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 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 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 */
1246 * Hooks for other subsystems that check memory allocations. In a typical
1247 * production configuration these hooks all should produce no code at all.
1249 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1251 kmemleak_alloc(ptr, size, 1, flags);
1252 kasan_kmalloc_large(ptr, size);
1255 static inline void kfree_hook(const void *x)
1258 kasan_kfree_large(x);
1261 static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
1264 flags &= gfp_allowed_mask;
1265 lockdep_trace_alloc(flags);
1266 might_sleep_if(flags & __GFP_WAIT);
1268 if (should_failslab(s->object_size, flags, s->flags))
1271 return memcg_kmem_get_cache(s, flags);
1274 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1275 gfp_t flags, void *object)
1277 flags &= gfp_allowed_mask;
1278 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1279 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1280 memcg_kmem_put_cache(s);
1281 kasan_slab_alloc(s, object);
1284 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1286 kmemleak_free_recursive(x, s->flags);
1289 * Trouble is that we may no longer disable interrupts in the fast path
1290 * So in order to make the debug calls that expect irqs to be
1291 * disabled we need to disable interrupts temporarily.
1293 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1295 unsigned long flags;
1297 local_irq_save(flags);
1298 kmemcheck_slab_free(s, x, s->object_size);
1299 debug_check_no_locks_freed(x, s->object_size);
1300 local_irq_restore(flags);
1303 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1304 debug_check_no_obj_freed(x, s->object_size);
1306 kasan_slab_free(s, x);
1309 static void setup_object(struct kmem_cache *s, struct page *page,
1312 setup_object_debug(s, page, object);
1313 if (unlikely(s->ctor)) {
1314 kasan_unpoison_object_data(s, object);
1316 kasan_poison_object_data(s, object);
1321 * Slab allocation and freeing
1323 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1324 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1327 int order = oo_order(oo);
1329 flags |= __GFP_NOTRACK;
1331 if (memcg_charge_slab(s, flags, order))
1334 if (node == NUMA_NO_NODE)
1335 page = alloc_pages(flags, order);
1337 page = __alloc_pages_node(node, flags, order);
1340 memcg_uncharge_slab(s, order);
1345 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1348 struct kmem_cache_order_objects oo = s->oo;
1353 flags &= gfp_allowed_mask;
1355 if (flags & __GFP_WAIT)
1358 flags |= s->allocflags;
1361 * Let the initial higher-order allocation fail under memory pressure
1362 * so we fall-back to the minimum order allocation.
1364 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1365 if ((alloc_gfp & __GFP_WAIT) && oo_order(oo) > oo_order(s->min))
1366 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_WAIT;
1368 page = alloc_slab_page(s, alloc_gfp, node, oo);
1369 if (unlikely(!page)) {
1373 * Allocation may have failed due to fragmentation.
1374 * Try a lower order alloc if possible
1376 page = alloc_slab_page(s, alloc_gfp, node, oo);
1377 if (unlikely(!page))
1379 stat(s, ORDER_FALLBACK);
1382 if (kmemcheck_enabled &&
1383 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1384 int pages = 1 << oo_order(oo);
1386 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1389 * Objects from caches that have a constructor don't get
1390 * cleared when they're allocated, so we need to do it here.
1393 kmemcheck_mark_uninitialized_pages(page, pages);
1395 kmemcheck_mark_unallocated_pages(page, pages);
1398 page->objects = oo_objects(oo);
1400 order = compound_order(page);
1401 page->slab_cache = s;
1402 __SetPageSlab(page);
1403 if (page_is_pfmemalloc(page))
1404 SetPageSlabPfmemalloc(page);
1406 start = page_address(page);
1408 if (unlikely(s->flags & SLAB_POISON))
1409 memset(start, POISON_INUSE, PAGE_SIZE << order);
1411 kasan_poison_slab(page);
1413 for_each_object_idx(p, idx, s, start, page->objects) {
1414 setup_object(s, page, p);
1415 if (likely(idx < page->objects))
1416 set_freepointer(s, p, p + s->size);
1418 set_freepointer(s, p, NULL);
1421 page->freelist = start;
1422 page->inuse = page->objects;
1426 if (flags & __GFP_WAIT)
1427 local_irq_disable();
1431 mod_zone_page_state(page_zone(page),
1432 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1433 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1436 inc_slabs_node(s, page_to_nid(page), page->objects);
1441 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1443 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1444 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1448 return allocate_slab(s,
1449 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1452 static void __free_slab(struct kmem_cache *s, struct page *page)
1454 int order = compound_order(page);
1455 int pages = 1 << order;
1457 if (kmem_cache_debug(s)) {
1460 slab_pad_check(s, page);
1461 for_each_object(p, s, page_address(page),
1463 check_object(s, page, p, SLUB_RED_INACTIVE);
1466 kmemcheck_free_shadow(page, compound_order(page));
1468 mod_zone_page_state(page_zone(page),
1469 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1470 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1473 __ClearPageSlabPfmemalloc(page);
1474 __ClearPageSlab(page);
1476 page_mapcount_reset(page);
1477 if (current->reclaim_state)
1478 current->reclaim_state->reclaimed_slab += pages;
1479 __free_pages(page, order);
1480 memcg_uncharge_slab(s, order);
1483 #define need_reserve_slab_rcu \
1484 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1486 static void rcu_free_slab(struct rcu_head *h)
1490 if (need_reserve_slab_rcu)
1491 page = virt_to_head_page(h);
1493 page = container_of((struct list_head *)h, struct page, lru);
1495 __free_slab(page->slab_cache, page);
1498 static void free_slab(struct kmem_cache *s, struct page *page)
1500 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1501 struct rcu_head *head;
1503 if (need_reserve_slab_rcu) {
1504 int order = compound_order(page);
1505 int offset = (PAGE_SIZE << order) - s->reserved;
1507 VM_BUG_ON(s->reserved != sizeof(*head));
1508 head = page_address(page) + offset;
1511 * RCU free overloads the RCU head over the LRU
1513 head = (void *)&page->lru;
1516 call_rcu(head, rcu_free_slab);
1518 __free_slab(s, page);
1521 static void discard_slab(struct kmem_cache *s, struct page *page)
1523 dec_slabs_node(s, page_to_nid(page), page->objects);
1528 * Management of partially allocated slabs.
1531 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1534 if (tail == DEACTIVATE_TO_TAIL)
1535 list_add_tail(&page->lru, &n->partial);
1537 list_add(&page->lru, &n->partial);
1540 static inline void add_partial(struct kmem_cache_node *n,
1541 struct page *page, int tail)
1543 lockdep_assert_held(&n->list_lock);
1544 __add_partial(n, page, tail);
1548 __remove_partial(struct kmem_cache_node *n, struct page *page)
1550 list_del(&page->lru);
1554 static inline void remove_partial(struct kmem_cache_node *n,
1557 lockdep_assert_held(&n->list_lock);
1558 __remove_partial(n, page);
1562 * Remove slab from the partial list, freeze it and
1563 * return the pointer to the freelist.
1565 * Returns a list of objects or NULL if it fails.
1567 static inline void *acquire_slab(struct kmem_cache *s,
1568 struct kmem_cache_node *n, struct page *page,
1569 int mode, int *objects)
1572 unsigned long counters;
1575 lockdep_assert_held(&n->list_lock);
1578 * Zap the freelist and set the frozen bit.
1579 * The old freelist is the list of objects for the
1580 * per cpu allocation list.
1582 freelist = page->freelist;
1583 counters = page->counters;
1584 new.counters = counters;
1585 *objects = new.objects - new.inuse;
1587 new.inuse = page->objects;
1588 new.freelist = NULL;
1590 new.freelist = freelist;
1593 VM_BUG_ON(new.frozen);
1596 if (!__cmpxchg_double_slab(s, page,
1598 new.freelist, new.counters,
1602 remove_partial(n, page);
1607 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1608 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1611 * Try to allocate a partial slab from a specific node.
1613 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1614 struct kmem_cache_cpu *c, gfp_t flags)
1616 struct page *page, *page2;
1617 void *object = NULL;
1622 * Racy check. If we mistakenly see no partial slabs then we
1623 * just allocate an empty slab. If we mistakenly try to get a
1624 * partial slab and there is none available then get_partials()
1627 if (!n || !n->nr_partial)
1630 spin_lock(&n->list_lock);
1631 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1634 if (!pfmemalloc_match(page, flags))
1637 t = acquire_slab(s, n, page, object == NULL, &objects);
1641 available += objects;
1644 stat(s, ALLOC_FROM_PARTIAL);
1647 put_cpu_partial(s, page, 0);
1648 stat(s, CPU_PARTIAL_NODE);
1650 if (!kmem_cache_has_cpu_partial(s)
1651 || available > s->cpu_partial / 2)
1655 spin_unlock(&n->list_lock);
1660 * Get a page from somewhere. Search in increasing NUMA distances.
1662 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1663 struct kmem_cache_cpu *c)
1666 struct zonelist *zonelist;
1669 enum zone_type high_zoneidx = gfp_zone(flags);
1671 unsigned int cpuset_mems_cookie;
1674 * The defrag ratio allows a configuration of the tradeoffs between
1675 * inter node defragmentation and node local allocations. A lower
1676 * defrag_ratio increases the tendency to do local allocations
1677 * instead of attempting to obtain partial slabs from other nodes.
1679 * If the defrag_ratio is set to 0 then kmalloc() always
1680 * returns node local objects. If the ratio is higher then kmalloc()
1681 * may return off node objects because partial slabs are obtained
1682 * from other nodes and filled up.
1684 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1685 * defrag_ratio = 1000) then every (well almost) allocation will
1686 * first attempt to defrag slab caches on other nodes. This means
1687 * scanning over all nodes to look for partial slabs which may be
1688 * expensive if we do it every time we are trying to find a slab
1689 * with available objects.
1691 if (!s->remote_node_defrag_ratio ||
1692 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1696 cpuset_mems_cookie = read_mems_allowed_begin();
1697 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1698 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1699 struct kmem_cache_node *n;
1701 n = get_node(s, zone_to_nid(zone));
1703 if (n && cpuset_zone_allowed(zone, flags) &&
1704 n->nr_partial > s->min_partial) {
1705 object = get_partial_node(s, n, c, flags);
1708 * Don't check read_mems_allowed_retry()
1709 * here - if mems_allowed was updated in
1710 * parallel, that was a harmless race
1711 * between allocation and the cpuset
1718 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1724 * Get a partial page, lock it and return it.
1726 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1727 struct kmem_cache_cpu *c)
1730 int searchnode = node;
1732 if (node == NUMA_NO_NODE)
1733 searchnode = numa_mem_id();
1734 else if (!node_present_pages(node))
1735 searchnode = node_to_mem_node(node);
1737 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1738 if (object || node != NUMA_NO_NODE)
1741 return get_any_partial(s, flags, c);
1744 #ifdef CONFIG_PREEMPT
1746 * Calculate the next globally unique transaction for disambiguiation
1747 * during cmpxchg. The transactions start with the cpu number and are then
1748 * incremented by CONFIG_NR_CPUS.
1750 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1753 * No preemption supported therefore also no need to check for
1759 static inline unsigned long next_tid(unsigned long tid)
1761 return tid + TID_STEP;
1764 static inline unsigned int tid_to_cpu(unsigned long tid)
1766 return tid % TID_STEP;
1769 static inline unsigned long tid_to_event(unsigned long tid)
1771 return tid / TID_STEP;
1774 static inline unsigned int init_tid(int cpu)
1779 static inline void note_cmpxchg_failure(const char *n,
1780 const struct kmem_cache *s, unsigned long tid)
1782 #ifdef SLUB_DEBUG_CMPXCHG
1783 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1785 pr_info("%s %s: cmpxchg redo ", n, s->name);
1787 #ifdef CONFIG_PREEMPT
1788 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1789 pr_warn("due to cpu change %d -> %d\n",
1790 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1793 if (tid_to_event(tid) != tid_to_event(actual_tid))
1794 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1795 tid_to_event(tid), tid_to_event(actual_tid));
1797 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1798 actual_tid, tid, next_tid(tid));
1800 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1803 static void init_kmem_cache_cpus(struct kmem_cache *s)
1807 for_each_possible_cpu(cpu)
1808 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1812 * Remove the cpu slab
1814 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1817 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1818 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1820 enum slab_modes l = M_NONE, m = M_NONE;
1822 int tail = DEACTIVATE_TO_HEAD;
1826 if (page->freelist) {
1827 stat(s, DEACTIVATE_REMOTE_FREES);
1828 tail = DEACTIVATE_TO_TAIL;
1832 * Stage one: Free all available per cpu objects back
1833 * to the page freelist while it is still frozen. Leave the
1836 * There is no need to take the list->lock because the page
1839 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1841 unsigned long counters;
1844 prior = page->freelist;
1845 counters = page->counters;
1846 set_freepointer(s, freelist, prior);
1847 new.counters = counters;
1849 VM_BUG_ON(!new.frozen);
1851 } while (!__cmpxchg_double_slab(s, page,
1853 freelist, new.counters,
1854 "drain percpu freelist"));
1856 freelist = nextfree;
1860 * Stage two: Ensure that the page is unfrozen while the
1861 * list presence reflects the actual number of objects
1864 * We setup the list membership and then perform a cmpxchg
1865 * with the count. If there is a mismatch then the page
1866 * is not unfrozen but the page is on the wrong list.
1868 * Then we restart the process which may have to remove
1869 * the page from the list that we just put it on again
1870 * because the number of objects in the slab may have
1875 old.freelist = page->freelist;
1876 old.counters = page->counters;
1877 VM_BUG_ON(!old.frozen);
1879 /* Determine target state of the slab */
1880 new.counters = old.counters;
1883 set_freepointer(s, freelist, old.freelist);
1884 new.freelist = freelist;
1886 new.freelist = old.freelist;
1890 if (!new.inuse && n->nr_partial >= s->min_partial)
1892 else if (new.freelist) {
1897 * Taking the spinlock removes the possiblity
1898 * that acquire_slab() will see a slab page that
1901 spin_lock(&n->list_lock);
1905 if (kmem_cache_debug(s) && !lock) {
1908 * This also ensures that the scanning of full
1909 * slabs from diagnostic functions will not see
1912 spin_lock(&n->list_lock);
1920 remove_partial(n, page);
1922 else if (l == M_FULL)
1924 remove_full(s, n, page);
1926 if (m == M_PARTIAL) {
1928 add_partial(n, page, tail);
1931 } else if (m == M_FULL) {
1933 stat(s, DEACTIVATE_FULL);
1934 add_full(s, n, page);
1940 if (!__cmpxchg_double_slab(s, page,
1941 old.freelist, old.counters,
1942 new.freelist, new.counters,
1947 spin_unlock(&n->list_lock);
1950 stat(s, DEACTIVATE_EMPTY);
1951 discard_slab(s, page);
1957 * Unfreeze all the cpu partial slabs.
1959 * This function must be called with interrupts disabled
1960 * for the cpu using c (or some other guarantee must be there
1961 * to guarantee no concurrent accesses).
1963 static void unfreeze_partials(struct kmem_cache *s,
1964 struct kmem_cache_cpu *c)
1966 #ifdef CONFIG_SLUB_CPU_PARTIAL
1967 struct kmem_cache_node *n = NULL, *n2 = NULL;
1968 struct page *page, *discard_page = NULL;
1970 while ((page = c->partial)) {
1974 c->partial = page->next;
1976 n2 = get_node(s, page_to_nid(page));
1979 spin_unlock(&n->list_lock);
1982 spin_lock(&n->list_lock);
1987 old.freelist = page->freelist;
1988 old.counters = page->counters;
1989 VM_BUG_ON(!old.frozen);
1991 new.counters = old.counters;
1992 new.freelist = old.freelist;
1996 } while (!__cmpxchg_double_slab(s, page,
1997 old.freelist, old.counters,
1998 new.freelist, new.counters,
1999 "unfreezing slab"));
2001 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2002 page->next = discard_page;
2003 discard_page = page;
2005 add_partial(n, page, DEACTIVATE_TO_TAIL);
2006 stat(s, FREE_ADD_PARTIAL);
2011 spin_unlock(&n->list_lock);
2013 while (discard_page) {
2014 page = discard_page;
2015 discard_page = discard_page->next;
2017 stat(s, DEACTIVATE_EMPTY);
2018 discard_slab(s, page);
2025 * Put a page that was just frozen (in __slab_free) into a partial page
2026 * slot if available. This is done without interrupts disabled and without
2027 * preemption disabled. The cmpxchg is racy and may put the partial page
2028 * onto a random cpus partial slot.
2030 * If we did not find a slot then simply move all the partials to the
2031 * per node partial list.
2033 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2035 #ifdef CONFIG_SLUB_CPU_PARTIAL
2036 struct page *oldpage;
2044 oldpage = this_cpu_read(s->cpu_slab->partial);
2047 pobjects = oldpage->pobjects;
2048 pages = oldpage->pages;
2049 if (drain && pobjects > s->cpu_partial) {
2050 unsigned long flags;
2052 * partial array is full. Move the existing
2053 * set to the per node partial list.
2055 local_irq_save(flags);
2056 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2057 local_irq_restore(flags);
2061 stat(s, CPU_PARTIAL_DRAIN);
2066 pobjects += page->objects - page->inuse;
2068 page->pages = pages;
2069 page->pobjects = pobjects;
2070 page->next = oldpage;
2072 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2074 if (unlikely(!s->cpu_partial)) {
2075 unsigned long flags;
2077 local_irq_save(flags);
2078 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2079 local_irq_restore(flags);
2085 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2087 stat(s, CPUSLAB_FLUSH);
2088 deactivate_slab(s, c->page, c->freelist);
2090 c->tid = next_tid(c->tid);
2098 * Called from IPI handler with interrupts disabled.
2100 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2102 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2108 unfreeze_partials(s, c);
2112 static void flush_cpu_slab(void *d)
2114 struct kmem_cache *s = d;
2116 __flush_cpu_slab(s, smp_processor_id());
2119 static bool has_cpu_slab(int cpu, void *info)
2121 struct kmem_cache *s = info;
2122 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2124 return c->page || c->partial;
2127 static void flush_all(struct kmem_cache *s)
2129 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2133 * Check if the objects in a per cpu structure fit numa
2134 * locality expectations.
2136 static inline int node_match(struct page *page, int node)
2139 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2145 #ifdef CONFIG_SLUB_DEBUG
2146 static int count_free(struct page *page)
2148 return page->objects - page->inuse;
2151 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2153 return atomic_long_read(&n->total_objects);
2155 #endif /* CONFIG_SLUB_DEBUG */
2157 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2158 static unsigned long count_partial(struct kmem_cache_node *n,
2159 int (*get_count)(struct page *))
2161 unsigned long flags;
2162 unsigned long x = 0;
2165 spin_lock_irqsave(&n->list_lock, flags);
2166 list_for_each_entry(page, &n->partial, lru)
2167 x += get_count(page);
2168 spin_unlock_irqrestore(&n->list_lock, flags);
2171 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2173 static noinline void
2174 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2176 #ifdef CONFIG_SLUB_DEBUG
2177 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2178 DEFAULT_RATELIMIT_BURST);
2180 struct kmem_cache_node *n;
2182 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2185 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2187 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2188 s->name, s->object_size, s->size, oo_order(s->oo),
2191 if (oo_order(s->min) > get_order(s->object_size))
2192 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2195 for_each_kmem_cache_node(s, node, n) {
2196 unsigned long nr_slabs;
2197 unsigned long nr_objs;
2198 unsigned long nr_free;
2200 nr_free = count_partial(n, count_free);
2201 nr_slabs = node_nr_slabs(n);
2202 nr_objs = node_nr_objs(n);
2204 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2205 node, nr_slabs, nr_objs, nr_free);
2210 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2211 int node, struct kmem_cache_cpu **pc)
2214 struct kmem_cache_cpu *c = *pc;
2217 freelist = get_partial(s, flags, node, c);
2222 page = new_slab(s, flags, node);
2224 c = raw_cpu_ptr(s->cpu_slab);
2229 * No other reference to the page yet so we can
2230 * muck around with it freely without cmpxchg
2232 freelist = page->freelist;
2233 page->freelist = NULL;
2235 stat(s, ALLOC_SLAB);
2244 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2246 if (unlikely(PageSlabPfmemalloc(page)))
2247 return gfp_pfmemalloc_allowed(gfpflags);
2253 * Check the page->freelist of a page and either transfer the freelist to the
2254 * per cpu freelist or deactivate the page.
2256 * The page is still frozen if the return value is not NULL.
2258 * If this function returns NULL then the page has been unfrozen.
2260 * This function must be called with interrupt disabled.
2262 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2265 unsigned long counters;
2269 freelist = page->freelist;
2270 counters = page->counters;
2272 new.counters = counters;
2273 VM_BUG_ON(!new.frozen);
2275 new.inuse = page->objects;
2276 new.frozen = freelist != NULL;
2278 } while (!__cmpxchg_double_slab(s, page,
2287 * Slow path. The lockless freelist is empty or we need to perform
2290 * Processing is still very fast if new objects have been freed to the
2291 * regular freelist. In that case we simply take over the regular freelist
2292 * as the lockless freelist and zap the regular freelist.
2294 * If that is not working then we fall back to the partial lists. We take the
2295 * first element of the freelist as the object to allocate now and move the
2296 * rest of the freelist to the lockless freelist.
2298 * And if we were unable to get a new slab from the partial slab lists then
2299 * we need to allocate a new slab. This is the slowest path since it involves
2300 * a call to the page allocator and the setup of a new slab.
2302 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2303 unsigned long addr, struct kmem_cache_cpu *c)
2307 unsigned long flags;
2309 local_irq_save(flags);
2310 #ifdef CONFIG_PREEMPT
2312 * We may have been preempted and rescheduled on a different
2313 * cpu before disabling interrupts. Need to reload cpu area
2316 c = this_cpu_ptr(s->cpu_slab);
2324 if (unlikely(!node_match(page, node))) {
2325 int searchnode = node;
2327 if (node != NUMA_NO_NODE && !node_present_pages(node))
2328 searchnode = node_to_mem_node(node);
2330 if (unlikely(!node_match(page, searchnode))) {
2331 stat(s, ALLOC_NODE_MISMATCH);
2332 deactivate_slab(s, page, c->freelist);
2340 * By rights, we should be searching for a slab page that was
2341 * PFMEMALLOC but right now, we are losing the pfmemalloc
2342 * information when the page leaves the per-cpu allocator
2344 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2345 deactivate_slab(s, page, c->freelist);
2351 /* must check again c->freelist in case of cpu migration or IRQ */
2352 freelist = c->freelist;
2356 freelist = get_freelist(s, page);
2360 stat(s, DEACTIVATE_BYPASS);
2364 stat(s, ALLOC_REFILL);
2368 * freelist is pointing to the list of objects to be used.
2369 * page is pointing to the page from which the objects are obtained.
2370 * That page must be frozen for per cpu allocations to work.
2372 VM_BUG_ON(!c->page->frozen);
2373 c->freelist = get_freepointer(s, freelist);
2374 c->tid = next_tid(c->tid);
2375 local_irq_restore(flags);
2381 page = c->page = c->partial;
2382 c->partial = page->next;
2383 stat(s, CPU_PARTIAL_ALLOC);
2388 freelist = new_slab_objects(s, gfpflags, node, &c);
2390 if (unlikely(!freelist)) {
2391 slab_out_of_memory(s, gfpflags, node);
2392 local_irq_restore(flags);
2397 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2400 /* Only entered in the debug case */
2401 if (kmem_cache_debug(s) &&
2402 !alloc_debug_processing(s, page, freelist, addr))
2403 goto new_slab; /* Slab failed checks. Next slab needed */
2405 deactivate_slab(s, page, get_freepointer(s, freelist));
2408 local_irq_restore(flags);
2413 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2414 * have the fastpath folded into their functions. So no function call
2415 * overhead for requests that can be satisfied on the fastpath.
2417 * The fastpath works by first checking if the lockless freelist can be used.
2418 * If not then __slab_alloc is called for slow processing.
2420 * Otherwise we can simply pick the next object from the lockless free list.
2422 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2423 gfp_t gfpflags, int node, unsigned long addr)
2426 struct kmem_cache_cpu *c;
2430 s = slab_pre_alloc_hook(s, gfpflags);
2435 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2436 * enabled. We may switch back and forth between cpus while
2437 * reading from one cpu area. That does not matter as long
2438 * as we end up on the original cpu again when doing the cmpxchg.
2440 * We should guarantee that tid and kmem_cache are retrieved on
2441 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2442 * to check if it is matched or not.
2445 tid = this_cpu_read(s->cpu_slab->tid);
2446 c = raw_cpu_ptr(s->cpu_slab);
2447 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2448 unlikely(tid != READ_ONCE(c->tid)));
2451 * Irqless object alloc/free algorithm used here depends on sequence
2452 * of fetching cpu_slab's data. tid should be fetched before anything
2453 * on c to guarantee that object and page associated with previous tid
2454 * won't be used with current tid. If we fetch tid first, object and
2455 * page could be one associated with next tid and our alloc/free
2456 * request will be failed. In this case, we will retry. So, no problem.
2461 * The transaction ids are globally unique per cpu and per operation on
2462 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2463 * occurs on the right processor and that there was no operation on the
2464 * linked list in between.
2467 object = c->freelist;
2469 if (unlikely(!object || !node_match(page, node))) {
2470 object = __slab_alloc(s, gfpflags, node, addr, c);
2471 stat(s, ALLOC_SLOWPATH);
2473 void *next_object = get_freepointer_safe(s, object);
2476 * The cmpxchg will only match if there was no additional
2477 * operation and if we are on the right processor.
2479 * The cmpxchg does the following atomically (without lock
2481 * 1. Relocate first pointer to the current per cpu area.
2482 * 2. Verify that tid and freelist have not been changed
2483 * 3. If they were not changed replace tid and freelist
2485 * Since this is without lock semantics the protection is only
2486 * against code executing on this cpu *not* from access by
2489 if (unlikely(!this_cpu_cmpxchg_double(
2490 s->cpu_slab->freelist, s->cpu_slab->tid,
2492 next_object, next_tid(tid)))) {
2494 note_cmpxchg_failure("slab_alloc", s, tid);
2497 prefetch_freepointer(s, next_object);
2498 stat(s, ALLOC_FASTPATH);
2501 if (unlikely(gfpflags & __GFP_ZERO) && object)
2502 memset(object, 0, s->object_size);
2504 slab_post_alloc_hook(s, gfpflags, object);
2509 static __always_inline void *slab_alloc(struct kmem_cache *s,
2510 gfp_t gfpflags, unsigned long addr)
2512 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2515 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2517 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2519 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2524 EXPORT_SYMBOL(kmem_cache_alloc);
2526 #ifdef CONFIG_TRACING
2527 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2529 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2530 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2531 kasan_kmalloc(s, ret, size);
2534 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2538 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2540 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2542 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2543 s->object_size, s->size, gfpflags, node);
2547 EXPORT_SYMBOL(kmem_cache_alloc_node);
2549 #ifdef CONFIG_TRACING
2550 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2552 int node, size_t size)
2554 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2556 trace_kmalloc_node(_RET_IP_, ret,
2557 size, s->size, gfpflags, node);
2559 kasan_kmalloc(s, ret, size);
2562 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2567 * Slow path handling. This may still be called frequently since objects
2568 * have a longer lifetime than the cpu slabs in most processing loads.
2570 * So we still attempt to reduce cache line usage. Just take the slab
2571 * lock and free the item. If there is no additional partial page
2572 * handling required then we can return immediately.
2574 static void __slab_free(struct kmem_cache *s, struct page *page,
2575 void *x, unsigned long addr)
2578 void **object = (void *)x;
2581 unsigned long counters;
2582 struct kmem_cache_node *n = NULL;
2583 unsigned long uninitialized_var(flags);
2585 stat(s, FREE_SLOWPATH);
2587 if (kmem_cache_debug(s) &&
2588 !(n = free_debug_processing(s, page, x, addr, &flags)))
2593 spin_unlock_irqrestore(&n->list_lock, flags);
2596 prior = page->freelist;
2597 counters = page->counters;
2598 set_freepointer(s, object, prior);
2599 new.counters = counters;
2600 was_frozen = new.frozen;
2602 if ((!new.inuse || !prior) && !was_frozen) {
2604 if (kmem_cache_has_cpu_partial(s) && !prior) {
2607 * Slab was on no list before and will be
2609 * We can defer the list move and instead
2614 } else { /* Needs to be taken off a list */
2616 n = get_node(s, page_to_nid(page));
2618 * Speculatively acquire the list_lock.
2619 * If the cmpxchg does not succeed then we may
2620 * drop the list_lock without any processing.
2622 * Otherwise the list_lock will synchronize with
2623 * other processors updating the list of slabs.
2625 spin_lock_irqsave(&n->list_lock, flags);
2630 } while (!cmpxchg_double_slab(s, page,
2632 object, new.counters,
2638 * If we just froze the page then put it onto the
2639 * per cpu partial list.
2641 if (new.frozen && !was_frozen) {
2642 put_cpu_partial(s, page, 1);
2643 stat(s, CPU_PARTIAL_FREE);
2646 * The list lock was not taken therefore no list
2647 * activity can be necessary.
2650 stat(s, FREE_FROZEN);
2654 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2658 * Objects left in the slab. If it was not on the partial list before
2661 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2662 if (kmem_cache_debug(s))
2663 remove_full(s, n, page);
2664 add_partial(n, page, DEACTIVATE_TO_TAIL);
2665 stat(s, FREE_ADD_PARTIAL);
2667 spin_unlock_irqrestore(&n->list_lock, flags);
2673 * Slab on the partial list.
2675 remove_partial(n, page);
2676 stat(s, FREE_REMOVE_PARTIAL);
2678 /* Slab must be on the full list */
2679 remove_full(s, n, page);
2682 spin_unlock_irqrestore(&n->list_lock, flags);
2684 discard_slab(s, page);
2688 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2689 * can perform fastpath freeing without additional function calls.
2691 * The fastpath is only possible if we are freeing to the current cpu slab
2692 * of this processor. This typically the case if we have just allocated
2695 * If fastpath is not possible then fall back to __slab_free where we deal
2696 * with all sorts of special processing.
2698 static __always_inline void slab_free(struct kmem_cache *s,
2699 struct page *page, void *x, unsigned long addr)
2701 void **object = (void *)x;
2702 struct kmem_cache_cpu *c;
2705 slab_free_hook(s, x);
2709 * Determine the currently cpus per cpu slab.
2710 * The cpu may change afterward. However that does not matter since
2711 * data is retrieved via this pointer. If we are on the same cpu
2712 * during the cmpxchg then the free will succeed.
2715 tid = this_cpu_read(s->cpu_slab->tid);
2716 c = raw_cpu_ptr(s->cpu_slab);
2717 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2718 unlikely(tid != READ_ONCE(c->tid)));
2720 /* Same with comment on barrier() in slab_alloc_node() */
2723 if (likely(page == c->page)) {
2724 set_freepointer(s, object, c->freelist);
2726 if (unlikely(!this_cpu_cmpxchg_double(
2727 s->cpu_slab->freelist, s->cpu_slab->tid,
2729 object, next_tid(tid)))) {
2731 note_cmpxchg_failure("slab_free", s, tid);
2734 stat(s, FREE_FASTPATH);
2736 __slab_free(s, page, x, addr);
2740 void kmem_cache_free(struct kmem_cache *s, void *x)
2742 s = cache_from_obj(s, x);
2745 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2746 trace_kmem_cache_free(_RET_IP_, x);
2748 EXPORT_SYMBOL(kmem_cache_free);
2750 /* Note that interrupts must be enabled when calling this function. */
2751 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2753 struct kmem_cache_cpu *c;
2757 local_irq_disable();
2758 c = this_cpu_ptr(s->cpu_slab);
2760 for (i = 0; i < size; i++) {
2761 void *object = p[i];
2764 /* kmem cache debug support */
2765 s = cache_from_obj(s, object);
2768 slab_free_hook(s, object);
2770 page = virt_to_head_page(object);
2772 if (c->page == page) {
2773 /* Fastpath: local CPU free */
2774 set_freepointer(s, object, c->freelist);
2775 c->freelist = object;
2777 c->tid = next_tid(c->tid);
2779 /* Slowpath: overhead locked cmpxchg_double_slab */
2780 __slab_free(s, page, object, _RET_IP_);
2781 local_irq_disable();
2782 c = this_cpu_ptr(s->cpu_slab);
2786 c->tid = next_tid(c->tid);
2789 EXPORT_SYMBOL(kmem_cache_free_bulk);
2791 /* Note that interrupts must be enabled when calling this function. */
2792 bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2795 struct kmem_cache_cpu *c;
2799 * Drain objects in the per cpu slab, while disabling local
2800 * IRQs, which protects against PREEMPT and interrupts
2801 * handlers invoking normal fastpath.
2803 local_irq_disable();
2804 c = this_cpu_ptr(s->cpu_slab);
2806 for (i = 0; i < size; i++) {
2807 void *object = c->freelist;
2809 if (unlikely(!object)) {
2812 * Invoking slow path likely have side-effect
2813 * of re-populating per CPU c->freelist
2815 p[i] = __slab_alloc(s, flags, NUMA_NO_NODE,
2817 if (unlikely(!p[i])) {
2818 __kmem_cache_free_bulk(s, i, p);
2821 local_irq_disable();
2822 c = this_cpu_ptr(s->cpu_slab);
2823 continue; /* goto for-loop */
2826 /* kmem_cache debug support */
2827 s = slab_pre_alloc_hook(s, flags);
2829 __kmem_cache_free_bulk(s, i, p);
2830 c->tid = next_tid(c->tid);
2835 c->freelist = get_freepointer(s, object);
2838 /* kmem_cache debug support */
2839 slab_post_alloc_hook(s, flags, object);
2841 c->tid = next_tid(c->tid);
2844 /* Clear memory outside IRQ disabled fastpath loop */
2845 if (unlikely(flags & __GFP_ZERO)) {
2848 for (j = 0; j < i; j++)
2849 memset(p[j], 0, s->object_size);
2854 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2858 * Object placement in a slab is made very easy because we always start at
2859 * offset 0. If we tune the size of the object to the alignment then we can
2860 * get the required alignment by putting one properly sized object after
2863 * Notice that the allocation order determines the sizes of the per cpu
2864 * caches. Each processor has always one slab available for allocations.
2865 * Increasing the allocation order reduces the number of times that slabs
2866 * must be moved on and off the partial lists and is therefore a factor in
2871 * Mininum / Maximum order of slab pages. This influences locking overhead
2872 * and slab fragmentation. A higher order reduces the number of partial slabs
2873 * and increases the number of allocations possible without having to
2874 * take the list_lock.
2876 static int slub_min_order;
2877 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2878 static int slub_min_objects;
2881 * Calculate the order of allocation given an slab object size.
2883 * The order of allocation has significant impact on performance and other
2884 * system components. Generally order 0 allocations should be preferred since
2885 * order 0 does not cause fragmentation in the page allocator. Larger objects
2886 * be problematic to put into order 0 slabs because there may be too much
2887 * unused space left. We go to a higher order if more than 1/16th of the slab
2890 * In order to reach satisfactory performance we must ensure that a minimum
2891 * number of objects is in one slab. Otherwise we may generate too much
2892 * activity on the partial lists which requires taking the list_lock. This is
2893 * less a concern for large slabs though which are rarely used.
2895 * slub_max_order specifies the order where we begin to stop considering the
2896 * number of objects in a slab as critical. If we reach slub_max_order then
2897 * we try to keep the page order as low as possible. So we accept more waste
2898 * of space in favor of a small page order.
2900 * Higher order allocations also allow the placement of more objects in a
2901 * slab and thereby reduce object handling overhead. If the user has
2902 * requested a higher mininum order then we start with that one instead of
2903 * the smallest order which will fit the object.
2905 static inline int slab_order(int size, int min_objects,
2906 int max_order, int fract_leftover, int reserved)
2910 int min_order = slub_min_order;
2912 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2913 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2915 for (order = max(min_order, get_order(min_objects * size + reserved));
2916 order <= max_order; order++) {
2918 unsigned long slab_size = PAGE_SIZE << order;
2920 rem = (slab_size - reserved) % size;
2922 if (rem <= slab_size / fract_leftover)
2929 static inline int calculate_order(int size, int reserved)
2937 * Attempt to find best configuration for a slab. This
2938 * works by first attempting to generate a layout with
2939 * the best configuration and backing off gradually.
2941 * First we increase the acceptable waste in a slab. Then
2942 * we reduce the minimum objects required in a slab.
2944 min_objects = slub_min_objects;
2946 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2947 max_objects = order_objects(slub_max_order, size, reserved);
2948 min_objects = min(min_objects, max_objects);
2950 while (min_objects > 1) {
2952 while (fraction >= 4) {
2953 order = slab_order(size, min_objects,
2954 slub_max_order, fraction, reserved);
2955 if (order <= slub_max_order)
2963 * We were unable to place multiple objects in a slab. Now
2964 * lets see if we can place a single object there.
2966 order = slab_order(size, 1, slub_max_order, 1, reserved);
2967 if (order <= slub_max_order)
2971 * Doh this slab cannot be placed using slub_max_order.
2973 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2974 if (order < MAX_ORDER)
2980 init_kmem_cache_node(struct kmem_cache_node *n)
2983 spin_lock_init(&n->list_lock);
2984 INIT_LIST_HEAD(&n->partial);
2985 #ifdef CONFIG_SLUB_DEBUG
2986 atomic_long_set(&n->nr_slabs, 0);
2987 atomic_long_set(&n->total_objects, 0);
2988 INIT_LIST_HEAD(&n->full);
2992 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2994 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2995 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2998 * Must align to double word boundary for the double cmpxchg
2999 * instructions to work; see __pcpu_double_call_return_bool().
3001 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3002 2 * sizeof(void *));
3007 init_kmem_cache_cpus(s);
3012 static struct kmem_cache *kmem_cache_node;
3015 * No kmalloc_node yet so do it by hand. We know that this is the first
3016 * slab on the node for this slabcache. There are no concurrent accesses
3019 * Note that this function only works on the kmem_cache_node
3020 * when allocating for the kmem_cache_node. This is used for bootstrapping
3021 * memory on a fresh node that has no slab structures yet.
3023 static void early_kmem_cache_node_alloc(int node)
3026 struct kmem_cache_node *n;
3028 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3030 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3033 if (page_to_nid(page) != node) {
3034 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3035 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3040 page->freelist = get_freepointer(kmem_cache_node, n);
3043 kmem_cache_node->node[node] = n;
3044 #ifdef CONFIG_SLUB_DEBUG
3045 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3046 init_tracking(kmem_cache_node, n);
3048 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3049 init_kmem_cache_node(n);
3050 inc_slabs_node(kmem_cache_node, node, page->objects);
3053 * No locks need to be taken here as it has just been
3054 * initialized and there is no concurrent access.
3056 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3059 static void free_kmem_cache_nodes(struct kmem_cache *s)
3062 struct kmem_cache_node *n;
3064 for_each_kmem_cache_node(s, node, n) {
3065 kmem_cache_free(kmem_cache_node, n);
3066 s->node[node] = NULL;
3070 static int init_kmem_cache_nodes(struct kmem_cache *s)
3074 for_each_node_state(node, N_NORMAL_MEMORY) {
3075 struct kmem_cache_node *n;
3077 if (slab_state == DOWN) {
3078 early_kmem_cache_node_alloc(node);
3081 n = kmem_cache_alloc_node(kmem_cache_node,
3085 free_kmem_cache_nodes(s);
3090 init_kmem_cache_node(n);
3095 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3097 if (min < MIN_PARTIAL)
3099 else if (min > MAX_PARTIAL)
3101 s->min_partial = min;
3105 * calculate_sizes() determines the order and the distribution of data within
3108 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3110 unsigned long flags = s->flags;
3111 unsigned long size = s->object_size;
3115 * Round up object size to the next word boundary. We can only
3116 * place the free pointer at word boundaries and this determines
3117 * the possible location of the free pointer.
3119 size = ALIGN(size, sizeof(void *));
3121 #ifdef CONFIG_SLUB_DEBUG
3123 * Determine if we can poison the object itself. If the user of
3124 * the slab may touch the object after free or before allocation
3125 * then we should never poison the object itself.
3127 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3129 s->flags |= __OBJECT_POISON;
3131 s->flags &= ~__OBJECT_POISON;
3135 * If we are Redzoning then check if there is some space between the
3136 * end of the object and the free pointer. If not then add an
3137 * additional word to have some bytes to store Redzone information.
3139 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3140 size += sizeof(void *);
3144 * With that we have determined the number of bytes in actual use
3145 * by the object. This is the potential offset to the free pointer.
3149 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3152 * Relocate free pointer after the object if it is not
3153 * permitted to overwrite the first word of the object on
3156 * This is the case if we do RCU, have a constructor or
3157 * destructor or are poisoning the objects.
3160 size += sizeof(void *);
3163 #ifdef CONFIG_SLUB_DEBUG
3164 if (flags & SLAB_STORE_USER)
3166 * Need to store information about allocs and frees after
3169 size += 2 * sizeof(struct track);
3171 if (flags & SLAB_RED_ZONE)
3173 * Add some empty padding so that we can catch
3174 * overwrites from earlier objects rather than let
3175 * tracking information or the free pointer be
3176 * corrupted if a user writes before the start
3179 size += sizeof(void *);
3183 * SLUB stores one object immediately after another beginning from
3184 * offset 0. In order to align the objects we have to simply size
3185 * each object to conform to the alignment.
3187 size = ALIGN(size, s->align);
3189 if (forced_order >= 0)
3190 order = forced_order;
3192 order = calculate_order(size, s->reserved);
3199 s->allocflags |= __GFP_COMP;
3201 if (s->flags & SLAB_CACHE_DMA)
3202 s->allocflags |= GFP_DMA;
3204 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3205 s->allocflags |= __GFP_RECLAIMABLE;
3208 * Determine the number of objects per slab
3210 s->oo = oo_make(order, size, s->reserved);
3211 s->min = oo_make(get_order(size), size, s->reserved);
3212 if (oo_objects(s->oo) > oo_objects(s->max))
3215 return !!oo_objects(s->oo);
3218 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3220 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3223 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3224 s->reserved = sizeof(struct rcu_head);
3226 if (!calculate_sizes(s, -1))
3228 if (disable_higher_order_debug) {
3230 * Disable debugging flags that store metadata if the min slab
3233 if (get_order(s->size) > get_order(s->object_size)) {
3234 s->flags &= ~DEBUG_METADATA_FLAGS;
3236 if (!calculate_sizes(s, -1))
3241 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3242 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3243 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3244 /* Enable fast mode */
3245 s->flags |= __CMPXCHG_DOUBLE;
3249 * The larger the object size is, the more pages we want on the partial
3250 * list to avoid pounding the page allocator excessively.
3252 set_min_partial(s, ilog2(s->size) / 2);
3255 * cpu_partial determined the maximum number of objects kept in the
3256 * per cpu partial lists of a processor.
3258 * Per cpu partial lists mainly contain slabs that just have one
3259 * object freed. If they are used for allocation then they can be
3260 * filled up again with minimal effort. The slab will never hit the
3261 * per node partial lists and therefore no locking will be required.
3263 * This setting also determines
3265 * A) The number of objects from per cpu partial slabs dumped to the
3266 * per node list when we reach the limit.
3267 * B) The number of objects in cpu partial slabs to extract from the
3268 * per node list when we run out of per cpu objects. We only fetch
3269 * 50% to keep some capacity around for frees.
3271 if (!kmem_cache_has_cpu_partial(s))
3273 else if (s->size >= PAGE_SIZE)
3275 else if (s->size >= 1024)
3277 else if (s->size >= 256)
3278 s->cpu_partial = 13;
3280 s->cpu_partial = 30;
3283 s->remote_node_defrag_ratio = 1000;
3285 if (!init_kmem_cache_nodes(s))
3288 if (alloc_kmem_cache_cpus(s))
3291 free_kmem_cache_nodes(s);
3293 if (flags & SLAB_PANIC)
3294 panic("Cannot create slab %s size=%lu realsize=%u "
3295 "order=%u offset=%u flags=%lx\n",
3296 s->name, (unsigned long)s->size, s->size,
3297 oo_order(s->oo), s->offset, flags);
3301 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3304 #ifdef CONFIG_SLUB_DEBUG
3305 void *addr = page_address(page);
3307 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3308 sizeof(long), GFP_ATOMIC);
3311 slab_err(s, page, text, s->name);
3314 get_map(s, page, map);
3315 for_each_object(p, s, addr, page->objects) {
3317 if (!test_bit(slab_index(p, s, addr), map)) {
3318 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3319 print_tracking(s, p);
3328 * Attempt to free all partial slabs on a node.
3329 * This is called from kmem_cache_close(). We must be the last thread
3330 * using the cache and therefore we do not need to lock anymore.
3332 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3334 struct page *page, *h;
3336 list_for_each_entry_safe(page, h, &n->partial, lru) {
3338 __remove_partial(n, page);
3339 discard_slab(s, page);
3341 list_slab_objects(s, page,
3342 "Objects remaining in %s on kmem_cache_close()");
3348 * Release all resources used by a slab cache.
3350 static inline int kmem_cache_close(struct kmem_cache *s)
3353 struct kmem_cache_node *n;
3356 /* Attempt to free all objects */
3357 for_each_kmem_cache_node(s, node, n) {
3359 if (n->nr_partial || slabs_node(s, node))
3362 free_percpu(s->cpu_slab);
3363 free_kmem_cache_nodes(s);
3367 int __kmem_cache_shutdown(struct kmem_cache *s)
3369 return kmem_cache_close(s);
3372 /********************************************************************
3374 *******************************************************************/
3376 static int __init setup_slub_min_order(char *str)
3378 get_option(&str, &slub_min_order);
3383 __setup("slub_min_order=", setup_slub_min_order);
3385 static int __init setup_slub_max_order(char *str)
3387 get_option(&str, &slub_max_order);
3388 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3393 __setup("slub_max_order=", setup_slub_max_order);
3395 static int __init setup_slub_min_objects(char *str)
3397 get_option(&str, &slub_min_objects);
3402 __setup("slub_min_objects=", setup_slub_min_objects);
3404 void *__kmalloc(size_t size, gfp_t flags)
3406 struct kmem_cache *s;
3409 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3410 return kmalloc_large(size, flags);
3412 s = kmalloc_slab(size, flags);
3414 if (unlikely(ZERO_OR_NULL_PTR(s)))
3417 ret = slab_alloc(s, flags, _RET_IP_);
3419 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3421 kasan_kmalloc(s, ret, size);
3425 EXPORT_SYMBOL(__kmalloc);
3428 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3433 flags |= __GFP_COMP | __GFP_NOTRACK;
3434 page = alloc_kmem_pages_node(node, flags, get_order(size));
3436 ptr = page_address(page);
3438 kmalloc_large_node_hook(ptr, size, flags);
3442 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3444 struct kmem_cache *s;
3447 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3448 ret = kmalloc_large_node(size, flags, node);
3450 trace_kmalloc_node(_RET_IP_, ret,
3451 size, PAGE_SIZE << get_order(size),
3457 s = kmalloc_slab(size, flags);
3459 if (unlikely(ZERO_OR_NULL_PTR(s)))
3462 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3464 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3466 kasan_kmalloc(s, ret, size);
3470 EXPORT_SYMBOL(__kmalloc_node);
3473 static size_t __ksize(const void *object)
3477 if (unlikely(object == ZERO_SIZE_PTR))
3480 page = virt_to_head_page(object);
3482 if (unlikely(!PageSlab(page))) {
3483 WARN_ON(!PageCompound(page));
3484 return PAGE_SIZE << compound_order(page);
3487 return slab_ksize(page->slab_cache);
3490 size_t ksize(const void *object)
3492 size_t size = __ksize(object);
3493 /* We assume that ksize callers could use whole allocated area,
3494 so we need unpoison this area. */
3495 kasan_krealloc(object, size);
3498 EXPORT_SYMBOL(ksize);
3500 void kfree(const void *x)
3503 void *object = (void *)x;
3505 trace_kfree(_RET_IP_, x);
3507 if (unlikely(ZERO_OR_NULL_PTR(x)))
3510 page = virt_to_head_page(x);
3511 if (unlikely(!PageSlab(page))) {
3512 BUG_ON(!PageCompound(page));
3514 __free_kmem_pages(page, compound_order(page));
3517 slab_free(page->slab_cache, page, object, _RET_IP_);
3519 EXPORT_SYMBOL(kfree);
3521 #define SHRINK_PROMOTE_MAX 32
3524 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3525 * up most to the head of the partial lists. New allocations will then
3526 * fill those up and thus they can be removed from the partial lists.
3528 * The slabs with the least items are placed last. This results in them
3529 * being allocated from last increasing the chance that the last objects
3530 * are freed in them.
3532 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3536 struct kmem_cache_node *n;
3539 struct list_head discard;
3540 struct list_head promote[SHRINK_PROMOTE_MAX];
3541 unsigned long flags;
3546 * Disable empty slabs caching. Used to avoid pinning offline
3547 * memory cgroups by kmem pages that can be freed.
3553 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3554 * so we have to make sure the change is visible.
3556 kick_all_cpus_sync();
3560 for_each_kmem_cache_node(s, node, n) {
3561 INIT_LIST_HEAD(&discard);
3562 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3563 INIT_LIST_HEAD(promote + i);
3565 spin_lock_irqsave(&n->list_lock, flags);
3568 * Build lists of slabs to discard or promote.
3570 * Note that concurrent frees may occur while we hold the
3571 * list_lock. page->inuse here is the upper limit.
3573 list_for_each_entry_safe(page, t, &n->partial, lru) {
3574 int free = page->objects - page->inuse;
3576 /* Do not reread page->inuse */
3579 /* We do not keep full slabs on the list */
3582 if (free == page->objects) {
3583 list_move(&page->lru, &discard);
3585 } else if (free <= SHRINK_PROMOTE_MAX)
3586 list_move(&page->lru, promote + free - 1);
3590 * Promote the slabs filled up most to the head of the
3593 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3594 list_splice(promote + i, &n->partial);
3596 spin_unlock_irqrestore(&n->list_lock, flags);
3598 /* Release empty slabs */
3599 list_for_each_entry_safe(page, t, &discard, lru)
3600 discard_slab(s, page);
3602 if (slabs_node(s, node))
3609 static int slab_mem_going_offline_callback(void *arg)
3611 struct kmem_cache *s;
3613 mutex_lock(&slab_mutex);
3614 list_for_each_entry(s, &slab_caches, list)
3615 __kmem_cache_shrink(s, false);
3616 mutex_unlock(&slab_mutex);
3621 static void slab_mem_offline_callback(void *arg)
3623 struct kmem_cache_node *n;
3624 struct kmem_cache *s;
3625 struct memory_notify *marg = arg;
3628 offline_node = marg->status_change_nid_normal;
3631 * If the node still has available memory. we need kmem_cache_node
3634 if (offline_node < 0)
3637 mutex_lock(&slab_mutex);
3638 list_for_each_entry(s, &slab_caches, list) {
3639 n = get_node(s, offline_node);
3642 * if n->nr_slabs > 0, slabs still exist on the node
3643 * that is going down. We were unable to free them,
3644 * and offline_pages() function shouldn't call this
3645 * callback. So, we must fail.
3647 BUG_ON(slabs_node(s, offline_node));
3649 s->node[offline_node] = NULL;
3650 kmem_cache_free(kmem_cache_node, n);
3653 mutex_unlock(&slab_mutex);
3656 static int slab_mem_going_online_callback(void *arg)
3658 struct kmem_cache_node *n;
3659 struct kmem_cache *s;
3660 struct memory_notify *marg = arg;
3661 int nid = marg->status_change_nid_normal;
3665 * If the node's memory is already available, then kmem_cache_node is
3666 * already created. Nothing to do.
3672 * We are bringing a node online. No memory is available yet. We must
3673 * allocate a kmem_cache_node structure in order to bring the node
3676 mutex_lock(&slab_mutex);
3677 list_for_each_entry(s, &slab_caches, list) {
3679 * XXX: kmem_cache_alloc_node will fallback to other nodes
3680 * since memory is not yet available from the node that
3683 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3688 init_kmem_cache_node(n);
3692 mutex_unlock(&slab_mutex);
3696 static int slab_memory_callback(struct notifier_block *self,
3697 unsigned long action, void *arg)
3702 case MEM_GOING_ONLINE:
3703 ret = slab_mem_going_online_callback(arg);
3705 case MEM_GOING_OFFLINE:
3706 ret = slab_mem_going_offline_callback(arg);
3709 case MEM_CANCEL_ONLINE:
3710 slab_mem_offline_callback(arg);
3713 case MEM_CANCEL_OFFLINE:
3717 ret = notifier_from_errno(ret);
3723 static struct notifier_block slab_memory_callback_nb = {
3724 .notifier_call = slab_memory_callback,
3725 .priority = SLAB_CALLBACK_PRI,
3728 /********************************************************************
3729 * Basic setup of slabs
3730 *******************************************************************/
3733 * Used for early kmem_cache structures that were allocated using
3734 * the page allocator. Allocate them properly then fix up the pointers
3735 * that may be pointing to the wrong kmem_cache structure.
3738 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3741 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3742 struct kmem_cache_node *n;
3744 memcpy(s, static_cache, kmem_cache->object_size);
3747 * This runs very early, and only the boot processor is supposed to be
3748 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3751 __flush_cpu_slab(s, smp_processor_id());
3752 for_each_kmem_cache_node(s, node, n) {
3755 list_for_each_entry(p, &n->partial, lru)
3758 #ifdef CONFIG_SLUB_DEBUG
3759 list_for_each_entry(p, &n->full, lru)
3763 slab_init_memcg_params(s);
3764 list_add(&s->list, &slab_caches);
3768 void __init kmem_cache_init(void)
3770 static __initdata struct kmem_cache boot_kmem_cache,
3771 boot_kmem_cache_node;
3773 if (debug_guardpage_minorder())
3776 kmem_cache_node = &boot_kmem_cache_node;
3777 kmem_cache = &boot_kmem_cache;
3779 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3780 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3782 register_hotmemory_notifier(&slab_memory_callback_nb);
3784 /* Able to allocate the per node structures */
3785 slab_state = PARTIAL;
3787 create_boot_cache(kmem_cache, "kmem_cache",
3788 offsetof(struct kmem_cache, node) +
3789 nr_node_ids * sizeof(struct kmem_cache_node *),
3790 SLAB_HWCACHE_ALIGN);
3792 kmem_cache = bootstrap(&boot_kmem_cache);
3795 * Allocate kmem_cache_node properly from the kmem_cache slab.
3796 * kmem_cache_node is separately allocated so no need to
3797 * update any list pointers.
3799 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3801 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3802 setup_kmalloc_cache_index_table();
3803 create_kmalloc_caches(0);
3806 register_cpu_notifier(&slab_notifier);
3809 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3811 slub_min_order, slub_max_order, slub_min_objects,
3812 nr_cpu_ids, nr_node_ids);
3815 void __init kmem_cache_init_late(void)
3820 __kmem_cache_alias(const char *name, size_t size, size_t align,
3821 unsigned long flags, void (*ctor)(void *))
3823 struct kmem_cache *s, *c;
3825 s = find_mergeable(size, align, flags, name, ctor);
3830 * Adjust the object sizes so that we clear
3831 * the complete object on kzalloc.
3833 s->object_size = max(s->object_size, (int)size);
3834 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3836 for_each_memcg_cache(c, s) {
3837 c->object_size = s->object_size;
3838 c->inuse = max_t(int, c->inuse,
3839 ALIGN(size, sizeof(void *)));
3842 if (sysfs_slab_alias(s, name)) {
3851 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3855 err = kmem_cache_open(s, flags);
3859 /* Mutex is not taken during early boot */
3860 if (slab_state <= UP)
3863 memcg_propagate_slab_attrs(s);
3864 err = sysfs_slab_add(s);
3866 kmem_cache_close(s);
3873 * Use the cpu notifier to insure that the cpu slabs are flushed when
3876 static int slab_cpuup_callback(struct notifier_block *nfb,
3877 unsigned long action, void *hcpu)
3879 long cpu = (long)hcpu;
3880 struct kmem_cache *s;
3881 unsigned long flags;
3884 case CPU_UP_CANCELED:
3885 case CPU_UP_CANCELED_FROZEN:
3887 case CPU_DEAD_FROZEN:
3888 mutex_lock(&slab_mutex);
3889 list_for_each_entry(s, &slab_caches, list) {
3890 local_irq_save(flags);
3891 __flush_cpu_slab(s, cpu);
3892 local_irq_restore(flags);
3894 mutex_unlock(&slab_mutex);
3902 static struct notifier_block slab_notifier = {
3903 .notifier_call = slab_cpuup_callback
3908 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3910 struct kmem_cache *s;
3913 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3914 return kmalloc_large(size, gfpflags);
3916 s = kmalloc_slab(size, gfpflags);
3918 if (unlikely(ZERO_OR_NULL_PTR(s)))
3921 ret = slab_alloc(s, gfpflags, caller);
3923 /* Honor the call site pointer we received. */
3924 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3930 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3931 int node, unsigned long caller)
3933 struct kmem_cache *s;
3936 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3937 ret = kmalloc_large_node(size, gfpflags, node);
3939 trace_kmalloc_node(caller, ret,
3940 size, PAGE_SIZE << get_order(size),
3946 s = kmalloc_slab(size, gfpflags);
3948 if (unlikely(ZERO_OR_NULL_PTR(s)))
3951 ret = slab_alloc_node(s, gfpflags, node, caller);
3953 /* Honor the call site pointer we received. */
3954 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3961 static int count_inuse(struct page *page)
3966 static int count_total(struct page *page)
3968 return page->objects;
3972 #ifdef CONFIG_SLUB_DEBUG
3973 static int validate_slab(struct kmem_cache *s, struct page *page,
3977 void *addr = page_address(page);
3979 if (!check_slab(s, page) ||
3980 !on_freelist(s, page, NULL))
3983 /* Now we know that a valid freelist exists */
3984 bitmap_zero(map, page->objects);
3986 get_map(s, page, map);
3987 for_each_object(p, s, addr, page->objects) {
3988 if (test_bit(slab_index(p, s, addr), map))
3989 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3993 for_each_object(p, s, addr, page->objects)
3994 if (!test_bit(slab_index(p, s, addr), map))
3995 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4000 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4004 validate_slab(s, page, map);
4008 static int validate_slab_node(struct kmem_cache *s,
4009 struct kmem_cache_node *n, unsigned long *map)
4011 unsigned long count = 0;
4013 unsigned long flags;
4015 spin_lock_irqsave(&n->list_lock, flags);
4017 list_for_each_entry(page, &n->partial, lru) {
4018 validate_slab_slab(s, page, map);
4021 if (count != n->nr_partial)
4022 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4023 s->name, count, n->nr_partial);
4025 if (!(s->flags & SLAB_STORE_USER))
4028 list_for_each_entry(page, &n->full, lru) {
4029 validate_slab_slab(s, page, map);
4032 if (count != atomic_long_read(&n->nr_slabs))
4033 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4034 s->name, count, atomic_long_read(&n->nr_slabs));
4037 spin_unlock_irqrestore(&n->list_lock, flags);
4041 static long validate_slab_cache(struct kmem_cache *s)
4044 unsigned long count = 0;
4045 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4046 sizeof(unsigned long), GFP_KERNEL);
4047 struct kmem_cache_node *n;
4053 for_each_kmem_cache_node(s, node, n)
4054 count += validate_slab_node(s, n, map);
4059 * Generate lists of code addresses where slabcache objects are allocated
4064 unsigned long count;
4071 DECLARE_BITMAP(cpus, NR_CPUS);
4077 unsigned long count;
4078 struct location *loc;
4081 static void free_loc_track(struct loc_track *t)
4084 free_pages((unsigned long)t->loc,
4085 get_order(sizeof(struct location) * t->max));
4088 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4093 order = get_order(sizeof(struct location) * max);
4095 l = (void *)__get_free_pages(flags, order);
4100 memcpy(l, t->loc, sizeof(struct location) * t->count);
4108 static int add_location(struct loc_track *t, struct kmem_cache *s,
4109 const struct track *track)
4111 long start, end, pos;
4113 unsigned long caddr;
4114 unsigned long age = jiffies - track->when;
4120 pos = start + (end - start + 1) / 2;
4123 * There is nothing at "end". If we end up there
4124 * we need to add something to before end.
4129 caddr = t->loc[pos].addr;
4130 if (track->addr == caddr) {
4136 if (age < l->min_time)
4138 if (age > l->max_time)
4141 if (track->pid < l->min_pid)
4142 l->min_pid = track->pid;
4143 if (track->pid > l->max_pid)
4144 l->max_pid = track->pid;
4146 cpumask_set_cpu(track->cpu,
4147 to_cpumask(l->cpus));
4149 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4153 if (track->addr < caddr)
4160 * Not found. Insert new tracking element.
4162 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4168 (t->count - pos) * sizeof(struct location));
4171 l->addr = track->addr;
4175 l->min_pid = track->pid;
4176 l->max_pid = track->pid;
4177 cpumask_clear(to_cpumask(l->cpus));
4178 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4179 nodes_clear(l->nodes);
4180 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4184 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4185 struct page *page, enum track_item alloc,
4188 void *addr = page_address(page);
4191 bitmap_zero(map, page->objects);
4192 get_map(s, page, map);
4194 for_each_object(p, s, addr, page->objects)
4195 if (!test_bit(slab_index(p, s, addr), map))
4196 add_location(t, s, get_track(s, p, alloc));
4199 static int list_locations(struct kmem_cache *s, char *buf,
4200 enum track_item alloc)
4204 struct loc_track t = { 0, 0, NULL };
4206 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4207 sizeof(unsigned long), GFP_KERNEL);
4208 struct kmem_cache_node *n;
4210 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4213 return sprintf(buf, "Out of memory\n");
4215 /* Push back cpu slabs */
4218 for_each_kmem_cache_node(s, node, n) {
4219 unsigned long flags;
4222 if (!atomic_long_read(&n->nr_slabs))
4225 spin_lock_irqsave(&n->list_lock, flags);
4226 list_for_each_entry(page, &n->partial, lru)
4227 process_slab(&t, s, page, alloc, map);
4228 list_for_each_entry(page, &n->full, lru)
4229 process_slab(&t, s, page, alloc, map);
4230 spin_unlock_irqrestore(&n->list_lock, flags);
4233 for (i = 0; i < t.count; i++) {
4234 struct location *l = &t.loc[i];
4236 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4238 len += sprintf(buf + len, "%7ld ", l->count);
4241 len += sprintf(buf + len, "%pS", (void *)l->addr);
4243 len += sprintf(buf + len, "<not-available>");
4245 if (l->sum_time != l->min_time) {
4246 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4248 (long)div_u64(l->sum_time, l->count),
4251 len += sprintf(buf + len, " age=%ld",
4254 if (l->min_pid != l->max_pid)
4255 len += sprintf(buf + len, " pid=%ld-%ld",
4256 l->min_pid, l->max_pid);
4258 len += sprintf(buf + len, " pid=%ld",
4261 if (num_online_cpus() > 1 &&
4262 !cpumask_empty(to_cpumask(l->cpus)) &&
4263 len < PAGE_SIZE - 60)
4264 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4266 cpumask_pr_args(to_cpumask(l->cpus)));
4268 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4269 len < PAGE_SIZE - 60)
4270 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4272 nodemask_pr_args(&l->nodes));
4274 len += sprintf(buf + len, "\n");
4280 len += sprintf(buf, "No data\n");
4285 #ifdef SLUB_RESILIENCY_TEST
4286 static void __init resiliency_test(void)
4290 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4292 pr_err("SLUB resiliency testing\n");
4293 pr_err("-----------------------\n");
4294 pr_err("A. Corruption after allocation\n");
4296 p = kzalloc(16, GFP_KERNEL);
4298 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4301 validate_slab_cache(kmalloc_caches[4]);
4303 /* Hmmm... The next two are dangerous */
4304 p = kzalloc(32, GFP_KERNEL);
4305 p[32 + sizeof(void *)] = 0x34;
4306 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4308 pr_err("If allocated object is overwritten then not detectable\n\n");
4310 validate_slab_cache(kmalloc_caches[5]);
4311 p = kzalloc(64, GFP_KERNEL);
4312 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4314 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4316 pr_err("If allocated object is overwritten then not detectable\n\n");
4317 validate_slab_cache(kmalloc_caches[6]);
4319 pr_err("\nB. Corruption after free\n");
4320 p = kzalloc(128, GFP_KERNEL);
4323 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4324 validate_slab_cache(kmalloc_caches[7]);
4326 p = kzalloc(256, GFP_KERNEL);
4329 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4330 validate_slab_cache(kmalloc_caches[8]);
4332 p = kzalloc(512, GFP_KERNEL);
4335 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4336 validate_slab_cache(kmalloc_caches[9]);
4340 static void resiliency_test(void) {};
4345 enum slab_stat_type {
4346 SL_ALL, /* All slabs */
4347 SL_PARTIAL, /* Only partially allocated slabs */
4348 SL_CPU, /* Only slabs used for cpu caches */
4349 SL_OBJECTS, /* Determine allocated objects not slabs */
4350 SL_TOTAL /* Determine object capacity not slabs */
4353 #define SO_ALL (1 << SL_ALL)
4354 #define SO_PARTIAL (1 << SL_PARTIAL)
4355 #define SO_CPU (1 << SL_CPU)
4356 #define SO_OBJECTS (1 << SL_OBJECTS)
4357 #define SO_TOTAL (1 << SL_TOTAL)
4359 static ssize_t show_slab_objects(struct kmem_cache *s,
4360 char *buf, unsigned long flags)
4362 unsigned long total = 0;
4365 unsigned long *nodes;
4367 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4371 if (flags & SO_CPU) {
4374 for_each_possible_cpu(cpu) {
4375 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4380 page = READ_ONCE(c->page);
4384 node = page_to_nid(page);
4385 if (flags & SO_TOTAL)
4387 else if (flags & SO_OBJECTS)
4395 page = READ_ONCE(c->partial);
4397 node = page_to_nid(page);
4398 if (flags & SO_TOTAL)
4400 else if (flags & SO_OBJECTS)
4411 #ifdef CONFIG_SLUB_DEBUG
4412 if (flags & SO_ALL) {
4413 struct kmem_cache_node *n;
4415 for_each_kmem_cache_node(s, node, n) {
4417 if (flags & SO_TOTAL)
4418 x = atomic_long_read(&n->total_objects);
4419 else if (flags & SO_OBJECTS)
4420 x = atomic_long_read(&n->total_objects) -
4421 count_partial(n, count_free);
4423 x = atomic_long_read(&n->nr_slabs);
4430 if (flags & SO_PARTIAL) {
4431 struct kmem_cache_node *n;
4433 for_each_kmem_cache_node(s, node, n) {
4434 if (flags & SO_TOTAL)
4435 x = count_partial(n, count_total);
4436 else if (flags & SO_OBJECTS)
4437 x = count_partial(n, count_inuse);
4444 x = sprintf(buf, "%lu", total);
4446 for (node = 0; node < nr_node_ids; node++)
4448 x += sprintf(buf + x, " N%d=%lu",
4453 return x + sprintf(buf + x, "\n");
4456 #ifdef CONFIG_SLUB_DEBUG
4457 static int any_slab_objects(struct kmem_cache *s)
4460 struct kmem_cache_node *n;
4462 for_each_kmem_cache_node(s, node, n)
4463 if (atomic_long_read(&n->total_objects))
4470 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4471 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4473 struct slab_attribute {
4474 struct attribute attr;
4475 ssize_t (*show)(struct kmem_cache *s, char *buf);
4476 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4479 #define SLAB_ATTR_RO(_name) \
4480 static struct slab_attribute _name##_attr = \
4481 __ATTR(_name, 0400, _name##_show, NULL)
4483 #define SLAB_ATTR(_name) \
4484 static struct slab_attribute _name##_attr = \
4485 __ATTR(_name, 0600, _name##_show, _name##_store)
4487 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4489 return sprintf(buf, "%d\n", s->size);
4491 SLAB_ATTR_RO(slab_size);
4493 static ssize_t align_show(struct kmem_cache *s, char *buf)
4495 return sprintf(buf, "%d\n", s->align);
4497 SLAB_ATTR_RO(align);
4499 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4501 return sprintf(buf, "%d\n", s->object_size);
4503 SLAB_ATTR_RO(object_size);
4505 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4507 return sprintf(buf, "%d\n", oo_objects(s->oo));
4509 SLAB_ATTR_RO(objs_per_slab);
4511 static ssize_t order_store(struct kmem_cache *s,
4512 const char *buf, size_t length)
4514 unsigned long order;
4517 err = kstrtoul(buf, 10, &order);
4521 if (order > slub_max_order || order < slub_min_order)
4524 calculate_sizes(s, order);
4528 static ssize_t order_show(struct kmem_cache *s, char *buf)
4530 return sprintf(buf, "%d\n", oo_order(s->oo));
4534 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4536 return sprintf(buf, "%lu\n", s->min_partial);
4539 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4545 err = kstrtoul(buf, 10, &min);
4549 set_min_partial(s, min);
4552 SLAB_ATTR(min_partial);
4554 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4556 return sprintf(buf, "%u\n", s->cpu_partial);
4559 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4562 unsigned long objects;
4565 err = kstrtoul(buf, 10, &objects);
4568 if (objects && !kmem_cache_has_cpu_partial(s))
4571 s->cpu_partial = objects;
4575 SLAB_ATTR(cpu_partial);
4577 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4581 return sprintf(buf, "%pS\n", s->ctor);
4585 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4587 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4589 SLAB_ATTR_RO(aliases);
4591 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4593 return show_slab_objects(s, buf, SO_PARTIAL);
4595 SLAB_ATTR_RO(partial);
4597 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4599 return show_slab_objects(s, buf, SO_CPU);
4601 SLAB_ATTR_RO(cpu_slabs);
4603 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4605 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4607 SLAB_ATTR_RO(objects);
4609 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4611 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4613 SLAB_ATTR_RO(objects_partial);
4615 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4622 for_each_online_cpu(cpu) {
4623 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4626 pages += page->pages;
4627 objects += page->pobjects;
4631 len = sprintf(buf, "%d(%d)", objects, pages);
4634 for_each_online_cpu(cpu) {
4635 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4637 if (page && len < PAGE_SIZE - 20)
4638 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4639 page->pobjects, page->pages);
4642 return len + sprintf(buf + len, "\n");
4644 SLAB_ATTR_RO(slabs_cpu_partial);
4646 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4648 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4651 static ssize_t reclaim_account_store(struct kmem_cache *s,
4652 const char *buf, size_t length)
4654 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4656 s->flags |= SLAB_RECLAIM_ACCOUNT;
4659 SLAB_ATTR(reclaim_account);
4661 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4663 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4665 SLAB_ATTR_RO(hwcache_align);
4667 #ifdef CONFIG_ZONE_DMA
4668 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4670 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4672 SLAB_ATTR_RO(cache_dma);
4675 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4677 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4679 SLAB_ATTR_RO(destroy_by_rcu);
4681 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4683 return sprintf(buf, "%d\n", s->reserved);
4685 SLAB_ATTR_RO(reserved);
4687 #ifdef CONFIG_SLUB_DEBUG
4688 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4690 return show_slab_objects(s, buf, SO_ALL);
4692 SLAB_ATTR_RO(slabs);
4694 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4696 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4698 SLAB_ATTR_RO(total_objects);
4700 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4702 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4705 static ssize_t sanity_checks_store(struct kmem_cache *s,
4706 const char *buf, size_t length)
4708 s->flags &= ~SLAB_DEBUG_FREE;
4709 if (buf[0] == '1') {
4710 s->flags &= ~__CMPXCHG_DOUBLE;
4711 s->flags |= SLAB_DEBUG_FREE;
4715 SLAB_ATTR(sanity_checks);
4717 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4719 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4722 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4726 * Tracing a merged cache is going to give confusing results
4727 * as well as cause other issues like converting a mergeable
4728 * cache into an umergeable one.
4730 if (s->refcount > 1)
4733 s->flags &= ~SLAB_TRACE;
4734 if (buf[0] == '1') {
4735 s->flags &= ~__CMPXCHG_DOUBLE;
4736 s->flags |= SLAB_TRACE;
4742 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4744 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4747 static ssize_t red_zone_store(struct kmem_cache *s,
4748 const char *buf, size_t length)
4750 if (any_slab_objects(s))
4753 s->flags &= ~SLAB_RED_ZONE;
4754 if (buf[0] == '1') {
4755 s->flags &= ~__CMPXCHG_DOUBLE;
4756 s->flags |= SLAB_RED_ZONE;
4758 calculate_sizes(s, -1);
4761 SLAB_ATTR(red_zone);
4763 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4765 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4768 static ssize_t poison_store(struct kmem_cache *s,
4769 const char *buf, size_t length)
4771 if (any_slab_objects(s))
4774 s->flags &= ~SLAB_POISON;
4775 if (buf[0] == '1') {
4776 s->flags &= ~__CMPXCHG_DOUBLE;
4777 s->flags |= SLAB_POISON;
4779 calculate_sizes(s, -1);
4784 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4786 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4789 static ssize_t store_user_store(struct kmem_cache *s,
4790 const char *buf, size_t length)
4792 if (any_slab_objects(s))
4795 s->flags &= ~SLAB_STORE_USER;
4796 if (buf[0] == '1') {
4797 s->flags &= ~__CMPXCHG_DOUBLE;
4798 s->flags |= SLAB_STORE_USER;
4800 calculate_sizes(s, -1);
4803 SLAB_ATTR(store_user);
4805 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4810 static ssize_t validate_store(struct kmem_cache *s,
4811 const char *buf, size_t length)
4815 if (buf[0] == '1') {
4816 ret = validate_slab_cache(s);
4822 SLAB_ATTR(validate);
4824 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4826 if (!(s->flags & SLAB_STORE_USER))
4828 return list_locations(s, buf, TRACK_ALLOC);
4830 SLAB_ATTR_RO(alloc_calls);
4832 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4834 if (!(s->flags & SLAB_STORE_USER))
4836 return list_locations(s, buf, TRACK_FREE);
4838 SLAB_ATTR_RO(free_calls);
4839 #endif /* CONFIG_SLUB_DEBUG */
4841 #ifdef CONFIG_FAILSLAB
4842 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4847 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4850 if (s->refcount > 1)
4853 s->flags &= ~SLAB_FAILSLAB;
4855 s->flags |= SLAB_FAILSLAB;
4858 SLAB_ATTR(failslab);
4861 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4866 static ssize_t shrink_store(struct kmem_cache *s,
4867 const char *buf, size_t length)
4870 kmem_cache_shrink(s);
4878 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4880 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4883 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4884 const char *buf, size_t length)
4886 unsigned long ratio;
4889 err = kstrtoul(buf, 10, &ratio);
4894 s->remote_node_defrag_ratio = ratio * 10;
4898 SLAB_ATTR(remote_node_defrag_ratio);
4901 #ifdef CONFIG_SLUB_STATS
4902 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4904 unsigned long sum = 0;
4907 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4912 for_each_online_cpu(cpu) {
4913 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4919 len = sprintf(buf, "%lu", sum);
4922 for_each_online_cpu(cpu) {
4923 if (data[cpu] && len < PAGE_SIZE - 20)
4924 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4928 return len + sprintf(buf + len, "\n");
4931 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4935 for_each_online_cpu(cpu)
4936 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4939 #define STAT_ATTR(si, text) \
4940 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4942 return show_stat(s, buf, si); \
4944 static ssize_t text##_store(struct kmem_cache *s, \
4945 const char *buf, size_t length) \
4947 if (buf[0] != '0') \
4949 clear_stat(s, si); \
4954 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4955 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4956 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4957 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4958 STAT_ATTR(FREE_FROZEN, free_frozen);
4959 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4960 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4961 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4962 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4963 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4964 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4965 STAT_ATTR(FREE_SLAB, free_slab);
4966 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4967 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4968 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4969 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4970 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4971 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4972 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4973 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4974 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4975 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4976 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4977 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4978 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4979 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4982 static struct attribute *slab_attrs[] = {
4983 &slab_size_attr.attr,
4984 &object_size_attr.attr,
4985 &objs_per_slab_attr.attr,
4987 &min_partial_attr.attr,
4988 &cpu_partial_attr.attr,
4990 &objects_partial_attr.attr,
4992 &cpu_slabs_attr.attr,
4996 &hwcache_align_attr.attr,
4997 &reclaim_account_attr.attr,
4998 &destroy_by_rcu_attr.attr,
5000 &reserved_attr.attr,
5001 &slabs_cpu_partial_attr.attr,
5002 #ifdef CONFIG_SLUB_DEBUG
5003 &total_objects_attr.attr,
5005 &sanity_checks_attr.attr,
5007 &red_zone_attr.attr,
5009 &store_user_attr.attr,
5010 &validate_attr.attr,
5011 &alloc_calls_attr.attr,
5012 &free_calls_attr.attr,
5014 #ifdef CONFIG_ZONE_DMA
5015 &cache_dma_attr.attr,
5018 &remote_node_defrag_ratio_attr.attr,
5020 #ifdef CONFIG_SLUB_STATS
5021 &alloc_fastpath_attr.attr,
5022 &alloc_slowpath_attr.attr,
5023 &free_fastpath_attr.attr,
5024 &free_slowpath_attr.attr,
5025 &free_frozen_attr.attr,
5026 &free_add_partial_attr.attr,
5027 &free_remove_partial_attr.attr,
5028 &alloc_from_partial_attr.attr,
5029 &alloc_slab_attr.attr,
5030 &alloc_refill_attr.attr,
5031 &alloc_node_mismatch_attr.attr,
5032 &free_slab_attr.attr,
5033 &cpuslab_flush_attr.attr,
5034 &deactivate_full_attr.attr,
5035 &deactivate_empty_attr.attr,
5036 &deactivate_to_head_attr.attr,
5037 &deactivate_to_tail_attr.attr,
5038 &deactivate_remote_frees_attr.attr,
5039 &deactivate_bypass_attr.attr,
5040 &order_fallback_attr.attr,
5041 &cmpxchg_double_fail_attr.attr,
5042 &cmpxchg_double_cpu_fail_attr.attr,
5043 &cpu_partial_alloc_attr.attr,
5044 &cpu_partial_free_attr.attr,
5045 &cpu_partial_node_attr.attr,
5046 &cpu_partial_drain_attr.attr,
5048 #ifdef CONFIG_FAILSLAB
5049 &failslab_attr.attr,
5055 static struct attribute_group slab_attr_group = {
5056 .attrs = slab_attrs,
5059 static ssize_t slab_attr_show(struct kobject *kobj,
5060 struct attribute *attr,
5063 struct slab_attribute *attribute;
5064 struct kmem_cache *s;
5067 attribute = to_slab_attr(attr);
5070 if (!attribute->show)
5073 err = attribute->show(s, buf);
5078 static ssize_t slab_attr_store(struct kobject *kobj,
5079 struct attribute *attr,
5080 const char *buf, size_t len)
5082 struct slab_attribute *attribute;
5083 struct kmem_cache *s;
5086 attribute = to_slab_attr(attr);
5089 if (!attribute->store)
5092 err = attribute->store(s, buf, len);
5093 #ifdef CONFIG_MEMCG_KMEM
5094 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5095 struct kmem_cache *c;
5097 mutex_lock(&slab_mutex);
5098 if (s->max_attr_size < len)
5099 s->max_attr_size = len;
5102 * This is a best effort propagation, so this function's return
5103 * value will be determined by the parent cache only. This is
5104 * basically because not all attributes will have a well
5105 * defined semantics for rollbacks - most of the actions will
5106 * have permanent effects.
5108 * Returning the error value of any of the children that fail
5109 * is not 100 % defined, in the sense that users seeing the
5110 * error code won't be able to know anything about the state of
5113 * Only returning the error code for the parent cache at least
5114 * has well defined semantics. The cache being written to
5115 * directly either failed or succeeded, in which case we loop
5116 * through the descendants with best-effort propagation.
5118 for_each_memcg_cache(c, s)
5119 attribute->store(c, buf, len);
5120 mutex_unlock(&slab_mutex);
5126 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5128 #ifdef CONFIG_MEMCG_KMEM
5130 char *buffer = NULL;
5131 struct kmem_cache *root_cache;
5133 if (is_root_cache(s))
5136 root_cache = s->memcg_params.root_cache;
5139 * This mean this cache had no attribute written. Therefore, no point
5140 * in copying default values around
5142 if (!root_cache->max_attr_size)
5145 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5148 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5150 if (!attr || !attr->store || !attr->show)
5154 * It is really bad that we have to allocate here, so we will
5155 * do it only as a fallback. If we actually allocate, though,
5156 * we can just use the allocated buffer until the end.
5158 * Most of the slub attributes will tend to be very small in
5159 * size, but sysfs allows buffers up to a page, so they can
5160 * theoretically happen.
5164 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5167 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5168 if (WARN_ON(!buffer))
5173 attr->show(root_cache, buf);
5174 attr->store(s, buf, strlen(buf));
5178 free_page((unsigned long)buffer);
5182 static void kmem_cache_release(struct kobject *k)
5184 slab_kmem_cache_release(to_slab(k));
5187 static const struct sysfs_ops slab_sysfs_ops = {
5188 .show = slab_attr_show,
5189 .store = slab_attr_store,
5192 static struct kobj_type slab_ktype = {
5193 .sysfs_ops = &slab_sysfs_ops,
5194 .release = kmem_cache_release,
5197 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5199 struct kobj_type *ktype = get_ktype(kobj);
5201 if (ktype == &slab_ktype)
5206 static const struct kset_uevent_ops slab_uevent_ops = {
5207 .filter = uevent_filter,
5210 static struct kset *slab_kset;
5212 static inline struct kset *cache_kset(struct kmem_cache *s)
5214 #ifdef CONFIG_MEMCG_KMEM
5215 if (!is_root_cache(s))
5216 return s->memcg_params.root_cache->memcg_kset;
5221 #define ID_STR_LENGTH 64
5223 /* Create a unique string id for a slab cache:
5225 * Format :[flags-]size
5227 static char *create_unique_id(struct kmem_cache *s)
5229 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5236 * First flags affecting slabcache operations. We will only
5237 * get here for aliasable slabs so we do not need to support
5238 * too many flags. The flags here must cover all flags that
5239 * are matched during merging to guarantee that the id is
5242 if (s->flags & SLAB_CACHE_DMA)
5244 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5246 if (s->flags & SLAB_DEBUG_FREE)
5248 if (!(s->flags & SLAB_NOTRACK))
5252 p += sprintf(p, "%07d", s->size);
5254 BUG_ON(p > name + ID_STR_LENGTH - 1);
5258 static int sysfs_slab_add(struct kmem_cache *s)
5262 int unmergeable = slab_unmergeable(s);
5266 * Slabcache can never be merged so we can use the name proper.
5267 * This is typically the case for debug situations. In that
5268 * case we can catch duplicate names easily.
5270 sysfs_remove_link(&slab_kset->kobj, s->name);
5274 * Create a unique name for the slab as a target
5277 name = create_unique_id(s);
5280 s->kobj.kset = cache_kset(s);
5281 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5285 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5289 #ifdef CONFIG_MEMCG_KMEM
5290 if (is_root_cache(s)) {
5291 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5292 if (!s->memcg_kset) {
5299 kobject_uevent(&s->kobj, KOBJ_ADD);
5301 /* Setup first alias */
5302 sysfs_slab_alias(s, s->name);
5309 kobject_del(&s->kobj);
5313 void sysfs_slab_remove(struct kmem_cache *s)
5315 if (slab_state < FULL)
5317 * Sysfs has not been setup yet so no need to remove the
5322 #ifdef CONFIG_MEMCG_KMEM
5323 kset_unregister(s->memcg_kset);
5325 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5326 kobject_del(&s->kobj);
5327 kobject_put(&s->kobj);
5331 * Need to buffer aliases during bootup until sysfs becomes
5332 * available lest we lose that information.
5334 struct saved_alias {
5335 struct kmem_cache *s;
5337 struct saved_alias *next;
5340 static struct saved_alias *alias_list;
5342 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5344 struct saved_alias *al;
5346 if (slab_state == FULL) {
5348 * If we have a leftover link then remove it.
5350 sysfs_remove_link(&slab_kset->kobj, name);
5351 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5354 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5360 al->next = alias_list;
5365 static int __init slab_sysfs_init(void)
5367 struct kmem_cache *s;
5370 mutex_lock(&slab_mutex);
5372 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5374 mutex_unlock(&slab_mutex);
5375 pr_err("Cannot register slab subsystem.\n");
5381 list_for_each_entry(s, &slab_caches, list) {
5382 err = sysfs_slab_add(s);
5384 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5388 while (alias_list) {
5389 struct saved_alias *al = alias_list;
5391 alias_list = alias_list->next;
5392 err = sysfs_slab_alias(al->s, al->name);
5394 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5399 mutex_unlock(&slab_mutex);
5404 __initcall(slab_sysfs_init);
5405 #endif /* CONFIG_SYSFS */
5408 * The /proc/slabinfo ABI
5410 #ifdef CONFIG_SLABINFO
5411 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5413 unsigned long nr_slabs = 0;
5414 unsigned long nr_objs = 0;
5415 unsigned long nr_free = 0;
5417 struct kmem_cache_node *n;
5419 for_each_kmem_cache_node(s, node, n) {
5420 nr_slabs += node_nr_slabs(n);
5421 nr_objs += node_nr_objs(n);
5422 nr_free += count_partial(n, count_free);
5425 sinfo->active_objs = nr_objs - nr_free;
5426 sinfo->num_objs = nr_objs;
5427 sinfo->active_slabs = nr_slabs;
5428 sinfo->num_slabs = nr_slabs;
5429 sinfo->objects_per_slab = oo_objects(s->oo);
5430 sinfo->cache_order = oo_order(s->oo);
5433 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5437 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5438 size_t count, loff_t *ppos)
5442 #endif /* CONFIG_SLABINFO */