Determine object capacity not slabs */
4309 #define SO_ALL (1 << SL_ALL)
4310 #define SO_PARTIAL (1 << SL_PARTIAL)
4311 #define SO_CPU (1 << SL_CPU)
4312 #define SO_OBJECTS (1 << SL_OBJECTS)
4313 #define SO_TOTAL (1 << SL_TOTAL)
4315 static ssize_t show_slab_objects(struct kmem_cache *s,
4316 char *buf, unsigned long flags)
4318 unsigned long total = 0;
4321 unsigned long *nodes;
4323 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4327 if (flags & SO_CPU) {
4330 for_each_possible_cpu(cpu) {
4331 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4336 page = READ_ONCE(c->page);
4340 node = page_to_nid(page);
4341 if (flags & SO_TOTAL)
4343 else if (flags & SO_OBJECTS)
4351 page = READ_ONCE(c->partial);
4353 node = page_to_nid(page);
4354 if (flags & SO_TOTAL)
4356 else if (flags & SO_OBJECTS)
4367 #ifdef CONFIG_SLUB_DEBUG
4368 if (flags & SO_ALL) {
4369 struct kmem_cache_node *n;
4371 for_each_kmem_cache_node(s, node, n) {
4373 if (flags & SO_TOTAL)
4374 x = atomic_long_read(&n->total_objects);
4375 else if (flags & SO_OBJECTS)
4376 x = atomic_long_read(&n->total_objects) -
4377 count_partial(n, count_free);
4379 x = atomic_long_read(&n->nr_slabs);
4386 if (flags & SO_PARTIAL) {
4387 struct kmem_cache_node *n;
4389 for_each_kmem_cache_node(s, node, n) {
4390 if (flags & SO_TOTAL)
4391 x = count_partial(n, count_total);
4392 else if (flags & SO_OBJECTS)
4393 x = count_partial(n, count_inuse);
4400 x = sprintf(buf, "%lu", total);
4402 for (node = 0; node < nr_node_ids; node++)
4404 x += sprintf(buf + x, " N%d=%lu",
4409 return x + sprintf(buf + x, "\n");
4412 #ifdef CONFIG_SLUB_DEBUG
4413 static int any_slab_objects(struct kmem_cache *s)
4416 struct kmem_cache_node *n;
4418 for_each_kmem_cache_node(s, node, n)
4419 if (atomic_long_read(&n->total_objects))
4426 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4427 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4429 struct slab_attribute {
4430 struct attribute attr;
4431 ssize_t (*show)(struct kmem_cache *s, char *buf);
4432 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4435 #define SLAB_ATTR_RO(_name) \
4436 static struct slab_attribute _name##_attr = \
4437 __ATTR(_name, 0400, _name##_show, NULL)
4439 #define SLAB_ATTR(_name) \
4440 static struct slab_attribute _name##_attr = \
4441 __ATTR(_name, 0600, _name##_show, _name##_store)
4443 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4445 return sprintf(buf, "%d\n", s->size);
4447 SLAB_ATTR_RO(slab_size);
4449 static ssize_t align_show(struct kmem_cache *s, char *buf)
4451 return sprintf(buf, "%d\n", s->align);
4453 SLAB_ATTR_RO(align);
4455 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4457 return sprintf(buf, "%d\n", s->object_size);
4459 SLAB_ATTR_RO(object_size);
4461 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4463 return sprintf(buf, "%d\n", oo_objects(s->oo));
4465 SLAB_ATTR_RO(objs_per_slab);
4467 static ssize_t order_store(struct kmem_cache *s,
4468 const char *buf, size_t length)
4470 unsigned long order;
4473 err = kstrtoul(buf, 10, &order);
4477 if (order > slub_max_order || order < slub_min_order)
4480 calculate_sizes(s, order);
4484 static ssize_t order_show(struct kmem_cache *s, char *buf)
4486 return sprintf(buf, "%d\n", oo_order(s->oo));
4490 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4492 return sprintf(buf, "%lu\n", s->min_partial);
4495 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4501 err = kstrtoul(buf, 10, &min);
4505 set_min_partial(s, min);
4508 SLAB_ATTR(min_partial);
4510 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4512 return sprintf(buf, "%u\n", s->cpu_partial);
4515 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4518 unsigned long objects;
4521 err = kstrtoul(buf, 10, &objects);
4524 if (objects && !kmem_cache_has_cpu_partial(s))
4527 s->cpu_partial = objects;
4531 SLAB_ATTR(cpu_partial);
4533 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4537 return sprintf(buf, "%pS\n", s->ctor);
4541 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4543 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4545 SLAB_ATTR_RO(aliases);
4547 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4549 return show_slab_objects(s, buf, SO_PARTIAL);
4551 SLAB_ATTR_RO(partial);
4553 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4555 return show_slab_objects(s, buf, SO_CPU);
4557 SLAB_ATTR_RO(cpu_slabs);
4559 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4561 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4563 SLAB_ATTR_RO(objects);
4565 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4567 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4569 SLAB_ATTR_RO(objects_partial);
4571 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4578 for_each_online_cpu(cpu) {
4579 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4582 pages += page->pages;
4583 objects += page->pobjects;
4587 len = sprintf(buf, "%d(%d)", objects, pages);
4590 for_each_online_cpu(cpu) {
4591 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4593 if (page && len < PAGE_SIZE - 20)
4594 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4595 page->pobjects, page->pages);
4598 return len + sprintf(buf + len, "\n");
4600 SLAB_ATTR_RO(slabs_cpu_partial);
4602 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4604 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4607 static ssize_t reclaim_account_store(struct kmem_cache *s,
4608 const char *buf, size_t length)
4610 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4612 s->flags |= SLAB_RECLAIM_ACCOUNT;
4615 SLAB_ATTR(reclaim_account);
4617 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4619 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4621 SLAB_ATTR_RO(hwcache_align);
4623 #ifdef CONFIG_ZONE_DMA
4624 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4626 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4628 SLAB_ATTR_RO(cache_dma);
4631 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4633 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4635 SLAB_ATTR_RO(destroy_by_rcu);
4637 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4639 return sprintf(buf, "%d\n", s->reserved);
4641 SLAB_ATTR_RO(reserved);
4643 #ifdef CONFIG_SLUB_DEBUG
4644 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4646 return show_slab_objects(s, buf, SO_ALL);
4648 SLAB_ATTR_RO(slabs);
4650 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4652 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4654 SLAB_ATTR_RO(total_objects);
4656 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4658 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4661 static ssize_t sanity_checks_store(struct kmem_cache *s,
4662 const char *buf, size_t length)
4664 s->flags &= ~SLAB_DEBUG_FREE;
4665 if (buf[0] == '1') {
4666 s->flags &= ~__CMPXCHG_DOUBLE;
4667 s->flags |= SLAB_DEBUG_FREE;
4671 SLAB_ATTR(sanity_checks);
4673 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4675 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4678 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4682 * Tracing a merged cache is going to give confusing results
4683 * as well as cause other issues like converting a mergeable
4684 * cache into an umergeable one.
4686 if (s->refcount > 1)
4689 s->flags &= ~SLAB_TRACE;
4690 if (buf[0] == '1') {
4691 s->flags &= ~__CMPXCHG_DOUBLE;
4692 s->flags |= SLAB_TRACE;
4698 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4700 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4703 static ssize_t red_zone_store(struct kmem_cache *s,
4704 const char *buf, size_t length)
4706 if (any_slab_objects(s))
4709 s->flags &= ~SLAB_RED_ZONE;
4710 if (buf[0] == '1') {
4711 s->flags &= ~__CMPXCHG_DOUBLE;
4712 s->flags |= SLAB_RED_ZONE;
4714 calculate_sizes(s, -1);
4717 SLAB_ATTR(red_zone);
4719 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4721 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4724 static ssize_t poison_store(struct kmem_cache *s,
4725 const char *buf, size_t length)
4727 if (any_slab_objects(s))
4730 s->flags &= ~SLAB_POISON;
4731 if (buf[0] == '1') {
4732 s->flags &= ~__CMPXCHG_DOUBLE;
4733 s->flags |= SLAB_POISON;
4735 calculate_sizes(s, -1);
4740 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4742 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4745 static ssize_t store_user_store(struct kmem_cache *s,
4746 const char *buf, size_t length)
4748 if (any_slab_objects(s))
4751 s->flags &= ~SLAB_STORE_USER;
4752 if (buf[0] == '1') {
4753 s->flags &= ~__CMPXCHG_DOUBLE;
4754 s->flags |= SLAB_STORE_USER;
4756 calculate_sizes(s, -1);
4759 SLAB_ATTR(store_user);
4761 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4766 static ssize_t validate_store(struct kmem_cache *s,
4767 const char *buf, size_t length)
4771 if (buf[0] == '1') {
4772 ret = validate_slab_cache(s);
4778 SLAB_ATTR(validate);
4780 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4782 if (!(s->flags & SLAB_STORE_USER))
4784 return list_locations(s, buf, TRACK_ALLOC);
4786 SLAB_ATTR_RO(alloc_calls);
4788 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4790 if (!(s->flags & SLAB_STORE_USER))
4792 return list_locations(s, buf, TRACK_FREE);
4794 SLAB_ATTR_RO(free_calls);
4795 #endif /* CONFIG_SLUB_DEBUG */
4797 #ifdef CONFIG_FAILSLAB
4798 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4800 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4803 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4806 if (s->refcount > 1)
4809 s->flags &= ~SLAB_FAILSLAB;
4811 s->flags |= SLAB_FAILSLAB;
4814 SLAB_ATTR(failslab);
4817 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4822 static ssize_t shrink_store(struct kmem_cache *s,
4823 const char *buf, size_t length)
4826 kmem_cache_shrink(s);
4834 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4836 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4839 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4840 const char *buf, size_t length)
4842 unsigned long ratio;
4845 err = kstrtoul(buf, 10, &ratio);
4850 s->remote_node_defrag_ratio = ratio * 10;
4854 SLAB_ATTR(remote_node_defrag_ratio);
4857 #ifdef CONFIG_SLUB_STATS
4858 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4860 unsigned long sum = 0;
4863 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4868 for_each_online_cpu(cpu) {
4869 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4875 len = sprintf(buf, "%lu", sum);
4878 for_each_online_cpu(cpu) {
4879 if (data[cpu] && len < PAGE_SIZE - 20)
4880 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4884 return len + sprintf(buf + len, "\n");
4887 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4891 for_each_online_cpu(cpu)
4892 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4895 #define STAT_ATTR(si, text) \
4896 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4898 return show_stat(s, buf, si); \
4900 static ssize_t text##_store(struct kmem_cache *s, \
4901 const char *buf, size_t length) \
4903 if (buf[0] != '0') \
4905 clear_stat(s, si); \
4910 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4911 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4912 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4913 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4914 STAT_ATTR(FREE_FROZEN, free_frozen);
4915 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4916 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4917 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4918 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4919 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4920 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4921 STAT_ATTR(FREE_SLAB, free_slab);
4922 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4923 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4924 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4925 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4926 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4927 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4928 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4929 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4930 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4931 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4932 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4933 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4934 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4935 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4938 static struct attribute *slab_attrs[] = {
4939 &slab_size_attr.attr,
4940 &object_size_attr.attr,
4941 &objs_per_slab_attr.attr,
4943 &min_partial_attr.attr,
4944 &cpu_partial_attr.attr,
4946 &objects_partial_attr.attr,
4948 &cpu_slabs_attr.attr,
4952 &hwcache_align_attr.attr,
4953 &reclaim_account_attr.attr,
4954 &destroy_by_rcu_attr.attr,
4956 &reserved_attr.attr,
4957 &slabs_cpu_partial_attr.attr,
4958 #ifdef CONFIG_SLUB_DEBUG
4959 &total_objects_attr.attr,
4961 &sanity_checks_attr.attr,
4963 &red_zone_attr.attr,
4965 &store_user_attr.attr,
4966 &validate_attr.attr,
4967 &alloc_calls_attr.attr,
4968 &free_calls_attr.attr,
4970 #ifdef CONFIG_ZONE_DMA
4971 &cache_dma_attr.attr,
4974 &remote_node_defrag_ratio_attr.attr,
4976 #ifdef CONFIG_SLUB_STATS
4977 &alloc_fastpath_attr.attr,
4978 &alloc_slowpath_attr.attr,
4979 &free_fastpath_attr.attr,
4980 &free_slowpath_attr.attr,
4981 &free_frozen_attr.attr,
4982 &free_add_partial_attr.attr,
4983 &free_remove_partial_attr.attr,
4984 &alloc_from_partial_attr.attr,
4985 &alloc_slab_attr.attr,
4986 &alloc_refill_attr.attr,
4987 &alloc_node_mismatch_attr.attr,
4988 &free_slab_attr.attr,
4989 &cpuslab_flush_attr.attr,
4990 &deactivate_full_attr.attr,
4991 &deactivate_empty_attr.attr,
4992 &deactivate_to_head_attr.attr,
4993 &deactivate_to_tail_attr.attr,
4994 &deactivate_remote_frees_attr.attr,
4995 &deactivate_bypass_attr.attr,
4996 &order_fallback_attr.attr,
4997 &cmpxchg_double_fail_attr.attr,
4998 &cmpxchg_double_cpu_fail_attr.attr,
4999 &cpu_partial_alloc_attr.attr,
5000 &cpu_partial_free_attr.attr,
5001 &cpu_partial_node_attr.attr,
5002 &cpu_partial_drain_attr.attr,
5004 #ifdef CONFIG_FAILSLAB
5005 &failslab_attr.attr,
5011 static struct attribute_group slab_attr_group = {
5012 .attrs = slab_attrs,
5015 static ssize_t slab_attr_show(struct kobject *kobj,
5016 struct attribute *attr,
5019 struct slab_attribute *attribute;
5020 struct kmem_cache *s;
5023 attribute = to_slab_attr(attr);
5026 if (!attribute->show)
5029 err = attribute->show(s, buf);
5034 static ssize_t slab_attr_store(struct kobject *kobj,
5035 struct attribute *attr,
5036 const char *buf, size_t len)
5038 struct slab_attribute *attribute;
5039 struct kmem_cache *s;
5042 attribute = to_slab_attr(attr);
5045 if (!attribute->store)
5048 err = attribute->store(s, buf, len);
5049 #ifdef CONFIG_MEMCG_KMEM
5050 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5051 struct kmem_cache *c;
5053 mutex_lock(&slab_mutex);
5054 if (s->max_attr_size < len)
5055 s->max_attr_size = len;
5058 * This is a best effort propagation, so this function's return
5059 * value will be determined by the parent cache only. This is
5060 * basically because not all attributes will have a well
5061 * defined semantics for rollbacks - most of the actions will
5062 * have permanent effects.
5064 * Returning the error value of any of the children that fail
5065 * is not 100 % defined, in the sense that users seeing the
5066 * error code won't be able to know anything about the state of
5069 * Only returning the error code for the parent cache at least
5070 * has well defined semantics. The cache being written to
5071 * directly either failed or succeeded, in which case we loop
5072 * through the descendants with best-effort propagation.
5074 for_each_memcg_cache(c, s)
5075 attribute->store(c, buf, len);
5076 mutex_unlock(&slab_mutex);
5082 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5084 #ifdef CONFIG_MEMCG_KMEM
5086 char *buffer = NULL;
5087 struct kmem_cache *root_cache;
5089 if (is_root_cache(s))
5092 root_cache = s->memcg_params.root_cache;
5095 * This mean this cache had no attribute written. Therefore, no point
5096 * in copying default values around
5098 if (!root_cache->max_attr_size)
5101 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5104 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5106 if (!attr || !attr->store || !attr->show)
5110 * It is really bad that we have to allocate here, so we will
5111 * do it only as a fallback. If we actually allocate, though,
5112 * we can just use the allocated buffer until the end.
5114 * Most of the slub attributes will tend to be very small in
5115 * size, but sysfs allows buffers up to a page, so they can
5116 * theoretically happen.
5120 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5123 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5124 if (WARN_ON(!buffer))
5129 attr->show(root_cache, buf);
5130 attr->store(s, buf, strlen(buf));
5134 free_page((unsigned long)buffer);
5138 static void kmem_cache_release(struct kobject *k)
5140 slab_kmem_cache_release(to_slab(k));
5143 static const struct sysfs_ops slab_sysfs_ops = {
5144 .show = slab_attr_show,
5145 .store = slab_attr_store,
5148 static struct kobj_type slab_ktype = {
5149 .sysfs_ops = &slab_sysfs_ops,
5150 .release = kmem_cache_release,
5153 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5155 struct kobj_type *ktype = get_ktype(kobj);
5157 if (ktype == &slab_ktype)
5162 static const struct kset_uevent_ops slab_uevent_ops = {
5163 .filter = uevent_filter,
5166 static struct kset *slab_kset;
5168 static inline struct kset *cache_kset(struct kmem_cache *s)
5170 #ifdef CONFIG_MEMCG_KMEM
5171 if (!is_root_cache(s))
5172 return s->memcg_params.root_cache->memcg_kset;
5177 #define ID_STR_LENGTH 64
5179 /* Create a unique string id for a slab cache:
5181 * Format :[flags-]size
5183 static char *create_unique_id(struct kmem_cache *s)
5185 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5192 * First flags affecting slabcache operations. We will only
5193 * get here for aliasable slabs so we do not need to support
5194 * too many flags. The flags here must cover all flags that
5195 * are matched during merging to guarantee that the id is
5198 if (s->flags & SLAB_CACHE_DMA)
5200 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5202 if (s->flags & SLAB_DEBUG_FREE)
5204 if (!(s->flags & SLAB_NOTRACK))
5208 p += sprintf(p, "%07d", s->size);
5210 BUG_ON(p > name + ID_STR_LENGTH - 1);
5214 static int sysfs_slab_add(struct kmem_cache *s)
5218 int unmergeable = slab_unmergeable(s);
5222 * Slabcache can never be merged so we can use the name proper.
5223 * This is typically the case for debug situations. In that
5224 * case we can catch duplicate names easily.
5226 sysfs_remove_link(&slab_kset->kobj, s->name);
5230 * Create a unique name for the slab as a target
5233 name = create_unique_id(s);
5236 s->kobj.kset = cache_kset(s);
5237 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5241 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5245 #ifdef CONFIG_MEMCG_KMEM
5246 if (is_root_cache(s)) {
5247 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5248 if (!s->memcg_kset) {
5255 kobject_uevent(&s->kobj, KOBJ_ADD);
5257 /* Setup first alias */
5258 sysfs_slab_alias(s, s->name);
5265 kobject_del(&s->kobj);
5267 kobject_put(&s->kobj);
5271 void sysfs_slab_remove(struct kmem_cache *s)
5273 if (slab_state < FULL)
5275 * Sysfs has not been setup yet so no need to remove the
5280 #ifdef CONFIG_MEMCG_KMEM
5281 kset_unregister(s->memcg_kset);
5283 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5284 kobject_del(&s->kobj);
5285 kobject_put(&s->kobj);
5289 * Need to buffer aliases during bootup until sysfs becomes
5290 * available lest we lose that information.
5292 struct saved_alias {
5293 struct kmem_cache *s;
5295 struct saved_alias *next;
5298 static struct saved_alias *alias_list;
5300 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5302 struct saved_alias *al;
5304 if (slab_state == FULL) {
5306 * If we have a leftover link then remove it.
5308 sysfs_remove_link(&slab_kset->kobj, name);
5309 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5312 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5318 al->next = alias_list;
5323 static int __init slab_sysfs_init(void)
5325 struct kmem_cache *s;
5328 mutex_lock(&slab_mutex);
5330 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5332 mutex_unlock(&slab_mutex);
5333 pr_err("Cannot register slab subsystem.\n");
5339 list_for_each_entry(s, &slab_caches, list) {
5340 err = sysfs_slab_add(s);
5342 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5346 while (alias_list) {
5347 struct saved_alias *al = alias_list;
5349 alias_list = alias_list->next;
5350 err = sysfs_slab_alias(al->s, al->name);
5352 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5357 mutex_unlock(&slab_mutex);
5362 __initcall(slab_sysfs_init);
5363 #endif /* CONFIG_SYSFS */
5366 * The /proc/slabinfo ABI
5368 #ifdef CONFIG_SLABINFO
5369 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5371 unsigned long nr_slabs = 0;
5372 unsigned long nr_objs = 0;
5373 unsigned long nr_free = 0;
5375 struct kmem_cache_node *n;
5377 for_each_kmem_cache_node(s, node, n) {
5378 nr_slabs += node_nr_slabs(n);
5379 nr_objs += node_nr_objs(n);
5380 nr_free += count_partial(n, count_free);
5383 sinfo->active_objs = nr_objs - nr_free;
5384 sinfo->num_objs = nr_objs;
5385 sinfo->active_slabs = nr_slabs;
5386 sinfo->num_slabs = nr_slabs;
5387 sinfo->objects_per_slab = oo_objects(s->oo);
5388 sinfo->cache_order = oo_order(s->oo);
5391 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5395 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5396 size_t count, loff_t *ppos)
5400 #endif /* CONFIG_SLABINFO */