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 (node == NUMA_NO_NODE)
1332 page = alloc_pages(flags, order);
1334 page = __alloc_pages_node(node, flags, order);
1336 if (page && memcg_charge_slab(page, flags, order, s)) {
1337 __free_pages(page, order);
1344 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1347 struct kmem_cache_order_objects oo = s->oo;
1352 flags &= gfp_allowed_mask;
1354 if (flags & __GFP_WAIT)
1357 flags |= s->allocflags;
1360 * Let the initial higher-order allocation fail under memory pressure
1361 * so we fall-back to the minimum order allocation.
1363 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1364 if ((alloc_gfp & __GFP_WAIT) && oo_order(oo) > oo_order(s->min))
1365 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_WAIT;
1367 page = alloc_slab_page(s, alloc_gfp, node, oo);
1368 if (unlikely(!page)) {
1372 * Allocation may have failed due to fragmentation.
1373 * Try a lower order alloc if possible
1375 page = alloc_slab_page(s, alloc_gfp, node, oo);
1376 if (unlikely(!page))
1378 stat(s, ORDER_FALLBACK);
1381 if (kmemcheck_enabled &&
1382 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1383 int pages = 1 << oo_order(oo);
1385 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1388 * Objects from caches that have a constructor don't get
1389 * cleared when they're allocated, so we need to do it here.
1392 kmemcheck_mark_uninitialized_pages(page, pages);
1394 kmemcheck_mark_unallocated_pages(page, pages);
1397 page->objects = oo_objects(oo);
1399 order = compound_order(page);
1400 page->slab_cache = s;
1401 __SetPageSlab(page);
1402 if (page_is_pfmemalloc(page))
1403 SetPageSlabPfmemalloc(page);
1405 start = page_address(page);
1407 if (unlikely(s->flags & SLAB_POISON))
1408 memset(start, POISON_INUSE, PAGE_SIZE << order);
1410 kasan_poison_slab(page);
1412 for_each_object_idx(p, idx, s, start, page->objects) {
1413 setup_object(s, page, p);
1414 if (likely(idx < page->objects))
1415 set_freepointer(s, p, p + s->size);
1417 set_freepointer(s, p, NULL);
1420 page->freelist = start;
1421 page->inuse = page->objects;
1425 if (flags & __GFP_WAIT)
1426 local_irq_disable();
1430 mod_zone_page_state(page_zone(page),
1431 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1432 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1435 inc_slabs_node(s, page_to_nid(page), page->objects);
1440 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1442 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1443 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1447 return allocate_slab(s,
1448 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1451 static void __free_slab(struct kmem_cache *s, struct page *page)
1453 int order = compound_order(page);
1454 int pages = 1 << order;
1456 if (kmem_cache_debug(s)) {
1459 slab_pad_check(s, page);
1460 for_each_object(p, s, page_address(page),
1462 check_object(s, page, p, SLUB_RED_INACTIVE);
1465 kmemcheck_free_shadow(page, compound_order(page));
1467 mod_zone_page_state(page_zone(page),
1468 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1469 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1472 __ClearPageSlabPfmemalloc(page);
1473 __ClearPageSlab(page);
1475 page_mapcount_reset(page);
1476 if (current->reclaim_state)
1477 current->reclaim_state->reclaimed_slab += pages;
1478 __free_kmem_pages(page, order);
1481 #define need_reserve_slab_rcu \
1482 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1484 static void rcu_free_slab(struct rcu_head *h)
1488 if (need_reserve_slab_rcu)
1489 page = virt_to_head_page(h);
1491 page = container_of((struct list_head *)h, struct page, lru);
1493 __free_slab(page->slab_cache, page);
1496 static void free_slab(struct kmem_cache *s, struct page *page)
1498 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1499 struct rcu_head *head;
1501 if (need_reserve_slab_rcu) {
1502 int order = compound_order(page);
1503 int offset = (PAGE_SIZE << order) - s->reserved;
1505 VM_BUG_ON(s->reserved != sizeof(*head));
1506 head = page_address(page) + offset;
1509 * RCU free overloads the RCU head over the LRU
1511 head = (void *)&page->lru;
1514 call_rcu(head, rcu_free_slab);
1516 __free_slab(s, page);
1519 static void discard_slab(struct kmem_cache *s, struct page *page)
1521 dec_slabs_node(s, page_to_nid(page), page->objects);
1526 * Management of partially allocated slabs.
1529 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1532 if (tail == DEACTIVATE_TO_TAIL)
1533 list_add_tail(&page->lru, &n->partial);
1535 list_add(&page->lru, &n->partial);
1538 static inline void add_partial(struct kmem_cache_node *n,
1539 struct page *page, int tail)
1541 lockdep_assert_held(&n->list_lock);
1542 __add_partial(n, page, tail);
1546 __remove_partial(struct kmem_cache_node *n, struct page *page)
1548 list_del(&page->lru);
1552 static inline void remove_partial(struct kmem_cache_node *n,
1555 lockdep_assert_held(&n->list_lock);
1556 __remove_partial(n, page);
1560 * Remove slab from the partial list, freeze it and
1561 * return the pointer to the freelist.
1563 * Returns a list of objects or NULL if it fails.
1565 static inline void *acquire_slab(struct kmem_cache *s,
1566 struct kmem_cache_node *n, struct page *page,
1567 int mode, int *objects)
1570 unsigned long counters;
1573 lockdep_assert_held(&n->list_lock);
1576 * Zap the freelist and set the frozen bit.
1577 * The old freelist is the list of objects for the
1578 * per cpu allocation list.
1580 freelist = page->freelist;
1581 counters = page->counters;
1582 new.counters = counters;
1583 *objects = new.objects - new.inuse;
1585 new.inuse = page->objects;
1586 new.freelist = NULL;
1588 new.freelist = freelist;
1591 VM_BUG_ON(new.frozen);
1594 if (!__cmpxchg_double_slab(s, page,
1596 new.freelist, new.counters,
1600 remove_partial(n, page);
1605 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1606 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1609 * Try to allocate a partial slab from a specific node.
1611 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1612 struct kmem_cache_cpu *c, gfp_t flags)
1614 struct page *page, *page2;
1615 void *object = NULL;
1620 * Racy check. If we mistakenly see no partial slabs then we
1621 * just allocate an empty slab. If we mistakenly try to get a
1622 * partial slab and there is none available then get_partials()
1625 if (!n || !n->nr_partial)
1628 spin_lock(&n->list_lock);
1629 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1632 if (!pfmemalloc_match(page, flags))
1635 t = acquire_slab(s, n, page, object == NULL, &objects);
1639 available += objects;
1642 stat(s, ALLOC_FROM_PARTIAL);
1645 put_cpu_partial(s, page, 0);
1646 stat(s, CPU_PARTIAL_NODE);
1648 if (!kmem_cache_has_cpu_partial(s)
1649 || available > s->cpu_partial / 2)
1653 spin_unlock(&n->list_lock);
1658 * Get a page from somewhere. Search in increasing NUMA distances.
1660 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1661 struct kmem_cache_cpu *c)
1664 struct zonelist *zonelist;
1667 enum zone_type high_zoneidx = gfp_zone(flags);
1669 unsigned int cpuset_mems_cookie;
1672 * The defrag ratio allows a configuration of the tradeoffs between
1673 * inter node defragmentation and node local allocations. A lower
1674 * defrag_ratio increases the tendency to do local allocations
1675 * instead of attempting to obtain partial slabs from other nodes.
1677 * If the defrag_ratio is set to 0 then kmalloc() always
1678 * returns node local objects. If the ratio is higher then kmalloc()
1679 * may return off node objects because partial slabs are obtained
1680 * from other nodes and filled up.
1682 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1683 * defrag_ratio = 1000) then every (well almost) allocation will
1684 * first attempt to defrag slab caches on other nodes. This means
1685 * scanning over all nodes to look for partial slabs which may be
1686 * expensive if we do it every time we are trying to find a slab
1687 * with available objects.
1689 if (!s->remote_node_defrag_ratio ||
1690 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1694 cpuset_mems_cookie = read_mems_allowed_begin();
1695 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1696 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1697 struct kmem_cache_node *n;
1699 n = get_node(s, zone_to_nid(zone));
1701 if (n && cpuset_zone_allowed(zone, flags) &&
1702 n->nr_partial > s->min_partial) {
1703 object = get_partial_node(s, n, c, flags);
1706 * Don't check read_mems_allowed_retry()
1707 * here - if mems_allowed was updated in
1708 * parallel, that was a harmless race
1709 * between allocation and the cpuset
1716 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1722 * Get a partial page, lock it and return it.
1724 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1725 struct kmem_cache_cpu *c)
1728 int searchnode = node;
1730 if (node == NUMA_NO_NODE)
1731 searchnode = numa_mem_id();
1732 else if (!node_present_pages(node))
1733 searchnode = node_to_mem_node(node);
1735 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1736 if (object || node != NUMA_NO_NODE)
1739 return get_any_partial(s, flags, c);
1742 #ifdef CONFIG_PREEMPT
1744 * Calculate the next globally unique transaction for disambiguiation
1745 * during cmpxchg. The transactions start with the cpu number and are then
1746 * incremented by CONFIG_NR_CPUS.
1748 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1751 * No preemption supported therefore also no need to check for
1757 static inline unsigned long next_tid(unsigned long tid)
1759 return tid + TID_STEP;
1762 static inline unsigned int tid_to_cpu(unsigned long tid)
1764 return tid % TID_STEP;
1767 static inline unsigned long tid_to_event(unsigned long tid)
1769 return tid / TID_STEP;
1772 static inline unsigned int init_tid(int cpu)
1777 static inline void note_cmpxchg_failure(const char *n,
1778 const struct kmem_cache *s, unsigned long tid)
1780 #ifdef SLUB_DEBUG_CMPXCHG
1781 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1783 pr_info("%s %s: cmpxchg redo ", n, s->name);
1785 #ifdef CONFIG_PREEMPT
1786 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1787 pr_warn("due to cpu change %d -> %d\n",
1788 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1791 if (tid_to_event(tid) != tid_to_event(actual_tid))
1792 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1793 tid_to_event(tid), tid_to_event(actual_tid));
1795 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1796 actual_tid, tid, next_tid(tid));
1798 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1801 static void init_kmem_cache_cpus(struct kmem_cache *s)
1805 for_each_possible_cpu(cpu)
1806 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1810 * Remove the cpu slab
1812 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1815 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1816 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1818 enum slab_modes l = M_NONE, m = M_NONE;
1820 int tail = DEACTIVATE_TO_HEAD;
1824 if (page->freelist) {
1825 stat(s, DEACTIVATE_REMOTE_FREES);
1826 tail = DEACTIVATE_TO_TAIL;
1830 * Stage one: Free all available per cpu objects back
1831 * to the page freelist while it is still frozen. Leave the
1834 * There is no need to take the list->lock because the page
1837 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1839 unsigned long counters;
1842 prior = page->freelist;
1843 counters = page->counters;
1844 set_freepointer(s, freelist, prior);
1845 new.counters = counters;
1847 VM_BUG_ON(!new.frozen);
1849 } while (!__cmpxchg_double_slab(s, page,
1851 freelist, new.counters,
1852 "drain percpu freelist"));
1854 freelist = nextfree;
1858 * Stage two: Ensure that the page is unfrozen while the
1859 * list presence reflects the actual number of objects
1862 * We setup the list membership and then perform a cmpxchg
1863 * with the count. If there is a mismatch then the page
1864 * is not unfrozen but the page is on the wrong list.
1866 * Then we restart the process which may have to remove
1867 * the page from the list that we just put it on again
1868 * because the number of objects in the slab may have
1873 old.freelist = page->freelist;
1874 old.counters = page->counters;
1875 VM_BUG_ON(!old.frozen);
1877 /* Determine target state of the slab */
1878 new.counters = old.counters;
1881 set_freepointer(s, freelist, old.freelist);
1882 new.freelist = freelist;
1884 new.freelist = old.freelist;
1888 if (!new.inuse && n->nr_partial >= s->min_partial)
1890 else if (new.freelist) {
1895 * Taking the spinlock removes the possiblity
1896 * that acquire_slab() will see a slab page that
1899 spin_lock(&n->list_lock);
1903 if (kmem_cache_debug(s) && !lock) {
1906 * This also ensures that the scanning of full
1907 * slabs from diagnostic functions will not see
1910 spin_lock(&n->list_lock);
1918 remove_partial(n, page);
1920 else if (l == M_FULL)
1922 remove_full(s, n, page);
1924 if (m == M_PARTIAL) {
1926 add_partial(n, page, tail);
1929 } else if (m == M_FULL) {
1931 stat(s, DEACTIVATE_FULL);
1932 add_full(s, n, page);
1938 if (!__cmpxchg_double_slab(s, page,
1939 old.freelist, old.counters,
1940 new.freelist, new.counters,
1945 spin_unlock(&n->list_lock);
1948 stat(s, DEACTIVATE_EMPTY);
1949 discard_slab(s, page);
1955 * Unfreeze all the cpu partial slabs.
1957 * This function must be called with interrupts disabled
1958 * for the cpu using c (or some other guarantee must be there
1959 * to guarantee no concurrent accesses).
1961 static void unfreeze_partials(struct kmem_cache *s,
1962 struct kmem_cache_cpu *c)
1964 #ifdef CONFIG_SLUB_CPU_PARTIAL
1965 struct kmem_cache_node *n = NULL, *n2 = NULL;
1966 struct page *page, *discard_page = NULL;
1968 while ((page = c->partial)) {
1972 c->partial = page->next;
1974 n2 = get_node(s, page_to_nid(page));
1977 spin_unlock(&n->list_lock);
1980 spin_lock(&n->list_lock);
1985 old.freelist = page->freelist;
1986 old.counters = page->counters;
1987 VM_BUG_ON(!old.frozen);
1989 new.counters = old.counters;
1990 new.freelist = old.freelist;
1994 } while (!__cmpxchg_double_slab(s, page,
1995 old.freelist, old.counters,
1996 new.freelist, new.counters,
1997 "unfreezing slab"));
1999 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2000 page->next = discard_page;
2001 discard_page = page;
2003 add_partial(n, page, DEACTIVATE_TO_TAIL);
2004 stat(s, FREE_ADD_PARTIAL);
2009 spin_unlock(&n->list_lock);
2011 while (discard_page) {
2012 page = discard_page;
2013 discard_page = discard_page->next;
2015 stat(s, DEACTIVATE_EMPTY);
2016 discard_slab(s, page);
2023 * Put a page that was just frozen (in __slab_free) into a partial page
2024 * slot if available. This is done without interrupts disabled and without
2025 * preemption disabled. The cmpxchg is racy and may put the partial page
2026 * onto a random cpus partial slot.
2028 * If we did not find a slot then simply move all the partials to the
2029 * per node partial list.
2031 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2033 #ifdef CONFIG_SLUB_CPU_PARTIAL
2034 struct page *oldpage;
2042 oldpage = this_cpu_read(s->cpu_slab->partial);
2045 pobjects = oldpage->pobjects;
2046 pages = oldpage->pages;
2047 if (drain && pobjects > s->cpu_partial) {
2048 unsigned long flags;
2050 * partial array is full. Move the existing
2051 * set to the per node partial list.
2053 local_irq_save(flags);
2054 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2055 local_irq_restore(flags);
2059 stat(s, CPU_PARTIAL_DRAIN);
2064 pobjects += page->objects - page->inuse;
2066 page->pages = pages;
2067 page->pobjects = pobjects;
2068 page->next = oldpage;
2070 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2072 if (unlikely(!s->cpu_partial)) {
2073 unsigned long flags;
2075 local_irq_save(flags);
2076 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2077 local_irq_restore(flags);
2083 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2085 stat(s, CPUSLAB_FLUSH);
2086 deactivate_slab(s, c->page, c->freelist);
2088 c->tid = next_tid(c->tid);
2096 * Called from IPI handler with interrupts disabled.
2098 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2100 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2106 unfreeze_partials(s, c);
2110 static void flush_cpu_slab(void *d)
2112 struct kmem_cache *s = d;
2114 __flush_cpu_slab(s, smp_processor_id());
2117 static bool has_cpu_slab(int cpu, void *info)
2119 struct kmem_cache *s = info;
2120 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2122 return c->page || c->partial;
2125 static void flush_all(struct kmem_cache *s)
2127 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2131 * Check if the objects in a per cpu structure fit numa
2132 * locality expectations.
2134 static inline int node_match(struct page *page, int node)
2137 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2143 #ifdef CONFIG_SLUB_DEBUG
2144 static int count_free(struct page *page)
2146 return page->objects - page->inuse;
2149 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2151 return atomic_long_read(&n->total_objects);
2153 #endif /* CONFIG_SLUB_DEBUG */
2155 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2156 static unsigned long count_partial(struct kmem_cache_node *n,
2157 int (*get_count)(struct page *))
2159 unsigned long flags;
2160 unsigned long x = 0;
2163 spin_lock_irqsave(&n->list_lock, flags);
2164 list_for_each_entry(page, &n->partial, lru)
2165 x += get_count(page);
2166 spin_unlock_irqrestore(&n->list_lock, flags);
2169 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2171 static noinline void
2172 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2174 #ifdef CONFIG_SLUB_DEBUG
2175 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2176 DEFAULT_RATELIMIT_BURST);
2178 struct kmem_cache_node *n;
2180 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2183 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2185 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2186 s->name, s->object_size, s->size, oo_order(s->oo),
2189 if (oo_order(s->min) > get_order(s->object_size))
2190 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2193 for_each_kmem_cache_node(s, node, n) {
2194 unsigned long nr_slabs;
2195 unsigned long nr_objs;
2196 unsigned long nr_free;
2198 nr_free = count_partial(n, count_free);
2199 nr_slabs = node_nr_slabs(n);
2200 nr_objs = node_nr_objs(n);
2202 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2203 node, nr_slabs, nr_objs, nr_free);
2208 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2209 int node, struct kmem_cache_cpu **pc)
2212 struct kmem_cache_cpu *c = *pc;
2215 freelist = get_partial(s, flags, node, c);
2220 page = new_slab(s, flags, node);
2222 c = raw_cpu_ptr(s->cpu_slab);
2227 * No other reference to the page yet so we can
2228 * muck around with it freely without cmpxchg
2230 freelist = page->freelist;
2231 page->freelist = NULL;
2233 stat(s, ALLOC_SLAB);
2242 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2244 if (unlikely(PageSlabPfmemalloc(page)))
2245 return gfp_pfmemalloc_allowed(gfpflags);
2251 * Check the page->freelist of a page and either transfer the freelist to the
2252 * per cpu freelist or deactivate the page.
2254 * The page is still frozen if the return value is not NULL.
2256 * If this function returns NULL then the page has been unfrozen.
2258 * This function must be called with interrupt disabled.
2260 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2263 unsigned long counters;
2267 freelist = page->freelist;
2268 counters = page->counters;
2270 new.counters = counters;
2271 VM_BUG_ON(!new.frozen);
2273 new.inuse = page->objects;
2274 new.frozen = freelist != NULL;
2276 } while (!__cmpxchg_double_slab(s, page,
2285 * Slow path. The lockless freelist is empty or we need to perform
2288 * Processing is still very fast if new objects have been freed to the
2289 * regular freelist. In that case we simply take over the regular freelist
2290 * as the lockless freelist and zap the regular freelist.
2292 * If that is not working then we fall back to the partial lists. We take the
2293 * first element of the freelist as the object to allocate now and move the
2294 * rest of the freelist to the lockless freelist.
2296 * And if we were unable to get a new slab from the partial slab lists then
2297 * we need to allocate a new slab. This is the slowest path since it involves
2298 * a call to the page allocator and the setup of a new slab.
2300 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2301 unsigned long addr, struct kmem_cache_cpu *c)
2305 unsigned long flags;
2307 local_irq_save(flags);
2308 #ifdef CONFIG_PREEMPT
2310 * We may have been preempted and rescheduled on a different
2311 * cpu before disabling interrupts. Need to reload cpu area
2314 c = this_cpu_ptr(s->cpu_slab);
2322 if (unlikely(!node_match(page, node))) {
2323 int searchnode = node;
2325 if (node != NUMA_NO_NODE && !node_present_pages(node))
2326 searchnode = node_to_mem_node(node);
2328 if (unlikely(!node_match(page, searchnode))) {
2329 stat(s, ALLOC_NODE_MISMATCH);
2330 deactivate_slab(s, page, c->freelist);
2338 * By rights, we should be searching for a slab page that was
2339 * PFMEMALLOC but right now, we are losing the pfmemalloc
2340 * information when the page leaves the per-cpu allocator
2342 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2343 deactivate_slab(s, page, c->freelist);
2349 /* must check again c->freelist in case of cpu migration or IRQ */
2350 freelist = c->freelist;
2354 freelist = get_freelist(s, page);
2358 stat(s, DEACTIVATE_BYPASS);
2362 stat(s, ALLOC_REFILL);
2366 * freelist is pointing to the list of objects to be used.
2367 * page is pointing to the page from which the objects are obtained.
2368 * That page must be frozen for per cpu allocations to work.
2370 VM_BUG_ON(!c->page->frozen);
2371 c->freelist = get_freepointer(s, freelist);
2372 c->tid = next_tid(c->tid);
2373 local_irq_restore(flags);
2379 page = c->page = c->partial;
2380 c->partial = page->next;
2381 stat(s, CPU_PARTIAL_ALLOC);
2386 freelist = new_slab_objects(s, gfpflags, node, &c);
2388 if (unlikely(!freelist)) {
2389 slab_out_of_memory(s, gfpflags, node);
2390 local_irq_restore(flags);
2395 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2398 /* Only entered in the debug case */
2399 if (kmem_cache_debug(s) &&
2400 !alloc_debug_processing(s, page, freelist, addr))
2401 goto new_slab; /* Slab failed checks. Next slab needed */
2403 deactivate_slab(s, page, get_freepointer(s, freelist));
2406 local_irq_restore(flags);
2411 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2412 * have the fastpath folded into their functions. So no function call
2413 * overhead for requests that can be satisfied on the fastpath.
2415 * The fastpath works by first checking if the lockless freelist can be used.
2416 * If not then __slab_alloc is called for slow processing.
2418 * Otherwise we can simply pick the next object from the lockless free list.
2420 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2421 gfp_t gfpflags, int node, unsigned long addr)
2424 struct kmem_cache_cpu *c;
2428 s = slab_pre_alloc_hook(s, gfpflags);
2433 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2434 * enabled. We may switch back and forth between cpus while
2435 * reading from one cpu area. That does not matter as long
2436 * as we end up on the original cpu again when doing the cmpxchg.
2438 * We should guarantee that tid and kmem_cache are retrieved on
2439 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2440 * to check if it is matched or not.
2443 tid = this_cpu_read(s->cpu_slab->tid);
2444 c = raw_cpu_ptr(s->cpu_slab);
2445 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2446 unlikely(tid != READ_ONCE(c->tid)));
2449 * Irqless object alloc/free algorithm used here depends on sequence
2450 * of fetching cpu_slab's data. tid should be fetched before anything
2451 * on c to guarantee that object and page associated with previous tid
2452 * won't be used with current tid. If we fetch tid first, object and
2453 * page could be one associated with next tid and our alloc/free
2454 * request will be failed. In this case, we will retry. So, no problem.
2459 * The transaction ids are globally unique per cpu and per operation on
2460 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2461 * occurs on the right processor and that there was no operation on the
2462 * linked list in between.
2465 object = c->freelist;
2467 if (unlikely(!object || !node_match(page, node))) {
2468 object = __slab_alloc(s, gfpflags, node, addr, c);
2469 stat(s, ALLOC_SLOWPATH);
2471 void *next_object = get_freepointer_safe(s, object);
2474 * The cmpxchg will only match if there was no additional
2475 * operation and if we are on the right processor.
2477 * The cmpxchg does the following atomically (without lock
2479 * 1. Relocate first pointer to the current per cpu area.
2480 * 2. Verify that tid and freelist have not been changed
2481 * 3. If they were not changed replace tid and freelist
2483 * Since this is without lock semantics the protection is only
2484 * against code executing on this cpu *not* from access by
2487 if (unlikely(!this_cpu_cmpxchg_double(
2488 s->cpu_slab->freelist, s->cpu_slab->tid,
2490 next_object, next_tid(tid)))) {
2492 note_cmpxchg_failure("slab_alloc", s, tid);
2495 prefetch_freepointer(s, next_object);
2496 stat(s, ALLOC_FASTPATH);
2499 if (unlikely(gfpflags & __GFP_ZERO) && object)
2500 memset(object, 0, s->object_size);
2502 slab_post_alloc_hook(s, gfpflags, object);
2507 static __always_inline void *slab_alloc(struct kmem_cache *s,
2508 gfp_t gfpflags, unsigned long addr)
2510 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2513 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2515 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2517 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2522 EXPORT_SYMBOL(kmem_cache_alloc);
2524 #ifdef CONFIG_TRACING
2525 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2527 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2528 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2529 kasan_kmalloc(s, ret, size);
2532 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2536 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2538 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2540 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2541 s->object_size, s->size, gfpflags, node);
2545 EXPORT_SYMBOL(kmem_cache_alloc_node);
2547 #ifdef CONFIG_TRACING
2548 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2550 int node, size_t size)
2552 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2554 trace_kmalloc_node(_RET_IP_, ret,
2555 size, s->size, gfpflags, node);
2557 kasan_kmalloc(s, ret, size);
2560 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2565 * Slow path handling. This may still be called frequently since objects
2566 * have a longer lifetime than the cpu slabs in most processing loads.
2568 * So we still attempt to reduce cache line usage. Just take the slab
2569 * lock and free the item. If there is no additional partial page
2570 * handling required then we can return immediately.
2572 static void __slab_free(struct kmem_cache *s, struct page *page,
2573 void *x, unsigned long addr)
2576 void **object = (void *)x;
2579 unsigned long counters;
2580 struct kmem_cache_node *n = NULL;
2581 unsigned long uninitialized_var(flags);
2583 stat(s, FREE_SLOWPATH);
2585 if (kmem_cache_debug(s) &&
2586 !(n = free_debug_processing(s, page, x, addr, &flags)))
2591 spin_unlock_irqrestore(&n->list_lock, flags);
2594 prior = page->freelist;
2595 counters = page->counters;
2596 set_freepointer(s, object, prior);
2597 new.counters = counters;
2598 was_frozen = new.frozen;
2600 if ((!new.inuse || !prior) && !was_frozen) {
2602 if (kmem_cache_has_cpu_partial(s) && !prior) {
2605 * Slab was on no list before and will be
2607 * We can defer the list move and instead
2612 } else { /* Needs to be taken off a list */
2614 n = get_node(s, page_to_nid(page));
2616 * Speculatively acquire the list_lock.
2617 * If the cmpxchg does not succeed then we may
2618 * drop the list_lock without any processing.
2620 * Otherwise the list_lock will synchronize with
2621 * other processors updating the list of slabs.
2623 spin_lock_irqsave(&n->list_lock, flags);
2628 } while (!cmpxchg_double_slab(s, page,
2630 object, new.counters,
2636 * If we just froze the page then put it onto the
2637 * per cpu partial list.
2639 if (new.frozen && !was_frozen) {
2640 put_cpu_partial(s, page, 1);
2641 stat(s, CPU_PARTIAL_FREE);
2644 * The list lock was not taken therefore no list
2645 * activity can be necessary.
2648 stat(s, FREE_FROZEN);
2652 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2656 * Objects left in the slab. If it was not on the partial list before
2659 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2660 if (kmem_cache_debug(s))
2661 remove_full(s, n, page);
2662 add_partial(n, page, DEACTIVATE_TO_TAIL);
2663 stat(s, FREE_ADD_PARTIAL);
2665 spin_unlock_irqrestore(&n->list_lock, flags);
2671 * Slab on the partial list.
2673 remove_partial(n, page);
2674 stat(s, FREE_REMOVE_PARTIAL);
2676 /* Slab must be on the full list */
2677 remove_full(s, n, page);
2680 spin_unlock_irqrestore(&n->list_lock, flags);
2682 discard_slab(s, page);
2686 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2687 * can perform fastpath freeing without additional function calls.
2689 * The fastpath is only possible if we are freeing to the current cpu slab
2690 * of this processor. This typically the case if we have just allocated
2693 * If fastpath is not possible then fall back to __slab_free where we deal
2694 * with all sorts of special processing.
2696 static __always_inline void slab_free(struct kmem_cache *s,
2697 struct page *page, void *x, unsigned long addr)
2699 void **object = (void *)x;
2700 struct kmem_cache_cpu *c;
2703 slab_free_hook(s, x);
2707 * Determine the currently cpus per cpu slab.
2708 * The cpu may change afterward. However that does not matter since
2709 * data is retrieved via this pointer. If we are on the same cpu
2710 * during the cmpxchg then the free will succeed.
2713 tid = this_cpu_read(s->cpu_slab->tid);
2714 c = raw_cpu_ptr(s->cpu_slab);
2715 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2716 unlikely(tid != READ_ONCE(c->tid)));
2718 /* Same with comment on barrier() in slab_alloc_node() */
2721 if (likely(page == c->page)) {
2722 set_freepointer(s, object, c->freelist);
2724 if (unlikely(!this_cpu_cmpxchg_double(
2725 s->cpu_slab->freelist, s->cpu_slab->tid,
2727 object, next_tid(tid)))) {
2729 note_cmpxchg_failure("slab_free", s, tid);
2732 stat(s, FREE_FASTPATH);
2734 __slab_free(s, page, x, addr);
2738 void kmem_cache_free(struct kmem_cache *s, void *x)
2740 s = cache_from_obj(s, x);
2743 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2744 trace_kmem_cache_free(_RET_IP_, x);
2746 EXPORT_SYMBOL(kmem_cache_free);
2748 /* Note that interrupts must be enabled when calling this function. */
2749 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2751 struct kmem_cache_cpu *c;
2755 local_irq_disable();
2756 c = this_cpu_ptr(s->cpu_slab);
2758 for (i = 0; i < size; i++) {
2759 void *object = p[i];
2762 /* kmem cache debug support */
2763 s = cache_from_obj(s, object);
2766 slab_free_hook(s, object);
2768 page = virt_to_head_page(object);
2770 if (c->page == page) {
2771 /* Fastpath: local CPU free */
2772 set_freepointer(s, object, c->freelist);
2773 c->freelist = object;
2775 c->tid = next_tid(c->tid);
2777 /* Slowpath: overhead locked cmpxchg_double_slab */
2778 __slab_free(s, page, object, _RET_IP_);
2779 local_irq_disable();
2780 c = this_cpu_ptr(s->cpu_slab);
2784 c->tid = next_tid(c->tid);
2787 EXPORT_SYMBOL(kmem_cache_free_bulk);
2789 /* Note that interrupts must be enabled when calling this function. */
2790 bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2793 struct kmem_cache_cpu *c;
2797 * Drain objects in the per cpu slab, while disabling local
2798 * IRQs, which protects against PREEMPT and interrupts
2799 * handlers invoking normal fastpath.
2801 local_irq_disable();
2802 c = this_cpu_ptr(s->cpu_slab);
2804 for (i = 0; i < size; i++) {
2805 void *object = c->freelist;
2807 if (unlikely(!object)) {
2810 * Invoking slow path likely have side-effect
2811 * of re-populating per CPU c->freelist
2813 p[i] = __slab_alloc(s, flags, NUMA_NO_NODE,
2815 if (unlikely(!p[i])) {
2816 __kmem_cache_free_bulk(s, i, p);
2819 local_irq_disable();
2820 c = this_cpu_ptr(s->cpu_slab);
2821 continue; /* goto for-loop */
2824 /* kmem_cache debug support */
2825 s = slab_pre_alloc_hook(s, flags);
2827 __kmem_cache_free_bulk(s, i, p);
2828 c->tid = next_tid(c->tid);
2833 c->freelist = get_freepointer(s, object);
2836 /* kmem_cache debug support */
2837 slab_post_alloc_hook(s, flags, object);
2839 c->tid = next_tid(c->tid);
2842 /* Clear memory outside IRQ disabled fastpath loop */
2843 if (unlikely(flags & __GFP_ZERO)) {
2846 for (j = 0; j < i; j++)
2847 memset(p[j], 0, s->object_size);
2852 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2856 * Object placement in a slab is made very easy because we always start at
2857 * offset 0. If we tune the size of the object to the alignment then we can
2858 * get the required alignment by putting one properly sized object after
2861 * Notice that the allocation order determines the sizes of the per cpu
2862 * caches. Each processor has always one slab available for allocations.
2863 * Increasing the allocation order reduces the number of times that slabs
2864 * must be moved on and off the partial lists and is therefore a factor in
2869 * Mininum / Maximum order of slab pages. This influences locking overhead
2870 * and slab fragmentation. A higher order reduces the number of partial slabs
2871 * and increases the number of allocations possible without having to
2872 * take the list_lock.
2874 static int slub_min_order;
2875 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2876 static int slub_min_objects;
2879 * Calculate the order of allocation given an slab object size.
2881 * The order of allocation has significant impact on performance and other
2882 * system components. Generally order 0 allocations should be preferred since
2883 * order 0 does not cause fragmentation in the page allocator. Larger objects
2884 * be problematic to put into order 0 slabs because there may be too much
2885 * unused space left. We go to a higher order if more than 1/16th of the slab
2888 * In order to reach satisfactory performance we must ensure that a minimum
2889 * number of objects is in one slab. Otherwise we may generate too much
2890 * activity on the partial lists which requires taking the list_lock. This is
2891 * less a concern for large slabs though which are rarely used.
2893 * slub_max_order specifies the order where we begin to stop considering the
2894 * number of objects in a slab as critical. If we reach slub_max_order then
2895 * we try to keep the page order as low as possible. So we accept more waste
2896 * of space in favor of a small page order.
2898 * Higher order allocations also allow the placement of more objects in a
2899 * slab and thereby reduce object handling overhead. If the user has
2900 * requested a higher mininum order then we start with that one instead of
2901 * the smallest order which will fit the object.
2903 static inline int slab_order(int size, int min_objects,
2904 int max_order, int fract_leftover, int reserved)
2908 int min_order = slub_min_order;
2910 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2911 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2913 for (order = max(min_order, get_order(min_objects * size + reserved));
2914 order <= max_order; order++) {
2916 unsigned long slab_size = PAGE_SIZE << order;
2918 rem = (slab_size - reserved) % size;
2920 if (rem <= slab_size / fract_leftover)
2927 static inline int calculate_order(int size, int reserved)
2935 * Attempt to find best configuration for a slab. This
2936 * works by first attempting to generate a layout with
2937 * the best configuration and backing off gradually.
2939 * First we increase the acceptable waste in a slab. Then
2940 * we reduce the minimum objects required in a slab.
2942 min_objects = slub_min_objects;
2944 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2945 max_objects = order_objects(slub_max_order, size, reserved);
2946 min_objects = min(min_objects, max_objects);
2948 while (min_objects > 1) {
2950 while (fraction >= 4) {
2951 order = slab_order(size, min_objects,
2952 slub_max_order, fraction, reserved);
2953 if (order <= slub_max_order)
2961 * We were unable to place multiple objects in a slab. Now
2962 * lets see if we can place a single object there.
2964 order = slab_order(size, 1, slub_max_order, 1, reserved);
2965 if (order <= slub_max_order)
2969 * Doh this slab cannot be placed using slub_max_order.
2971 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2972 if (order < MAX_ORDER)
2978 init_kmem_cache_node(struct kmem_cache_node *n)
2981 spin_lock_init(&n->list_lock);
2982 INIT_LIST_HEAD(&n->partial);
2983 #ifdef CONFIG_SLUB_DEBUG
2984 atomic_long_set(&n->nr_slabs, 0);
2985 atomic_long_set(&n->total_objects, 0);
2986 INIT_LIST_HEAD(&n->full);
2990 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2992 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2993 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2996 * Must align to double word boundary for the double cmpxchg
2997 * instructions to work; see __pcpu_double_call_return_bool().
2999 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3000 2 * sizeof(void *));
3005 init_kmem_cache_cpus(s);
3010 static struct kmem_cache *kmem_cache_node;
3013 * No kmalloc_node yet so do it by hand. We know that this is the first
3014 * slab on the node for this slabcache. There are no concurrent accesses
3017 * Note that this function only works on the kmem_cache_node
3018 * when allocating for the kmem_cache_node. This is used for bootstrapping
3019 * memory on a fresh node that has no slab structures yet.
3021 static void early_kmem_cache_node_alloc(int node)
3024 struct kmem_cache_node *n;
3026 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3028 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3031 if (page_to_nid(page) != node) {
3032 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3033 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3038 page->freelist = get_freepointer(kmem_cache_node, n);
3041 kmem_cache_node->node[node] = n;
3042 #ifdef CONFIG_SLUB_DEBUG
3043 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3044 init_tracking(kmem_cache_node, n);
3046 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3047 init_kmem_cache_node(n);
3048 inc_slabs_node(kmem_cache_node, node, page->objects);
3051 * No locks need to be taken here as it has just been
3052 * initialized and there is no concurrent access.
3054 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3057 static void free_kmem_cache_nodes(struct kmem_cache *s)
3060 struct kmem_cache_node *n;
3062 for_each_kmem_cache_node(s, node, n) {
3063 kmem_cache_free(kmem_cache_node, n);
3064 s->node[node] = NULL;
3068 static int init_kmem_cache_nodes(struct kmem_cache *s)
3072 for_each_node_state(node, N_NORMAL_MEMORY) {
3073 struct kmem_cache_node *n;
3075 if (slab_state == DOWN) {
3076 early_kmem_cache_node_alloc(node);
3079 n = kmem_cache_alloc_node(kmem_cache_node,
3083 free_kmem_cache_nodes(s);
3088 init_kmem_cache_node(n);
3093 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3095 if (min < MIN_PARTIAL)
3097 else if (min > MAX_PARTIAL)
3099 s->min_partial = min;
3103 * calculate_sizes() determines the order and the distribution of data within
3106 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3108 unsigned long flags = s->flags;
3109 unsigned long size = s->object_size;
3113 * Round up object size to the next word boundary. We can only
3114 * place the free pointer at word boundaries and this determines
3115 * the possible location of the free pointer.
3117 size = ALIGN(size, sizeof(void *));
3119 #ifdef CONFIG_SLUB_DEBUG
3121 * Determine if we can poison the object itself. If the user of
3122 * the slab may touch the object after free or before allocation
3123 * then we should never poison the object itself.
3125 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3127 s->flags |= __OBJECT_POISON;
3129 s->flags &= ~__OBJECT_POISON;
3133 * If we are Redzoning then check if there is some space between the
3134 * end of the object and the free pointer. If not then add an
3135 * additional word to have some bytes to store Redzone information.
3137 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3138 size += sizeof(void *);
3142 * With that we have determined the number of bytes in actual use
3143 * by the object. This is the potential offset to the free pointer.
3147 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3150 * Relocate free pointer after the object if it is not
3151 * permitted to overwrite the first word of the object on
3154 * This is the case if we do RCU, have a constructor or
3155 * destructor or are poisoning the objects.
3158 size += sizeof(void *);
3161 #ifdef CONFIG_SLUB_DEBUG
3162 if (flags & SLAB_STORE_USER)
3164 * Need to store information about allocs and frees after
3167 size += 2 * sizeof(struct track);
3169 if (flags & SLAB_RED_ZONE)
3171 * Add some empty padding so that we can catch
3172 * overwrites from earlier objects rather than let
3173 * tracking information or the free pointer be
3174 * corrupted if a user writes before the start
3177 size += sizeof(void *);
3181 * SLUB stores one object immediately after another beginning from
3182 * offset 0. In order to align the objects we have to simply size
3183 * each object to conform to the alignment.
3185 size = ALIGN(size, s->align);
3187 if (forced_order >= 0)
3188 order = forced_order;
3190 order = calculate_order(size, s->reserved);
3197 s->allocflags |= __GFP_COMP;
3199 if (s->flags & SLAB_CACHE_DMA)
3200 s->allocflags |= GFP_DMA;
3202 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3203 s->allocflags |= __GFP_RECLAIMABLE;
3206 * Determine the number of objects per slab
3208 s->oo = oo_make(order, size, s->reserved);
3209 s->min = oo_make(get_order(size), size, s->reserved);
3210 if (oo_objects(s->oo) > oo_objects(s->max))
3213 return !!oo_objects(s->oo);
3216 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3218 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3221 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3222 s->reserved = sizeof(struct rcu_head);
3224 if (!calculate_sizes(s, -1))
3226 if (disable_higher_order_debug) {
3228 * Disable debugging flags that store metadata if the min slab
3231 if (get_order(s->size) > get_order(s->object_size)) {
3232 s->flags &= ~DEBUG_METADATA_FLAGS;
3234 if (!calculate_sizes(s, -1))
3239 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3240 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3241 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3242 /* Enable fast mode */
3243 s->flags |= __CMPXCHG_DOUBLE;
3247 * The larger the object size is, the more pages we want on the partial
3248 * list to avoid pounding the page allocator excessively.
3250 set_min_partial(s, ilog2(s->size) / 2);
3253 * cpu_partial determined the maximum number of objects kept in the
3254 * per cpu partial lists of a processor.
3256 * Per cpu partial lists mainly contain slabs that just have one
3257 * object freed. If they are used for allocation then they can be
3258 * filled up again with minimal effort. The slab will never hit the
3259 * per node partial lists and therefore no locking will be required.
3261 * This setting also determines
3263 * A) The number of objects from per cpu partial slabs dumped to the
3264 * per node list when we reach the limit.
3265 * B) The number of objects in cpu partial slabs to extract from the
3266 * per node list when we run out of per cpu objects. We only fetch
3267 * 50% to keep some capacity around for frees.
3269 if (!kmem_cache_has_cpu_partial(s))
3271 else if (s->size >= PAGE_SIZE)
3273 else if (s->size >= 1024)
3275 else if (s->size >= 256)
3276 s->cpu_partial = 13;
3278 s->cpu_partial = 30;
3281 s->remote_node_defrag_ratio = 1000;
3283 if (!init_kmem_cache_nodes(s))
3286 if (alloc_kmem_cache_cpus(s))
3289 free_kmem_cache_nodes(s);
3291 if (flags & SLAB_PANIC)
3292 panic("Cannot create slab %s size=%lu realsize=%u "
3293 "order=%u offset=%u flags=%lx\n",
3294 s->name, (unsigned long)s->size, s->size,
3295 oo_order(s->oo), s->offset, flags);
3299 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3302 #ifdef CONFIG_SLUB_DEBUG
3303 void *addr = page_address(page);
3305 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3306 sizeof(long), GFP_ATOMIC);
3309 slab_err(s, page, text, s->name);
3312 get_map(s, page, map);
3313 for_each_object(p, s, addr, page->objects) {
3315 if (!test_bit(slab_index(p, s, addr), map)) {
3316 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3317 print_tracking(s, p);
3326 * Attempt to free all partial slabs on a node.
3327 * This is called from kmem_cache_close(). We must be the last thread
3328 * using the cache and therefore we do not need to lock anymore.
3330 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3332 struct page *page, *h;
3334 list_for_each_entry_safe(page, h, &n->partial, lru) {
3336 __remove_partial(n, page);
3337 discard_slab(s, page);
3339 list_slab_objects(s, page,
3340 "Objects remaining in %s on kmem_cache_close()");
3346 * Release all resources used by a slab cache.
3348 static inline int kmem_cache_close(struct kmem_cache *s)
3351 struct kmem_cache_node *n;
3354 /* Attempt to free all objects */
3355 for_each_kmem_cache_node(s, node, n) {
3357 if (n->nr_partial || slabs_node(s, node))
3360 free_percpu(s->cpu_slab);
3361 free_kmem_cache_nodes(s);
3365 int __kmem_cache_shutdown(struct kmem_cache *s)
3367 return kmem_cache_close(s);
3370 /********************************************************************
3372 *******************************************************************/
3374 static int __init setup_slub_min_order(char *str)
3376 get_option(&str, &slub_min_order);
3381 __setup("slub_min_order=", setup_slub_min_order);
3383 static int __init setup_slub_max_order(char *str)
3385 get_option(&str, &slub_max_order);
3386 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3391 __setup("slub_max_order=", setup_slub_max_order);
3393 static int __init setup_slub_min_objects(char *str)
3395 get_option(&str, &slub_min_objects);
3400 __setup("slub_min_objects=", setup_slub_min_objects);
3402 void *__kmalloc(size_t size, gfp_t flags)
3404 struct kmem_cache *s;
3407 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3408 return kmalloc_large(size, flags);
3410 s = kmalloc_slab(size, flags);
3412 if (unlikely(ZERO_OR_NULL_PTR(s)))
3415 ret = slab_alloc(s, flags, _RET_IP_);
3417 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3419 kasan_kmalloc(s, ret, size);
3423 EXPORT_SYMBOL(__kmalloc);
3426 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3431 flags |= __GFP_COMP | __GFP_NOTRACK;
3432 page = alloc_kmem_pages_node(node, flags, get_order(size));
3434 ptr = page_address(page);
3436 kmalloc_large_node_hook(ptr, size, flags);
3440 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3442 struct kmem_cache *s;
3445 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3446 ret = kmalloc_large_node(size, flags, node);
3448 trace_kmalloc_node(_RET_IP_, ret,
3449 size, PAGE_SIZE << get_order(size),
3455 s = kmalloc_slab(size, flags);
3457 if (unlikely(ZERO_OR_NULL_PTR(s)))
3460 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3462 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3464 kasan_kmalloc(s, ret, size);
3468 EXPORT_SYMBOL(__kmalloc_node);
3471 static size_t __ksize(const void *object)
3475 if (unlikely(object == ZERO_SIZE_PTR))
3478 page = virt_to_head_page(object);
3480 if (unlikely(!PageSlab(page))) {
3481 WARN_ON(!PageCompound(page));
3482 return PAGE_SIZE << compound_order(page);
3485 return slab_ksize(page->slab_cache);
3488 size_t ksize(const void *object)
3490 size_t size = __ksize(object);
3491 /* We assume that ksize callers could use whole allocated area,
3492 so we need unpoison this area. */
3493 kasan_krealloc(object, size);
3496 EXPORT_SYMBOL(ksize);
3498 void kfree(const void *x)
3501 void *object = (void *)x;
3503 trace_kfree(_RET_IP_, x);
3505 if (unlikely(ZERO_OR_NULL_PTR(x)))
3508 page = virt_to_head_page(x);
3509 if (unlikely(!PageSlab(page))) {
3510 BUG_ON(!PageCompound(page));
3512 __free_kmem_pages(page, compound_order(page));
3515 slab_free(page->slab_cache, page, object, _RET_IP_);
3517 EXPORT_SYMBOL(kfree);
3519 #define SHRINK_PROMOTE_MAX 32
3522 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3523 * up most to the head of the partial lists. New allocations will then
3524 * fill those up and thus they can be removed from the partial lists.
3526 * The slabs with the least items are placed last. This results in them
3527 * being allocated from last increasing the chance that the last objects
3528 * are freed in them.
3530 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3534 struct kmem_cache_node *n;
3537 struct list_head discard;
3538 struct list_head promote[SHRINK_PROMOTE_MAX];
3539 unsigned long flags;
3544 * Disable empty slabs caching. Used to avoid pinning offline
3545 * memory cgroups by kmem pages that can be freed.
3551 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3552 * so we have to make sure the change is visible.
3554 kick_all_cpus_sync();
3558 for_each_kmem_cache_node(s, node, n) {
3559 INIT_LIST_HEAD(&discard);
3560 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3561 INIT_LIST_HEAD(promote + i);
3563 spin_lock_irqsave(&n->list_lock, flags);
3566 * Build lists of slabs to discard or promote.
3568 * Note that concurrent frees may occur while we hold the
3569 * list_lock. page->inuse here is the upper limit.
3571 list_for_each_entry_safe(page, t, &n->partial, lru) {
3572 int free = page->objects - page->inuse;
3574 /* Do not reread page->inuse */
3577 /* We do not keep full slabs on the list */
3580 if (free == page->objects) {
3581 list_move(&page->lru, &discard);
3583 } else if (free <= SHRINK_PROMOTE_MAX)
3584 list_move(&page->lru, promote + free - 1);
3588 * Promote the slabs filled up most to the head of the
3591 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3592 list_splice(promote + i, &n->partial);
3594 spin_unlock_irqrestore(&n->list_lock, flags);
3596 /* Release empty slabs */
3597 list_for_each_entry_safe(page, t, &discard, lru)
3598 discard_slab(s, page);
3600 if (slabs_node(s, node))
3607 static int slab_mem_going_offline_callback(void *arg)
3609 struct kmem_cache *s;
3611 mutex_lock(&slab_mutex);
3612 list_for_each_entry(s, &slab_caches, list)
3613 __kmem_cache_shrink(s, false);
3614 mutex_unlock(&slab_mutex);
3619 static void slab_mem_offline_callback(void *arg)
3621 struct kmem_cache_node *n;
3622 struct kmem_cache *s;
3623 struct memory_notify *marg = arg;
3626 offline_node = marg->status_change_nid_normal;
3629 * If the node still has available memory. we need kmem_cache_node
3632 if (offline_node < 0)
3635 mutex_lock(&slab_mutex);
3636 list_for_each_entry(s, &slab_caches, list) {
3637 n = get_node(s, offline_node);
3640 * if n->nr_slabs > 0, slabs still exist on the node
3641 * that is going down. We were unable to free them,
3642 * and offline_pages() function shouldn't call this
3643 * callback. So, we must fail.
3645 BUG_ON(slabs_node(s, offline_node));
3647 s->node[offline_node] = NULL;
3648 kmem_cache_free(kmem_cache_node, n);
3651 mutex_unlock(&slab_mutex);
3654 static int slab_mem_going_online_callback(void *arg)
3656 struct kmem_cache_node *n;
3657 struct kmem_cache *s;
3658 struct memory_notify *marg = arg;
3659 int nid = marg->status_change_nid_normal;
3663 * If the node's memory is already available, then kmem_cache_node is
3664 * already created. Nothing to do.
3670 * We are bringing a node online. No memory is available yet. We must
3671 * allocate a kmem_cache_node structure in order to bring the node
3674 mutex_lock(&slab_mutex);
3675 list_for_each_entry(s, &slab_caches, list) {
3677 * XXX: kmem_cache_alloc_node will fallback to other nodes
3678 * since memory is not yet available from the node that
3681 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3686 init_kmem_cache_node(n);
3690 mutex_unlock(&slab_mutex);
3694 static int slab_memory_callback(struct notifier_block *self,
3695 unsigned long action, void *arg)
3700 case MEM_GOING_ONLINE:
3701 ret = slab_mem_going_online_callback(arg);
3703 case MEM_GOING_OFFLINE:
3704 ret = slab_mem_going_offline_callback(arg);
3707 case MEM_CANCEL_ONLINE:
3708 slab_mem_offline_callback(arg);
3711 case MEM_CANCEL_OFFLINE:
3715 ret = notifier_from_errno(ret);
3721 static struct notifier_block slab_memory_callback_nb = {
3722 .notifier_call = slab_memory_callback,
3723 .priority = SLAB_CALLBACK_PRI,
3726 /********************************************************************
3727 * Basic setup of slabs
3728 *******************************************************************/
3731 * Used for early kmem_cache structures that were allocated using
3732 * the page allocator. Allocate them properly then fix up the pointers
3733 * that may be pointing to the wrong kmem_cache structure.
3736 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3739 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3740 struct kmem_cache_node *n;
3742 memcpy(s, static_cache, kmem_cache->object_size);
3745 * This runs very early, and only the boot processor is supposed to be
3746 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3749 __flush_cpu_slab(s, smp_processor_id());
3750 for_each_kmem_cache_node(s, node, n) {
3753 list_for_each_entry(p, &n->partial, lru)
3756 #ifdef CONFIG_SLUB_DEBUG
3757 list_for_each_entry(p, &n->full, lru)
3761 slab_init_memcg_params(s);
3762 list_add(&s->list, &slab_caches);
3766 void __init kmem_cache_init(void)
3768 static __initdata struct kmem_cache boot_kmem_cache,
3769 boot_kmem_cache_node;
3771 if (debug_guardpage_minorder())
3774 kmem_cache_node = &boot_kmem_cache_node;
3775 kmem_cache = &boot_kmem_cache;
3777 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3778 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3780 register_hotmemory_notifier(&slab_memory_callback_nb);
3782 /* Able to allocate the per node structures */
3783 slab_state = PARTIAL;
3785 create_boot_cache(kmem_cache, "kmem_cache",
3786 offsetof(struct kmem_cache, node) +
3787 nr_node_ids * sizeof(struct kmem_cache_node *),
3788 SLAB_HWCACHE_ALIGN);
3790 kmem_cache = bootstrap(&boot_kmem_cache);
3793 * Allocate kmem_cache_node properly from the kmem_cache slab.
3794 * kmem_cache_node is separately allocated so no need to
3795 * update any list pointers.
3797 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3799 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3800 setup_kmalloc_cache_index_table();
3801 create_kmalloc_caches(0);
3804 register_cpu_notifier(&slab_notifier);
3807 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3809 slub_min_order, slub_max_order, slub_min_objects,
3810 nr_cpu_ids, nr_node_ids);
3813 void __init kmem_cache_init_late(void)
3818 __kmem_cache_alias(const char *name, size_t size, size_t align,
3819 unsigned long flags, void (*ctor)(void *))
3821 struct kmem_cache *s, *c;
3823 s = find_mergeable(size, align, flags, name, ctor);
3828 * Adjust the object sizes so that we clear
3829 * the complete object on kzalloc.
3831 s->object_size = max(s->object_size, (int)size);
3832 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3834 for_each_memcg_cache(c, s) {
3835 c->object_size = s->object_size;
3836 c->inuse = max_t(int, c->inuse,
3837 ALIGN(size, sizeof(void *)));
3840 if (sysfs_slab_alias(s, name)) {
3849 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3853 err = kmem_cache_open(s, flags);
3857 /* Mutex is not taken during early boot */
3858 if (slab_state <= UP)
3861 memcg_propagate_slab_attrs(s);
3862 err = sysfs_slab_add(s);
3864 kmem_cache_close(s);
3871 * Use the cpu notifier to insure that the cpu slabs are flushed when
3874 static int slab_cpuup_callback(struct notifier_block *nfb,
3875 unsigned long action, void *hcpu)
3877 long cpu = (long)hcpu;
3878 struct kmem_cache *s;
3879 unsigned long flags;
3882 case CPU_UP_CANCELED:
3883 case CPU_UP_CANCELED_FROZEN:
3885 case CPU_DEAD_FROZEN:
3886 mutex_lock(&slab_mutex);
3887 list_for_each_entry(s, &slab_caches, list) {
3888 local_irq_save(flags);
3889 __flush_cpu_slab(s, cpu);
3890 local_irq_restore(flags);
3892 mutex_unlock(&slab_mutex);
3900 static struct notifier_block slab_notifier = {
3901 .notifier_call = slab_cpuup_callback
3906 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3908 struct kmem_cache *s;
3911 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3912 return kmalloc_large(size, gfpflags);
3914 s = kmalloc_slab(size, gfpflags);
3916 if (unlikely(ZERO_OR_NULL_PTR(s)))
3919 ret = slab_alloc(s, gfpflags, caller);
3921 /* Honor the call site pointer we received. */
3922 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3928 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3929 int node, unsigned long caller)
3931 struct kmem_cache *s;
3934 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3935 ret = kmalloc_large_node(size, gfpflags, node);
3937 trace_kmalloc_node(caller, ret,
3938 size, PAGE_SIZE << get_order(size),
3944 s = kmalloc_slab(size, gfpflags);
3946 if (unlikely(ZERO_OR_NULL_PTR(s)))
3949 ret = slab_alloc_node(s, gfpflags, node, caller);
3951 /* Honor the call site pointer we received. */
3952 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3959 static int count_inuse(struct page *page)
3964 static int count_total(struct page *page)
3966 return page->objects;
3970 #ifdef CONFIG_SLUB_DEBUG
3971 static int validate_slab(struct kmem_cache *s, struct page *page,
3975 void *addr = page_address(page);
3977 if (!check_slab(s, page) ||
3978 !on_freelist(s, page, NULL))
3981 /* Now we know that a valid freelist exists */
3982 bitmap_zero(map, page->objects);
3984 get_map(s, page, map);
3985 for_each_object(p, s, addr, page->objects) {
3986 if (test_bit(slab_index(p, s, addr), map))
3987 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3991 for_each_object(p, s, addr, page->objects)
3992 if (!test_bit(slab_index(p, s, addr), map))
3993 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3998 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4002 validate_slab(s, page, map);
4006 static int validate_slab_node(struct kmem_cache *s,
4007 struct kmem_cache_node *n, unsigned long *map)
4009 unsigned long count = 0;
4011 unsigned long flags;
4013 spin_lock_irqsave(&n->list_lock, flags);
4015 list_for_each_entry(page, &n->partial, lru) {
4016 validate_slab_slab(s, page, map);
4019 if (count != n->nr_partial)
4020 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4021 s->name, count, n->nr_partial);
4023 if (!(s->flags & SLAB_STORE_USER))
4026 list_for_each_entry(page, &n->full, lru) {
4027 validate_slab_slab(s, page, map);
4030 if (count != atomic_long_read(&n->nr_slabs))
4031 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4032 s->name, count, atomic_long_read(&n->nr_slabs));
4035 spin_unlock_irqrestore(&n->list_lock, flags);
4039 static long validate_slab_cache(struct kmem_cache *s)
4042 unsigned long count = 0;
4043 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4044 sizeof(unsigned long), GFP_KERNEL);
4045 struct kmem_cache_node *n;
4051 for_each_kmem_cache_node(s, node, n)
4052 count += validate_slab_node(s, n, map);
4057 * Generate lists of code addresses where slabcache objects are allocated
4062 unsigned long count;
4069 DECLARE_BITMAP(cpus, NR_CPUS);
4075 unsigned long count;
4076 struct location *loc;
4079 static void free_loc_track(struct loc_track *t)
4082 free_pages((unsigned long)t->loc,
4083 get_order(sizeof(struct location) * t->max));
4086 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4091 order = get_order(sizeof(struct location) * max);
4093 l = (void *)__get_free_pages(flags, order);
4098 memcpy(l, t->loc, sizeof(struct location) * t->count);
4106 static int add_location(struct loc_track *t, struct kmem_cache *s,
4107 const struct track *track)
4109 long start, end, pos;
4111 unsigned long caddr;
4112 unsigned long age = jiffies - track->when;
4118 pos = start + (end - start + 1) / 2;
4121 * There is nothing at "end". If we end up there
4122 * we need to add something to before end.
4127 caddr = t->loc[pos].addr;
4128 if (track->addr == caddr) {
4134 if (age < l->min_time)
4136 if (age > l->max_time)
4139 if (track->pid < l->min_pid)
4140 l->min_pid = track->pid;
4141 if (track->pid > l->max_pid)
4142 l->max_pid = track->pid;
4144 cpumask_set_cpu(track->cpu,
4145 to_cpumask(l->cpus));
4147 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4151 if (track->addr < caddr)
4158 * Not found. Insert new tracking element.
4160 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4166 (t->count - pos) * sizeof(struct location));
4169 l->addr = track->addr;
4173 l->min_pid = track->pid;
4174 l->max_pid = track->pid;
4175 cpumask_clear(to_cpumask(l->cpus));
4176 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4177 nodes_clear(l->nodes);
4178 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4182 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4183 struct page *page, enum track_item alloc,
4186 void *addr = page_address(page);
4189 bitmap_zero(map, page->objects);
4190 get_map(s, page, map);
4192 for_each_object(p, s, addr, page->objects)
4193 if (!test_bit(slab_index(p, s, addr), map))
4194 add_location(t, s, get_track(s, p, alloc));
4197 static int list_locations(struct kmem_cache *s, char *buf,
4198 enum track_item alloc)
4202 struct loc_track t = { 0, 0, NULL };
4204 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4205 sizeof(unsigned long), GFP_KERNEL);
4206 struct kmem_cache_node *n;
4208 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4211 return sprintf(buf, "Out of memory\n");
4213 /* Push back cpu slabs */
4216 for_each_kmem_cache_node(s, node, n) {
4217 unsigned long flags;
4220 if (!atomic_long_read(&n->nr_slabs))
4223 spin_lock_irqsave(&n->list_lock, flags);
4224 list_for_each_entry(page, &n->partial, lru)
4225 process_slab(&t, s, page, alloc, map);
4226 list_for_each_entry(page, &n->full, lru)
4227 process_slab(&t, s, page, alloc, map);
4228 spin_unlock_irqrestore(&n->list_lock, flags);
4231 for (i = 0; i < t.count; i++) {
4232 struct location *l = &t.loc[i];
4234 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4236 len += sprintf(buf + len, "%7ld ", l->count);
4239 len += sprintf(buf + len, "%pS", (void *)l->addr);
4241 len += sprintf(buf + len, "<not-available>");
4243 if (l->sum_time != l->min_time) {
4244 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4246 (long)div_u64(l->sum_time, l->count),
4249 len += sprintf(buf + len, " age=%ld",
4252 if (l->min_pid != l->max_pid)
4253 len += sprintf(buf + len, " pid=%ld-%ld",
4254 l->min_pid, l->max_pid);
4256 len += sprintf(buf + len, " pid=%ld",
4259 if (num_online_cpus() > 1 &&
4260 !cpumask_empty(to_cpumask(l->cpus)) &&
4261 len < PAGE_SIZE - 60)
4262 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4264 cpumask_pr_args(to_cpumask(l->cpus)));
4266 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4267 len < PAGE_SIZE - 60)
4268 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4270 nodemask_pr_args(&l->nodes));
4272 len += sprintf(buf + len, "\n");
4278 len += sprintf(buf, "No data\n");
4283 #ifdef SLUB_RESILIENCY_TEST
4284 static void __init resiliency_test(void)
4288 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4290 pr_err("SLUB resiliency testing\n");
4291 pr_err("-----------------------\n");
4292 pr_err("A. Corruption after allocation\n");
4294 p = kzalloc(16, GFP_KERNEL);
4296 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4299 validate_slab_cache(kmalloc_caches[4]);
4301 /* Hmmm... The next two are dangerous */
4302 p = kzalloc(32, GFP_KERNEL);
4303 p[32 + sizeof(void *)] = 0x34;
4304 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4306 pr_err("If allocated object is overwritten then not detectable\n\n");
4308 validate_slab_cache(kmalloc_caches[5]);
4309 p = kzalloc(64, GFP_KERNEL);
4310 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4312 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4314 pr_err("If allocated object is overwritten then not detectable\n\n");
4315 validate_slab_cache(kmalloc_caches[6]);
4317 pr_err("\nB. Corruption after free\n");
4318 p = kzalloc(128, GFP_KERNEL);
4321 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4322 validate_slab_cache(kmalloc_caches[7]);
4324 p = kzalloc(256, GFP_KERNEL);
4327 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4328 validate_slab_cache(kmalloc_caches[8]);
4330 p = kzalloc(512, GFP_KERNEL);
4333 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4334 validate_slab_cache(kmalloc_caches[9]);
4338 static void resiliency_test(void) {};
4343 enum slab_stat_type {
4344 SL_ALL, /* All slabs */
4345 SL_PARTIAL, /* Only partially allocated slabs */
4346 SL_CPU, /* Only slabs used for cpu caches */
4347 SL_OBJECTS, /* Determine allocated objects not slabs */
4348 SL_TOTAL /* Determine object capacity not slabs */
4351 #define SO_ALL (1 << SL_ALL)
4352 #define SO_PARTIAL (1 << SL_PARTIAL)
4353 #define SO_CPU (1 << SL_CPU)
4354 #define SO_OBJECTS (1 << SL_OBJECTS)
4355 #define SO_TOTAL (1 << SL_TOTAL)
4357 static ssize_t show_slab_objects(struct kmem_cache *s,
4358 char *buf, unsigned long flags)
4360 unsigned long total = 0;
4363 unsigned long *nodes;
4365 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4369 if (flags & SO_CPU) {
4372 for_each_possible_cpu(cpu) {
4373 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4378 page = READ_ONCE(c->page);
4382 node = page_to_nid(page);
4383 if (flags & SO_TOTAL)
4385 else if (flags & SO_OBJECTS)
4393 page = READ_ONCE(c->partial);
4395 node = page_to_nid(page);
4396 if (flags & SO_TOTAL)
4398 else if (flags & SO_OBJECTS)
4409 #ifdef CONFIG_SLUB_DEBUG
4410 if (flags & SO_ALL) {
4411 struct kmem_cache_node *n;
4413 for_each_kmem_cache_node(s, node, n) {
4415 if (flags & SO_TOTAL)
4416 x = atomic_long_read(&n->total_objects);
4417 else if (flags & SO_OBJECTS)
4418 x = atomic_long_read(&n->total_objects) -
4419 count_partial(n, count_free);
4421 x = atomic_long_read(&n->nr_slabs);
4428 if (flags & SO_PARTIAL) {
4429 struct kmem_cache_node *n;
4431 for_each_kmem_cache_node(s, node, n) {
4432 if (flags & SO_TOTAL)
4433 x = count_partial(n, count_total);
4434 else if (flags & SO_OBJECTS)
4435 x = count_partial(n, count_inuse);
4442 x = sprintf(buf, "%lu", total);
4444 for (node = 0; node < nr_node_ids; node++)
4446 x += sprintf(buf + x, " N%d=%lu",
4451 return x + sprintf(buf + x, "\n");
4454 #ifdef CONFIG_SLUB_DEBUG
4455 static int any_slab_objects(struct kmem_cache *s)
4458 struct kmem_cache_node *n;
4460 for_each_kmem_cache_node(s, node, n)
4461 if (atomic_long_read(&n->total_objects))
4468 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4469 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4471 struct slab_attribute {
4472 struct attribute attr;
4473 ssize_t (*show)(struct kmem_cache *s, char *buf);
4474 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4477 #define SLAB_ATTR_RO(_name) \
4478 static struct slab_attribute _name##_attr = \
4479 __ATTR(_name, 0400, _name##_show, NULL)
4481 #define SLAB_ATTR(_name) \
4482 static struct slab_attribute _name##_attr = \
4483 __ATTR(_name, 0600, _name##_show, _name##_store)
4485 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4487 return sprintf(buf, "%d\n", s->size);
4489 SLAB_ATTR_RO(slab_size);
4491 static ssize_t align_show(struct kmem_cache *s, char *buf)
4493 return sprintf(buf, "%d\n", s->align);
4495 SLAB_ATTR_RO(align);
4497 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4499 return sprintf(buf, "%d\n", s->object_size);
4501 SLAB_ATTR_RO(object_size);
4503 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4505 return sprintf(buf, "%d\n", oo_objects(s->oo));
4507 SLAB_ATTR_RO(objs_per_slab);
4509 static ssize_t order_store(struct kmem_cache *s,
4510 const char *buf, size_t length)
4512 unsigned long order;
4515 err = kstrtoul(buf, 10, &order);
4519 if (order > slub_max_order || order < slub_min_order)
4522 calculate_sizes(s, order);
4526 static ssize_t order_show(struct kmem_cache *s, char *buf)
4528 return sprintf(buf, "%d\n", oo_order(s->oo));
4532 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4534 return sprintf(buf, "%lu\n", s->min_partial);
4537 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4543 err = kstrtoul(buf, 10, &min);
4547 set_min_partial(s, min);
4550 SLAB_ATTR(min_partial);
4552 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4554 return sprintf(buf, "%u\n", s->cpu_partial);
4557 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4560 unsigned long objects;
4563 err = kstrtoul(buf, 10, &objects);
4566 if (objects && !kmem_cache_has_cpu_partial(s))
4569 s->cpu_partial = objects;
4573 SLAB_ATTR(cpu_partial);
4575 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4579 return sprintf(buf, "%pS\n", s->ctor);
4583 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4585 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4587 SLAB_ATTR_RO(aliases);
4589 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4591 return show_slab_objects(s, buf, SO_PARTIAL);
4593 SLAB_ATTR_RO(partial);
4595 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4597 return show_slab_objects(s, buf, SO_CPU);
4599 SLAB_ATTR_RO(cpu_slabs);
4601 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4603 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4605 SLAB_ATTR_RO(objects);
4607 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4609 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4611 SLAB_ATTR_RO(objects_partial);
4613 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4620 for_each_online_cpu(cpu) {
4621 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4624 pages += page->pages;
4625 objects += page->pobjects;
4629 len = sprintf(buf, "%d(%d)", objects, pages);
4632 for_each_online_cpu(cpu) {
4633 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4635 if (page && len < PAGE_SIZE - 20)
4636 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4637 page->pobjects, page->pages);
4640 return len + sprintf(buf + len, "\n");
4642 SLAB_ATTR_RO(slabs_cpu_partial);
4644 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4646 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4649 static ssize_t reclaim_account_store(struct kmem_cache *s,
4650 const char *buf, size_t length)
4652 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4654 s->flags |= SLAB_RECLAIM_ACCOUNT;
4657 SLAB_ATTR(reclaim_account);
4659 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4661 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4663 SLAB_ATTR_RO(hwcache_align);
4665 #ifdef CONFIG_ZONE_DMA
4666 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4668 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4670 SLAB_ATTR_RO(cache_dma);
4673 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4675 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4677 SLAB_ATTR_RO(destroy_by_rcu);
4679 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4681 return sprintf(buf, "%d\n", s->reserved);
4683 SLAB_ATTR_RO(reserved);
4685 #ifdef CONFIG_SLUB_DEBUG
4686 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4688 return show_slab_objects(s, buf, SO_ALL);
4690 SLAB_ATTR_RO(slabs);
4692 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4694 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4696 SLAB_ATTR_RO(total_objects);
4698 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4700 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4703 static ssize_t sanity_checks_store(struct kmem_cache *s,
4704 const char *buf, size_t length)
4706 s->flags &= ~SLAB_DEBUG_FREE;
4707 if (buf[0] == '1') {
4708 s->flags &= ~__CMPXCHG_DOUBLE;
4709 s->flags |= SLAB_DEBUG_FREE;
4713 SLAB_ATTR(sanity_checks);
4715 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4717 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4720 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4724 * Tracing a merged cache is going to give confusing results
4725 * as well as cause other issues like converting a mergeable
4726 * cache into an umergeable one.
4728 if (s->refcount > 1)
4731 s->flags &= ~SLAB_TRACE;
4732 if (buf[0] == '1') {
4733 s->flags &= ~__CMPXCHG_DOUBLE;
4734 s->flags |= SLAB_TRACE;
4740 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4742 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4745 static ssize_t red_zone_store(struct kmem_cache *s,
4746 const char *buf, size_t length)
4748 if (any_slab_objects(s))
4751 s->flags &= ~SLAB_RED_ZONE;
4752 if (buf[0] == '1') {
4753 s->flags &= ~__CMPXCHG_DOUBLE;
4754 s->flags |= SLAB_RED_ZONE;
4756 calculate_sizes(s, -1);
4759 SLAB_ATTR(red_zone);
4761 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4763 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4766 static ssize_t poison_store(struct kmem_cache *s,
4767 const char *buf, size_t length)
4769 if (any_slab_objects(s))
4772 s->flags &= ~SLAB_POISON;
4773 if (buf[0] == '1') {
4774 s->flags &= ~__CMPXCHG_DOUBLE;
4775 s->flags |= SLAB_POISON;
4777 calculate_sizes(s, -1);
4782 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4784 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4787 static ssize_t store_user_store(struct kmem_cache *s,
4788 const char *buf, size_t length)
4790 if (any_slab_objects(s))
4793 s->flags &= ~SLAB_STORE_USER;
4794 if (buf[0] == '1') {
4795 s->flags &= ~__CMPXCHG_DOUBLE;
4796 s->flags |= SLAB_STORE_USER;
4798 calculate_sizes(s, -1);
4801 SLAB_ATTR(store_user);
4803 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4808 static ssize_t validate_store(struct kmem_cache *s,
4809 const char *buf, size_t length)
4813 if (buf[0] == '1') {
4814 ret = validate_slab_cache(s);
4820 SLAB_ATTR(validate);
4822 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4824 if (!(s->flags & SLAB_STORE_USER))
4826 return list_locations(s, buf, TRACK_ALLOC);
4828 SLAB_ATTR_RO(alloc_calls);
4830 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4832 if (!(s->flags & SLAB_STORE_USER))
4834 return list_locations(s, buf, TRACK_FREE);
4836 SLAB_ATTR_RO(free_calls);
4837 #endif /* CONFIG_SLUB_DEBUG */
4839 #ifdef CONFIG_FAILSLAB
4840 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4842 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4845 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4848 if (s->refcount > 1)
4851 s->flags &= ~SLAB_FAILSLAB;
4853 s->flags |= SLAB_FAILSLAB;
4856 SLAB_ATTR(failslab);
4859 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4864 static ssize_t shrink_store(struct kmem_cache *s,
4865 const char *buf, size_t length)
4868 kmem_cache_shrink(s);
4876 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4878 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4881 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4882 const char *buf, size_t length)
4884 unsigned long ratio;
4887 err = kstrtoul(buf, 10, &ratio);
4892 s->remote_node_defrag_ratio = ratio * 10;
4896 SLAB_ATTR(remote_node_defrag_ratio);
4899 #ifdef CONFIG_SLUB_STATS
4900 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4902 unsigned long sum = 0;
4905 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4910 for_each_online_cpu(cpu) {
4911 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4917 len = sprintf(buf, "%lu", sum);
4920 for_each_online_cpu(cpu) {
4921 if (data[cpu] && len < PAGE_SIZE - 20)
4922 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4926 return len + sprintf(buf + len, "\n");
4929 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4933 for_each_online_cpu(cpu)
4934 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4937 #define STAT_ATTR(si, text) \
4938 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4940 return show_stat(s, buf, si); \
4942 static ssize_t text##_store(struct kmem_cache *s, \
4943 const char *buf, size_t length) \
4945 if (buf[0] != '0') \
4947 clear_stat(s, si); \
4952 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4953 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4954 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4955 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4956 STAT_ATTR(FREE_FROZEN, free_frozen);
4957 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4958 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4959 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4960 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4961 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4962 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4963 STAT_ATTR(FREE_SLAB, free_slab);
4964 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4965 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4966 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4967 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4968 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4969 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4970 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4971 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4972 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4973 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4974 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4975 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4976 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4977 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4980 static struct attribute *slab_attrs[] = {
4981 &slab_size_attr.attr,
4982 &object_size_attr.attr,
4983 &objs_per_slab_attr.attr,
4985 &min_partial_attr.attr,
4986 &cpu_partial_attr.attr,
4988 &objects_partial_attr.attr,
4990 &cpu_slabs_attr.attr,
4994 &hwcache_align_attr.attr,
4995 &reclaim_account_attr.attr,
4996 &destroy_by_rcu_attr.attr,
4998 &reserved_attr.attr,
4999 &slabs_cpu_partial_attr.attr,
5000 #ifdef CONFIG_SLUB_DEBUG
5001 &total_objects_attr.attr,
5003 &sanity_checks_attr.attr,
5005 &red_zone_attr.attr,
5007 &store_user_attr.attr,
5008 &validate_attr.attr,
5009 &alloc_calls_attr.attr,
5010 &free_calls_attr.attr,
5012 #ifdef CONFIG_ZONE_DMA
5013 &cache_dma_attr.attr,
5016 &remote_node_defrag_ratio_attr.attr,
5018 #ifdef CONFIG_SLUB_STATS
5019 &alloc_fastpath_attr.attr,
5020 &alloc_slowpath_attr.attr,
5021 &free_fastpath_attr.attr,
5022 &free_slowpath_attr.attr,
5023 &free_frozen_attr.attr,
5024 &free_add_partial_attr.attr,
5025 &free_remove_partial_attr.attr,
5026 &alloc_from_partial_attr.attr,
5027 &alloc_slab_attr.attr,
5028 &alloc_refill_attr.attr,
5029 &alloc_node_mismatch_attr.attr,
5030 &free_slab_attr.attr,
5031 &cpuslab_flush_attr.attr,
5032 &deactivate_full_attr.attr,
5033 &deactivate_empty_attr.attr,
5034 &deactivate_to_head_attr.attr,
5035 &deactivate_to_tail_attr.attr,
5036 &deactivate_remote_frees_attr.attr,
5037 &deactivate_bypass_attr.attr,
5038 &order_fallback_attr.attr,
5039 &cmpxchg_double_fail_attr.attr,
5040 &cmpxchg_double_cpu_fail_attr.attr,
5041 &cpu_partial_alloc_attr.attr,
5042 &cpu_partial_free_attr.attr,
5043 &cpu_partial_node_attr.attr,
5044 &cpu_partial_drain_attr.attr,
5046 #ifdef CONFIG_FAILSLAB
5047 &failslab_attr.attr,
5053 static struct attribute_group slab_attr_group = {
5054 .attrs = slab_attrs,
5057 static ssize_t slab_attr_show(struct kobject *kobj,
5058 struct attribute *attr,
5061 struct slab_attribute *attribute;
5062 struct kmem_cache *s;
5065 attribute = to_slab_attr(attr);
5068 if (!attribute->show)
5071 err = attribute->show(s, buf);
5076 static ssize_t slab_attr_store(struct kobject *kobj,
5077 struct attribute *attr,
5078 const char *buf, size_t len)
5080 struct slab_attribute *attribute;
5081 struct kmem_cache *s;
5084 attribute = to_slab_attr(attr);
5087 if (!attribute->store)
5090 err = attribute->store(s, buf, len);
5091 #ifdef CONFIG_MEMCG_KMEM
5092 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5093 struct kmem_cache *c;
5095 mutex_lock(&slab_mutex);
5096 if (s->max_attr_size < len)
5097 s->max_attr_size = len;
5100 * This is a best effort propagation, so this function's return
5101 * value will be determined by the parent cache only. This is
5102 * basically because not all attributes will have a well
5103 * defined semantics for rollbacks - most of the actions will
5104 * have permanent effects.
5106 * Returning the error value of any of the children that fail
5107 * is not 100 % defined, in the sense that users seeing the
5108 * error code won't be able to know anything about the state of
5111 * Only returning the error code for the parent cache at least
5112 * has well defined semantics. The cache being written to
5113 * directly either failed or succeeded, in which case we loop
5114 * through the descendants with best-effort propagation.
5116 for_each_memcg_cache(c, s)
5117 attribute->store(c, buf, len);
5118 mutex_unlock(&slab_mutex);
5124 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5126 #ifdef CONFIG_MEMCG_KMEM
5128 char *buffer = NULL;
5129 struct kmem_cache *root_cache;
5131 if (is_root_cache(s))
5134 root_cache = s->memcg_params.root_cache;
5137 * This mean this cache had no attribute written. Therefore, no point
5138 * in copying default values around
5140 if (!root_cache->max_attr_size)
5143 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5146 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5148 if (!attr || !attr->store || !attr->show)
5152 * It is really bad that we have to allocate here, so we will
5153 * do it only as a fallback. If we actually allocate, though,
5154 * we can just use the allocated buffer until the end.
5156 * Most of the slub attributes will tend to be very small in
5157 * size, but sysfs allows buffers up to a page, so they can
5158 * theoretically happen.
5162 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5165 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5166 if (WARN_ON(!buffer))
5171 attr->show(root_cache, buf);
5172 attr->store(s, buf, strlen(buf));
5176 free_page((unsigned long)buffer);
5180 static void kmem_cache_release(struct kobject *k)
5182 slab_kmem_cache_release(to_slab(k));
5185 static const struct sysfs_ops slab_sysfs_ops = {
5186 .show = slab_attr_show,
5187 .store = slab_attr_store,
5190 static struct kobj_type slab_ktype = {
5191 .sysfs_ops = &slab_sysfs_ops,
5192 .release = kmem_cache_release,
5195 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5197 struct kobj_type *ktype = get_ktype(kobj);
5199 if (ktype == &slab_ktype)
5204 static const struct kset_uevent_ops slab_uevent_ops = {
5205 .filter = uevent_filter,
5208 static struct kset *slab_kset;
5210 static inline struct kset *cache_kset(struct kmem_cache *s)
5212 #ifdef CONFIG_MEMCG_KMEM
5213 if (!is_root_cache(s))
5214 return s->memcg_params.root_cache->memcg_kset;
5219 #define ID_STR_LENGTH 64
5221 /* Create a unique string id for a slab cache:
5223 * Format :[flags-]size
5225 static char *create_unique_id(struct kmem_cache *s)
5227 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5234 * First flags affecting slabcache operations. We will only
5235 * get here for aliasable slabs so we do not need to support
5236 * too many flags. The flags here must cover all flags that
5237 * are matched during merging to guarantee that the id is
5240 if (s->flags & SLAB_CACHE_DMA)
5242 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5244 if (s->flags & SLAB_DEBUG_FREE)
5246 if (!(s->flags & SLAB_NOTRACK))
5250 p += sprintf(p, "%07d", s->size);
5252 BUG_ON(p > name + ID_STR_LENGTH - 1);
5256 static int sysfs_slab_add(struct kmem_cache *s)
5260 int unmergeable = slab_unmergeable(s);
5264 * Slabcache can never be merged so we can use the name proper.
5265 * This is typically the case for debug situations. In that
5266 * case we can catch duplicate names easily.
5268 sysfs_remove_link(&slab_kset->kobj, s->name);
5272 * Create a unique name for the slab as a target
5275 name = create_unique_id(s);
5278 s->kobj.kset = cache_kset(s);
5279 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5283 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5287 #ifdef CONFIG_MEMCG_KMEM
5288 if (is_root_cache(s)) {
5289 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5290 if (!s->memcg_kset) {
5297 kobject_uevent(&s->kobj, KOBJ_ADD);
5299 /* Setup first alias */
5300 sysfs_slab_alias(s, s->name);
5307 kobject_del(&s->kobj);
5311 void sysfs_slab_remove(struct kmem_cache *s)
5313 if (slab_state < FULL)
5315 * Sysfs has not been setup yet so no need to remove the
5320 #ifdef CONFIG_MEMCG_KMEM
5321 kset_unregister(s->memcg_kset);
5323 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5324 kobject_del(&s->kobj);
5325 kobject_put(&s->kobj);
5329 * Need to buffer aliases during bootup until sysfs becomes
5330 * available lest we lose that information.
5332 struct saved_alias {
5333 struct kmem_cache *s;
5335 struct saved_alias *next;
5338 static struct saved_alias *alias_list;
5340 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5342 struct saved_alias *al;
5344 if (slab_state == FULL) {
5346 * If we have a leftover link then remove it.
5348 sysfs_remove_link(&slab_kset->kobj, name);
5349 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5352 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5358 al->next = alias_list;
5363 static int __init slab_sysfs_init(void)
5365 struct kmem_cache *s;
5368 mutex_lock(&slab_mutex);
5370 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5372 mutex_unlock(&slab_mutex);
5373 pr_err("Cannot register slab subsystem.\n");
5379 list_for_each_entry(s, &slab_caches, list) {
5380 err = sysfs_slab_add(s);
5382 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5386 while (alias_list) {
5387 struct saved_alias *al = alias_list;
5389 alias_list = alias_list->next;
5390 err = sysfs_slab_alias(al->s, al->name);
5392 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5397 mutex_unlock(&slab_mutex);
5402 __initcall(slab_sysfs_init);
5403 #endif /* CONFIG_SYSFS */
5406 * The /proc/slabinfo ABI
5408 #ifdef CONFIG_SLABINFO
5409 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5411 unsigned long nr_slabs = 0;
5412 unsigned long nr_objs = 0;
5413 unsigned long nr_free = 0;
5415 struct kmem_cache_node *n;
5417 for_each_kmem_cache_node(s, node, n) {
5418 nr_slabs += node_nr_slabs(n);
5419 nr_objs += node_nr_objs(n);
5420 nr_free += count_partial(n, count_free);
5423 sinfo->active_objs = nr_objs - nr_free;
5424 sinfo->num_objs = nr_objs;
5425 sinfo->active_slabs = nr_slabs;
5426 sinfo->num_slabs = nr_slabs;
5427 sinfo->objects_per_slab = oo_objects(s->oo);
5428 sinfo->cache_order = oo_order(s->oo);
5431 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5435 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5436 size_t count, loff_t *ppos)
5440 #endif /* CONFIG_SLABINFO */