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/kmemcheck.h>
24 #include <linux/cpu.h>
25 #include <linux/cpuset.h>
26 #include <linux/mempolicy.h>
27 #include <linux/ctype.h>
28 #include <linux/debugobjects.h>
29 #include <linux/kallsyms.h>
30 #include <linux/memory.h>
31 #include <linux/math64.h>
32 #include <linux/fault-inject.h>
33 #include <linux/stacktrace.h>
34 #include <linux/prefetch.h>
35 #include <linux/memcontrol.h>
37 #include <trace/events/kmem.h>
43 * 1. slab_mutex (Global Mutex)
45 * 3. slab_lock(page) (Only on some arches and for debugging)
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
94 * Overloading of page flags that are otherwise used for LRU management.
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
117 static inline int kmem_cache_debug(struct kmem_cache *s)
119 #ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
136 * Issues still to be resolved:
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
140 * - Variable sizing of the per node arrays
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
153 #define MIN_PARTIAL 5
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
160 #define MAX_PARTIAL 10
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
173 * Set of flags that will prevent slab merging
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
183 #define OO_MASK ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON 0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
191 static struct notifier_block slab_notifier;
195 * Tracking user of a slab.
197 #define TRACK_ADDRS_COUNT 16
199 unsigned long addr; /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
208 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 static int sysfs_slab_add(struct kmem_cache *);
212 static int sysfs_slab_alias(struct kmem_cache *, const char *);
213 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
215 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
218 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
221 static inline void stat(const struct kmem_cache *s, enum stat_item si)
223 #ifdef CONFIG_SLUB_STATS
225 * The rmw is racy on a preemptible kernel but this is acceptable, so
226 * avoid this_cpu_add()'s irq-disable overhead.
228 raw_cpu_inc(s->cpu_slab->stat[si]);
232 /********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
236 /* Verify that a pointer has an address that is valid within a slab page */
237 static inline int check_valid_pointer(struct kmem_cache *s,
238 struct page *page, const void *object)
245 base = page_address(page);
246 if (object < base || object >= base + page->objects * s->size ||
247 (object - base) % s->size) {
254 static inline void *get_freepointer(struct kmem_cache *s, void *object)
256 return *(void **)(object + s->offset);
259 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
261 prefetch(object + s->offset);
264 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
268 #ifdef CONFIG_DEBUG_PAGEALLOC
269 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
271 p = get_freepointer(s, object);
276 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
278 *(void **)(object + s->offset) = fp;
281 /* Loop over all objects in a slab */
282 #define for_each_object(__p, __s, __addr, __objects) \
283 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
286 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
287 for (__p = (__addr), __idx = 1; __idx <= __objects;\
288 __p += (__s)->size, __idx++)
290 /* Determine object index from a given position */
291 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 return (p - addr) / s->size;
296 static inline size_t slab_ksize(const struct kmem_cache *s)
298 #ifdef CONFIG_SLUB_DEBUG
300 * Debugging requires use of the padding between object
301 * and whatever may come after it.
303 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
304 return s->object_size;
308 * If we have the need to store the freelist pointer
309 * back there or track user information then we can
310 * only use the space before that information.
312 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
315 * Else we can use all the padding etc for the allocation
320 static inline int order_objects(int order, unsigned long size, int reserved)
322 return ((PAGE_SIZE << order) - reserved) / size;
325 static inline struct kmem_cache_order_objects oo_make(int order,
326 unsigned long size, int reserved)
328 struct kmem_cache_order_objects x = {
329 (order << OO_SHIFT) + order_objects(order, size, reserved)
335 static inline int oo_order(struct kmem_cache_order_objects x)
337 return x.x >> OO_SHIFT;
340 static inline int oo_objects(struct kmem_cache_order_objects x)
342 return x.x & OO_MASK;
346 * Per slab locking using the pagelock
348 static __always_inline void slab_lock(struct page *page)
350 bit_spin_lock(PG_locked, &page->flags);
353 static __always_inline void slab_unlock(struct page *page)
355 __bit_spin_unlock(PG_locked, &page->flags);
358 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
361 tmp.counters = counters_new;
363 * page->counters can cover frozen/inuse/objects as well
364 * as page->_count. If we assign to ->counters directly
365 * we run the risk of losing updates to page->_count, so
366 * be careful and only assign to the fields we need.
368 page->frozen = tmp.frozen;
369 page->inuse = tmp.inuse;
370 page->objects = tmp.objects;
373 /* Interrupts must be disabled (for the fallback code to work right) */
374 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
375 void *freelist_old, unsigned long counters_old,
376 void *freelist_new, unsigned long counters_new,
379 VM_BUG_ON(!irqs_disabled());
380 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
381 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
382 if (s->flags & __CMPXCHG_DOUBLE) {
383 if (cmpxchg_double(&page->freelist, &page->counters,
384 freelist_old, counters_old,
385 freelist_new, counters_new))
391 if (page->freelist == freelist_old &&
392 page->counters == counters_old) {
393 page->freelist = freelist_new;
394 set_page_slub_counters(page, counters_new);
402 stat(s, CMPXCHG_DOUBLE_FAIL);
404 #ifdef SLUB_DEBUG_CMPXCHG
405 pr_info("%s %s: cmpxchg double redo ", n, s->name);
411 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
412 void *freelist_old, unsigned long counters_old,
413 void *freelist_new, unsigned long counters_new,
416 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
417 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
418 if (s->flags & __CMPXCHG_DOUBLE) {
419 if (cmpxchg_double(&page->freelist, &page->counters,
420 freelist_old, counters_old,
421 freelist_new, counters_new))
428 local_irq_save(flags);
430 if (page->freelist == freelist_old &&
431 page->counters == counters_old) {
432 page->freelist = freelist_new;
433 set_page_slub_counters(page, counters_new);
435 local_irq_restore(flags);
439 local_irq_restore(flags);
443 stat(s, CMPXCHG_DOUBLE_FAIL);
445 #ifdef SLUB_DEBUG_CMPXCHG
446 pr_info("%s %s: cmpxchg double redo ", n, s->name);
452 #ifdef CONFIG_SLUB_DEBUG
454 * Determine a map of object in use on a page.
456 * Node listlock must be held to guarantee that the page does
457 * not vanish from under us.
459 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
462 void *addr = page_address(page);
464 for (p = page->freelist; p; p = get_freepointer(s, p))
465 set_bit(slab_index(p, s, addr), map);
471 #ifdef CONFIG_SLUB_DEBUG_ON
472 static int slub_debug = DEBUG_DEFAULT_FLAGS;
474 static int slub_debug;
477 static char *slub_debug_slabs;
478 static int disable_higher_order_debug;
483 static void print_section(char *text, u8 *addr, unsigned int length)
485 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
489 static struct track *get_track(struct kmem_cache *s, void *object,
490 enum track_item alloc)
495 p = object + s->offset + sizeof(void *);
497 p = object + s->inuse;
502 static void set_track(struct kmem_cache *s, void *object,
503 enum track_item alloc, unsigned long addr)
505 struct track *p = get_track(s, object, alloc);
508 #ifdef CONFIG_STACKTRACE
509 struct stack_trace trace;
512 trace.nr_entries = 0;
513 trace.max_entries = TRACK_ADDRS_COUNT;
514 trace.entries = p->addrs;
516 save_stack_trace(&trace);
518 /* See rant in lockdep.c */
519 if (trace.nr_entries != 0 &&
520 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
523 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
527 p->cpu = smp_processor_id();
528 p->pid = current->pid;
531 memset(p, 0, sizeof(struct track));
534 static void init_tracking(struct kmem_cache *s, void *object)
536 if (!(s->flags & SLAB_STORE_USER))
539 set_track(s, object, TRACK_FREE, 0UL);
540 set_track(s, object, TRACK_ALLOC, 0UL);
543 static void print_track(const char *s, struct track *t)
548 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
549 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
550 #ifdef CONFIG_STACKTRACE
553 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
555 pr_err("\t%pS\n", (void *)t->addrs[i]);
562 static void print_tracking(struct kmem_cache *s, void *object)
564 if (!(s->flags & SLAB_STORE_USER))
567 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
568 print_track("Freed", get_track(s, object, TRACK_FREE));
571 static void print_page_info(struct page *page)
573 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
574 page, page->objects, page->inuse, page->freelist, page->flags);
578 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
580 struct va_format vaf;
586 pr_err("=============================================================================\n");
587 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
588 pr_err("-----------------------------------------------------------------------------\n\n");
590 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
594 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
596 struct va_format vaf;
602 pr_err("FIX %s: %pV\n", s->name, &vaf);
606 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
608 unsigned int off; /* Offset of last byte */
609 u8 *addr = page_address(page);
611 print_tracking(s, p);
613 print_page_info(page);
615 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
616 p, p - addr, get_freepointer(s, p));
619 print_section("Bytes b4 ", p - 16, 16);
621 print_section("Object ", p, min_t(unsigned long, s->object_size,
623 if (s->flags & SLAB_RED_ZONE)
624 print_section("Redzone ", p + s->object_size,
625 s->inuse - s->object_size);
628 off = s->offset + sizeof(void *);
632 if (s->flags & SLAB_STORE_USER)
633 off += 2 * sizeof(struct track);
636 /* Beginning of the filler is the free pointer */
637 print_section("Padding ", p + off, s->size - off);
642 static void object_err(struct kmem_cache *s, struct page *page,
643 u8 *object, char *reason)
645 slab_bug(s, "%s", reason);
646 print_trailer(s, page, object);
649 static void slab_err(struct kmem_cache *s, struct page *page,
650 const char *fmt, ...)
656 vsnprintf(buf, sizeof(buf), fmt, args);
658 slab_bug(s, "%s", buf);
659 print_page_info(page);
663 static void init_object(struct kmem_cache *s, void *object, u8 val)
667 if (s->flags & __OBJECT_POISON) {
668 memset(p, POISON_FREE, s->object_size - 1);
669 p[s->object_size - 1] = POISON_END;
672 if (s->flags & SLAB_RED_ZONE)
673 memset(p + s->object_size, val, s->inuse - s->object_size);
676 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
677 void *from, void *to)
679 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
680 memset(from, data, to - from);
683 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
684 u8 *object, char *what,
685 u8 *start, unsigned int value, unsigned int bytes)
690 fault = memchr_inv(start, value, bytes);
695 while (end > fault && end[-1] == value)
698 slab_bug(s, "%s overwritten", what);
699 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
700 fault, end - 1, fault[0], value);
701 print_trailer(s, page, object);
703 restore_bytes(s, what, value, fault, end);
711 * Bytes of the object to be managed.
712 * If the freepointer may overlay the object then the free
713 * pointer is the first word of the object.
715 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
718 * object + s->object_size
719 * Padding to reach word boundary. This is also used for Redzoning.
720 * Padding is extended by another word if Redzoning is enabled and
721 * object_size == inuse.
723 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
724 * 0xcc (RED_ACTIVE) for objects in use.
727 * Meta data starts here.
729 * A. Free pointer (if we cannot overwrite object on free)
730 * B. Tracking data for SLAB_STORE_USER
731 * C. Padding to reach required alignment boundary or at mininum
732 * one word if debugging is on to be able to detect writes
733 * before the word boundary.
735 * Padding is done using 0x5a (POISON_INUSE)
738 * Nothing is used beyond s->size.
740 * If slabcaches are merged then the object_size and inuse boundaries are mostly
741 * ignored. And therefore no slab options that rely on these boundaries
742 * may be used with merged slabcaches.
745 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
747 unsigned long off = s->inuse; /* The end of info */
750 /* Freepointer is placed after the object. */
751 off += sizeof(void *);
753 if (s->flags & SLAB_STORE_USER)
754 /* We also have user information there */
755 off += 2 * sizeof(struct track);
760 return check_bytes_and_report(s, page, p, "Object padding",
761 p + off, POISON_INUSE, s->size - off);
764 /* Check the pad bytes at the end of a slab page */
765 static int slab_pad_check(struct kmem_cache *s, struct page *page)
773 if (!(s->flags & SLAB_POISON))
776 start = page_address(page);
777 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
778 end = start + length;
779 remainder = length % s->size;
783 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
786 while (end > fault && end[-1] == POISON_INUSE)
789 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
790 print_section("Padding ", end - remainder, remainder);
792 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
796 static int check_object(struct kmem_cache *s, struct page *page,
797 void *object, u8 val)
800 u8 *endobject = object + s->object_size;
802 if (s->flags & SLAB_RED_ZONE) {
803 if (!check_bytes_and_report(s, page, object, "Redzone",
804 endobject, val, s->inuse - s->object_size))
807 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
808 check_bytes_and_report(s, page, p, "Alignment padding",
809 endobject, POISON_INUSE,
810 s->inuse - s->object_size);
814 if (s->flags & SLAB_POISON) {
815 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
816 (!check_bytes_and_report(s, page, p, "Poison", p,
817 POISON_FREE, s->object_size - 1) ||
818 !check_bytes_and_report(s, page, p, "Poison",
819 p + s->object_size - 1, POISON_END, 1)))
822 * check_pad_bytes cleans up on its own.
824 check_pad_bytes(s, page, p);
827 if (!s->offset && val == SLUB_RED_ACTIVE)
829 * Object and freepointer overlap. Cannot check
830 * freepointer while object is allocated.
834 /* Check free pointer validity */
835 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
836 object_err(s, page, p, "Freepointer corrupt");
838 * No choice but to zap it and thus lose the remainder
839 * of the free objects in this slab. May cause
840 * another error because the object count is now wrong.
842 set_freepointer(s, p, NULL);
848 static int check_slab(struct kmem_cache *s, struct page *page)
852 VM_BUG_ON(!irqs_disabled());
854 if (!PageSlab(page)) {
855 slab_err(s, page, "Not a valid slab page");
859 maxobj = order_objects(compound_order(page), s->size, s->reserved);
860 if (page->objects > maxobj) {
861 slab_err(s, page, "objects %u > max %u",
862 s->name, page->objects, maxobj);
865 if (page->inuse > page->objects) {
866 slab_err(s, page, "inuse %u > max %u",
867 s->name, page->inuse, page->objects);
870 /* Slab_pad_check fixes things up after itself */
871 slab_pad_check(s, page);
876 * Determine if a certain object on a page is on the freelist. Must hold the
877 * slab lock to guarantee that the chains are in a consistent state.
879 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
884 unsigned long max_objects;
887 while (fp && nr <= page->objects) {
890 if (!check_valid_pointer(s, page, fp)) {
892 object_err(s, page, object,
893 "Freechain corrupt");
894 set_freepointer(s, object, NULL);
896 slab_err(s, page, "Freepointer corrupt");
897 page->freelist = NULL;
898 page->inuse = page->objects;
899 slab_fix(s, "Freelist cleared");
905 fp = get_freepointer(s, object);
909 max_objects = order_objects(compound_order(page), s->size, s->reserved);
910 if (max_objects > MAX_OBJS_PER_PAGE)
911 max_objects = MAX_OBJS_PER_PAGE;
913 if (page->objects != max_objects) {
914 slab_err(s, page, "Wrong number of objects. Found %d but "
915 "should be %d", page->objects, max_objects);
916 page->objects = max_objects;
917 slab_fix(s, "Number of objects adjusted.");
919 if (page->inuse != page->objects - nr) {
920 slab_err(s, page, "Wrong object count. Counter is %d but "
921 "counted were %d", page->inuse, page->objects - nr);
922 page->inuse = page->objects - nr;
923 slab_fix(s, "Object count adjusted.");
925 return search == NULL;
928 static void trace(struct kmem_cache *s, struct page *page, void *object,
931 if (s->flags & SLAB_TRACE) {
932 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
934 alloc ? "alloc" : "free",
939 print_section("Object ", (void *)object,
947 * Tracking of fully allocated slabs for debugging purposes.
949 static void add_full(struct kmem_cache *s,
950 struct kmem_cache_node *n, struct page *page)
952 if (!(s->flags & SLAB_STORE_USER))
955 lockdep_assert_held(&n->list_lock);
956 list_add(&page->lru, &n->full);
959 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
961 if (!(s->flags & SLAB_STORE_USER))
964 lockdep_assert_held(&n->list_lock);
965 list_del(&page->lru);
968 /* Tracking of the number of slabs for debugging purposes */
969 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
971 struct kmem_cache_node *n = get_node(s, node);
973 return atomic_long_read(&n->nr_slabs);
976 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
978 return atomic_long_read(&n->nr_slabs);
981 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
983 struct kmem_cache_node *n = get_node(s, node);
986 * May be called early in order to allocate a slab for the
987 * kmem_cache_node structure. Solve the chicken-egg
988 * dilemma by deferring the increment of the count during
989 * bootstrap (see early_kmem_cache_node_alloc).
992 atomic_long_inc(&n->nr_slabs);
993 atomic_long_add(objects, &n->total_objects);
996 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
998 struct kmem_cache_node *n = get_node(s, node);
1000 atomic_long_dec(&n->nr_slabs);
1001 atomic_long_sub(objects, &n->total_objects);
1004 /* Object debug checks for alloc/free paths */
1005 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1008 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1011 init_object(s, object, SLUB_RED_INACTIVE);
1012 init_tracking(s, object);
1015 static noinline int alloc_debug_processing(struct kmem_cache *s,
1017 void *object, unsigned long addr)
1019 if (!check_slab(s, page))
1022 if (!check_valid_pointer(s, page, object)) {
1023 object_err(s, page, object, "Freelist Pointer check fails");
1027 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1030 /* Success perform special debug activities for allocs */
1031 if (s->flags & SLAB_STORE_USER)
1032 set_track(s, object, TRACK_ALLOC, addr);
1033 trace(s, page, object, 1);
1034 init_object(s, object, SLUB_RED_ACTIVE);
1038 if (PageSlab(page)) {
1040 * If this is a slab page then lets do the best we can
1041 * to avoid issues in the future. Marking all objects
1042 * as used avoids touching the remaining objects.
1044 slab_fix(s, "Marking all objects used");
1045 page->inuse = page->objects;
1046 page->freelist = NULL;
1051 static noinline struct kmem_cache_node *free_debug_processing(
1052 struct kmem_cache *s, struct page *page, void *object,
1053 unsigned long addr, unsigned long *flags)
1055 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1057 spin_lock_irqsave(&n->list_lock, *flags);
1060 if (!check_slab(s, page))
1063 if (!check_valid_pointer(s, page, object)) {
1064 slab_err(s, page, "Invalid object pointer 0x%p", object);
1068 if (on_freelist(s, page, object)) {
1069 object_err(s, page, object, "Object already free");
1073 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1076 if (unlikely(s != page->slab_cache)) {
1077 if (!PageSlab(page)) {
1078 slab_err(s, page, "Attempt to free object(0x%p) "
1079 "outside of slab", object);
1080 } else if (!page->slab_cache) {
1081 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1085 object_err(s, page, object,
1086 "page slab pointer corrupt.");
1090 if (s->flags & SLAB_STORE_USER)
1091 set_track(s, object, TRACK_FREE, addr);
1092 trace(s, page, object, 0);
1093 init_object(s, object, SLUB_RED_INACTIVE);
1097 * Keep node_lock to preserve integrity
1098 * until the object is actually freed
1104 spin_unlock_irqrestore(&n->list_lock, *flags);
1105 slab_fix(s, "Object at 0x%p not freed", object);
1109 static int __init setup_slub_debug(char *str)
1111 slub_debug = DEBUG_DEFAULT_FLAGS;
1112 if (*str++ != '=' || !*str)
1114 * No options specified. Switch on full debugging.
1120 * No options but restriction on slabs. This means full
1121 * debugging for slabs matching a pattern.
1125 if (tolower(*str) == 'o') {
1127 * Avoid enabling debugging on caches if its minimum order
1128 * would increase as a result.
1130 disable_higher_order_debug = 1;
1137 * Switch off all debugging measures.
1142 * Determine which debug features should be switched on
1144 for (; *str && *str != ','; str++) {
1145 switch (tolower(*str)) {
1147 slub_debug |= SLAB_DEBUG_FREE;
1150 slub_debug |= SLAB_RED_ZONE;
1153 slub_debug |= SLAB_POISON;
1156 slub_debug |= SLAB_STORE_USER;
1159 slub_debug |= SLAB_TRACE;
1162 slub_debug |= SLAB_FAILSLAB;
1165 pr_err("slub_debug option '%c' unknown. skipped\n",
1172 slub_debug_slabs = str + 1;
1177 __setup("slub_debug", setup_slub_debug);
1179 static unsigned long kmem_cache_flags(unsigned long object_size,
1180 unsigned long flags, const char *name,
1181 void (*ctor)(void *))
1184 * Enable debugging if selected on the kernel commandline.
1186 if (slub_debug && (!slub_debug_slabs || (name &&
1187 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1188 flags |= slub_debug;
1193 static inline void setup_object_debug(struct kmem_cache *s,
1194 struct page *page, void *object) {}
1196 static inline int alloc_debug_processing(struct kmem_cache *s,
1197 struct page *page, void *object, unsigned long addr) { return 0; }
1199 static inline struct kmem_cache_node *free_debug_processing(
1200 struct kmem_cache *s, struct page *page, void *object,
1201 unsigned long addr, unsigned long *flags) { return NULL; }
1203 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1205 static inline int check_object(struct kmem_cache *s, struct page *page,
1206 void *object, u8 val) { return 1; }
1207 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1208 struct page *page) {}
1209 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1210 struct page *page) {}
1211 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1212 unsigned long flags, const char *name,
1213 void (*ctor)(void *))
1217 #define slub_debug 0
1219 #define disable_higher_order_debug 0
1221 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1223 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1225 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1227 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1230 #endif /* CONFIG_SLUB_DEBUG */
1233 * Hooks for other subsystems that check memory allocations. In a typical
1234 * production configuration these hooks all should produce no code at all.
1236 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1238 kmemleak_alloc(ptr, size, 1, flags);
1241 static inline void kfree_hook(const void *x)
1246 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1248 flags &= gfp_allowed_mask;
1249 lockdep_trace_alloc(flags);
1250 might_sleep_if(flags & __GFP_WAIT);
1252 return should_failslab(s->object_size, flags, s->flags);
1255 static inline void slab_post_alloc_hook(struct kmem_cache *s,
1256 gfp_t flags, void *object)
1258 flags &= gfp_allowed_mask;
1259 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1260 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
1263 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1265 kmemleak_free_recursive(x, s->flags);
1268 * Trouble is that we may no longer disable interrupts in the fast path
1269 * So in order to make the debug calls that expect irqs to be
1270 * disabled we need to disable interrupts temporarily.
1272 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1274 unsigned long flags;
1276 local_irq_save(flags);
1277 kmemcheck_slab_free(s, x, s->object_size);
1278 debug_check_no_locks_freed(x, s->object_size);
1279 local_irq_restore(flags);
1282 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1283 debug_check_no_obj_freed(x, s->object_size);
1287 * Slab allocation and freeing
1289 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1290 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1293 int order = oo_order(oo);
1295 flags |= __GFP_NOTRACK;
1297 if (memcg_charge_slab(s, flags, order))
1300 if (node == NUMA_NO_NODE)
1301 page = alloc_pages(flags, order);
1303 page = alloc_pages_exact_node(node, flags, order);
1306 memcg_uncharge_slab(s, order);
1311 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1314 struct kmem_cache_order_objects oo = s->oo;
1317 flags &= gfp_allowed_mask;
1319 if (flags & __GFP_WAIT)
1322 flags |= s->allocflags;
1325 * Let the initial higher-order allocation fail under memory pressure
1326 * so we fall-back to the minimum order allocation.
1328 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1330 page = alloc_slab_page(s, alloc_gfp, node, oo);
1331 if (unlikely(!page)) {
1335 * Allocation may have failed due to fragmentation.
1336 * Try a lower order alloc if possible
1338 page = alloc_slab_page(s, alloc_gfp, node, oo);
1341 stat(s, ORDER_FALLBACK);
1344 if (kmemcheck_enabled && page
1345 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1346 int pages = 1 << oo_order(oo);
1348 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1351 * Objects from caches that have a constructor don't get
1352 * cleared when they're allocated, so we need to do it here.
1355 kmemcheck_mark_uninitialized_pages(page, pages);
1357 kmemcheck_mark_unallocated_pages(page, pages);
1360 if (flags & __GFP_WAIT)
1361 local_irq_disable();
1365 page->objects = oo_objects(oo);
1366 mod_zone_page_state(page_zone(page),
1367 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1368 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1374 static void setup_object(struct kmem_cache *s, struct page *page,
1377 setup_object_debug(s, page, object);
1378 if (unlikely(s->ctor))
1382 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1390 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1392 page = allocate_slab(s,
1393 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1397 order = compound_order(page);
1398 inc_slabs_node(s, page_to_nid(page), page->objects);
1399 page->slab_cache = s;
1400 __SetPageSlab(page);
1401 if (page->pfmemalloc)
1402 SetPageSlabPfmemalloc(page);
1404 start = page_address(page);
1406 if (unlikely(s->flags & SLAB_POISON))
1407 memset(start, POISON_INUSE, PAGE_SIZE << order);
1409 for_each_object_idx(p, idx, s, start, page->objects) {
1410 setup_object(s, page, p);
1411 if (likely(idx < page->objects))
1412 set_freepointer(s, p, p + s->size);
1414 set_freepointer(s, p, NULL);
1417 page->freelist = start;
1418 page->inuse = page->objects;
1424 static void __free_slab(struct kmem_cache *s, struct page *page)
1426 int order = compound_order(page);
1427 int pages = 1 << order;
1429 if (kmem_cache_debug(s)) {
1432 slab_pad_check(s, page);
1433 for_each_object(p, s, page_address(page),
1435 check_object(s, page, p, SLUB_RED_INACTIVE);
1438 kmemcheck_free_shadow(page, compound_order(page));
1440 mod_zone_page_state(page_zone(page),
1441 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1442 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1445 __ClearPageSlabPfmemalloc(page);
1446 __ClearPageSlab(page);
1448 page_mapcount_reset(page);
1449 if (current->reclaim_state)
1450 current->reclaim_state->reclaimed_slab += pages;
1451 __free_pages(page, order);
1452 memcg_uncharge_slab(s, order);
1455 #define need_reserve_slab_rcu \
1456 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1458 static void rcu_free_slab(struct rcu_head *h)
1462 if (need_reserve_slab_rcu)
1463 page = virt_to_head_page(h);
1465 page = container_of((struct list_head *)h, struct page, lru);
1467 __free_slab(page->slab_cache, page);
1470 static void free_slab(struct kmem_cache *s, struct page *page)
1472 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1473 struct rcu_head *head;
1475 if (need_reserve_slab_rcu) {
1476 int order = compound_order(page);
1477 int offset = (PAGE_SIZE << order) - s->reserved;
1479 VM_BUG_ON(s->reserved != sizeof(*head));
1480 head = page_address(page) + offset;
1483 * RCU free overloads the RCU head over the LRU
1485 head = (void *)&page->lru;
1488 call_rcu(head, rcu_free_slab);
1490 __free_slab(s, page);
1493 static void discard_slab(struct kmem_cache *s, struct page *page)
1495 dec_slabs_node(s, page_to_nid(page), page->objects);
1500 * Management of partially allocated slabs.
1503 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1506 if (tail == DEACTIVATE_TO_TAIL)
1507 list_add_tail(&page->lru, &n->partial);
1509 list_add(&page->lru, &n->partial);
1512 static inline void add_partial(struct kmem_cache_node *n,
1513 struct page *page, int tail)
1515 lockdep_assert_held(&n->list_lock);
1516 __add_partial(n, page, tail);
1520 __remove_partial(struct kmem_cache_node *n, struct page *page)
1522 list_del(&page->lru);
1526 static inline void remove_partial(struct kmem_cache_node *n,
1529 lockdep_assert_held(&n->list_lock);
1530 __remove_partial(n, page);
1534 * Remove slab from the partial list, freeze it and
1535 * return the pointer to the freelist.
1537 * Returns a list of objects or NULL if it fails.
1539 static inline void *acquire_slab(struct kmem_cache *s,
1540 struct kmem_cache_node *n, struct page *page,
1541 int mode, int *objects)
1544 unsigned long counters;
1547 lockdep_assert_held(&n->list_lock);
1550 * Zap the freelist and set the frozen bit.
1551 * The old freelist is the list of objects for the
1552 * per cpu allocation list.
1554 freelist = page->freelist;
1555 counters = page->counters;
1556 new.counters = counters;
1557 *objects = new.objects - new.inuse;
1559 new.inuse = page->objects;
1560 new.freelist = NULL;
1562 new.freelist = freelist;
1565 VM_BUG_ON(new.frozen);
1568 if (!__cmpxchg_double_slab(s, page,
1570 new.freelist, new.counters,
1574 remove_partial(n, page);
1579 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1580 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1583 * Try to allocate a partial slab from a specific node.
1585 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1586 struct kmem_cache_cpu *c, gfp_t flags)
1588 struct page *page, *page2;
1589 void *object = NULL;
1594 * Racy check. If we mistakenly see no partial slabs then we
1595 * just allocate an empty slab. If we mistakenly try to get a
1596 * partial slab and there is none available then get_partials()
1599 if (!n || !n->nr_partial)
1602 spin_lock(&n->list_lock);
1603 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1606 if (!pfmemalloc_match(page, flags))
1609 t = acquire_slab(s, n, page, object == NULL, &objects);
1613 available += objects;
1616 stat(s, ALLOC_FROM_PARTIAL);
1619 put_cpu_partial(s, page, 0);
1620 stat(s, CPU_PARTIAL_NODE);
1622 if (!kmem_cache_has_cpu_partial(s)
1623 || available > s->cpu_partial / 2)
1627 spin_unlock(&n->list_lock);
1632 * Get a page from somewhere. Search in increasing NUMA distances.
1634 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1635 struct kmem_cache_cpu *c)
1638 struct zonelist *zonelist;
1641 enum zone_type high_zoneidx = gfp_zone(flags);
1643 unsigned int cpuset_mems_cookie;
1646 * The defrag ratio allows a configuration of the tradeoffs between
1647 * inter node defragmentation and node local allocations. A lower
1648 * defrag_ratio increases the tendency to do local allocations
1649 * instead of attempting to obtain partial slabs from other nodes.
1651 * If the defrag_ratio is set to 0 then kmalloc() always
1652 * returns node local objects. If the ratio is higher then kmalloc()
1653 * may return off node objects because partial slabs are obtained
1654 * from other nodes and filled up.
1656 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1657 * defrag_ratio = 1000) then every (well almost) allocation will
1658 * first attempt to defrag slab caches on other nodes. This means
1659 * scanning over all nodes to look for partial slabs which may be
1660 * expensive if we do it every time we are trying to find a slab
1661 * with available objects.
1663 if (!s->remote_node_defrag_ratio ||
1664 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1668 cpuset_mems_cookie = read_mems_allowed_begin();
1669 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1670 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1671 struct kmem_cache_node *n;
1673 n = get_node(s, zone_to_nid(zone));
1675 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1676 n->nr_partial > s->min_partial) {
1677 object = get_partial_node(s, n, c, flags);
1680 * Don't check read_mems_allowed_retry()
1681 * here - if mems_allowed was updated in
1682 * parallel, that was a harmless race
1683 * between allocation and the cpuset
1690 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1696 * Get a partial page, lock it and return it.
1698 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1699 struct kmem_cache_cpu *c)
1702 int searchnode = (node == NUMA_NO_NODE) ? numa_mem_id() : node;
1704 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1705 if (object || node != NUMA_NO_NODE)
1708 return get_any_partial(s, flags, c);
1711 #ifdef CONFIG_PREEMPT
1713 * Calculate the next globally unique transaction for disambiguiation
1714 * during cmpxchg. The transactions start with the cpu number and are then
1715 * incremented by CONFIG_NR_CPUS.
1717 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1720 * No preemption supported therefore also no need to check for
1726 static inline unsigned long next_tid(unsigned long tid)
1728 return tid + TID_STEP;
1731 static inline unsigned int tid_to_cpu(unsigned long tid)
1733 return tid % TID_STEP;
1736 static inline unsigned long tid_to_event(unsigned long tid)
1738 return tid / TID_STEP;
1741 static inline unsigned int init_tid(int cpu)
1746 static inline void note_cmpxchg_failure(const char *n,
1747 const struct kmem_cache *s, unsigned long tid)
1749 #ifdef SLUB_DEBUG_CMPXCHG
1750 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1752 pr_info("%s %s: cmpxchg redo ", n, s->name);
1754 #ifdef CONFIG_PREEMPT
1755 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1756 pr_warn("due to cpu change %d -> %d\n",
1757 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1760 if (tid_to_event(tid) != tid_to_event(actual_tid))
1761 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1762 tid_to_event(tid), tid_to_event(actual_tid));
1764 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1765 actual_tid, tid, next_tid(tid));
1767 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1770 static void init_kmem_cache_cpus(struct kmem_cache *s)
1774 for_each_possible_cpu(cpu)
1775 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1779 * Remove the cpu slab
1781 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1784 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1785 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1787 enum slab_modes l = M_NONE, m = M_NONE;
1789 int tail = DEACTIVATE_TO_HEAD;
1793 if (page->freelist) {
1794 stat(s, DEACTIVATE_REMOTE_FREES);
1795 tail = DEACTIVATE_TO_TAIL;
1799 * Stage one: Free all available per cpu objects back
1800 * to the page freelist while it is still frozen. Leave the
1803 * There is no need to take the list->lock because the page
1806 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1808 unsigned long counters;
1811 prior = page->freelist;
1812 counters = page->counters;
1813 set_freepointer(s, freelist, prior);
1814 new.counters = counters;
1816 VM_BUG_ON(!new.frozen);
1818 } while (!__cmpxchg_double_slab(s, page,
1820 freelist, new.counters,
1821 "drain percpu freelist"));
1823 freelist = nextfree;
1827 * Stage two: Ensure that the page is unfrozen while the
1828 * list presence reflects the actual number of objects
1831 * We setup the list membership and then perform a cmpxchg
1832 * with the count. If there is a mismatch then the page
1833 * is not unfrozen but the page is on the wrong list.
1835 * Then we restart the process which may have to remove
1836 * the page from the list that we just put it on again
1837 * because the number of objects in the slab may have
1842 old.freelist = page->freelist;
1843 old.counters = page->counters;
1844 VM_BUG_ON(!old.frozen);
1846 /* Determine target state of the slab */
1847 new.counters = old.counters;
1850 set_freepointer(s, freelist, old.freelist);
1851 new.freelist = freelist;
1853 new.freelist = old.freelist;
1857 if (!new.inuse && n->nr_partial >= s->min_partial)
1859 else if (new.freelist) {
1864 * Taking the spinlock removes the possiblity
1865 * that acquire_slab() will see a slab page that
1868 spin_lock(&n->list_lock);
1872 if (kmem_cache_debug(s) && !lock) {
1875 * This also ensures that the scanning of full
1876 * slabs from diagnostic functions will not see
1879 spin_lock(&n->list_lock);
1887 remove_partial(n, page);
1889 else if (l == M_FULL)
1891 remove_full(s, n, page);
1893 if (m == M_PARTIAL) {
1895 add_partial(n, page, tail);
1898 } else if (m == M_FULL) {
1900 stat(s, DEACTIVATE_FULL);
1901 add_full(s, n, page);
1907 if (!__cmpxchg_double_slab(s, page,
1908 old.freelist, old.counters,
1909 new.freelist, new.counters,
1914 spin_unlock(&n->list_lock);
1917 stat(s, DEACTIVATE_EMPTY);
1918 discard_slab(s, page);
1924 * Unfreeze all the cpu partial slabs.
1926 * This function must be called with interrupts disabled
1927 * for the cpu using c (or some other guarantee must be there
1928 * to guarantee no concurrent accesses).
1930 static void unfreeze_partials(struct kmem_cache *s,
1931 struct kmem_cache_cpu *c)
1933 #ifdef CONFIG_SLUB_CPU_PARTIAL
1934 struct kmem_cache_node *n = NULL, *n2 = NULL;
1935 struct page *page, *discard_page = NULL;
1937 while ((page = c->partial)) {
1941 c->partial = page->next;
1943 n2 = get_node(s, page_to_nid(page));
1946 spin_unlock(&n->list_lock);
1949 spin_lock(&n->list_lock);
1954 old.freelist = page->freelist;
1955 old.counters = page->counters;
1956 VM_BUG_ON(!old.frozen);
1958 new.counters = old.counters;
1959 new.freelist = old.freelist;
1963 } while (!__cmpxchg_double_slab(s, page,
1964 old.freelist, old.counters,
1965 new.freelist, new.counters,
1966 "unfreezing slab"));
1968 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1969 page->next = discard_page;
1970 discard_page = page;
1972 add_partial(n, page, DEACTIVATE_TO_TAIL);
1973 stat(s, FREE_ADD_PARTIAL);
1978 spin_unlock(&n->list_lock);
1980 while (discard_page) {
1981 page = discard_page;
1982 discard_page = discard_page->next;
1984 stat(s, DEACTIVATE_EMPTY);
1985 discard_slab(s, page);
1992 * Put a page that was just frozen (in __slab_free) into a partial page
1993 * slot if available. This is done without interrupts disabled and without
1994 * preemption disabled. The cmpxchg is racy and may put the partial page
1995 * onto a random cpus partial slot.
1997 * If we did not find a slot then simply move all the partials to the
1998 * per node partial list.
2000 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2002 #ifdef CONFIG_SLUB_CPU_PARTIAL
2003 struct page *oldpage;
2010 oldpage = this_cpu_read(s->cpu_slab->partial);
2013 pobjects = oldpage->pobjects;
2014 pages = oldpage->pages;
2015 if (drain && pobjects > s->cpu_partial) {
2016 unsigned long flags;
2018 * partial array is full. Move the existing
2019 * set to the per node partial list.
2021 local_irq_save(flags);
2022 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2023 local_irq_restore(flags);
2027 stat(s, CPU_PARTIAL_DRAIN);
2032 pobjects += page->objects - page->inuse;
2034 page->pages = pages;
2035 page->pobjects = pobjects;
2036 page->next = oldpage;
2038 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2043 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2045 stat(s, CPUSLAB_FLUSH);
2046 deactivate_slab(s, c->page, c->freelist);
2048 c->tid = next_tid(c->tid);
2056 * Called from IPI handler with interrupts disabled.
2058 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2060 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2066 unfreeze_partials(s, c);
2070 static void flush_cpu_slab(void *d)
2072 struct kmem_cache *s = d;
2074 __flush_cpu_slab(s, smp_processor_id());
2077 static bool has_cpu_slab(int cpu, void *info)
2079 struct kmem_cache *s = info;
2080 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2082 return c->page || c->partial;
2085 static void flush_all(struct kmem_cache *s)
2087 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2091 * Check if the objects in a per cpu structure fit numa
2092 * locality expectations.
2094 static inline int node_match(struct page *page, int node)
2097 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2103 #ifdef CONFIG_SLUB_DEBUG
2104 static int count_free(struct page *page)
2106 return page->objects - page->inuse;
2109 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2111 return atomic_long_read(&n->total_objects);
2113 #endif /* CONFIG_SLUB_DEBUG */
2115 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2116 static unsigned long count_partial(struct kmem_cache_node *n,
2117 int (*get_count)(struct page *))
2119 unsigned long flags;
2120 unsigned long x = 0;
2123 spin_lock_irqsave(&n->list_lock, flags);
2124 list_for_each_entry(page, &n->partial, lru)
2125 x += get_count(page);
2126 spin_unlock_irqrestore(&n->list_lock, flags);
2129 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2131 static noinline void
2132 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2134 #ifdef CONFIG_SLUB_DEBUG
2135 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2136 DEFAULT_RATELIMIT_BURST);
2138 struct kmem_cache_node *n;
2140 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2143 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2145 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2146 s->name, s->object_size, s->size, oo_order(s->oo),
2149 if (oo_order(s->min) > get_order(s->object_size))
2150 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2153 for_each_kmem_cache_node(s, node, n) {
2154 unsigned long nr_slabs;
2155 unsigned long nr_objs;
2156 unsigned long nr_free;
2158 nr_free = count_partial(n, count_free);
2159 nr_slabs = node_nr_slabs(n);
2160 nr_objs = node_nr_objs(n);
2162 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2163 node, nr_slabs, nr_objs, nr_free);
2168 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2169 int node, struct kmem_cache_cpu **pc)
2172 struct kmem_cache_cpu *c = *pc;
2175 freelist = get_partial(s, flags, node, c);
2180 page = new_slab(s, flags, node);
2182 c = raw_cpu_ptr(s->cpu_slab);
2187 * No other reference to the page yet so we can
2188 * muck around with it freely without cmpxchg
2190 freelist = page->freelist;
2191 page->freelist = NULL;
2193 stat(s, ALLOC_SLAB);
2202 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2204 if (unlikely(PageSlabPfmemalloc(page)))
2205 return gfp_pfmemalloc_allowed(gfpflags);
2211 * Check the page->freelist of a page and either transfer the freelist to the
2212 * per cpu freelist or deactivate the page.
2214 * The page is still frozen if the return value is not NULL.
2216 * If this function returns NULL then the page has been unfrozen.
2218 * This function must be called with interrupt disabled.
2220 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2223 unsigned long counters;
2227 freelist = page->freelist;
2228 counters = page->counters;
2230 new.counters = counters;
2231 VM_BUG_ON(!new.frozen);
2233 new.inuse = page->objects;
2234 new.frozen = freelist != NULL;
2236 } while (!__cmpxchg_double_slab(s, page,
2245 * Slow path. The lockless freelist is empty or we need to perform
2248 * Processing is still very fast if new objects have been freed to the
2249 * regular freelist. In that case we simply take over the regular freelist
2250 * as the lockless freelist and zap the regular freelist.
2252 * If that is not working then we fall back to the partial lists. We take the
2253 * first element of the freelist as the object to allocate now and move the
2254 * rest of the freelist to the lockless freelist.
2256 * And if we were unable to get a new slab from the partial slab lists then
2257 * we need to allocate a new slab. This is the slowest path since it involves
2258 * a call to the page allocator and the setup of a new slab.
2260 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2261 unsigned long addr, struct kmem_cache_cpu *c)
2265 unsigned long flags;
2267 local_irq_save(flags);
2268 #ifdef CONFIG_PREEMPT
2270 * We may have been preempted and rescheduled on a different
2271 * cpu before disabling interrupts. Need to reload cpu area
2274 c = this_cpu_ptr(s->cpu_slab);
2282 if (unlikely(!node_match(page, node))) {
2283 stat(s, ALLOC_NODE_MISMATCH);
2284 deactivate_slab(s, page, c->freelist);
2291 * By rights, we should be searching for a slab page that was
2292 * PFMEMALLOC but right now, we are losing the pfmemalloc
2293 * information when the page leaves the per-cpu allocator
2295 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2296 deactivate_slab(s, page, c->freelist);
2302 /* must check again c->freelist in case of cpu migration or IRQ */
2303 freelist = c->freelist;
2307 freelist = get_freelist(s, page);
2311 stat(s, DEACTIVATE_BYPASS);
2315 stat(s, ALLOC_REFILL);
2319 * freelist is pointing to the list of objects to be used.
2320 * page is pointing to the page from which the objects are obtained.
2321 * That page must be frozen for per cpu allocations to work.
2323 VM_BUG_ON(!c->page->frozen);
2324 c->freelist = get_freepointer(s, freelist);
2325 c->tid = next_tid(c->tid);
2326 local_irq_restore(flags);
2332 page = c->page = c->partial;
2333 c->partial = page->next;
2334 stat(s, CPU_PARTIAL_ALLOC);
2339 freelist = new_slab_objects(s, gfpflags, node, &c);
2341 if (unlikely(!freelist)) {
2342 slab_out_of_memory(s, gfpflags, node);
2343 local_irq_restore(flags);
2348 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2351 /* Only entered in the debug case */
2352 if (kmem_cache_debug(s) &&
2353 !alloc_debug_processing(s, page, freelist, addr))
2354 goto new_slab; /* Slab failed checks. Next slab needed */
2356 deactivate_slab(s, page, get_freepointer(s, freelist));
2359 local_irq_restore(flags);
2364 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2365 * have the fastpath folded into their functions. So no function call
2366 * overhead for requests that can be satisfied on the fastpath.
2368 * The fastpath works by first checking if the lockless freelist can be used.
2369 * If not then __slab_alloc is called for slow processing.
2371 * Otherwise we can simply pick the next object from the lockless free list.
2373 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2374 gfp_t gfpflags, int node, unsigned long addr)
2377 struct kmem_cache_cpu *c;
2381 if (slab_pre_alloc_hook(s, gfpflags))
2384 s = memcg_kmem_get_cache(s, gfpflags);
2387 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2388 * enabled. We may switch back and forth between cpus while
2389 * reading from one cpu area. That does not matter as long
2390 * as we end up on the original cpu again when doing the cmpxchg.
2392 * Preemption is disabled for the retrieval of the tid because that
2393 * must occur from the current processor. We cannot allow rescheduling
2394 * on a different processor between the determination of the pointer
2395 * and the retrieval of the tid.
2398 c = this_cpu_ptr(s->cpu_slab);
2401 * The transaction ids are globally unique per cpu and per operation on
2402 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2403 * occurs on the right processor and that there was no operation on the
2404 * linked list in between.
2409 object = c->freelist;
2411 if (unlikely(!object || !node_match(page, node))) {
2412 object = __slab_alloc(s, gfpflags, node, addr, c);
2413 stat(s, ALLOC_SLOWPATH);
2415 void *next_object = get_freepointer_safe(s, object);
2418 * The cmpxchg will only match if there was no additional
2419 * operation and if we are on the right processor.
2421 * The cmpxchg does the following atomically (without lock
2423 * 1. Relocate first pointer to the current per cpu area.
2424 * 2. Verify that tid and freelist have not been changed
2425 * 3. If they were not changed replace tid and freelist
2427 * Since this is without lock semantics the protection is only
2428 * against code executing on this cpu *not* from access by
2431 if (unlikely(!this_cpu_cmpxchg_double(
2432 s->cpu_slab->freelist, s->cpu_slab->tid,
2434 next_object, next_tid(tid)))) {
2436 note_cmpxchg_failure("slab_alloc", s, tid);
2439 prefetch_freepointer(s, next_object);
2440 stat(s, ALLOC_FASTPATH);
2443 if (unlikely(gfpflags & __GFP_ZERO) && object)
2444 memset(object, 0, s->object_size);
2446 slab_post_alloc_hook(s, gfpflags, object);
2451 static __always_inline void *slab_alloc(struct kmem_cache *s,
2452 gfp_t gfpflags, unsigned long addr)
2454 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2457 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2459 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2461 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2466 EXPORT_SYMBOL(kmem_cache_alloc);
2468 #ifdef CONFIG_TRACING
2469 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2471 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2472 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2475 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2479 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2481 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2483 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2484 s->object_size, s->size, gfpflags, node);
2488 EXPORT_SYMBOL(kmem_cache_alloc_node);
2490 #ifdef CONFIG_TRACING
2491 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2493 int node, size_t size)
2495 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2497 trace_kmalloc_node(_RET_IP_, ret,
2498 size, s->size, gfpflags, node);
2501 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2506 * Slow patch handling. This may still be called frequently since objects
2507 * have a longer lifetime than the cpu slabs in most processing loads.
2509 * So we still attempt to reduce cache line usage. Just take the slab
2510 * lock and free the item. If there is no additional partial page
2511 * handling required then we can return immediately.
2513 static void __slab_free(struct kmem_cache *s, struct page *page,
2514 void *x, unsigned long addr)
2517 void **object = (void *)x;
2520 unsigned long counters;
2521 struct kmem_cache_node *n = NULL;
2522 unsigned long uninitialized_var(flags);
2524 stat(s, FREE_SLOWPATH);
2526 if (kmem_cache_debug(s) &&
2527 !(n = free_debug_processing(s, page, x, addr, &flags)))
2532 spin_unlock_irqrestore(&n->list_lock, flags);
2535 prior = page->freelist;
2536 counters = page->counters;
2537 set_freepointer(s, object, prior);
2538 new.counters = counters;
2539 was_frozen = new.frozen;
2541 if ((!new.inuse || !prior) && !was_frozen) {
2543 if (kmem_cache_has_cpu_partial(s) && !prior) {
2546 * Slab was on no list before and will be
2548 * We can defer the list move and instead
2553 } else { /* Needs to be taken off a list */
2555 n = get_node(s, page_to_nid(page));
2557 * Speculatively acquire the list_lock.
2558 * If the cmpxchg does not succeed then we may
2559 * drop the list_lock without any processing.
2561 * Otherwise the list_lock will synchronize with
2562 * other processors updating the list of slabs.
2564 spin_lock_irqsave(&n->list_lock, flags);
2569 } while (!cmpxchg_double_slab(s, page,
2571 object, new.counters,
2577 * If we just froze the page then put it onto the
2578 * per cpu partial list.
2580 if (new.frozen && !was_frozen) {
2581 put_cpu_partial(s, page, 1);
2582 stat(s, CPU_PARTIAL_FREE);
2585 * The list lock was not taken therefore no list
2586 * activity can be necessary.
2589 stat(s, FREE_FROZEN);
2593 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2597 * Objects left in the slab. If it was not on the partial list before
2600 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2601 if (kmem_cache_debug(s))
2602 remove_full(s, n, page);
2603 add_partial(n, page, DEACTIVATE_TO_TAIL);
2604 stat(s, FREE_ADD_PARTIAL);
2606 spin_unlock_irqrestore(&n->list_lock, flags);
2612 * Slab on the partial list.
2614 remove_partial(n, page);
2615 stat(s, FREE_REMOVE_PARTIAL);
2617 /* Slab must be on the full list */
2618 remove_full(s, n, page);
2621 spin_unlock_irqrestore(&n->list_lock, flags);
2623 discard_slab(s, page);
2627 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2628 * can perform fastpath freeing without additional function calls.
2630 * The fastpath is only possible if we are freeing to the current cpu slab
2631 * of this processor. This typically the case if we have just allocated
2634 * If fastpath is not possible then fall back to __slab_free where we deal
2635 * with all sorts of special processing.
2637 static __always_inline void slab_free(struct kmem_cache *s,
2638 struct page *page, void *x, unsigned long addr)
2640 void **object = (void *)x;
2641 struct kmem_cache_cpu *c;
2644 slab_free_hook(s, x);
2648 * Determine the currently cpus per cpu slab.
2649 * The cpu may change afterward. However that does not matter since
2650 * data is retrieved via this pointer. If we are on the same cpu
2651 * during the cmpxchg then the free will succedd.
2654 c = this_cpu_ptr(s->cpu_slab);
2659 if (likely(page == c->page)) {
2660 set_freepointer(s, object, c->freelist);
2662 if (unlikely(!this_cpu_cmpxchg_double(
2663 s->cpu_slab->freelist, s->cpu_slab->tid,
2665 object, next_tid(tid)))) {
2667 note_cmpxchg_failure("slab_free", s, tid);
2670 stat(s, FREE_FASTPATH);
2672 __slab_free(s, page, x, addr);
2676 void kmem_cache_free(struct kmem_cache *s, void *x)
2678 s = cache_from_obj(s, x);
2681 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2682 trace_kmem_cache_free(_RET_IP_, x);
2684 EXPORT_SYMBOL(kmem_cache_free);
2687 * Object placement in a slab is made very easy because we always start at
2688 * offset 0. If we tune the size of the object to the alignment then we can
2689 * get the required alignment by putting one properly sized object after
2692 * Notice that the allocation order determines the sizes of the per cpu
2693 * caches. Each processor has always one slab available for allocations.
2694 * Increasing the allocation order reduces the number of times that slabs
2695 * must be moved on and off the partial lists and is therefore a factor in
2700 * Mininum / Maximum order of slab pages. This influences locking overhead
2701 * and slab fragmentation. A higher order reduces the number of partial slabs
2702 * and increases the number of allocations possible without having to
2703 * take the list_lock.
2705 static int slub_min_order;
2706 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2707 static int slub_min_objects;
2710 * Merge control. If this is set then no merging of slab caches will occur.
2711 * (Could be removed. This was introduced to pacify the merge skeptics.)
2713 static int slub_nomerge;
2716 * Calculate the order of allocation given an slab object size.
2718 * The order of allocation has significant impact on performance and other
2719 * system components. Generally order 0 allocations should be preferred since
2720 * order 0 does not cause fragmentation in the page allocator. Larger objects
2721 * be problematic to put into order 0 slabs because there may be too much
2722 * unused space left. We go to a higher order if more than 1/16th of the slab
2725 * In order to reach satisfactory performance we must ensure that a minimum
2726 * number of objects is in one slab. Otherwise we may generate too much
2727 * activity on the partial lists which requires taking the list_lock. This is
2728 * less a concern for large slabs though which are rarely used.
2730 * slub_max_order specifies the order where we begin to stop considering the
2731 * number of objects in a slab as critical. If we reach slub_max_order then
2732 * we try to keep the page order as low as possible. So we accept more waste
2733 * of space in favor of a small page order.
2735 * Higher order allocations also allow the placement of more objects in a
2736 * slab and thereby reduce object handling overhead. If the user has
2737 * requested a higher mininum order then we start with that one instead of
2738 * the smallest order which will fit the object.
2740 static inline int slab_order(int size, int min_objects,
2741 int max_order, int fract_leftover, int reserved)
2745 int min_order = slub_min_order;
2747 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2748 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2750 for (order = max(min_order,
2751 fls(min_objects * size - 1) - PAGE_SHIFT);
2752 order <= max_order; order++) {
2754 unsigned long slab_size = PAGE_SIZE << order;
2756 if (slab_size < min_objects * size + reserved)
2759 rem = (slab_size - reserved) % size;
2761 if (rem <= slab_size / fract_leftover)
2769 static inline int calculate_order(int size, int reserved)
2777 * Attempt to find best configuration for a slab. This
2778 * works by first attempting to generate a layout with
2779 * the best configuration and backing off gradually.
2781 * First we reduce the acceptable waste in a slab. Then
2782 * we reduce the minimum objects required in a slab.
2784 min_objects = slub_min_objects;
2786 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2787 max_objects = order_objects(slub_max_order, size, reserved);
2788 min_objects = min(min_objects, max_objects);
2790 while (min_objects > 1) {
2792 while (fraction >= 4) {
2793 order = slab_order(size, min_objects,
2794 slub_max_order, fraction, reserved);
2795 if (order <= slub_max_order)
2803 * We were unable to place multiple objects in a slab. Now
2804 * lets see if we can place a single object there.
2806 order = slab_order(size, 1, slub_max_order, 1, reserved);
2807 if (order <= slub_max_order)
2811 * Doh this slab cannot be placed using slub_max_order.
2813 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2814 if (order < MAX_ORDER)
2820 init_kmem_cache_node(struct kmem_cache_node *n)
2823 spin_lock_init(&n->list_lock);
2824 INIT_LIST_HEAD(&n->partial);
2825 #ifdef CONFIG_SLUB_DEBUG
2826 atomic_long_set(&n->nr_slabs, 0);
2827 atomic_long_set(&n->total_objects, 0);
2828 INIT_LIST_HEAD(&n->full);
2832 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2834 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2835 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2838 * Must align to double word boundary for the double cmpxchg
2839 * instructions to work; see __pcpu_double_call_return_bool().
2841 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2842 2 * sizeof(void *));
2847 init_kmem_cache_cpus(s);
2852 static struct kmem_cache *kmem_cache_node;
2855 * No kmalloc_node yet so do it by hand. We know that this is the first
2856 * slab on the node for this slabcache. There are no concurrent accesses
2859 * Note that this function only works on the kmem_cache_node
2860 * when allocating for the kmem_cache_node. This is used for bootstrapping
2861 * memory on a fresh node that has no slab structures yet.
2863 static void early_kmem_cache_node_alloc(int node)
2866 struct kmem_cache_node *n;
2868 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2870 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2873 if (page_to_nid(page) != node) {
2874 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2875 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2880 page->freelist = get_freepointer(kmem_cache_node, n);
2883 kmem_cache_node->node[node] = n;
2884 #ifdef CONFIG_SLUB_DEBUG
2885 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2886 init_tracking(kmem_cache_node, n);
2888 init_kmem_cache_node(n);
2889 inc_slabs_node(kmem_cache_node, node, page->objects);
2892 * No locks need to be taken here as it has just been
2893 * initialized and there is no concurrent access.
2895 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2898 static void free_kmem_cache_nodes(struct kmem_cache *s)
2901 struct kmem_cache_node *n;
2903 for_each_kmem_cache_node(s, node, n) {
2904 kmem_cache_free(kmem_cache_node, n);
2905 s->node[node] = NULL;
2909 static int init_kmem_cache_nodes(struct kmem_cache *s)
2913 for_each_node_state(node, N_NORMAL_MEMORY) {
2914 struct kmem_cache_node *n;
2916 if (slab_state == DOWN) {
2917 early_kmem_cache_node_alloc(node);
2920 n = kmem_cache_alloc_node(kmem_cache_node,
2924 free_kmem_cache_nodes(s);
2929 init_kmem_cache_node(n);
2934 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2936 if (min < MIN_PARTIAL)
2938 else if (min > MAX_PARTIAL)
2940 s->min_partial = min;
2944 * calculate_sizes() determines the order and the distribution of data within
2947 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2949 unsigned long flags = s->flags;
2950 unsigned long size = s->object_size;
2954 * Round up object size to the next word boundary. We can only
2955 * place the free pointer at word boundaries and this determines
2956 * the possible location of the free pointer.
2958 size = ALIGN(size, sizeof(void *));
2960 #ifdef CONFIG_SLUB_DEBUG
2962 * Determine if we can poison the object itself. If the user of
2963 * the slab may touch the object after free or before allocation
2964 * then we should never poison the object itself.
2966 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2968 s->flags |= __OBJECT_POISON;
2970 s->flags &= ~__OBJECT_POISON;
2974 * If we are Redzoning then check if there is some space between the
2975 * end of the object and the free pointer. If not then add an
2976 * additional word to have some bytes to store Redzone information.
2978 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
2979 size += sizeof(void *);
2983 * With that we have determined the number of bytes in actual use
2984 * by the object. This is the potential offset to the free pointer.
2988 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2991 * Relocate free pointer after the object if it is not
2992 * permitted to overwrite the first word of the object on
2995 * This is the case if we do RCU, have a constructor or
2996 * destructor or are poisoning the objects.
2999 size += sizeof(void *);
3002 #ifdef CONFIG_SLUB_DEBUG
3003 if (flags & SLAB_STORE_USER)
3005 * Need to store information about allocs and frees after
3008 size += 2 * sizeof(struct track);
3010 if (flags & SLAB_RED_ZONE)
3012 * Add some empty padding so that we can catch
3013 * overwrites from earlier objects rather than let
3014 * tracking information or the free pointer be
3015 * corrupted if a user writes before the start
3018 size += sizeof(void *);
3022 * SLUB stores one object immediately after another beginning from
3023 * offset 0. In order to align the objects we have to simply size
3024 * each object to conform to the alignment.
3026 size = ALIGN(size, s->align);
3028 if (forced_order >= 0)
3029 order = forced_order;
3031 order = calculate_order(size, s->reserved);
3038 s->allocflags |= __GFP_COMP;
3040 if (s->flags & SLAB_CACHE_DMA)
3041 s->allocflags |= GFP_DMA;
3043 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3044 s->allocflags |= __GFP_RECLAIMABLE;
3047 * Determine the number of objects per slab
3049 s->oo = oo_make(order, size, s->reserved);
3050 s->min = oo_make(get_order(size), size, s->reserved);
3051 if (oo_objects(s->oo) > oo_objects(s->max))
3054 return !!oo_objects(s->oo);
3057 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3059 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3062 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3063 s->reserved = sizeof(struct rcu_head);
3065 if (!calculate_sizes(s, -1))
3067 if (disable_higher_order_debug) {
3069 * Disable debugging flags that store metadata if the min slab
3072 if (get_order(s->size) > get_order(s->object_size)) {
3073 s->flags &= ~DEBUG_METADATA_FLAGS;
3075 if (!calculate_sizes(s, -1))
3080 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3081 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3082 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3083 /* Enable fast mode */
3084 s->flags |= __CMPXCHG_DOUBLE;
3088 * The larger the object size is, the more pages we want on the partial
3089 * list to avoid pounding the page allocator excessively.
3091 set_min_partial(s, ilog2(s->size) / 2);
3094 * cpu_partial determined the maximum number of objects kept in the
3095 * per cpu partial lists of a processor.
3097 * Per cpu partial lists mainly contain slabs that just have one
3098 * object freed. If they are used for allocation then they can be
3099 * filled up again with minimal effort. The slab will never hit the
3100 * per node partial lists and therefore no locking will be required.
3102 * This setting also determines
3104 * A) The number of objects from per cpu partial slabs dumped to the
3105 * per node list when we reach the limit.
3106 * B) The number of objects in cpu partial slabs to extract from the
3107 * per node list when we run out of per cpu objects. We only fetch
3108 * 50% to keep some capacity around for frees.
3110 if (!kmem_cache_has_cpu_partial(s))
3112 else if (s->size >= PAGE_SIZE)
3114 else if (s->size >= 1024)
3116 else if (s->size >= 256)
3117 s->cpu_partial = 13;
3119 s->cpu_partial = 30;
3122 s->remote_node_defrag_ratio = 1000;
3124 if (!init_kmem_cache_nodes(s))
3127 if (alloc_kmem_cache_cpus(s))
3130 free_kmem_cache_nodes(s);
3132 if (flags & SLAB_PANIC)
3133 panic("Cannot create slab %s size=%lu realsize=%u "
3134 "order=%u offset=%u flags=%lx\n",
3135 s->name, (unsigned long)s->size, s->size,
3136 oo_order(s->oo), s->offset, flags);
3140 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3143 #ifdef CONFIG_SLUB_DEBUG
3144 void *addr = page_address(page);
3146 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3147 sizeof(long), GFP_ATOMIC);
3150 slab_err(s, page, text, s->name);
3153 get_map(s, page, map);
3154 for_each_object(p, s, addr, page->objects) {
3156 if (!test_bit(slab_index(p, s, addr), map)) {
3157 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3158 print_tracking(s, p);
3167 * Attempt to free all partial slabs on a node.
3168 * This is called from kmem_cache_close(). We must be the last thread
3169 * using the cache and therefore we do not need to lock anymore.
3171 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3173 struct page *page, *h;
3175 list_for_each_entry_safe(page, h, &n->partial, lru) {
3177 __remove_partial(n, page);
3178 discard_slab(s, page);
3180 list_slab_objects(s, page,
3181 "Objects remaining in %s on kmem_cache_close()");
3187 * Release all resources used by a slab cache.
3189 static inline int kmem_cache_close(struct kmem_cache *s)
3192 struct kmem_cache_node *n;
3195 /* Attempt to free all objects */
3196 for_each_kmem_cache_node(s, node, n) {
3198 if (n->nr_partial || slabs_node(s, node))
3201 free_percpu(s->cpu_slab);
3202 free_kmem_cache_nodes(s);
3206 int __kmem_cache_shutdown(struct kmem_cache *s)
3208 return kmem_cache_close(s);
3211 /********************************************************************
3213 *******************************************************************/
3215 static int __init setup_slub_min_order(char *str)
3217 get_option(&str, &slub_min_order);
3222 __setup("slub_min_order=", setup_slub_min_order);
3224 static int __init setup_slub_max_order(char *str)
3226 get_option(&str, &slub_max_order);
3227 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3232 __setup("slub_max_order=", setup_slub_max_order);
3234 static int __init setup_slub_min_objects(char *str)
3236 get_option(&str, &slub_min_objects);
3241 __setup("slub_min_objects=", setup_slub_min_objects);
3243 static int __init setup_slub_nomerge(char *str)
3249 __setup("slub_nomerge", setup_slub_nomerge);
3251 void *__kmalloc(size_t size, gfp_t flags)
3253 struct kmem_cache *s;
3256 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3257 return kmalloc_large(size, flags);
3259 s = kmalloc_slab(size, flags);
3261 if (unlikely(ZERO_OR_NULL_PTR(s)))
3264 ret = slab_alloc(s, flags, _RET_IP_);
3266 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3270 EXPORT_SYMBOL(__kmalloc);
3273 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3278 flags |= __GFP_COMP | __GFP_NOTRACK;
3279 page = alloc_kmem_pages_node(node, flags, get_order(size));
3281 ptr = page_address(page);
3283 kmalloc_large_node_hook(ptr, size, flags);
3287 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3289 struct kmem_cache *s;
3292 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3293 ret = kmalloc_large_node(size, flags, node);
3295 trace_kmalloc_node(_RET_IP_, ret,
3296 size, PAGE_SIZE << get_order(size),
3302 s = kmalloc_slab(size, flags);
3304 if (unlikely(ZERO_OR_NULL_PTR(s)))
3307 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3309 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3313 EXPORT_SYMBOL(__kmalloc_node);
3316 size_t ksize(const void *object)
3320 if (unlikely(object == ZERO_SIZE_PTR))
3323 page = virt_to_head_page(object);
3325 if (unlikely(!PageSlab(page))) {
3326 WARN_ON(!PageCompound(page));
3327 return PAGE_SIZE << compound_order(page);
3330 return slab_ksize(page->slab_cache);
3332 EXPORT_SYMBOL(ksize);
3334 void kfree(const void *x)
3337 void *object = (void *)x;
3339 trace_kfree(_RET_IP_, x);
3341 if (unlikely(ZERO_OR_NULL_PTR(x)))
3344 page = virt_to_head_page(x);
3345 if (unlikely(!PageSlab(page))) {
3346 BUG_ON(!PageCompound(page));
3348 __free_kmem_pages(page, compound_order(page));
3351 slab_free(page->slab_cache, page, object, _RET_IP_);
3353 EXPORT_SYMBOL(kfree);
3356 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3357 * the remaining slabs by the number of items in use. The slabs with the
3358 * most items in use come first. New allocations will then fill those up
3359 * and thus they can be removed from the partial lists.
3361 * The slabs with the least items are placed last. This results in them
3362 * being allocated from last increasing the chance that the last objects
3363 * are freed in them.
3365 int __kmem_cache_shrink(struct kmem_cache *s)
3369 struct kmem_cache_node *n;
3372 int objects = oo_objects(s->max);
3373 struct list_head *slabs_by_inuse =
3374 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3375 unsigned long flags;
3377 if (!slabs_by_inuse)
3381 for_each_kmem_cache_node(s, node, n) {
3385 for (i = 0; i < objects; i++)
3386 INIT_LIST_HEAD(slabs_by_inuse + i);
3388 spin_lock_irqsave(&n->list_lock, flags);
3391 * Build lists indexed by the items in use in each slab.
3393 * Note that concurrent frees may occur while we hold the
3394 * list_lock. page->inuse here is the upper limit.
3396 list_for_each_entry_safe(page, t, &n->partial, lru) {
3397 list_move(&page->lru, slabs_by_inuse + page->inuse);
3403 * Rebuild the partial list with the slabs filled up most
3404 * first and the least used slabs at the end.
3406 for (i = objects - 1; i > 0; i--)
3407 list_splice(slabs_by_inuse + i, n->partial.prev);
3409 spin_unlock_irqrestore(&n->list_lock, flags);
3411 /* Release empty slabs */
3412 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3413 discard_slab(s, page);
3416 kfree(slabs_by_inuse);
3420 static int slab_mem_going_offline_callback(void *arg)
3422 struct kmem_cache *s;
3424 mutex_lock(&slab_mutex);
3425 list_for_each_entry(s, &slab_caches, list)
3426 __kmem_cache_shrink(s);
3427 mutex_unlock(&slab_mutex);
3432 static void slab_mem_offline_callback(void *arg)
3434 struct kmem_cache_node *n;
3435 struct kmem_cache *s;
3436 struct memory_notify *marg = arg;
3439 offline_node = marg->status_change_nid_normal;
3442 * If the node still has available memory. we need kmem_cache_node
3445 if (offline_node < 0)
3448 mutex_lock(&slab_mutex);
3449 list_for_each_entry(s, &slab_caches, list) {
3450 n = get_node(s, offline_node);
3453 * if n->nr_slabs > 0, slabs still exist on the node
3454 * that is going down. We were unable to free them,
3455 * and offline_pages() function shouldn't call this
3456 * callback. So, we must fail.
3458 BUG_ON(slabs_node(s, offline_node));
3460 s->node[offline_node] = NULL;
3461 kmem_cache_free(kmem_cache_node, n);
3464 mutex_unlock(&slab_mutex);
3467 static int slab_mem_going_online_callback(void *arg)
3469 struct kmem_cache_node *n;
3470 struct kmem_cache *s;
3471 struct memory_notify *marg = arg;
3472 int nid = marg->status_change_nid_normal;
3476 * If the node's memory is already available, then kmem_cache_node is
3477 * already created. Nothing to do.
3483 * We are bringing a node online. No memory is available yet. We must
3484 * allocate a kmem_cache_node structure in order to bring the node
3487 mutex_lock(&slab_mutex);
3488 list_for_each_entry(s, &slab_caches, list) {
3490 * XXX: kmem_cache_alloc_node will fallback to other nodes
3491 * since memory is not yet available from the node that
3494 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3499 init_kmem_cache_node(n);
3503 mutex_unlock(&slab_mutex);
3507 static int slab_memory_callback(struct notifier_block *self,
3508 unsigned long action, void *arg)
3513 case MEM_GOING_ONLINE:
3514 ret = slab_mem_going_online_callback(arg);
3516 case MEM_GOING_OFFLINE:
3517 ret = slab_mem_going_offline_callback(arg);
3520 case MEM_CANCEL_ONLINE:
3521 slab_mem_offline_callback(arg);
3524 case MEM_CANCEL_OFFLINE:
3528 ret = notifier_from_errno(ret);
3534 static struct notifier_block slab_memory_callback_nb = {
3535 .notifier_call = slab_memory_callback,
3536 .priority = SLAB_CALLBACK_PRI,
3539 /********************************************************************
3540 * Basic setup of slabs
3541 *******************************************************************/
3544 * Used for early kmem_cache structures that were allocated using
3545 * the page allocator. Allocate them properly then fix up the pointers
3546 * that may be pointing to the wrong kmem_cache structure.
3549 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3552 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3553 struct kmem_cache_node *n;
3555 memcpy(s, static_cache, kmem_cache->object_size);
3558 * This runs very early, and only the boot processor is supposed to be
3559 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3562 __flush_cpu_slab(s, smp_processor_id());
3563 for_each_kmem_cache_node(s, node, n) {
3566 list_for_each_entry(p, &n->partial, lru)
3569 #ifdef CONFIG_SLUB_DEBUG
3570 list_for_each_entry(p, &n->full, lru)
3574 list_add(&s->list, &slab_caches);
3578 void __init kmem_cache_init(void)
3580 static __initdata struct kmem_cache boot_kmem_cache,
3581 boot_kmem_cache_node;
3583 if (debug_guardpage_minorder())
3586 kmem_cache_node = &boot_kmem_cache_node;
3587 kmem_cache = &boot_kmem_cache;
3589 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3590 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3592 register_hotmemory_notifier(&slab_memory_callback_nb);
3594 /* Able to allocate the per node structures */
3595 slab_state = PARTIAL;
3597 create_boot_cache(kmem_cache, "kmem_cache",
3598 offsetof(struct kmem_cache, node) +
3599 nr_node_ids * sizeof(struct kmem_cache_node *),
3600 SLAB_HWCACHE_ALIGN);
3602 kmem_cache = bootstrap(&boot_kmem_cache);
3605 * Allocate kmem_cache_node properly from the kmem_cache slab.
3606 * kmem_cache_node is separately allocated so no need to
3607 * update any list pointers.
3609 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3611 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3612 create_kmalloc_caches(0);
3615 register_cpu_notifier(&slab_notifier);
3618 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3620 slub_min_order, slub_max_order, slub_min_objects,
3621 nr_cpu_ids, nr_node_ids);
3624 void __init kmem_cache_init_late(void)
3629 * Find a mergeable slab cache
3631 static int slab_unmergeable(struct kmem_cache *s)
3633 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3636 if (!is_root_cache(s))
3643 * We may have set a slab to be unmergeable during bootstrap.
3645 if (s->refcount < 0)
3651 static struct kmem_cache *find_mergeable(size_t size, size_t align,
3652 unsigned long flags, const char *name, void (*ctor)(void *))
3654 struct kmem_cache *s;
3656 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3662 size = ALIGN(size, sizeof(void *));
3663 align = calculate_alignment(flags, align, size);
3664 size = ALIGN(size, align);
3665 flags = kmem_cache_flags(size, flags, name, NULL);
3667 list_for_each_entry(s, &slab_caches, list) {
3668 if (slab_unmergeable(s))
3674 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3677 * Check if alignment is compatible.
3678 * Courtesy of Adrian Drzewiecki
3680 if ((s->size & ~(align - 1)) != s->size)
3683 if (s->size - size >= sizeof(void *))
3692 __kmem_cache_alias(const char *name, size_t size, size_t align,
3693 unsigned long flags, void (*ctor)(void *))
3695 struct kmem_cache *s;
3697 s = find_mergeable(size, align, flags, name, ctor);
3700 struct kmem_cache *c;
3705 * Adjust the object sizes so that we clear
3706 * the complete object on kzalloc.
3708 s->object_size = max(s->object_size, (int)size);
3709 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3711 for_each_memcg_cache_index(i) {
3712 c = cache_from_memcg_idx(s, i);
3715 c->object_size = s->object_size;
3716 c->inuse = max_t(int, c->inuse,
3717 ALIGN(size, sizeof(void *)));
3720 if (sysfs_slab_alias(s, name)) {
3729 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3733 err = kmem_cache_open(s, flags);
3737 /* Mutex is not taken during early boot */
3738 if (slab_state <= UP)
3741 memcg_propagate_slab_attrs(s);
3742 err = sysfs_slab_add(s);
3744 kmem_cache_close(s);
3751 * Use the cpu notifier to insure that the cpu slabs are flushed when
3754 static int slab_cpuup_callback(struct notifier_block *nfb,
3755 unsigned long action, void *hcpu)
3757 long cpu = (long)hcpu;
3758 struct kmem_cache *s;
3759 unsigned long flags;
3762 case CPU_UP_CANCELED:
3763 case CPU_UP_CANCELED_FROZEN:
3765 case CPU_DEAD_FROZEN:
3766 mutex_lock(&slab_mutex);
3767 list_for_each_entry(s, &slab_caches, list) {
3768 local_irq_save(flags);
3769 __flush_cpu_slab(s, cpu);
3770 local_irq_restore(flags);
3772 mutex_unlock(&slab_mutex);
3780 static struct notifier_block slab_notifier = {
3781 .notifier_call = slab_cpuup_callback
3786 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3788 struct kmem_cache *s;
3791 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3792 return kmalloc_large(size, gfpflags);
3794 s = kmalloc_slab(size, gfpflags);
3796 if (unlikely(ZERO_OR_NULL_PTR(s)))
3799 ret = slab_alloc(s, gfpflags, caller);
3801 /* Honor the call site pointer we received. */
3802 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3808 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3809 int node, unsigned long caller)
3811 struct kmem_cache *s;
3814 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3815 ret = kmalloc_large_node(size, gfpflags, node);
3817 trace_kmalloc_node(caller, ret,
3818 size, PAGE_SIZE << get_order(size),
3824 s = kmalloc_slab(size, gfpflags);
3826 if (unlikely(ZERO_OR_NULL_PTR(s)))
3829 ret = slab_alloc_node(s, gfpflags, node, caller);
3831 /* Honor the call site pointer we received. */
3832 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3839 static int count_inuse(struct page *page)
3844 static int count_total(struct page *page)
3846 return page->objects;
3850 #ifdef CONFIG_SLUB_DEBUG
3851 static int validate_slab(struct kmem_cache *s, struct page *page,
3855 void *addr = page_address(page);
3857 if (!check_slab(s, page) ||
3858 !on_freelist(s, page, NULL))
3861 /* Now we know that a valid freelist exists */
3862 bitmap_zero(map, page->objects);
3864 get_map(s, page, map);
3865 for_each_object(p, s, addr, page->objects) {
3866 if (test_bit(slab_index(p, s, addr), map))
3867 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3871 for_each_object(p, s, addr, page->objects)
3872 if (!test_bit(slab_index(p, s, addr), map))
3873 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3878 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3882 validate_slab(s, page, map);
3886 static int validate_slab_node(struct kmem_cache *s,
3887 struct kmem_cache_node *n, unsigned long *map)
3889 unsigned long count = 0;
3891 unsigned long flags;
3893 spin_lock_irqsave(&n->list_lock, flags);
3895 list_for_each_entry(page, &n->partial, lru) {
3896 validate_slab_slab(s, page, map);
3899 if (count != n->nr_partial)
3900 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3901 s->name, count, n->nr_partial);
3903 if (!(s->flags & SLAB_STORE_USER))
3906 list_for_each_entry(page, &n->full, lru) {
3907 validate_slab_slab(s, page, map);
3910 if (count != atomic_long_read(&n->nr_slabs))
3911 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3912 s->name, count, atomic_long_read(&n->nr_slabs));
3915 spin_unlock_irqrestore(&n->list_lock, flags);
3919 static long validate_slab_cache(struct kmem_cache *s)
3922 unsigned long count = 0;
3923 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3924 sizeof(unsigned long), GFP_KERNEL);
3925 struct kmem_cache_node *n;
3931 for_each_kmem_cache_node(s, node, n)
3932 count += validate_slab_node(s, n, map);
3937 * Generate lists of code addresses where slabcache objects are allocated
3942 unsigned long count;
3949 DECLARE_BITMAP(cpus, NR_CPUS);
3955 unsigned long count;
3956 struct location *loc;
3959 static void free_loc_track(struct loc_track *t)
3962 free_pages((unsigned long)t->loc,
3963 get_order(sizeof(struct location) * t->max));
3966 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3971 order = get_order(sizeof(struct location) * max);
3973 l = (void *)__get_free_pages(flags, order);
3978 memcpy(l, t->loc, sizeof(struct location) * t->count);
3986 static int add_location(struct loc_track *t, struct kmem_cache *s,
3987 const struct track *track)
3989 long start, end, pos;
3991 unsigned long caddr;
3992 unsigned long age = jiffies - track->when;
3998 pos = start + (end - start + 1) / 2;
4001 * There is nothing at "end". If we end up there
4002 * we need to add something to before end.
4007 caddr = t->loc[pos].addr;
4008 if (track->addr == caddr) {
4014 if (age < l->min_time)
4016 if (age > l->max_time)
4019 if (track->pid < l->min_pid)
4020 l->min_pid = track->pid;
4021 if (track->pid > l->max_pid)
4022 l->max_pid = track->pid;
4024 cpumask_set_cpu(track->cpu,
4025 to_cpumask(l->cpus));
4027 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4031 if (track->addr < caddr)
4038 * Not found. Insert new tracking element.
4040 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4046 (t->count - pos) * sizeof(struct location));
4049 l->addr = track->addr;
4053 l->min_pid = track->pid;
4054 l->max_pid = track->pid;
4055 cpumask_clear(to_cpumask(l->cpus));
4056 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4057 nodes_clear(l->nodes);
4058 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4062 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4063 struct page *page, enum track_item alloc,
4066 void *addr = page_address(page);
4069 bitmap_zero(map, page->objects);
4070 get_map(s, page, map);
4072 for_each_object(p, s, addr, page->objects)
4073 if (!test_bit(slab_index(p, s, addr), map))
4074 add_location(t, s, get_track(s, p, alloc));
4077 static int list_locations(struct kmem_cache *s, char *buf,
4078 enum track_item alloc)
4082 struct loc_track t = { 0, 0, NULL };
4084 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4085 sizeof(unsigned long), GFP_KERNEL);
4086 struct kmem_cache_node *n;
4088 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4091 return sprintf(buf, "Out of memory\n");
4093 /* Push back cpu slabs */
4096 for_each_kmem_cache_node(s, node, n) {
4097 unsigned long flags;
4100 if (!atomic_long_read(&n->nr_slabs))
4103 spin_lock_irqsave(&n->list_lock, flags);
4104 list_for_each_entry(page, &n->partial, lru)
4105 process_slab(&t, s, page, alloc, map);
4106 list_for_each_entry(page, &n->full, lru)
4107 process_slab(&t, s, page, alloc, map);
4108 spin_unlock_irqrestore(&n->list_lock, flags);
4111 for (i = 0; i < t.count; i++) {
4112 struct location *l = &t.loc[i];
4114 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4116 len += sprintf(buf + len, "%7ld ", l->count);
4119 len += sprintf(buf + len, "%pS", (void *)l->addr);
4121 len += sprintf(buf + len, "<not-available>");
4123 if (l->sum_time != l->min_time) {
4124 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4126 (long)div_u64(l->sum_time, l->count),
4129 len += sprintf(buf + len, " age=%ld",
4132 if (l->min_pid != l->max_pid)
4133 len += sprintf(buf + len, " pid=%ld-%ld",
4134 l->min_pid, l->max_pid);
4136 len += sprintf(buf + len, " pid=%ld",
4139 if (num_online_cpus() > 1 &&
4140 !cpumask_empty(to_cpumask(l->cpus)) &&
4141 len < PAGE_SIZE - 60) {
4142 len += sprintf(buf + len, " cpus=");
4143 len += cpulist_scnprintf(buf + len,
4144 PAGE_SIZE - len - 50,
4145 to_cpumask(l->cpus));
4148 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4149 len < PAGE_SIZE - 60) {
4150 len += sprintf(buf + len, " nodes=");
4151 len += nodelist_scnprintf(buf + len,
4152 PAGE_SIZE - len - 50,
4156 len += sprintf(buf + len, "\n");
4162 len += sprintf(buf, "No data\n");
4167 #ifdef SLUB_RESILIENCY_TEST
4168 static void __init resiliency_test(void)
4172 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4174 pr_err("SLUB resiliency testing\n");
4175 pr_err("-----------------------\n");
4176 pr_err("A. Corruption after allocation\n");
4178 p = kzalloc(16, GFP_KERNEL);
4180 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4183 validate_slab_cache(kmalloc_caches[4]);
4185 /* Hmmm... The next two are dangerous */
4186 p = kzalloc(32, GFP_KERNEL);
4187 p[32 + sizeof(void *)] = 0x34;
4188 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4190 pr_err("If allocated object is overwritten then not detectable\n\n");
4192 validate_slab_cache(kmalloc_caches[5]);
4193 p = kzalloc(64, GFP_KERNEL);
4194 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4196 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4198 pr_err("If allocated object is overwritten then not detectable\n\n");
4199 validate_slab_cache(kmalloc_caches[6]);
4201 pr_err("\nB. Corruption after free\n");
4202 p = kzalloc(128, GFP_KERNEL);
4205 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4206 validate_slab_cache(kmalloc_caches[7]);
4208 p = kzalloc(256, GFP_KERNEL);
4211 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4212 validate_slab_cache(kmalloc_caches[8]);
4214 p = kzalloc(512, GFP_KERNEL);
4217 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4218 validate_slab_cache(kmalloc_caches[9]);
4222 static void resiliency_test(void) {};
4227 enum slab_stat_type {
4228 SL_ALL, /* All slabs */
4229 SL_PARTIAL, /* Only partially allocated slabs */
4230 SL_CPU, /* Only slabs used for cpu caches */
4231 SL_OBJECTS, /* Determine allocated objects not slabs */
4232 SL_TOTAL /* Determine object capacity not slabs */
4235 #define SO_ALL (1 << SL_ALL)
4236 #define SO_PARTIAL (1 << SL_PARTIAL)
4237 #define SO_CPU (1 << SL_CPU)
4238 #define SO_OBJECTS (1 << SL_OBJECTS)
4239 #define SO_TOTAL (1 << SL_TOTAL)
4241 static ssize_t show_slab_objects(struct kmem_cache *s,
4242 char *buf, unsigned long flags)
4244 unsigned long total = 0;
4247 unsigned long *nodes;
4249 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4253 if (flags & SO_CPU) {
4256 for_each_possible_cpu(cpu) {
4257 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4262 page = ACCESS_ONCE(c->page);
4266 node = page_to_nid(page);
4267 if (flags & SO_TOTAL)
4269 else if (flags & SO_OBJECTS)
4277 page = ACCESS_ONCE(c->partial);
4279 node = page_to_nid(page);
4280 if (flags & SO_TOTAL)
4282 else if (flags & SO_OBJECTS)
4293 #ifdef CONFIG_SLUB_DEBUG
4294 if (flags & SO_ALL) {
4295 struct kmem_cache_node *n;
4297 for_each_kmem_cache_node(s, node, n) {
4299 if (flags & SO_TOTAL)
4300 x = atomic_long_read(&n->total_objects);
4301 else if (flags & SO_OBJECTS)
4302 x = atomic_long_read(&n->total_objects) -
4303 count_partial(n, count_free);
4305 x = atomic_long_read(&n->nr_slabs);
4312 if (flags & SO_PARTIAL) {
4313 struct kmem_cache_node *n;
4315 for_each_kmem_cache_node(s, node, n) {
4316 if (flags & SO_TOTAL)
4317 x = count_partial(n, count_total);
4318 else if (flags & SO_OBJECTS)
4319 x = count_partial(n, count_inuse);
4326 x = sprintf(buf, "%lu", total);
4328 for (node = 0; node < nr_node_ids; node++)
4330 x += sprintf(buf + x, " N%d=%lu",
4335 return x + sprintf(buf + x, "\n");
4338 #ifdef CONFIG_SLUB_DEBUG
4339 static int any_slab_objects(struct kmem_cache *s)
4342 struct kmem_cache_node *n;
4344 for_each_kmem_cache_node(s, node, n)
4345 if (atomic_long_read(&n->total_objects))
4352 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4353 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4355 struct slab_attribute {
4356 struct attribute attr;
4357 ssize_t (*show)(struct kmem_cache *s, char *buf);
4358 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4361 #define SLAB_ATTR_RO(_name) \
4362 static struct slab_attribute _name##_attr = \
4363 __ATTR(_name, 0400, _name##_show, NULL)
4365 #define SLAB_ATTR(_name) \
4366 static struct slab_attribute _name##_attr = \
4367 __ATTR(_name, 0600, _name##_show, _name##_store)
4369 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4371 return sprintf(buf, "%d\n", s->size);
4373 SLAB_ATTR_RO(slab_size);
4375 static ssize_t align_show(struct kmem_cache *s, char *buf)
4377 return sprintf(buf, "%d\n", s->align);
4379 SLAB_ATTR_RO(align);
4381 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4383 return sprintf(buf, "%d\n", s->object_size);
4385 SLAB_ATTR_RO(object_size);
4387 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4389 return sprintf(buf, "%d\n", oo_objects(s->oo));
4391 SLAB_ATTR_RO(objs_per_slab);
4393 static ssize_t order_store(struct kmem_cache *s,
4394 const char *buf, size_t length)
4396 unsigned long order;
4399 err = kstrtoul(buf, 10, &order);
4403 if (order > slub_max_order || order < slub_min_order)
4406 calculate_sizes(s, order);
4410 static ssize_t order_show(struct kmem_cache *s, char *buf)
4412 return sprintf(buf, "%d\n", oo_order(s->oo));
4416 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4418 return sprintf(buf, "%lu\n", s->min_partial);
4421 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4427 err = kstrtoul(buf, 10, &min);
4431 set_min_partial(s, min);
4434 SLAB_ATTR(min_partial);
4436 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4438 return sprintf(buf, "%u\n", s->cpu_partial);
4441 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4444 unsigned long objects;
4447 err = kstrtoul(buf, 10, &objects);
4450 if (objects && !kmem_cache_has_cpu_partial(s))
4453 s->cpu_partial = objects;
4457 SLAB_ATTR(cpu_partial);
4459 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4463 return sprintf(buf, "%pS\n", s->ctor);
4467 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4469 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4471 SLAB_ATTR_RO(aliases);
4473 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4475 return show_slab_objects(s, buf, SO_PARTIAL);
4477 SLAB_ATTR_RO(partial);
4479 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4481 return show_slab_objects(s, buf, SO_CPU);
4483 SLAB_ATTR_RO(cpu_slabs);
4485 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4487 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4489 SLAB_ATTR_RO(objects);
4491 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4493 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4495 SLAB_ATTR_RO(objects_partial);
4497 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4504 for_each_online_cpu(cpu) {
4505 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4508 pages += page->pages;
4509 objects += page->pobjects;
4513 len = sprintf(buf, "%d(%d)", objects, pages);
4516 for_each_online_cpu(cpu) {
4517 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4519 if (page && len < PAGE_SIZE - 20)
4520 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4521 page->pobjects, page->pages);
4524 return len + sprintf(buf + len, "\n");
4526 SLAB_ATTR_RO(slabs_cpu_partial);
4528 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4530 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4533 static ssize_t reclaim_account_store(struct kmem_cache *s,
4534 const char *buf, size_t length)
4536 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4538 s->flags |= SLAB_RECLAIM_ACCOUNT;
4541 SLAB_ATTR(reclaim_account);
4543 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4545 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4547 SLAB_ATTR_RO(hwcache_align);
4549 #ifdef CONFIG_ZONE_DMA
4550 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4552 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4554 SLAB_ATTR_RO(cache_dma);
4557 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4559 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4561 SLAB_ATTR_RO(destroy_by_rcu);
4563 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4565 return sprintf(buf, "%d\n", s->reserved);
4567 SLAB_ATTR_RO(reserved);
4569 #ifdef CONFIG_SLUB_DEBUG
4570 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4572 return show_slab_objects(s, buf, SO_ALL);
4574 SLAB_ATTR_RO(slabs);
4576 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4578 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4580 SLAB_ATTR_RO(total_objects);
4582 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4584 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4587 static ssize_t sanity_checks_store(struct kmem_cache *s,
4588 const char *buf, size_t length)
4590 s->flags &= ~SLAB_DEBUG_FREE;
4591 if (buf[0] == '1') {
4592 s->flags &= ~__CMPXCHG_DOUBLE;
4593 s->flags |= SLAB_DEBUG_FREE;
4597 SLAB_ATTR(sanity_checks);
4599 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4601 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4604 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4607 s->flags &= ~SLAB_TRACE;
4608 if (buf[0] == '1') {
4609 s->flags &= ~__CMPXCHG_DOUBLE;
4610 s->flags |= SLAB_TRACE;
4616 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4618 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4621 static ssize_t red_zone_store(struct kmem_cache *s,
4622 const char *buf, size_t length)
4624 if (any_slab_objects(s))
4627 s->flags &= ~SLAB_RED_ZONE;
4628 if (buf[0] == '1') {
4629 s->flags &= ~__CMPXCHG_DOUBLE;
4630 s->flags |= SLAB_RED_ZONE;
4632 calculate_sizes(s, -1);
4635 SLAB_ATTR(red_zone);
4637 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4639 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4642 static ssize_t poison_store(struct kmem_cache *s,
4643 const char *buf, size_t length)
4645 if (any_slab_objects(s))
4648 s->flags &= ~SLAB_POISON;
4649 if (buf[0] == '1') {
4650 s->flags &= ~__CMPXCHG_DOUBLE;
4651 s->flags |= SLAB_POISON;
4653 calculate_sizes(s, -1);
4658 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4660 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4663 static ssize_t store_user_store(struct kmem_cache *s,
4664 const char *buf, size_t length)
4666 if (any_slab_objects(s))
4669 s->flags &= ~SLAB_STORE_USER;
4670 if (buf[0] == '1') {
4671 s->flags &= ~__CMPXCHG_DOUBLE;
4672 s->flags |= SLAB_STORE_USER;
4674 calculate_sizes(s, -1);
4677 SLAB_ATTR(store_user);
4679 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4684 static ssize_t validate_store(struct kmem_cache *s,
4685 const char *buf, size_t length)
4689 if (buf[0] == '1') {
4690 ret = validate_slab_cache(s);
4696 SLAB_ATTR(validate);
4698 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4700 if (!(s->flags & SLAB_STORE_USER))
4702 return list_locations(s, buf, TRACK_ALLOC);
4704 SLAB_ATTR_RO(alloc_calls);
4706 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4708 if (!(s->flags & SLAB_STORE_USER))
4710 return list_locations(s, buf, TRACK_FREE);
4712 SLAB_ATTR_RO(free_calls);
4713 #endif /* CONFIG_SLUB_DEBUG */
4715 #ifdef CONFIG_FAILSLAB
4716 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4718 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4721 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4724 s->flags &= ~SLAB_FAILSLAB;
4726 s->flags |= SLAB_FAILSLAB;
4729 SLAB_ATTR(failslab);
4732 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4737 static ssize_t shrink_store(struct kmem_cache *s,
4738 const char *buf, size_t length)
4740 if (buf[0] == '1') {
4741 int rc = kmem_cache_shrink(s);
4752 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4754 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4757 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4758 const char *buf, size_t length)
4760 unsigned long ratio;
4763 err = kstrtoul(buf, 10, &ratio);
4768 s->remote_node_defrag_ratio = ratio * 10;
4772 SLAB_ATTR(remote_node_defrag_ratio);
4775 #ifdef CONFIG_SLUB_STATS
4776 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4778 unsigned long sum = 0;
4781 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4786 for_each_online_cpu(cpu) {
4787 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4793 len = sprintf(buf, "%lu", sum);
4796 for_each_online_cpu(cpu) {
4797 if (data[cpu] && len < PAGE_SIZE - 20)
4798 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4802 return len + sprintf(buf + len, "\n");
4805 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4809 for_each_online_cpu(cpu)
4810 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4813 #define STAT_ATTR(si, text) \
4814 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4816 return show_stat(s, buf, si); \
4818 static ssize_t text##_store(struct kmem_cache *s, \
4819 const char *buf, size_t length) \
4821 if (buf[0] != '0') \
4823 clear_stat(s, si); \
4828 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4829 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4830 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4831 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4832 STAT_ATTR(FREE_FROZEN, free_frozen);
4833 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4834 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4835 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4836 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4837 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4838 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4839 STAT_ATTR(FREE_SLAB, free_slab);
4840 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4841 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4842 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4843 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4844 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4845 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4846 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4847 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4848 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4849 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4850 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4851 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4852 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4853 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4856 static struct attribute *slab_attrs[] = {
4857 &slab_size_attr.attr,
4858 &object_size_attr.attr,
4859 &objs_per_slab_attr.attr,
4861 &min_partial_attr.attr,
4862 &cpu_partial_attr.attr,
4864 &objects_partial_attr.attr,
4866 &cpu_slabs_attr.attr,
4870 &hwcache_align_attr.attr,
4871 &reclaim_account_attr.attr,
4872 &destroy_by_rcu_attr.attr,
4874 &reserved_attr.attr,
4875 &slabs_cpu_partial_attr.attr,
4876 #ifdef CONFIG_SLUB_DEBUG
4877 &total_objects_attr.attr,
4879 &sanity_checks_attr.attr,
4881 &red_zone_attr.attr,
4883 &store_user_attr.attr,
4884 &validate_attr.attr,
4885 &alloc_calls_attr.attr,
4886 &free_calls_attr.attr,
4888 #ifdef CONFIG_ZONE_DMA
4889 &cache_dma_attr.attr,
4892 &remote_node_defrag_ratio_attr.attr,
4894 #ifdef CONFIG_SLUB_STATS
4895 &alloc_fastpath_attr.attr,
4896 &alloc_slowpath_attr.attr,
4897 &free_fastpath_attr.attr,
4898 &free_slowpath_attr.attr,
4899 &free_frozen_attr.attr,
4900 &free_add_partial_attr.attr,
4901 &free_remove_partial_attr.attr,
4902 &alloc_from_partial_attr.attr,
4903 &alloc_slab_attr.attr,
4904 &alloc_refill_attr.attr,
4905 &alloc_node_mismatch_attr.attr,
4906 &free_slab_attr.attr,
4907 &cpuslab_flush_attr.attr,
4908 &deactivate_full_attr.attr,
4909 &deactivate_empty_attr.attr,
4910 &deactivate_to_head_attr.attr,
4911 &deactivate_to_tail_attr.attr,
4912 &deactivate_remote_frees_attr.attr,
4913 &deactivate_bypass_attr.attr,
4914 &order_fallback_attr.attr,
4915 &cmpxchg_double_fail_attr.attr,
4916 &cmpxchg_double_cpu_fail_attr.attr,
4917 &cpu_partial_alloc_attr.attr,
4918 &cpu_partial_free_attr.attr,
4919 &cpu_partial_node_attr.attr,
4920 &cpu_partial_drain_attr.attr,
4922 #ifdef CONFIG_FAILSLAB
4923 &failslab_attr.attr,
4929 static struct attribute_group slab_attr_group = {
4930 .attrs = slab_attrs,
4933 static ssize_t slab_attr_show(struct kobject *kobj,
4934 struct attribute *attr,
4937 struct slab_attribute *attribute;
4938 struct kmem_cache *s;
4941 attribute = to_slab_attr(attr);
4944 if (!attribute->show)
4947 err = attribute->show(s, buf);
4952 static ssize_t slab_attr_store(struct kobject *kobj,
4953 struct attribute *attr,
4954 const char *buf, size_t len)
4956 struct slab_attribute *attribute;
4957 struct kmem_cache *s;
4960 attribute = to_slab_attr(attr);
4963 if (!attribute->store)
4966 err = attribute->store(s, buf, len);
4967 #ifdef CONFIG_MEMCG_KMEM
4968 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
4971 mutex_lock(&slab_mutex);
4972 if (s->max_attr_size < len)
4973 s->max_attr_size = len;
4976 * This is a best effort propagation, so this function's return
4977 * value will be determined by the parent cache only. This is
4978 * basically because not all attributes will have a well
4979 * defined semantics for rollbacks - most of the actions will
4980 * have permanent effects.
4982 * Returning the error value of any of the children that fail
4983 * is not 100 % defined, in the sense that users seeing the
4984 * error code won't be able to know anything about the state of
4987 * Only returning the error code for the parent cache at least
4988 * has well defined semantics. The cache being written to
4989 * directly either failed or succeeded, in which case we loop
4990 * through the descendants with best-effort propagation.
4992 for_each_memcg_cache_index(i) {
4993 struct kmem_cache *c = cache_from_memcg_idx(s, i);
4995 attribute->store(c, buf, len);
4997 mutex_unlock(&slab_mutex);
5003 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5005 #ifdef CONFIG_MEMCG_KMEM
5007 char *buffer = NULL;
5008 struct kmem_cache *root_cache;
5010 if (is_root_cache(s))
5013 root_cache = s->memcg_params->root_cache;
5016 * This mean this cache had no attribute written. Therefore, no point
5017 * in copying default values around
5019 if (!root_cache->max_attr_size)
5022 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5025 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5027 if (!attr || !attr->store || !attr->show)
5031 * It is really bad that we have to allocate here, so we will
5032 * do it only as a fallback. If we actually allocate, though,
5033 * we can just use the allocated buffer until the end.
5035 * Most of the slub attributes will tend to be very small in
5036 * size, but sysfs allows buffers up to a page, so they can
5037 * theoretically happen.
5041 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5044 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5045 if (WARN_ON(!buffer))
5050 attr->show(root_cache, buf);
5051 attr->store(s, buf, strlen(buf));
5055 free_page((unsigned long)buffer);
5059 static void kmem_cache_release(struct kobject *k)
5061 slab_kmem_cache_release(to_slab(k));
5064 static const struct sysfs_ops slab_sysfs_ops = {
5065 .show = slab_attr_show,
5066 .store = slab_attr_store,
5069 static struct kobj_type slab_ktype = {
5070 .sysfs_ops = &slab_sysfs_ops,
5071 .release = kmem_cache_release,
5074 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5076 struct kobj_type *ktype = get_ktype(kobj);
5078 if (ktype == &slab_ktype)
5083 static const struct kset_uevent_ops slab_uevent_ops = {
5084 .filter = uevent_filter,
5087 static struct kset *slab_kset;
5089 static inline struct kset *cache_kset(struct kmem_cache *s)
5091 #ifdef CONFIG_MEMCG_KMEM
5092 if (!is_root_cache(s))
5093 return s->memcg_params->root_cache->memcg_kset;
5098 #define ID_STR_LENGTH 64
5100 /* Create a unique string id for a slab cache:
5102 * Format :[flags-]size
5104 static char *create_unique_id(struct kmem_cache *s)
5106 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5113 * First flags affecting slabcache operations. We will only
5114 * get here for aliasable slabs so we do not need to support
5115 * too many flags. The flags here must cover all flags that
5116 * are matched during merging to guarantee that the id is
5119 if (s->flags & SLAB_CACHE_DMA)
5121 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5123 if (s->flags & SLAB_DEBUG_FREE)
5125 if (!(s->flags & SLAB_NOTRACK))
5129 p += sprintf(p, "%07d", s->size);
5131 BUG_ON(p > name + ID_STR_LENGTH - 1);
5135 static int sysfs_slab_add(struct kmem_cache *s)
5139 int unmergeable = slab_unmergeable(s);
5143 * Slabcache can never be merged so we can use the name proper.
5144 * This is typically the case for debug situations. In that
5145 * case we can catch duplicate names easily.
5147 sysfs_remove_link(&slab_kset->kobj, s->name);
5151 * Create a unique name for the slab as a target
5154 name = create_unique_id(s);
5157 s->kobj.kset = cache_kset(s);
5158 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5162 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5166 #ifdef CONFIG_MEMCG_KMEM
5167 if (is_root_cache(s)) {
5168 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5169 if (!s->memcg_kset) {
5176 kobject_uevent(&s->kobj, KOBJ_ADD);
5178 /* Setup first alias */
5179 sysfs_slab_alias(s, s->name);
5186 kobject_del(&s->kobj);
5188 kobject_put(&s->kobj);
5192 void sysfs_slab_remove(struct kmem_cache *s)
5194 if (slab_state < FULL)
5196 * Sysfs has not been setup yet so no need to remove the
5201 #ifdef CONFIG_MEMCG_KMEM
5202 kset_unregister(s->memcg_kset);
5204 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5205 kobject_del(&s->kobj);
5206 kobject_put(&s->kobj);
5210 * Need to buffer aliases during bootup until sysfs becomes
5211 * available lest we lose that information.
5213 struct saved_alias {
5214 struct kmem_cache *s;
5216 struct saved_alias *next;
5219 static struct saved_alias *alias_list;
5221 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5223 struct saved_alias *al;
5225 if (slab_state == FULL) {
5227 * If we have a leftover link then remove it.
5229 sysfs_remove_link(&slab_kset->kobj, name);
5230 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5233 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5239 al->next = alias_list;
5244 static int __init slab_sysfs_init(void)
5246 struct kmem_cache *s;
5249 mutex_lock(&slab_mutex);
5251 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5253 mutex_unlock(&slab_mutex);
5254 pr_err("Cannot register slab subsystem.\n");
5260 list_for_each_entry(s, &slab_caches, list) {
5261 err = sysfs_slab_add(s);
5263 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5267 while (alias_list) {
5268 struct saved_alias *al = alias_list;
5270 alias_list = alias_list->next;
5271 err = sysfs_slab_alias(al->s, al->name);
5273 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5278 mutex_unlock(&slab_mutex);
5283 __initcall(slab_sysfs_init);
5284 #endif /* CONFIG_SYSFS */
5287 * The /proc/slabinfo ABI
5289 #ifdef CONFIG_SLABINFO
5290 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5292 unsigned long nr_slabs = 0;
5293 unsigned long nr_objs = 0;
5294 unsigned long nr_free = 0;
5296 struct kmem_cache_node *n;
5298 for_each_kmem_cache_node(s, node, n) {
5299 nr_slabs += node_nr_slabs(n);
5300 nr_objs += node_nr_objs(n);
5301 nr_free += count_partial(n, count_free);
5304 sinfo->active_objs = nr_objs - nr_free;
5305 sinfo->num_objs = nr_objs;
5306 sinfo->active_slabs = nr_slabs;
5307 sinfo->num_slabs = nr_slabs;
5308 sinfo->objects_per_slab = oo_objects(s->oo);
5309 sinfo->cache_order = oo_order(s->oo);
5312 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5316 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5317 size_t count, loff_t *ppos)
5321 #endif /* CONFIG_SLABINFO */