2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
129 #ifdef CONFIG_SLUB_CPU_PARTIAL
130 return !kmem_cache_debug(s);
137 * Issues still to be resolved:
139 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
141 * - Variable sizing of the per node arrays
144 /* Enable to test recovery from slab corruption on boot */
145 #undef SLUB_RESILIENCY_TEST
147 /* Enable to log cmpxchg failures */
148 #undef SLUB_DEBUG_CMPXCHG
151 * Mininum number of partial slabs. These will be left on the partial
152 * lists even if they are empty. kmem_cache_shrink may reclaim them.
154 #define MIN_PARTIAL 5
157 * Maximum number of desirable partial slabs.
158 * The existence of more partial slabs makes kmem_cache_shrink
159 * sort the partial list by the number of objects in use.
161 #define MAX_PARTIAL 10
163 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
164 SLAB_POISON | SLAB_STORE_USER)
167 * These debug flags cannot use CMPXCHG because there might be consistency
168 * issues when checking or reading debug information
170 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
175 * Debugging flags that require metadata to be stored in the slab. These get
176 * disabled when slub_debug=O is used and a cache's min order increases with
179 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
182 #define OO_MASK ((1 << OO_SHIFT) - 1)
183 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
185 /* Internal SLUB flags */
186 #define __OBJECT_POISON 0x80000000UL /* Poison object */
187 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
190 static struct notifier_block slab_notifier;
194 * Tracking user of a slab.
196 #define TRACK_ADDRS_COUNT 16
198 unsigned long addr; /* Called from address */
199 #ifdef CONFIG_STACKTRACE
200 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202 int cpu; /* Was running on cpu */
203 int pid; /* Pid context */
204 unsigned long when; /* When did the operation occur */
207 enum track_item { TRACK_ALLOC, TRACK_FREE };
210 static int sysfs_slab_add(struct kmem_cache *);
211 static int sysfs_slab_alias(struct kmem_cache *, const char *);
212 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
214 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
215 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
217 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
220 static inline void stat(const struct kmem_cache *s, enum stat_item si)
222 #ifdef CONFIG_SLUB_STATS
224 * The rmw is racy on a preemptible kernel but this is acceptable, so
225 * avoid this_cpu_add()'s irq-disable overhead.
227 raw_cpu_inc(s->cpu_slab->stat[si]);
231 /********************************************************************
232 * Core slab cache functions
233 *******************************************************************/
235 /* Verify that a pointer has an address that is valid within a slab page */
236 static inline int check_valid_pointer(struct kmem_cache *s,
237 struct page *page, const void *object)
244 base = page_address(page);
245 if (object < base || object >= base + page->objects * s->size ||
246 (object - base) % s->size) {
253 static inline void *get_freepointer(struct kmem_cache *s, void *object)
255 return *(void **)(object + s->offset);
258 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
260 prefetch(object + s->offset);
263 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
267 #ifdef CONFIG_DEBUG_PAGEALLOC
268 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
270 p = get_freepointer(s, object);
275 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
277 *(void **)(object + s->offset) = fp;
280 /* Loop over all objects in a slab */
281 #define for_each_object(__p, __s, __addr, __objects) \
282 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
285 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
286 for (__p = (__addr), __idx = 1; __idx <= __objects;\
287 __p += (__s)->size, __idx++)
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 static inline int order_objects(int order, unsigned long size, int reserved)
297 return ((PAGE_SIZE << order) - reserved) / size;
300 static inline struct kmem_cache_order_objects oo_make(int order,
301 unsigned long size, int reserved)
303 struct kmem_cache_order_objects x = {
304 (order << OO_SHIFT) + order_objects(order, size, reserved)
310 static inline int oo_order(struct kmem_cache_order_objects x)
312 return x.x >> OO_SHIFT;
315 static inline int oo_objects(struct kmem_cache_order_objects x)
317 return x.x & OO_MASK;
321 * Per slab locking using the pagelock
323 static __always_inline void slab_lock(struct page *page)
325 VM_BUG_ON_PAGE(PageTail(page), page);
326 bit_spin_lock(PG_locked, &page->flags);
329 static __always_inline void slab_unlock(struct page *page)
331 VM_BUG_ON_PAGE(PageTail(page), page);
332 __bit_spin_unlock(PG_locked, &page->flags);
335 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
338 tmp.counters = counters_new;
340 * page->counters can cover frozen/inuse/objects as well
341 * as page->_count. If we assign to ->counters directly
342 * we run the risk of losing updates to page->_count, so
343 * be careful and only assign to the fields we need.
345 page->frozen = tmp.frozen;
346 page->inuse = tmp.inuse;
347 page->objects = tmp.objects;
350 /* Interrupts must be disabled (for the fallback code to work right) */
351 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
352 void *freelist_old, unsigned long counters_old,
353 void *freelist_new, unsigned long counters_new,
356 VM_BUG_ON(!irqs_disabled());
357 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
358 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
359 if (s->flags & __CMPXCHG_DOUBLE) {
360 if (cmpxchg_double(&page->freelist, &page->counters,
361 freelist_old, counters_old,
362 freelist_new, counters_new))
368 if (page->freelist == freelist_old &&
369 page->counters == counters_old) {
370 page->freelist = freelist_new;
371 set_page_slub_counters(page, counters_new);
379 stat(s, CMPXCHG_DOUBLE_FAIL);
381 #ifdef SLUB_DEBUG_CMPXCHG
382 pr_info("%s %s: cmpxchg double redo ", n, s->name);
388 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
389 void *freelist_old, unsigned long counters_old,
390 void *freelist_new, unsigned long counters_new,
393 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
394 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
395 if (s->flags & __CMPXCHG_DOUBLE) {
396 if (cmpxchg_double(&page->freelist, &page->counters,
397 freelist_old, counters_old,
398 freelist_new, counters_new))
405 local_irq_save(flags);
407 if (page->freelist == freelist_old &&
408 page->counters == counters_old) {
409 page->freelist = freelist_new;
410 set_page_slub_counters(page, counters_new);
412 local_irq_restore(flags);
416 local_irq_restore(flags);
420 stat(s, CMPXCHG_DOUBLE_FAIL);
422 #ifdef SLUB_DEBUG_CMPXCHG
423 pr_info("%s %s: cmpxchg double redo ", n, s->name);
429 #ifdef CONFIG_SLUB_DEBUG
431 * Determine a map of object in use on a page.
433 * Node listlock must be held to guarantee that the page does
434 * not vanish from under us.
436 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
439 void *addr = page_address(page);
441 for (p = page->freelist; p; p = get_freepointer(s, p))
442 set_bit(slab_index(p, s, addr), map);
448 #if defined(CONFIG_SLUB_DEBUG_ON)
449 static int slub_debug = DEBUG_DEFAULT_FLAGS;
450 #elif defined(CONFIG_KASAN)
451 static int slub_debug = SLAB_STORE_USER;
453 static int slub_debug;
456 static char *slub_debug_slabs;
457 static int disable_higher_order_debug;
460 * slub is about to manipulate internal object metadata. This memory lies
461 * outside the range of the allocated object, so accessing it would normally
462 * be reported by kasan as a bounds error. metadata_access_enable() is used
463 * to tell kasan that these accesses are OK.
465 static inline void metadata_access_enable(void)
467 kasan_disable_current();
470 static inline void metadata_access_disable(void)
472 kasan_enable_current();
478 static void print_section(char *text, u8 *addr, unsigned int length)
480 metadata_access_enable();
481 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
483 metadata_access_disable();
486 static struct track *get_track(struct kmem_cache *s, void *object,
487 enum track_item alloc)
492 p = object + s->offset + sizeof(void *);
494 p = object + s->inuse;
499 static void set_track(struct kmem_cache *s, void *object,
500 enum track_item alloc, unsigned long addr)
502 struct track *p = get_track(s, object, alloc);
505 #ifdef CONFIG_STACKTRACE
506 struct stack_trace trace;
509 trace.nr_entries = 0;
510 trace.max_entries = TRACK_ADDRS_COUNT;
511 trace.entries = p->addrs;
513 metadata_access_enable();
514 save_stack_trace(&trace);
515 metadata_access_disable();
517 /* See rant in lockdep.c */
518 if (trace.nr_entries != 0 &&
519 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
522 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
526 p->cpu = smp_processor_id();
527 p->pid = current->pid;
530 memset(p, 0, sizeof(struct track));
533 static void init_tracking(struct kmem_cache *s, void *object)
535 if (!(s->flags & SLAB_STORE_USER))
538 set_track(s, object, TRACK_FREE, 0UL);
539 set_track(s, object, TRACK_ALLOC, 0UL);
542 static void print_track(const char *s, struct track *t)
547 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
548 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
549 #ifdef CONFIG_STACKTRACE
552 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
554 pr_err("\t%pS\n", (void *)t->addrs[i]);
561 static void print_tracking(struct kmem_cache *s, void *object)
563 if (!(s->flags & SLAB_STORE_USER))
566 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
567 print_track("Freed", get_track(s, object, TRACK_FREE));
570 static void print_page_info(struct page *page)
572 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
573 page, page->objects, page->inuse, page->freelist, page->flags);
577 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
579 struct va_format vaf;
585 pr_err("=============================================================================\n");
586 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
587 pr_err("-----------------------------------------------------------------------------\n\n");
589 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
593 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
595 struct va_format vaf;
601 pr_err("FIX %s: %pV\n", s->name, &vaf);
605 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
607 unsigned int off; /* Offset of last byte */
608 u8 *addr = page_address(page);
610 print_tracking(s, p);
612 print_page_info(page);
614 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
615 p, p - addr, get_freepointer(s, p));
618 print_section("Bytes b4 ", p - 16, 16);
620 print_section("Object ", p, min_t(unsigned long, s->object_size,
622 if (s->flags & SLAB_RED_ZONE)
623 print_section("Redzone ", p + s->object_size,
624 s->inuse - s->object_size);
627 off = s->offset + sizeof(void *);
631 if (s->flags & SLAB_STORE_USER)
632 off += 2 * sizeof(struct track);
635 /* Beginning of the filler is the free pointer */
636 print_section("Padding ", p + off, s->size - off);
641 void object_err(struct kmem_cache *s, struct page *page,
642 u8 *object, char *reason)
644 slab_bug(s, "%s", reason);
645 print_trailer(s, page, object);
648 static void slab_err(struct kmem_cache *s, struct page *page,
649 const char *fmt, ...)
655 vsnprintf(buf, sizeof(buf), fmt, args);
657 slab_bug(s, "%s", buf);
658 print_page_info(page);
662 static void init_object(struct kmem_cache *s, void *object, u8 val)
666 if (s->flags & __OBJECT_POISON) {
667 memset(p, POISON_FREE, s->object_size - 1);
668 p[s->object_size - 1] = POISON_END;
671 if (s->flags & SLAB_RED_ZONE)
672 memset(p + s->object_size, val, s->inuse - s->object_size);
675 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
676 void *from, void *to)
678 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
679 memset(from, data, to - from);
682 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
683 u8 *object, char *what,
684 u8 *start, unsigned int value, unsigned int bytes)
689 metadata_access_enable();
690 fault = memchr_inv(start, value, bytes);
691 metadata_access_disable();
696 while (end > fault && end[-1] == value)
699 slab_bug(s, "%s overwritten", what);
700 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
701 fault, end - 1, fault[0], value);
702 print_trailer(s, page, object);
704 restore_bytes(s, what, value, fault, end);
712 * Bytes of the object to be managed.
713 * If the freepointer may overlay the object then the free
714 * pointer is the first word of the object.
716 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
719 * object + s->object_size
720 * Padding to reach word boundary. This is also used for Redzoning.
721 * Padding is extended by another word if Redzoning is enabled and
722 * object_size == inuse.
724 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
725 * 0xcc (RED_ACTIVE) for objects in use.
728 * Meta data starts here.
730 * A. Free pointer (if we cannot overwrite object on free)
731 * B. Tracking data for SLAB_STORE_USER
732 * C. Padding to reach required alignment boundary or at mininum
733 * one word if debugging is on to be able to detect writes
734 * before the word boundary.
736 * Padding is done using 0x5a (POISON_INUSE)
739 * Nothing is used beyond s->size.
741 * If slabcaches are merged then the object_size and inuse boundaries are mostly
742 * ignored. And therefore no slab options that rely on these boundaries
743 * may be used with merged slabcaches.
746 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
748 unsigned long off = s->inuse; /* The end of info */
751 /* Freepointer is placed after the object. */
752 off += sizeof(void *);
754 if (s->flags & SLAB_STORE_USER)
755 /* We also have user information there */
756 off += 2 * sizeof(struct track);
761 return check_bytes_and_report(s, page, p, "Object padding",
762 p + off, POISON_INUSE, s->size - off);
765 /* Check the pad bytes at the end of a slab page */
766 static int slab_pad_check(struct kmem_cache *s, struct page *page)
774 if (!(s->flags & SLAB_POISON))
777 start = page_address(page);
778 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
779 end = start + length;
780 remainder = length % s->size;
784 metadata_access_enable();
785 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
786 metadata_access_disable();
789 while (end > fault && end[-1] == POISON_INUSE)
792 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
793 print_section("Padding ", end - remainder, remainder);
795 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
799 static int check_object(struct kmem_cache *s, struct page *page,
800 void *object, u8 val)
803 u8 *endobject = object + s->object_size;
805 if (s->flags & SLAB_RED_ZONE) {
806 if (!check_bytes_and_report(s, page, object, "Redzone",
807 endobject, val, s->inuse - s->object_size))
810 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
811 check_bytes_and_report(s, page, p, "Alignment padding",
812 endobject, POISON_INUSE,
813 s->inuse - s->object_size);
817 if (s->flags & SLAB_POISON) {
818 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
819 (!check_bytes_and_report(s, page, p, "Poison", p,
820 POISON_FREE, s->object_size - 1) ||
821 !check_bytes_and_report(s, page, p, "Poison",
822 p + s->object_size - 1, POISON_END, 1)))
825 * check_pad_bytes cleans up on its own.
827 check_pad_bytes(s, page, p);
830 if (!s->offset && val == SLUB_RED_ACTIVE)
832 * Object and freepointer overlap. Cannot check
833 * freepointer while object is allocated.
837 /* Check free pointer validity */
838 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
839 object_err(s, page, p, "Freepointer corrupt");
841 * No choice but to zap it and thus lose the remainder
842 * of the free objects in this slab. May cause
843 * another error because the object count is now wrong.
845 set_freepointer(s, p, NULL);
851 static int check_slab(struct kmem_cache *s, struct page *page)
855 VM_BUG_ON(!irqs_disabled());
857 if (!PageSlab(page)) {
858 slab_err(s, page, "Not a valid slab page");
862 maxobj = order_objects(compound_order(page), s->size, s->reserved);
863 if (page->objects > maxobj) {
864 slab_err(s, page, "objects %u > max %u",
865 page->objects, maxobj);
868 if (page->inuse > page->objects) {
869 slab_err(s, page, "inuse %u > max %u",
870 page->inuse, page->objects);
873 /* Slab_pad_check fixes things up after itself */
874 slab_pad_check(s, page);
879 * Determine if a certain object on a page is on the freelist. Must hold the
880 * slab lock to guarantee that the chains are in a consistent state.
882 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
890 while (fp && nr <= page->objects) {
893 if (!check_valid_pointer(s, page, fp)) {
895 object_err(s, page, object,
896 "Freechain corrupt");
897 set_freepointer(s, object, NULL);
899 slab_err(s, page, "Freepointer corrupt");
900 page->freelist = NULL;
901 page->inuse = page->objects;
902 slab_fix(s, "Freelist cleared");
908 fp = get_freepointer(s, object);
912 max_objects = order_objects(compound_order(page), s->size, s->reserved);
913 if (max_objects > MAX_OBJS_PER_PAGE)
914 max_objects = MAX_OBJS_PER_PAGE;
916 if (page->objects != max_objects) {
917 slab_err(s, page, "Wrong number of objects. Found %d but "
918 "should be %d", page->objects, max_objects);
919 page->objects = max_objects;
920 slab_fix(s, "Number of objects adjusted.");
922 if (page->inuse != page->objects - nr) {
923 slab_err(s, page, "Wrong object count. Counter is %d but "
924 "counted were %d", page->inuse, page->objects - nr);
925 page->inuse = page->objects - nr;
926 slab_fix(s, "Object count adjusted.");
928 return search == NULL;
931 static void trace(struct kmem_cache *s, struct page *page, void *object,
934 if (s->flags & SLAB_TRACE) {
935 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
937 alloc ? "alloc" : "free",
942 print_section("Object ", (void *)object,
950 * Tracking of fully allocated slabs for debugging purposes.
952 static void add_full(struct kmem_cache *s,
953 struct kmem_cache_node *n, struct page *page)
955 if (!(s->flags & SLAB_STORE_USER))
958 lockdep_assert_held(&n->list_lock);
959 list_add(&page->lru, &n->full);
962 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
964 if (!(s->flags & SLAB_STORE_USER))
967 lockdep_assert_held(&n->list_lock);
968 list_del(&page->lru);
971 /* Tracking of the number of slabs for debugging purposes */
972 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
974 struct kmem_cache_node *n = get_node(s, node);
976 return atomic_long_read(&n->nr_slabs);
979 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
981 return atomic_long_read(&n->nr_slabs);
984 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
986 struct kmem_cache_node *n = get_node(s, node);
989 * May be called early in order to allocate a slab for the
990 * kmem_cache_node structure. Solve the chicken-egg
991 * dilemma by deferring the increment of the count during
992 * bootstrap (see early_kmem_cache_node_alloc).
995 atomic_long_inc(&n->nr_slabs);
996 atomic_long_add(objects, &n->total_objects);
999 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1001 struct kmem_cache_node *n = get_node(s, node);
1003 atomic_long_dec(&n->nr_slabs);
1004 atomic_long_sub(objects, &n->total_objects);
1007 /* Object debug checks for alloc/free paths */
1008 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1011 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1014 init_object(s, object, SLUB_RED_INACTIVE);
1015 init_tracking(s, object);
1018 static inline int alloc_consistency_checks(struct kmem_cache *s,
1020 void *object, unsigned long addr)
1022 if (!check_slab(s, page))
1025 if (!check_valid_pointer(s, page, object)) {
1026 object_err(s, page, object, "Freelist Pointer check fails");
1030 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1036 static noinline int alloc_debug_processing(struct kmem_cache *s,
1038 void *object, unsigned long addr)
1040 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1041 if (!alloc_consistency_checks(s, page, object, addr))
1045 /* Success perform special debug activities for allocs */
1046 if (s->flags & SLAB_STORE_USER)
1047 set_track(s, object, TRACK_ALLOC, addr);
1048 trace(s, page, object, 1);
1049 init_object(s, object, SLUB_RED_ACTIVE);
1053 if (PageSlab(page)) {
1055 * If this is a slab page then lets do the best we can
1056 * to avoid issues in the future. Marking all objects
1057 * as used avoids touching the remaining objects.
1059 slab_fix(s, "Marking all objects used");
1060 page->inuse = page->objects;
1061 page->freelist = NULL;
1066 static inline int free_consistency_checks(struct kmem_cache *s,
1067 struct page *page, void *object, unsigned long addr)
1069 if (!check_valid_pointer(s, page, object)) {
1070 slab_err(s, page, "Invalid object pointer 0x%p", object);
1074 if (on_freelist(s, page, object)) {
1075 object_err(s, page, object, "Object already free");
1079 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1082 if (unlikely(s != page->slab_cache)) {
1083 if (!PageSlab(page)) {
1084 slab_err(s, page, "Attempt to free object(0x%p) "
1085 "outside of slab", object);
1086 } else if (!page->slab_cache) {
1087 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1091 object_err(s, page, object,
1092 "page slab pointer corrupt.");
1098 /* Supports checking bulk free of a constructed freelist */
1099 static noinline int free_debug_processing(
1100 struct kmem_cache *s, struct page *page,
1101 void *head, void *tail, int bulk_cnt,
1104 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1105 void *object = head;
1107 unsigned long uninitialized_var(flags);
1110 spin_lock_irqsave(&n->list_lock, flags);
1113 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1114 if (!check_slab(s, page))
1121 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1122 if (!free_consistency_checks(s, page, object, addr))
1126 if (s->flags & SLAB_STORE_USER)
1127 set_track(s, object, TRACK_FREE, addr);
1128 trace(s, page, object, 0);
1129 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1130 init_object(s, object, SLUB_RED_INACTIVE);
1132 /* Reached end of constructed freelist yet? */
1133 if (object != tail) {
1134 object = get_freepointer(s, object);
1140 if (cnt != bulk_cnt)
1141 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1145 spin_unlock_irqrestore(&n->list_lock, flags);
1147 slab_fix(s, "Object at 0x%p not freed", object);
1151 static int __init setup_slub_debug(char *str)
1153 slub_debug = DEBUG_DEFAULT_FLAGS;
1154 if (*str++ != '=' || !*str)
1156 * No options specified. Switch on full debugging.
1162 * No options but restriction on slabs. This means full
1163 * debugging for slabs matching a pattern.
1170 * Switch off all debugging measures.
1175 * Determine which debug features should be switched on
1177 for (; *str && *str != ','; str++) {
1178 switch (tolower(*str)) {
1180 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1183 slub_debug |= SLAB_RED_ZONE;
1186 slub_debug |= SLAB_POISON;
1189 slub_debug |= SLAB_STORE_USER;
1192 slub_debug |= SLAB_TRACE;
1195 slub_debug |= SLAB_FAILSLAB;
1199 * Avoid enabling debugging on caches if its minimum
1200 * order would increase as a result.
1202 disable_higher_order_debug = 1;
1205 pr_err("slub_debug option '%c' unknown. skipped\n",
1212 slub_debug_slabs = str + 1;
1217 __setup("slub_debug", setup_slub_debug);
1219 unsigned long kmem_cache_flags(unsigned long object_size,
1220 unsigned long flags, const char *name,
1221 void (*ctor)(void *))
1224 * Enable debugging if selected on the kernel commandline.
1226 if (slub_debug && (!slub_debug_slabs || (name &&
1227 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1228 flags |= slub_debug;
1232 #else /* !CONFIG_SLUB_DEBUG */
1233 static inline void setup_object_debug(struct kmem_cache *s,
1234 struct page *page, void *object) {}
1236 static inline int alloc_debug_processing(struct kmem_cache *s,
1237 struct page *page, void *object, unsigned long addr) { return 0; }
1239 static inline int free_debug_processing(
1240 struct kmem_cache *s, struct page *page,
1241 void *head, void *tail, int bulk_cnt,
1242 unsigned long addr) { return 0; }
1244 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1246 static inline int check_object(struct kmem_cache *s, struct page *page,
1247 void *object, u8 val) { return 1; }
1248 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1249 struct page *page) {}
1250 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1251 struct page *page) {}
1252 unsigned long kmem_cache_flags(unsigned long object_size,
1253 unsigned long flags, const char *name,
1254 void (*ctor)(void *))
1258 #define slub_debug 0
1260 #define disable_higher_order_debug 0
1262 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1264 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1266 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1268 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1271 #endif /* CONFIG_SLUB_DEBUG */
1274 * Hooks for other subsystems that check memory allocations. In a typical
1275 * production configuration these hooks all should produce no code at all.
1277 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1279 kmemleak_alloc(ptr, size, 1, flags);
1280 kasan_kmalloc_large(ptr, size);
1283 static inline void kfree_hook(const void *x)
1286 kasan_kfree_large(x);
1289 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1291 kmemleak_free_recursive(x, s->flags);
1294 * Trouble is that we may no longer disable interrupts in the fast path
1295 * So in order to make the debug calls that expect irqs to be
1296 * disabled we need to disable interrupts temporarily.
1298 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1300 unsigned long flags;
1302 local_irq_save(flags);
1303 kmemcheck_slab_free(s, x, s->object_size);
1304 debug_check_no_locks_freed(x, s->object_size);
1305 local_irq_restore(flags);
1308 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1309 debug_check_no_obj_freed(x, s->object_size);
1311 kasan_slab_free(s, x);
1314 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1315 void *head, void *tail)
1318 * Compiler cannot detect this function can be removed if slab_free_hook()
1319 * evaluates to nothing. Thus, catch all relevant config debug options here.
1321 #if defined(CONFIG_KMEMCHECK) || \
1322 defined(CONFIG_LOCKDEP) || \
1323 defined(CONFIG_DEBUG_KMEMLEAK) || \
1324 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1325 defined(CONFIG_KASAN)
1327 void *object = head;
1328 void *tail_obj = tail ? : head;
1331 slab_free_hook(s, object);
1332 } while ((object != tail_obj) &&
1333 (object = get_freepointer(s, object)));
1337 static void setup_object(struct kmem_cache *s, struct page *page,
1340 setup_object_debug(s, page, object);
1341 if (unlikely(s->ctor)) {
1342 kasan_unpoison_object_data(s, object);
1344 kasan_poison_object_data(s, object);
1349 * Slab allocation and freeing
1351 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1352 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1355 int order = oo_order(oo);
1357 flags |= __GFP_NOTRACK;
1359 if (node == NUMA_NO_NODE)
1360 page = alloc_pages(flags, order);
1362 page = __alloc_pages_node(node, flags, order);
1364 if (page && memcg_charge_slab(page, flags, order, s)) {
1365 __free_pages(page, order);
1372 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1375 struct kmem_cache_order_objects oo = s->oo;
1380 flags &= gfp_allowed_mask;
1382 if (gfpflags_allow_blocking(flags))
1385 flags |= s->allocflags;
1388 * Let the initial higher-order allocation fail under memory pressure
1389 * so we fall-back to the minimum order allocation.
1391 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1392 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1393 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1395 page = alloc_slab_page(s, alloc_gfp, node, oo);
1396 if (unlikely(!page)) {
1400 * Allocation may have failed due to fragmentation.
1401 * Try a lower order alloc if possible
1403 page = alloc_slab_page(s, alloc_gfp, node, oo);
1404 if (unlikely(!page))
1406 stat(s, ORDER_FALLBACK);
1409 if (kmemcheck_enabled &&
1410 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1411 int pages = 1 << oo_order(oo);
1413 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1416 * Objects from caches that have a constructor don't get
1417 * cleared when they're allocated, so we need to do it here.
1420 kmemcheck_mark_uninitialized_pages(page, pages);
1422 kmemcheck_mark_unallocated_pages(page, pages);
1425 page->objects = oo_objects(oo);
1427 order = compound_order(page);
1428 page->slab_cache = s;
1429 __SetPageSlab(page);
1430 if (page_is_pfmemalloc(page))
1431 SetPageSlabPfmemalloc(page);
1433 start = page_address(page);
1435 if (unlikely(s->flags & SLAB_POISON))
1436 memset(start, POISON_INUSE, PAGE_SIZE << order);
1438 kasan_poison_slab(page);
1440 for_each_object_idx(p, idx, s, start, page->objects) {
1441 setup_object(s, page, p);
1442 if (likely(idx < page->objects))
1443 set_freepointer(s, p, p + s->size);
1445 set_freepointer(s, p, NULL);
1448 page->freelist = start;
1449 page->inuse = page->objects;
1453 if (gfpflags_allow_blocking(flags))
1454 local_irq_disable();
1458 mod_zone_page_state(page_zone(page),
1459 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1460 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1463 inc_slabs_node(s, page_to_nid(page), page->objects);
1468 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1470 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1471 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1475 return allocate_slab(s,
1476 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1479 static void __free_slab(struct kmem_cache *s, struct page *page)
1481 int order = compound_order(page);
1482 int pages = 1 << order;
1484 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1487 slab_pad_check(s, page);
1488 for_each_object(p, s, page_address(page),
1490 check_object(s, page, p, SLUB_RED_INACTIVE);
1493 kmemcheck_free_shadow(page, compound_order(page));
1495 mod_zone_page_state(page_zone(page),
1496 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1497 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1500 __ClearPageSlabPfmemalloc(page);
1501 __ClearPageSlab(page);
1503 page_mapcount_reset(page);
1504 if (current->reclaim_state)
1505 current->reclaim_state->reclaimed_slab += pages;
1506 __free_kmem_pages(page, order);
1509 #define need_reserve_slab_rcu \
1510 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1512 static void rcu_free_slab(struct rcu_head *h)
1516 if (need_reserve_slab_rcu)
1517 page = virt_to_head_page(h);
1519 page = container_of((struct list_head *)h, struct page, lru);
1521 __free_slab(page->slab_cache, page);
1524 static void free_slab(struct kmem_cache *s, struct page *page)
1526 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1527 struct rcu_head *head;
1529 if (need_reserve_slab_rcu) {
1530 int order = compound_order(page);
1531 int offset = (PAGE_SIZE << order) - s->reserved;
1533 VM_BUG_ON(s->reserved != sizeof(*head));
1534 head = page_address(page) + offset;
1536 head = &page->rcu_head;
1539 call_rcu(head, rcu_free_slab);
1541 __free_slab(s, page);
1544 static void discard_slab(struct kmem_cache *s, struct page *page)
1546 dec_slabs_node(s, page_to_nid(page), page->objects);
1551 * Management of partially allocated slabs.
1554 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1557 if (tail == DEACTIVATE_TO_TAIL)
1558 list_add_tail(&page->lru, &n->partial);
1560 list_add(&page->lru, &n->partial);
1563 static inline void add_partial(struct kmem_cache_node *n,
1564 struct page *page, int tail)
1566 lockdep_assert_held(&n->list_lock);
1567 __add_partial(n, page, tail);
1570 static inline void remove_partial(struct kmem_cache_node *n,
1573 lockdep_assert_held(&n->list_lock);
1574 list_del(&page->lru);
1579 * Remove slab from the partial list, freeze it and
1580 * return the pointer to the freelist.
1582 * Returns a list of objects or NULL if it fails.
1584 static inline void *acquire_slab(struct kmem_cache *s,
1585 struct kmem_cache_node *n, struct page *page,
1586 int mode, int *objects)
1589 unsigned long counters;
1592 lockdep_assert_held(&n->list_lock);
1595 * Zap the freelist and set the frozen bit.
1596 * The old freelist is the list of objects for the
1597 * per cpu allocation list.
1599 freelist = page->freelist;
1600 counters = page->counters;
1601 new.counters = counters;
1602 *objects = new.objects - new.inuse;
1604 new.inuse = page->objects;
1605 new.freelist = NULL;
1607 new.freelist = freelist;
1610 VM_BUG_ON(new.frozen);
1613 if (!__cmpxchg_double_slab(s, page,
1615 new.freelist, new.counters,
1619 remove_partial(n, page);
1624 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1625 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1628 * Try to allocate a partial slab from a specific node.
1630 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1631 struct kmem_cache_cpu *c, gfp_t flags)
1633 struct page *page, *page2;
1634 void *object = NULL;
1639 * Racy check. If we mistakenly see no partial slabs then we
1640 * just allocate an empty slab. If we mistakenly try to get a
1641 * partial slab and there is none available then get_partials()
1644 if (!n || !n->nr_partial)
1647 spin_lock(&n->list_lock);
1648 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1651 if (!pfmemalloc_match(page, flags))
1654 t = acquire_slab(s, n, page, object == NULL, &objects);
1658 available += objects;
1661 stat(s, ALLOC_FROM_PARTIAL);
1664 put_cpu_partial(s, page, 0);
1665 stat(s, CPU_PARTIAL_NODE);
1667 if (!kmem_cache_has_cpu_partial(s)
1668 || available > s->cpu_partial / 2)
1672 spin_unlock(&n->list_lock);
1677 * Get a page from somewhere. Search in increasing NUMA distances.
1679 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1680 struct kmem_cache_cpu *c)
1683 struct zonelist *zonelist;
1686 enum zone_type high_zoneidx = gfp_zone(flags);
1688 unsigned int cpuset_mems_cookie;
1691 * The defrag ratio allows a configuration of the tradeoffs between
1692 * inter node defragmentation and node local allocations. A lower
1693 * defrag_ratio increases the tendency to do local allocations
1694 * instead of attempting to obtain partial slabs from other nodes.
1696 * If the defrag_ratio is set to 0 then kmalloc() always
1697 * returns node local objects. If the ratio is higher then kmalloc()
1698 * may return off node objects because partial slabs are obtained
1699 * from other nodes and filled up.
1701 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1702 * defrag_ratio = 1000) then every (well almost) allocation will
1703 * first attempt to defrag slab caches on other nodes. This means
1704 * scanning over all nodes to look for partial slabs which may be
1705 * expensive if we do it every time we are trying to find a slab
1706 * with available objects.
1708 if (!s->remote_node_defrag_ratio ||
1709 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1713 cpuset_mems_cookie = read_mems_allowed_begin();
1714 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1715 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1716 struct kmem_cache_node *n;
1718 n = get_node(s, zone_to_nid(zone));
1720 if (n && cpuset_zone_allowed(zone, flags) &&
1721 n->nr_partial > s->min_partial) {
1722 object = get_partial_node(s, n, c, flags);
1725 * Don't check read_mems_allowed_retry()
1726 * here - if mems_allowed was updated in
1727 * parallel, that was a harmless race
1728 * between allocation and the cpuset
1735 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1741 * Get a partial page, lock it and return it.
1743 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1744 struct kmem_cache_cpu *c)
1747 int searchnode = node;
1749 if (node == NUMA_NO_NODE)
1750 searchnode = numa_mem_id();
1751 else if (!node_present_pages(node))
1752 searchnode = node_to_mem_node(node);
1754 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1755 if (object || node != NUMA_NO_NODE)
1758 return get_any_partial(s, flags, c);
1761 #ifdef CONFIG_PREEMPT
1763 * Calculate the next globally unique transaction for disambiguiation
1764 * during cmpxchg. The transactions start with the cpu number and are then
1765 * incremented by CONFIG_NR_CPUS.
1767 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1770 * No preemption supported therefore also no need to check for
1776 static inline unsigned long next_tid(unsigned long tid)
1778 return tid + TID_STEP;
1781 static inline unsigned int tid_to_cpu(unsigned long tid)
1783 return tid % TID_STEP;
1786 static inline unsigned long tid_to_event(unsigned long tid)
1788 return tid / TID_STEP;
1791 static inline unsigned int init_tid(int cpu)
1796 static inline void note_cmpxchg_failure(const char *n,
1797 const struct kmem_cache *s, unsigned long tid)
1799 #ifdef SLUB_DEBUG_CMPXCHG
1800 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1802 pr_info("%s %s: cmpxchg redo ", n, s->name);
1804 #ifdef CONFIG_PREEMPT
1805 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1806 pr_warn("due to cpu change %d -> %d\n",
1807 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1810 if (tid_to_event(tid) != tid_to_event(actual_tid))
1811 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1812 tid_to_event(tid), tid_to_event(actual_tid));
1814 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1815 actual_tid, tid, next_tid(tid));
1817 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1820 static void init_kmem_cache_cpus(struct kmem_cache *s)
1824 for_each_possible_cpu(cpu)
1825 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1829 * Remove the cpu slab
1831 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1834 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1835 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1837 enum slab_modes l = M_NONE, m = M_NONE;
1839 int tail = DEACTIVATE_TO_HEAD;
1843 if (page->freelist) {
1844 stat(s, DEACTIVATE_REMOTE_FREES);
1845 tail = DEACTIVATE_TO_TAIL;
1849 * Stage one: Free all available per cpu objects back
1850 * to the page freelist while it is still frozen. Leave the
1853 * There is no need to take the list->lock because the page
1856 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1858 unsigned long counters;
1861 prior = page->freelist;
1862 counters = page->counters;
1863 set_freepointer(s, freelist, prior);
1864 new.counters = counters;
1866 VM_BUG_ON(!new.frozen);
1868 } while (!__cmpxchg_double_slab(s, page,
1870 freelist, new.counters,
1871 "drain percpu freelist"));
1873 freelist = nextfree;
1877 * Stage two: Ensure that the page is unfrozen while the
1878 * list presence reflects the actual number of objects
1881 * We setup the list membership and then perform a cmpxchg
1882 * with the count. If there is a mismatch then the page
1883 * is not unfrozen but the page is on the wrong list.
1885 * Then we restart the process which may have to remove
1886 * the page from the list that we just put it on again
1887 * because the number of objects in the slab may have
1892 old.freelist = page->freelist;
1893 old.counters = page->counters;
1894 VM_BUG_ON(!old.frozen);
1896 /* Determine target state of the slab */
1897 new.counters = old.counters;
1900 set_freepointer(s, freelist, old.freelist);
1901 new.freelist = freelist;
1903 new.freelist = old.freelist;
1907 if (!new.inuse && n->nr_partial >= s->min_partial)
1909 else if (new.freelist) {
1914 * Taking the spinlock removes the possiblity
1915 * that acquire_slab() will see a slab page that
1918 spin_lock(&n->list_lock);
1922 if (kmem_cache_debug(s) && !lock) {
1925 * This also ensures that the scanning of full
1926 * slabs from diagnostic functions will not see
1929 spin_lock(&n->list_lock);
1937 remove_partial(n, page);
1939 else if (l == M_FULL)
1941 remove_full(s, n, page);
1943 if (m == M_PARTIAL) {
1945 add_partial(n, page, tail);
1948 } else if (m == M_FULL) {
1950 stat(s, DEACTIVATE_FULL);
1951 add_full(s, n, page);
1957 if (!__cmpxchg_double_slab(s, page,
1958 old.freelist, old.counters,
1959 new.freelist, new.counters,
1964 spin_unlock(&n->list_lock);
1967 stat(s, DEACTIVATE_EMPTY);
1968 discard_slab(s, page);
1974 * Unfreeze all the cpu partial slabs.
1976 * This function must be called with interrupts disabled
1977 * for the cpu using c (or some other guarantee must be there
1978 * to guarantee no concurrent accesses).
1980 static void unfreeze_partials(struct kmem_cache *s,
1981 struct kmem_cache_cpu *c)
1983 #ifdef CONFIG_SLUB_CPU_PARTIAL
1984 struct kmem_cache_node *n = NULL, *n2 = NULL;
1985 struct page *page, *discard_page = NULL;
1987 while ((page = c->partial)) {
1991 c->partial = page->next;
1993 n2 = get_node(s, page_to_nid(page));
1996 spin_unlock(&n->list_lock);
1999 spin_lock(&n->list_lock);
2004 old.freelist = page->freelist;
2005 old.counters = page->counters;
2006 VM_BUG_ON(!old.frozen);
2008 new.counters = old.counters;
2009 new.freelist = old.freelist;
2013 } while (!__cmpxchg_double_slab(s, page,
2014 old.freelist, old.counters,
2015 new.freelist, new.counters,
2016 "unfreezing slab"));
2018 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2019 page->next = discard_page;
2020 discard_page = page;
2022 add_partial(n, page, DEACTIVATE_TO_TAIL);
2023 stat(s, FREE_ADD_PARTIAL);
2028 spin_unlock(&n->list_lock);
2030 while (discard_page) {
2031 page = discard_page;
2032 discard_page = discard_page->next;
2034 stat(s, DEACTIVATE_EMPTY);
2035 discard_slab(s, page);
2042 * Put a page that was just frozen (in __slab_free) into a partial page
2043 * slot if available. This is done without interrupts disabled and without
2044 * preemption disabled. The cmpxchg is racy and may put the partial page
2045 * onto a random cpus partial slot.
2047 * If we did not find a slot then simply move all the partials to the
2048 * per node partial list.
2050 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2052 #ifdef CONFIG_SLUB_CPU_PARTIAL
2053 struct page *oldpage;
2061 oldpage = this_cpu_read(s->cpu_slab->partial);
2064 pobjects = oldpage->pobjects;
2065 pages = oldpage->pages;
2066 if (drain && pobjects > s->cpu_partial) {
2067 unsigned long flags;
2069 * partial array is full. Move the existing
2070 * set to the per node partial list.
2072 local_irq_save(flags);
2073 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2074 local_irq_restore(flags);
2078 stat(s, CPU_PARTIAL_DRAIN);
2083 pobjects += page->objects - page->inuse;
2085 page->pages = pages;
2086 page->pobjects = pobjects;
2087 page->next = oldpage;
2089 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2091 if (unlikely(!s->cpu_partial)) {
2092 unsigned long flags;
2094 local_irq_save(flags);
2095 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2096 local_irq_restore(flags);
2102 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2104 stat(s, CPUSLAB_FLUSH);
2105 deactivate_slab(s, c->page, c->freelist);
2107 c->tid = next_tid(c->tid);
2115 * Called from IPI handler with interrupts disabled.
2117 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2119 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2125 unfreeze_partials(s, c);
2129 static void flush_cpu_slab(void *d)
2131 struct kmem_cache *s = d;
2133 __flush_cpu_slab(s, smp_processor_id());
2136 static bool has_cpu_slab(int cpu, void *info)
2138 struct kmem_cache *s = info;
2139 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2141 return c->page || c->partial;
2144 static void flush_all(struct kmem_cache *s)
2146 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2150 * Check if the objects in a per cpu structure fit numa
2151 * locality expectations.
2153 static inline int node_match(struct page *page, int node)
2156 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2162 #ifdef CONFIG_SLUB_DEBUG
2163 static int count_free(struct page *page)
2165 return page->objects - page->inuse;
2168 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2170 return atomic_long_read(&n->total_objects);
2172 #endif /* CONFIG_SLUB_DEBUG */
2174 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2175 static unsigned long count_partial(struct kmem_cache_node *n,
2176 int (*get_count)(struct page *))
2178 unsigned long flags;
2179 unsigned long x = 0;
2182 spin_lock_irqsave(&n->list_lock, flags);
2183 list_for_each_entry(page, &n->partial, lru)
2184 x += get_count(page);
2185 spin_unlock_irqrestore(&n->list_lock, flags);
2188 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2190 static noinline void
2191 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2193 #ifdef CONFIG_SLUB_DEBUG
2194 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2195 DEFAULT_RATELIMIT_BURST);
2197 struct kmem_cache_node *n;
2199 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2202 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2204 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2205 s->name, s->object_size, s->size, oo_order(s->oo),
2208 if (oo_order(s->min) > get_order(s->object_size))
2209 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2212 for_each_kmem_cache_node(s, node, n) {
2213 unsigned long nr_slabs;
2214 unsigned long nr_objs;
2215 unsigned long nr_free;
2217 nr_free = count_partial(n, count_free);
2218 nr_slabs = node_nr_slabs(n);
2219 nr_objs = node_nr_objs(n);
2221 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2222 node, nr_slabs, nr_objs, nr_free);
2227 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2228 int node, struct kmem_cache_cpu **pc)
2231 struct kmem_cache_cpu *c = *pc;
2234 freelist = get_partial(s, flags, node, c);
2239 page = new_slab(s, flags, node);
2241 c = raw_cpu_ptr(s->cpu_slab);
2246 * No other reference to the page yet so we can
2247 * muck around with it freely without cmpxchg
2249 freelist = page->freelist;
2250 page->freelist = NULL;
2252 stat(s, ALLOC_SLAB);
2261 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2263 if (unlikely(PageSlabPfmemalloc(page)))
2264 return gfp_pfmemalloc_allowed(gfpflags);
2270 * Check the page->freelist of a page and either transfer the freelist to the
2271 * per cpu freelist or deactivate the page.
2273 * The page is still frozen if the return value is not NULL.
2275 * If this function returns NULL then the page has been unfrozen.
2277 * This function must be called with interrupt disabled.
2279 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2282 unsigned long counters;
2286 freelist = page->freelist;
2287 counters = page->counters;
2289 new.counters = counters;
2290 VM_BUG_ON(!new.frozen);
2292 new.inuse = page->objects;
2293 new.frozen = freelist != NULL;
2295 } while (!__cmpxchg_double_slab(s, page,
2304 * Slow path. The lockless freelist is empty or we need to perform
2307 * Processing is still very fast if new objects have been freed to the
2308 * regular freelist. In that case we simply take over the regular freelist
2309 * as the lockless freelist and zap the regular freelist.
2311 * If that is not working then we fall back to the partial lists. We take the
2312 * first element of the freelist as the object to allocate now and move the
2313 * rest of the freelist to the lockless freelist.
2315 * And if we were unable to get a new slab from the partial slab lists then
2316 * we need to allocate a new slab. This is the slowest path since it involves
2317 * a call to the page allocator and the setup of a new slab.
2319 * Version of __slab_alloc to use when we know that interrupts are
2320 * already disabled (which is the case for bulk allocation).
2322 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2323 unsigned long addr, struct kmem_cache_cpu *c)
2333 if (unlikely(!node_match(page, node))) {
2334 int searchnode = node;
2336 if (node != NUMA_NO_NODE && !node_present_pages(node))
2337 searchnode = node_to_mem_node(node);
2339 if (unlikely(!node_match(page, searchnode))) {
2340 stat(s, ALLOC_NODE_MISMATCH);
2341 deactivate_slab(s, page, c->freelist);
2349 * By rights, we should be searching for a slab page that was
2350 * PFMEMALLOC but right now, we are losing the pfmemalloc
2351 * information when the page leaves the per-cpu allocator
2353 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2354 deactivate_slab(s, page, c->freelist);
2360 /* must check again c->freelist in case of cpu migration or IRQ */
2361 freelist = c->freelist;
2365 freelist = get_freelist(s, page);
2369 stat(s, DEACTIVATE_BYPASS);
2373 stat(s, ALLOC_REFILL);
2377 * freelist is pointing to the list of objects to be used.
2378 * page is pointing to the page from which the objects are obtained.
2379 * That page must be frozen for per cpu allocations to work.
2381 VM_BUG_ON(!c->page->frozen);
2382 c->freelist = get_freepointer(s, freelist);
2383 c->tid = next_tid(c->tid);
2389 page = c->page = c->partial;
2390 c->partial = page->next;
2391 stat(s, CPU_PARTIAL_ALLOC);
2396 freelist = new_slab_objects(s, gfpflags, node, &c);
2398 if (unlikely(!freelist)) {
2399 slab_out_of_memory(s, gfpflags, node);
2404 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2407 /* Only entered in the debug case */
2408 if (kmem_cache_debug(s) &&
2409 !alloc_debug_processing(s, page, freelist, addr))
2410 goto new_slab; /* Slab failed checks. Next slab needed */
2412 deactivate_slab(s, page, get_freepointer(s, freelist));
2419 * Another one that disabled interrupt and compensates for possible
2420 * cpu changes by refetching the per cpu area pointer.
2422 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2423 unsigned long addr, struct kmem_cache_cpu *c)
2426 unsigned long flags;
2428 local_irq_save(flags);
2429 #ifdef CONFIG_PREEMPT
2431 * We may have been preempted and rescheduled on a different
2432 * cpu before disabling interrupts. Need to reload cpu area
2435 c = this_cpu_ptr(s->cpu_slab);
2438 p = ___slab_alloc(s, gfpflags, node, addr, c);
2439 local_irq_restore(flags);
2444 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2445 * have the fastpath folded into their functions. So no function call
2446 * overhead for requests that can be satisfied on the fastpath.
2448 * The fastpath works by first checking if the lockless freelist can be used.
2449 * If not then __slab_alloc is called for slow processing.
2451 * Otherwise we can simply pick the next object from the lockless free list.
2453 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2454 gfp_t gfpflags, int node, unsigned long addr)
2457 struct kmem_cache_cpu *c;
2461 s = slab_pre_alloc_hook(s, gfpflags);
2466 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2467 * enabled. We may switch back and forth between cpus while
2468 * reading from one cpu area. That does not matter as long
2469 * as we end up on the original cpu again when doing the cmpxchg.
2471 * We should guarantee that tid and kmem_cache are retrieved on
2472 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2473 * to check if it is matched or not.
2476 tid = this_cpu_read(s->cpu_slab->tid);
2477 c = raw_cpu_ptr(s->cpu_slab);
2478 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2479 unlikely(tid != READ_ONCE(c->tid)));
2482 * Irqless object alloc/free algorithm used here depends on sequence
2483 * of fetching cpu_slab's data. tid should be fetched before anything
2484 * on c to guarantee that object and page associated with previous tid
2485 * won't be used with current tid. If we fetch tid first, object and
2486 * page could be one associated with next tid and our alloc/free
2487 * request will be failed. In this case, we will retry. So, no problem.
2492 * The transaction ids are globally unique per cpu and per operation on
2493 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2494 * occurs on the right processor and that there was no operation on the
2495 * linked list in between.
2498 object = c->freelist;
2500 if (unlikely(!object || !node_match(page, node))) {
2501 object = __slab_alloc(s, gfpflags, node, addr, c);
2502 stat(s, ALLOC_SLOWPATH);
2504 void *next_object = get_freepointer_safe(s, object);
2507 * The cmpxchg will only match if there was no additional
2508 * operation and if we are on the right processor.
2510 * The cmpxchg does the following atomically (without lock
2512 * 1. Relocate first pointer to the current per cpu area.
2513 * 2. Verify that tid and freelist have not been changed
2514 * 3. If they were not changed replace tid and freelist
2516 * Since this is without lock semantics the protection is only
2517 * against code executing on this cpu *not* from access by
2520 if (unlikely(!this_cpu_cmpxchg_double(
2521 s->cpu_slab->freelist, s->cpu_slab->tid,
2523 next_object, next_tid(tid)))) {
2525 note_cmpxchg_failure("slab_alloc", s, tid);
2528 prefetch_freepointer(s, next_object);
2529 stat(s, ALLOC_FASTPATH);
2532 if (unlikely(gfpflags & __GFP_ZERO) && object)
2533 memset(object, 0, s->object_size);
2535 slab_post_alloc_hook(s, gfpflags, 1, &object);
2540 static __always_inline void *slab_alloc(struct kmem_cache *s,
2541 gfp_t gfpflags, unsigned long addr)
2543 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2546 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2548 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2550 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2555 EXPORT_SYMBOL(kmem_cache_alloc);
2557 #ifdef CONFIG_TRACING
2558 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2560 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2561 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2562 kasan_kmalloc(s, ret, size);
2565 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2569 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2571 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2573 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2574 s->object_size, s->size, gfpflags, node);
2578 EXPORT_SYMBOL(kmem_cache_alloc_node);
2580 #ifdef CONFIG_TRACING
2581 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2583 int node, size_t size)
2585 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2587 trace_kmalloc_node(_RET_IP_, ret,
2588 size, s->size, gfpflags, node);
2590 kasan_kmalloc(s, ret, size);
2593 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2598 * Slow path handling. This may still be called frequently since objects
2599 * have a longer lifetime than the cpu slabs in most processing loads.
2601 * So we still attempt to reduce cache line usage. Just take the slab
2602 * lock and free the item. If there is no additional partial page
2603 * handling required then we can return immediately.
2605 static void __slab_free(struct kmem_cache *s, struct page *page,
2606 void *head, void *tail, int cnt,
2613 unsigned long counters;
2614 struct kmem_cache_node *n = NULL;
2615 unsigned long uninitialized_var(flags);
2617 stat(s, FREE_SLOWPATH);
2619 if (kmem_cache_debug(s) &&
2620 !free_debug_processing(s, page, head, tail, cnt, addr))
2625 spin_unlock_irqrestore(&n->list_lock, flags);
2628 prior = page->freelist;
2629 counters = page->counters;
2630 set_freepointer(s, tail, prior);
2631 new.counters = counters;
2632 was_frozen = new.frozen;
2634 if ((!new.inuse || !prior) && !was_frozen) {
2636 if (kmem_cache_has_cpu_partial(s) && !prior) {
2639 * Slab was on no list before and will be
2641 * We can defer the list move and instead
2646 } else { /* Needs to be taken off a list */
2648 n = get_node(s, page_to_nid(page));
2650 * Speculatively acquire the list_lock.
2651 * If the cmpxchg does not succeed then we may
2652 * drop the list_lock without any processing.
2654 * Otherwise the list_lock will synchronize with
2655 * other processors updating the list of slabs.
2657 spin_lock_irqsave(&n->list_lock, flags);
2662 } while (!cmpxchg_double_slab(s, page,
2670 * If we just froze the page then put it onto the
2671 * per cpu partial list.
2673 if (new.frozen && !was_frozen) {
2674 put_cpu_partial(s, page, 1);
2675 stat(s, CPU_PARTIAL_FREE);
2678 * The list lock was not taken therefore no list
2679 * activity can be necessary.
2682 stat(s, FREE_FROZEN);
2686 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2690 * Objects left in the slab. If it was not on the partial list before
2693 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2694 if (kmem_cache_debug(s))
2695 remove_full(s, n, page);
2696 add_partial(n, page, DEACTIVATE_TO_TAIL);
2697 stat(s, FREE_ADD_PARTIAL);
2699 spin_unlock_irqrestore(&n->list_lock, flags);
2705 * Slab on the partial list.
2707 remove_partial(n, page);
2708 stat(s, FREE_REMOVE_PARTIAL);
2710 /* Slab must be on the full list */
2711 remove_full(s, n, page);
2714 spin_unlock_irqrestore(&n->list_lock, flags);
2716 discard_slab(s, page);
2720 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2721 * can perform fastpath freeing without additional function calls.
2723 * The fastpath is only possible if we are freeing to the current cpu slab
2724 * of this processor. This typically the case if we have just allocated
2727 * If fastpath is not possible then fall back to __slab_free where we deal
2728 * with all sorts of special processing.
2730 * Bulk free of a freelist with several objects (all pointing to the
2731 * same page) possible by specifying head and tail ptr, plus objects
2732 * count (cnt). Bulk free indicated by tail pointer being set.
2734 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2735 void *head, void *tail, int cnt,
2738 void *tail_obj = tail ? : head;
2739 struct kmem_cache_cpu *c;
2742 slab_free_freelist_hook(s, head, tail);
2746 * Determine the currently cpus per cpu slab.
2747 * The cpu may change afterward. However that does not matter since
2748 * data is retrieved via this pointer. If we are on the same cpu
2749 * during the cmpxchg then the free will succeed.
2752 tid = this_cpu_read(s->cpu_slab->tid);
2753 c = raw_cpu_ptr(s->cpu_slab);
2754 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2755 unlikely(tid != READ_ONCE(c->tid)));
2757 /* Same with comment on barrier() in slab_alloc_node() */
2760 if (likely(page == c->page)) {
2761 set_freepointer(s, tail_obj, c->freelist);
2763 if (unlikely(!this_cpu_cmpxchg_double(
2764 s->cpu_slab->freelist, s->cpu_slab->tid,
2766 head, next_tid(tid)))) {
2768 note_cmpxchg_failure("slab_free", s, tid);
2771 stat(s, FREE_FASTPATH);
2773 __slab_free(s, page, head, tail_obj, cnt, addr);
2777 void kmem_cache_free(struct kmem_cache *s, void *x)
2779 s = cache_from_obj(s, x);
2782 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2783 trace_kmem_cache_free(_RET_IP_, x);
2785 EXPORT_SYMBOL(kmem_cache_free);
2787 struct detached_freelist {
2792 struct kmem_cache *s;
2796 * This function progressively scans the array with free objects (with
2797 * a limited look ahead) and extract objects belonging to the same
2798 * page. It builds a detached freelist directly within the given
2799 * page/objects. This can happen without any need for
2800 * synchronization, because the objects are owned by running process.
2801 * The freelist is build up as a single linked list in the objects.
2802 * The idea is, that this detached freelist can then be bulk
2803 * transferred to the real freelist(s), but only requiring a single
2804 * synchronization primitive. Look ahead in the array is limited due
2805 * to performance reasons.
2808 int build_detached_freelist(struct kmem_cache *s, size_t size,
2809 void **p, struct detached_freelist *df)
2811 size_t first_skipped_index = 0;
2816 /* Always re-init detached_freelist */
2821 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2822 } while (!object && size);
2827 page = virt_to_head_page(object);
2829 /* Handle kalloc'ed objects */
2830 if (unlikely(!PageSlab(page))) {
2831 BUG_ON(!PageCompound(page));
2833 __free_kmem_pages(page, compound_order(page));
2834 p[size] = NULL; /* mark object processed */
2837 /* Derive kmem_cache from object */
2838 df->s = page->slab_cache;
2840 df->s = cache_from_obj(s, object); /* Support for memcg */
2843 /* Start new detached freelist */
2845 set_freepointer(df->s, object, NULL);
2847 df->freelist = object;
2848 p[size] = NULL; /* mark object processed */
2854 continue; /* Skip processed objects */
2856 /* df->page is always set at this point */
2857 if (df->page == virt_to_head_page(object)) {
2858 /* Opportunity build freelist */
2859 set_freepointer(df->s, object, df->freelist);
2860 df->freelist = object;
2862 p[size] = NULL; /* mark object processed */
2867 /* Limit look ahead search */
2871 if (!first_skipped_index)
2872 first_skipped_index = size + 1;
2875 return first_skipped_index;
2878 /* Note that interrupts must be enabled when calling this function. */
2879 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2885 struct detached_freelist df;
2887 size = build_detached_freelist(s, size, p, &df);
2888 if (unlikely(!df.page))
2891 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2892 } while (likely(size));
2894 EXPORT_SYMBOL(kmem_cache_free_bulk);
2896 /* Note that interrupts must be enabled when calling this function. */
2897 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2900 struct kmem_cache_cpu *c;
2903 /* memcg and kmem_cache debug support */
2904 s = slab_pre_alloc_hook(s, flags);
2908 * Drain objects in the per cpu slab, while disabling local
2909 * IRQs, which protects against PREEMPT and interrupts
2910 * handlers invoking normal fastpath.
2912 local_irq_disable();
2913 c = this_cpu_ptr(s->cpu_slab);
2915 for (i = 0; i < size; i++) {
2916 void *object = c->freelist;
2918 if (unlikely(!object)) {
2920 * Invoking slow path likely have side-effect
2921 * of re-populating per CPU c->freelist
2923 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2925 if (unlikely(!p[i]))
2928 c = this_cpu_ptr(s->cpu_slab);
2929 continue; /* goto for-loop */
2931 c->freelist = get_freepointer(s, object);
2934 c->tid = next_tid(c->tid);
2937 /* Clear memory outside IRQ disabled fastpath loop */
2938 if (unlikely(flags & __GFP_ZERO)) {
2941 for (j = 0; j < i; j++)
2942 memset(p[j], 0, s->object_size);
2945 /* memcg and kmem_cache debug support */
2946 slab_post_alloc_hook(s, flags, size, p);
2950 slab_post_alloc_hook(s, flags, i, p);
2951 __kmem_cache_free_bulk(s, i, p);
2954 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2958 * Object placement in a slab is made very easy because we always start at
2959 * offset 0. If we tune the size of the object to the alignment then we can
2960 * get the required alignment by putting one properly sized object after
2963 * Notice that the allocation order determines the sizes of the per cpu
2964 * caches. Each processor has always one slab available for allocations.
2965 * Increasing the allocation order reduces the number of times that slabs
2966 * must be moved on and off the partial lists and is therefore a factor in
2971 * Mininum / Maximum order of slab pages. This influences locking overhead
2972 * and slab fragmentation. A higher order reduces the number of partial slabs
2973 * and increases the number of allocations possible without having to
2974 * take the list_lock.
2976 static int slub_min_order;
2977 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2978 static int slub_min_objects;
2981 * Calculate the order of allocation given an slab object size.
2983 * The order of allocation has significant impact on performance and other
2984 * system components. Generally order 0 allocations should be preferred since
2985 * order 0 does not cause fragmentation in the page allocator. Larger objects
2986 * be problematic to put into order 0 slabs because there may be too much
2987 * unused space left. We go to a higher order if more than 1/16th of the slab
2990 * In order to reach satisfactory performance we must ensure that a minimum
2991 * number of objects is in one slab. Otherwise we may generate too much
2992 * activity on the partial lists which requires taking the list_lock. This is
2993 * less a concern for large slabs though which are rarely used.
2995 * slub_max_order specifies the order where we begin to stop considering the
2996 * number of objects in a slab as critical. If we reach slub_max_order then
2997 * we try to keep the page order as low as possible. So we accept more waste
2998 * of space in favor of a small page order.
3000 * Higher order allocations also allow the placement of more objects in a
3001 * slab and thereby reduce object handling overhead. If the user has
3002 * requested a higher mininum order then we start with that one instead of
3003 * the smallest order which will fit the object.
3005 static inline int slab_order(int size, int min_objects,
3006 int max_order, int fract_leftover, int reserved)
3010 int min_order = slub_min_order;
3012 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3013 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3015 for (order = max(min_order, get_order(min_objects * size + reserved));
3016 order <= max_order; order++) {
3018 unsigned long slab_size = PAGE_SIZE << order;
3020 rem = (slab_size - reserved) % size;
3022 if (rem <= slab_size / fract_leftover)
3029 static inline int calculate_order(int size, int reserved)
3037 * Attempt to find best configuration for a slab. This
3038 * works by first attempting to generate a layout with
3039 * the best configuration and backing off gradually.
3041 * First we increase the acceptable waste in a slab. Then
3042 * we reduce the minimum objects required in a slab.
3044 min_objects = slub_min_objects;
3046 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3047 max_objects = order_objects(slub_max_order, size, reserved);
3048 min_objects = min(min_objects, max_objects);
3050 while (min_objects > 1) {
3052 while (fraction >= 4) {
3053 order = slab_order(size, min_objects,
3054 slub_max_order, fraction, reserved);
3055 if (order <= slub_max_order)
3063 * We were unable to place multiple objects in a slab. Now
3064 * lets see if we can place a single object there.
3066 order = slab_order(size, 1, slub_max_order, 1, reserved);
3067 if (order <= slub_max_order)
3071 * Doh this slab cannot be placed using slub_max_order.
3073 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3074 if (order < MAX_ORDER)
3080 init_kmem_cache_node(struct kmem_cache_node *n)
3083 spin_lock_init(&n->list_lock);
3084 INIT_LIST_HEAD(&n->partial);
3085 #ifdef CONFIG_SLUB_DEBUG
3086 atomic_long_set(&n->nr_slabs, 0);
3087 atomic_long_set(&n->total_objects, 0);
3088 INIT_LIST_HEAD(&n->full);
3092 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3094 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3095 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3098 * Must align to double word boundary for the double cmpxchg
3099 * instructions to work; see __pcpu_double_call_return_bool().
3101 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3102 2 * sizeof(void *));
3107 init_kmem_cache_cpus(s);
3112 static struct kmem_cache *kmem_cache_node;
3115 * No kmalloc_node yet so do it by hand. We know that this is the first
3116 * slab on the node for this slabcache. There are no concurrent accesses
3119 * Note that this function only works on the kmem_cache_node
3120 * when allocating for the kmem_cache_node. This is used for bootstrapping
3121 * memory on a fresh node that has no slab structures yet.
3123 static void early_kmem_cache_node_alloc(int node)
3126 struct kmem_cache_node *n;
3128 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3130 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3133 if (page_to_nid(page) != node) {
3134 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3135 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3140 page->freelist = get_freepointer(kmem_cache_node, n);
3143 kmem_cache_node->node[node] = n;
3144 #ifdef CONFIG_SLUB_DEBUG
3145 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3146 init_tracking(kmem_cache_node, n);
3148 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3149 init_kmem_cache_node(n);
3150 inc_slabs_node(kmem_cache_node, node, page->objects);
3153 * No locks need to be taken here as it has just been
3154 * initialized and there is no concurrent access.
3156 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3159 static void free_kmem_cache_nodes(struct kmem_cache *s)
3162 struct kmem_cache_node *n;
3164 for_each_kmem_cache_node(s, node, n) {
3165 kmem_cache_free(kmem_cache_node, n);
3166 s->node[node] = NULL;
3170 void __kmem_cache_release(struct kmem_cache *s)
3172 free_percpu(s->cpu_slab);
3173 free_kmem_cache_nodes(s);
3176 static int init_kmem_cache_nodes(struct kmem_cache *s)
3180 for_each_node_state(node, N_NORMAL_MEMORY) {
3181 struct kmem_cache_node *n;
3183 if (slab_state == DOWN) {
3184 early_kmem_cache_node_alloc(node);
3187 n = kmem_cache_alloc_node(kmem_cache_node,
3191 free_kmem_cache_nodes(s);
3196 init_kmem_cache_node(n);
3201 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3203 if (min < MIN_PARTIAL)
3205 else if (min > MAX_PARTIAL)
3207 s->min_partial = min;
3211 * calculate_sizes() determines the order and the distribution of data within
3214 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3216 unsigned long flags = s->flags;
3217 unsigned long size = s->object_size;
3221 * Round up object size to the next word boundary. We can only
3222 * place the free pointer at word boundaries and this determines
3223 * the possible location of the free pointer.
3225 size = ALIGN(size, sizeof(void *));
3227 #ifdef CONFIG_SLUB_DEBUG
3229 * Determine if we can poison the object itself. If the user of
3230 * the slab may touch the object after free or before allocation
3231 * then we should never poison the object itself.
3233 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3235 s->flags |= __OBJECT_POISON;
3237 s->flags &= ~__OBJECT_POISON;
3241 * If we are Redzoning then check if there is some space between the
3242 * end of the object and the free pointer. If not then add an
3243 * additional word to have some bytes to store Redzone information.
3245 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3246 size += sizeof(void *);
3250 * With that we have determined the number of bytes in actual use
3251 * by the object. This is the potential offset to the free pointer.
3255 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3258 * Relocate free pointer after the object if it is not
3259 * permitted to overwrite the first word of the object on
3262 * This is the case if we do RCU, have a constructor or
3263 * destructor or are poisoning the objects.
3266 size += sizeof(void *);
3269 #ifdef CONFIG_SLUB_DEBUG
3270 if (flags & SLAB_STORE_USER)
3272 * Need to store information about allocs and frees after
3275 size += 2 * sizeof(struct track);
3277 if (flags & SLAB_RED_ZONE)
3279 * Add some empty padding so that we can catch
3280 * overwrites from earlier objects rather than let
3281 * tracking information or the free pointer be
3282 * corrupted if a user writes before the start
3285 size += sizeof(void *);
3289 * SLUB stores one object immediately after another beginning from
3290 * offset 0. In order to align the objects we have to simply size
3291 * each object to conform to the alignment.
3293 size = ALIGN(size, s->align);
3295 if (forced_order >= 0)
3296 order = forced_order;
3298 order = calculate_order(size, s->reserved);
3305 s->allocflags |= __GFP_COMP;
3307 if (s->flags & SLAB_CACHE_DMA)
3308 s->allocflags |= GFP_DMA;
3310 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3311 s->allocflags |= __GFP_RECLAIMABLE;
3314 * Determine the number of objects per slab
3316 s->oo = oo_make(order, size, s->reserved);
3317 s->min = oo_make(get_order(size), size, s->reserved);
3318 if (oo_objects(s->oo) > oo_objects(s->max))
3321 return !!oo_objects(s->oo);
3324 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3326 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3329 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3330 s->reserved = sizeof(struct rcu_head);
3332 if (!calculate_sizes(s, -1))
3334 if (disable_higher_order_debug) {
3336 * Disable debugging flags that store metadata if the min slab
3339 if (get_order(s->size) > get_order(s->object_size)) {
3340 s->flags &= ~DEBUG_METADATA_FLAGS;
3342 if (!calculate_sizes(s, -1))
3347 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3348 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3349 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3350 /* Enable fast mode */
3351 s->flags |= __CMPXCHG_DOUBLE;
3355 * The larger the object size is, the more pages we want on the partial
3356 * list to avoid pounding the page allocator excessively.
3358 set_min_partial(s, ilog2(s->size) / 2);
3361 * cpu_partial determined the maximum number of objects kept in the
3362 * per cpu partial lists of a processor.
3364 * Per cpu partial lists mainly contain slabs that just have one
3365 * object freed. If they are used for allocation then they can be
3366 * filled up again with minimal effort. The slab will never hit the
3367 * per node partial lists and therefore no locking will be required.
3369 * This setting also determines
3371 * A) The number of objects from per cpu partial slabs dumped to the
3372 * per node list when we reach the limit.
3373 * B) The number of objects in cpu partial slabs to extract from the
3374 * per node list when we run out of per cpu objects. We only fetch
3375 * 50% to keep some capacity around for frees.
3377 if (!kmem_cache_has_cpu_partial(s))
3379 else if (s->size >= PAGE_SIZE)
3381 else if (s->size >= 1024)
3383 else if (s->size >= 256)
3384 s->cpu_partial = 13;
3386 s->cpu_partial = 30;
3389 s->remote_node_defrag_ratio = 1000;
3391 if (!init_kmem_cache_nodes(s))
3394 if (alloc_kmem_cache_cpus(s))
3397 free_kmem_cache_nodes(s);
3399 if (flags & SLAB_PANIC)
3400 panic("Cannot create slab %s size=%lu realsize=%u "
3401 "order=%u offset=%u flags=%lx\n",
3402 s->name, (unsigned long)s->size, s->size,
3403 oo_order(s->oo), s->offset, flags);
3407 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3410 #ifdef CONFIG_SLUB_DEBUG
3411 void *addr = page_address(page);
3413 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3414 sizeof(long), GFP_ATOMIC);
3417 slab_err(s, page, text, s->name);
3420 get_map(s, page, map);
3421 for_each_object(p, s, addr, page->objects) {
3423 if (!test_bit(slab_index(p, s, addr), map)) {
3424 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3425 print_tracking(s, p);
3434 * Attempt to free all partial slabs on a node.
3435 * This is called from __kmem_cache_shutdown(). We must take list_lock
3436 * because sysfs file might still access partial list after the shutdowning.
3438 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3440 struct page *page, *h;
3442 BUG_ON(irqs_disabled());
3443 spin_lock_irq(&n->list_lock);
3444 list_for_each_entry_safe(page, h, &n->partial, lru) {
3446 remove_partial(n, page);
3447 discard_slab(s, page);
3449 list_slab_objects(s, page,
3450 "Objects remaining in %s on __kmem_cache_shutdown()");
3453 spin_unlock_irq(&n->list_lock);
3457 * Release all resources used by a slab cache.
3459 int __kmem_cache_shutdown(struct kmem_cache *s)
3462 struct kmem_cache_node *n;
3465 /* Attempt to free all objects */
3466 for_each_kmem_cache_node(s, node, n) {
3468 if (n->nr_partial || slabs_node(s, node))
3474 /********************************************************************
3476 *******************************************************************/
3478 static int __init setup_slub_min_order(char *str)
3480 get_option(&str, &slub_min_order);
3485 __setup("slub_min_order=", setup_slub_min_order);
3487 static int __init setup_slub_max_order(char *str)
3489 get_option(&str, &slub_max_order);
3490 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3495 __setup("slub_max_order=", setup_slub_max_order);
3497 static int __init setup_slub_min_objects(char *str)
3499 get_option(&str, &slub_min_objects);
3504 __setup("slub_min_objects=", setup_slub_min_objects);
3506 void *__kmalloc(size_t size, gfp_t flags)
3508 struct kmem_cache *s;
3511 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3512 return kmalloc_large(size, flags);
3514 s = kmalloc_slab(size, flags);
3516 if (unlikely(ZERO_OR_NULL_PTR(s)))
3519 ret = slab_alloc(s, flags, _RET_IP_);
3521 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3523 kasan_kmalloc(s, ret, size);
3527 EXPORT_SYMBOL(__kmalloc);
3530 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3535 flags |= __GFP_COMP | __GFP_NOTRACK;
3536 page = alloc_kmem_pages_node(node, flags, get_order(size));
3538 ptr = page_address(page);
3540 kmalloc_large_node_hook(ptr, size, flags);
3544 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3546 struct kmem_cache *s;
3549 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3550 ret = kmalloc_large_node(size, flags, node);
3552 trace_kmalloc_node(_RET_IP_, ret,
3553 size, PAGE_SIZE << get_order(size),
3559 s = kmalloc_slab(size, flags);
3561 if (unlikely(ZERO_OR_NULL_PTR(s)))
3564 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3566 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3568 kasan_kmalloc(s, ret, size);
3572 EXPORT_SYMBOL(__kmalloc_node);
3575 static size_t __ksize(const void *object)
3579 if (unlikely(object == ZERO_SIZE_PTR))
3582 page = virt_to_head_page(object);
3584 if (unlikely(!PageSlab(page))) {
3585 WARN_ON(!PageCompound(page));
3586 return PAGE_SIZE << compound_order(page);
3589 return slab_ksize(page->slab_cache);
3592 size_t ksize(const void *object)
3594 size_t size = __ksize(object);
3595 /* We assume that ksize callers could use whole allocated area,
3596 so we need unpoison this area. */
3597 kasan_krealloc(object, size);
3600 EXPORT_SYMBOL(ksize);
3602 void kfree(const void *x)
3605 void *object = (void *)x;
3607 trace_kfree(_RET_IP_, x);
3609 if (unlikely(ZERO_OR_NULL_PTR(x)))
3612 page = virt_to_head_page(x);
3613 if (unlikely(!PageSlab(page))) {
3614 BUG_ON(!PageCompound(page));
3616 __free_kmem_pages(page, compound_order(page));
3619 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3621 EXPORT_SYMBOL(kfree);
3623 #define SHRINK_PROMOTE_MAX 32
3626 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3627 * up most to the head of the partial lists. New allocations will then
3628 * fill those up and thus they can be removed from the partial lists.
3630 * The slabs with the least items are placed last. This results in them
3631 * being allocated from last increasing the chance that the last objects
3632 * are freed in them.
3634 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3638 struct kmem_cache_node *n;
3641 struct list_head discard;
3642 struct list_head promote[SHRINK_PROMOTE_MAX];
3643 unsigned long flags;
3648 * Disable empty slabs caching. Used to avoid pinning offline
3649 * memory cgroups by kmem pages that can be freed.
3655 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3656 * so we have to make sure the change is visible.
3658 kick_all_cpus_sync();
3662 for_each_kmem_cache_node(s, node, n) {
3663 INIT_LIST_HEAD(&discard);
3664 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3665 INIT_LIST_HEAD(promote + i);
3667 spin_lock_irqsave(&n->list_lock, flags);
3670 * Build lists of slabs to discard or promote.
3672 * Note that concurrent frees may occur while we hold the
3673 * list_lock. page->inuse here is the upper limit.
3675 list_for_each_entry_safe(page, t, &n->partial, lru) {
3676 int free = page->objects - page->inuse;
3678 /* Do not reread page->inuse */
3681 /* We do not keep full slabs on the list */
3684 if (free == page->objects) {
3685 list_move(&page->lru, &discard);
3687 } else if (free <= SHRINK_PROMOTE_MAX)
3688 list_move(&page->lru, promote + free - 1);
3692 * Promote the slabs filled up most to the head of the
3695 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3696 list_splice(promote + i, &n->partial);
3698 spin_unlock_irqrestore(&n->list_lock, flags);
3700 /* Release empty slabs */
3701 list_for_each_entry_safe(page, t, &discard, lru)
3702 discard_slab(s, page);
3704 if (slabs_node(s, node))
3711 static int slab_mem_going_offline_callback(void *arg)
3713 struct kmem_cache *s;
3715 mutex_lock(&slab_mutex);
3716 list_for_each_entry(s, &slab_caches, list)
3717 __kmem_cache_shrink(s, false);
3718 mutex_unlock(&slab_mutex);
3723 static void slab_mem_offline_callback(void *arg)
3725 struct kmem_cache_node *n;
3726 struct kmem_cache *s;
3727 struct memory_notify *marg = arg;
3730 offline_node = marg->status_change_nid_normal;
3733 * If the node still has available memory. we need kmem_cache_node
3736 if (offline_node < 0)
3739 mutex_lock(&slab_mutex);
3740 list_for_each_entry(s, &slab_caches, list) {
3741 n = get_node(s, offline_node);
3744 * if n->nr_slabs > 0, slabs still exist on the node
3745 * that is going down. We were unable to free them,
3746 * and offline_pages() function shouldn't call this
3747 * callback. So, we must fail.
3749 BUG_ON(slabs_node(s, offline_node));
3751 s->node[offline_node] = NULL;
3752 kmem_cache_free(kmem_cache_node, n);
3755 mutex_unlock(&slab_mutex);
3758 static int slab_mem_going_online_callback(void *arg)
3760 struct kmem_cache_node *n;
3761 struct kmem_cache *s;
3762 struct memory_notify *marg = arg;
3763 int nid = marg->status_change_nid_normal;
3767 * If the node's memory is already available, then kmem_cache_node is
3768 * already created. Nothing to do.
3774 * We are bringing a node online. No memory is available yet. We must
3775 * allocate a kmem_cache_node structure in order to bring the node
3778 mutex_lock(&slab_mutex);
3779 list_for_each_entry(s, &slab_caches, list) {
3781 * XXX: kmem_cache_alloc_node will fallback to other nodes
3782 * since memory is not yet available from the node that
3785 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3790 init_kmem_cache_node(n);
3794 mutex_unlock(&slab_mutex);
3798 static int slab_memory_callback(struct notifier_block *self,
3799 unsigned long action, void *arg)
3804 case MEM_GOING_ONLINE:
3805 ret = slab_mem_going_online_callback(arg);
3807 case MEM_GOING_OFFLINE:
3808 ret = slab_mem_going_offline_callback(arg);
3811 case MEM_CANCEL_ONLINE:
3812 slab_mem_offline_callback(arg);
3815 case MEM_CANCEL_OFFLINE:
3819 ret = notifier_from_errno(ret);
3825 static struct notifier_block slab_memory_callback_nb = {
3826 .notifier_call = slab_memory_callback,
3827 .priority = SLAB_CALLBACK_PRI,
3830 /********************************************************************
3831 * Basic setup of slabs
3832 *******************************************************************/
3835 * Used for early kmem_cache structures that were allocated using
3836 * the page allocator. Allocate them properly then fix up the pointers
3837 * that may be pointing to the wrong kmem_cache structure.
3840 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3843 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3844 struct kmem_cache_node *n;
3846 memcpy(s, static_cache, kmem_cache->object_size);
3849 * This runs very early, and only the boot processor is supposed to be
3850 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3853 __flush_cpu_slab(s, smp_processor_id());
3854 for_each_kmem_cache_node(s, node, n) {
3857 list_for_each_entry(p, &n->partial, lru)
3860 #ifdef CONFIG_SLUB_DEBUG
3861 list_for_each_entry(p, &n->full, lru)
3865 slab_init_memcg_params(s);
3866 list_add(&s->list, &slab_caches);
3870 void __init kmem_cache_init(void)
3872 static __initdata struct kmem_cache boot_kmem_cache,
3873 boot_kmem_cache_node;
3875 if (debug_guardpage_minorder())
3878 kmem_cache_node = &boot_kmem_cache_node;
3879 kmem_cache = &boot_kmem_cache;
3881 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3882 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3884 register_hotmemory_notifier(&slab_memory_callback_nb);
3886 /* Able to allocate the per node structures */
3887 slab_state = PARTIAL;
3889 create_boot_cache(kmem_cache, "kmem_cache",
3890 offsetof(struct kmem_cache, node) +
3891 nr_node_ids * sizeof(struct kmem_cache_node *),
3892 SLAB_HWCACHE_ALIGN);
3894 kmem_cache = bootstrap(&boot_kmem_cache);
3897 * Allocate kmem_cache_node properly from the kmem_cache slab.
3898 * kmem_cache_node is separately allocated so no need to
3899 * update any list pointers.
3901 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3903 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3904 setup_kmalloc_cache_index_table();
3905 create_kmalloc_caches(0);
3908 register_cpu_notifier(&slab_notifier);
3911 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3913 slub_min_order, slub_max_order, slub_min_objects,
3914 nr_cpu_ids, nr_node_ids);
3917 void __init kmem_cache_init_late(void)
3922 __kmem_cache_alias(const char *name, size_t size, size_t align,
3923 unsigned long flags, void (*ctor)(void *))
3925 struct kmem_cache *s, *c;
3927 s = find_mergeable(size, align, flags, name, ctor);
3932 * Adjust the object sizes so that we clear
3933 * the complete object on kzalloc.
3935 s->object_size = max(s->object_size, (int)size);
3936 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3938 for_each_memcg_cache(c, s) {
3939 c->object_size = s->object_size;
3940 c->inuse = max_t(int, c->inuse,
3941 ALIGN(size, sizeof(void *)));
3944 if (sysfs_slab_alias(s, name)) {
3953 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3957 err = kmem_cache_open(s, flags);
3961 /* Mutex is not taken during early boot */
3962 if (slab_state <= UP)
3965 memcg_propagate_slab_attrs(s);
3966 err = sysfs_slab_add(s);
3968 __kmem_cache_release(s);
3975 * Use the cpu notifier to insure that the cpu slabs are flushed when
3978 static int slab_cpuup_callback(struct notifier_block *nfb,
3979 unsigned long action, void *hcpu)
3981 long cpu = (long)hcpu;
3982 struct kmem_cache *s;
3983 unsigned long flags;
3986 case CPU_UP_CANCELED:
3987 case CPU_UP_CANCELED_FROZEN:
3989 case CPU_DEAD_FROZEN:
3990 mutex_lock(&slab_mutex);
3991 list_for_each_entry(s, &slab_caches, list) {
3992 local_irq_save(flags);
3993 __flush_cpu_slab(s, cpu);
3994 local_irq_restore(flags);
3996 mutex_unlock(&slab_mutex);
4004 static struct notifier_block slab_notifier = {
4005 .notifier_call = slab_cpuup_callback
4010 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4012 struct kmem_cache *s;
4015 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4016 return kmalloc_large(size, gfpflags);
4018 s = kmalloc_slab(size, gfpflags);
4020 if (unlikely(ZERO_OR_NULL_PTR(s)))
4023 ret = slab_alloc(s, gfpflags, caller);
4025 /* Honor the call site pointer we received. */
4026 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4032 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4033 int node, unsigned long caller)
4035 struct kmem_cache *s;
4038 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4039 ret = kmalloc_large_node(size, gfpflags, node);
4041 trace_kmalloc_node(caller, ret,
4042 size, PAGE_SIZE << get_order(size),
4048 s = kmalloc_slab(size, gfpflags);
4050 if (unlikely(ZERO_OR_NULL_PTR(s)))
4053 ret = slab_alloc_node(s, gfpflags, node, caller);
4055 /* Honor the call site pointer we received. */
4056 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4063 static int count_inuse(struct page *page)
4068 static int count_total(struct page *page)
4070 return page->objects;
4074 #ifdef CONFIG_SLUB_DEBUG
4075 static int validate_slab(struct kmem_cache *s, struct page *page,
4079 void *addr = page_address(page);
4081 if (!check_slab(s, page) ||
4082 !on_freelist(s, page, NULL))
4085 /* Now we know that a valid freelist exists */
4086 bitmap_zero(map, page->objects);
4088 get_map(s, page, map);
4089 for_each_object(p, s, addr, page->objects) {
4090 if (test_bit(slab_index(p, s, addr), map))
4091 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4095 for_each_object(p, s, addr, page->objects)
4096 if (!test_bit(slab_index(p, s, addr), map))
4097 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4102 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4106 validate_slab(s, page, map);
4110 static int validate_slab_node(struct kmem_cache *s,
4111 struct kmem_cache_node *n, unsigned long *map)
4113 unsigned long count = 0;
4115 unsigned long flags;
4117 spin_lock_irqsave(&n->list_lock, flags);
4119 list_for_each_entry(page, &n->partial, lru) {
4120 validate_slab_slab(s, page, map);
4123 if (count != n->nr_partial)
4124 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4125 s->name, count, n->nr_partial);
4127 if (!(s->flags & SLAB_STORE_USER))
4130 list_for_each_entry(page, &n->full, lru) {
4131 validate_slab_slab(s, page, map);
4134 if (count != atomic_long_read(&n->nr_slabs))
4135 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4136 s->name, count, atomic_long_read(&n->nr_slabs));
4139 spin_unlock_irqrestore(&n->list_lock, flags);
4143 static long validate_slab_cache(struct kmem_cache *s)
4146 unsigned long count = 0;
4147 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4148 sizeof(unsigned long), GFP_KERNEL);
4149 struct kmem_cache_node *n;
4155 for_each_kmem_cache_node(s, node, n)
4156 count += validate_slab_node(s, n, map);
4161 * Generate lists of code addresses where slabcache objects are allocated
4166 unsigned long count;
4173 DECLARE_BITMAP(cpus, NR_CPUS);
4179 unsigned long count;
4180 struct location *loc;
4183 static void free_loc_track(struct loc_track *t)
4186 free_pages((unsigned long)t->loc,
4187 get_order(sizeof(struct location) * t->max));
4190 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4195 order = get_order(sizeof(struct location) * max);
4197 l = (void *)__get_free_pages(flags, order);
4202 memcpy(l, t->loc, sizeof(struct location) * t->count);
4210 static int add_location(struct loc_track *t, struct kmem_cache *s,
4211 const struct track *track)
4213 long start, end, pos;
4215 unsigned long caddr;
4216 unsigned long age = jiffies - track->when;
4222 pos = start + (end - start + 1) / 2;
4225 * There is nothing at "end". If we end up there
4226 * we need to add something to before end.
4231 caddr = t->loc[pos].addr;
4232 if (track->addr == caddr) {
4238 if (age < l->min_time)
4240 if (age > l->max_time)
4243 if (track->pid < l->min_pid)
4244 l->min_pid = track->pid;
4245 if (track->pid > l->max_pid)
4246 l->max_pid = track->pid;
4248 cpumask_set_cpu(track->cpu,
4249 to_cpumask(l->cpus));
4251 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4255 if (track->addr < caddr)
4262 * Not found. Insert new tracking element.
4264 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4270 (t->count - pos) * sizeof(struct location));
4273 l->addr = track->addr;
4277 l->min_pid = track->pid;
4278 l->max_pid = track->pid;
4279 cpumask_clear(to_cpumask(l->cpus));
4280 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4281 nodes_clear(l->nodes);
4282 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4286 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4287 struct page *page, enum track_item alloc,
4290 void *addr = page_address(page);
4293 bitmap_zero(map, page->objects);
4294 get_map(s, page, map);
4296 for_each_object(p, s, addr, page->objects)
4297 if (!test_bit(slab_index(p, s, addr), map))
4298 add_location(t, s, get_track(s, p, alloc));
4301 static int list_locations(struct kmem_cache *s, char *buf,
4302 enum track_item alloc)
4306 struct loc_track t = { 0, 0, NULL };
4308 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4309 sizeof(unsigned long), GFP_KERNEL);
4310 struct kmem_cache_node *n;
4312 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4315 return sprintf(buf, "Out of memory\n");
4317 /* Push back cpu slabs */
4320 for_each_kmem_cache_node(s, node, n) {
4321 unsigned long flags;
4324 if (!atomic_long_read(&n->nr_slabs))
4327 spin_lock_irqsave(&n->list_lock, flags);
4328 list_for_each_entry(page, &n->partial, lru)
4329 process_slab(&t, s, page, alloc, map);
4330 list_for_each_entry(page, &n->full, lru)
4331 process_slab(&t, s, page, alloc, map);
4332 spin_unlock_irqrestore(&n->list_lock, flags);
4335 for (i = 0; i < t.count; i++) {
4336 struct location *l = &t.loc[i];
4338 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4340 len += sprintf(buf + len, "%7ld ", l->count);
4343 len += sprintf(buf + len, "%pS", (void *)l->addr);
4345 len += sprintf(buf + len, "<not-available>");
4347 if (l->sum_time != l->min_time) {
4348 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4350 (long)div_u64(l->sum_time, l->count),
4353 len += sprintf(buf + len, " age=%ld",
4356 if (l->min_pid != l->max_pid)
4357 len += sprintf(buf + len, " pid=%ld-%ld",
4358 l->min_pid, l->max_pid);
4360 len += sprintf(buf + len, " pid=%ld",
4363 if (num_online_cpus() > 1 &&
4364 !cpumask_empty(to_cpumask(l->cpus)) &&
4365 len < PAGE_SIZE - 60)
4366 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4368 cpumask_pr_args(to_cpumask(l->cpus)));
4370 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4371 len < PAGE_SIZE - 60)
4372 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4374 nodemask_pr_args(&l->nodes));
4376 len += sprintf(buf + len, "\n");
4382 len += sprintf(buf, "No data\n");
4387 #ifdef SLUB_RESILIENCY_TEST
4388 static void __init resiliency_test(void)
4392 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4394 pr_err("SLUB resiliency testing\n");
4395 pr_err("-----------------------\n");
4396 pr_err("A. Corruption after allocation\n");
4398 p = kzalloc(16, GFP_KERNEL);
4400 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4403 validate_slab_cache(kmalloc_caches[4]);
4405 /* Hmmm... The next two are dangerous */
4406 p = kzalloc(32, GFP_KERNEL);
4407 p[32 + sizeof(void *)] = 0x34;
4408 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4410 pr_err("If allocated object is overwritten then not detectable\n\n");
4412 validate_slab_cache(kmalloc_caches[5]);
4413 p = kzalloc(64, GFP_KERNEL);
4414 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4416 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4418 pr_err("If allocated object is overwritten then not detectable\n\n");
4419 validate_slab_cache(kmalloc_caches[6]);
4421 pr_err("\nB. Corruption after free\n");
4422 p = kzalloc(128, GFP_KERNEL);
4425 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4426 validate_slab_cache(kmalloc_caches[7]);
4428 p = kzalloc(256, GFP_KERNEL);
4431 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4432 validate_slab_cache(kmalloc_caches[8]);
4434 p = kzalloc(512, GFP_KERNEL);
4437 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4438 validate_slab_cache(kmalloc_caches[9]);
4442 static void resiliency_test(void) {};
4447 enum slab_stat_type {
4448 SL_ALL, /* All slabs */
4449 SL_PARTIAL, /* Only partially allocated slabs */
4450 SL_CPU, /* Only slabs used for cpu caches */
4451 SL_OBJECTS, /* Determine allocated objects not slabs */
4452 SL_TOTAL /* Determine object capacity not slabs */
4455 #define SO_ALL (1 << SL_ALL)
4456 #define SO_PARTIAL (1 << SL_PARTIAL)
4457 #define SO_CPU (1 << SL_CPU)
4458 #define SO_OBJECTS (1 << SL_OBJECTS)
4459 #define SO_TOTAL (1 << SL_TOTAL)
4461 static ssize_t show_slab_objects(struct kmem_cache *s,
4462 char *buf, unsigned long flags)
4464 unsigned long total = 0;
4467 unsigned long *nodes;
4469 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4473 if (flags & SO_CPU) {
4476 for_each_possible_cpu(cpu) {
4477 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4482 page = READ_ONCE(c->page);
4486 node = page_to_nid(page);
4487 if (flags & SO_TOTAL)
4489 else if (flags & SO_OBJECTS)
4497 page = READ_ONCE(c->partial);
4499 node = page_to_nid(page);
4500 if (flags & SO_TOTAL)
4502 else if (flags & SO_OBJECTS)
4513 #ifdef CONFIG_SLUB_DEBUG
4514 if (flags & SO_ALL) {
4515 struct kmem_cache_node *n;
4517 for_each_kmem_cache_node(s, node, n) {
4519 if (flags & SO_TOTAL)
4520 x = atomic_long_read(&n->total_objects);
4521 else if (flags & SO_OBJECTS)
4522 x = atomic_long_read(&n->total_objects) -
4523 count_partial(n, count_free);
4525 x = atomic_long_read(&n->nr_slabs);
4532 if (flags & SO_PARTIAL) {
4533 struct kmem_cache_node *n;
4535 for_each_kmem_cache_node(s, node, n) {
4536 if (flags & SO_TOTAL)
4537 x = count_partial(n, count_total);
4538 else if (flags & SO_OBJECTS)
4539 x = count_partial(n, count_inuse);
4546 x = sprintf(buf, "%lu", total);
4548 for (node = 0; node < nr_node_ids; node++)
4550 x += sprintf(buf + x, " N%d=%lu",
4555 return x + sprintf(buf + x, "\n");
4558 #ifdef CONFIG_SLUB_DEBUG
4559 static int any_slab_objects(struct kmem_cache *s)
4562 struct kmem_cache_node *n;
4564 for_each_kmem_cache_node(s, node, n)
4565 if (atomic_long_read(&n->total_objects))
4572 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4573 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4575 struct slab_attribute {
4576 struct attribute attr;
4577 ssize_t (*show)(struct kmem_cache *s, char *buf);
4578 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4581 #define SLAB_ATTR_RO(_name) \
4582 static struct slab_attribute _name##_attr = \
4583 __ATTR(_name, 0400, _name##_show, NULL)
4585 #define SLAB_ATTR(_name) \
4586 static struct slab_attribute _name##_attr = \
4587 __ATTR(_name, 0600, _name##_show, _name##_store)
4589 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4591 return sprintf(buf, "%d\n", s->size);
4593 SLAB_ATTR_RO(slab_size);
4595 static ssize_t align_show(struct kmem_cache *s, char *buf)
4597 return sprintf(buf, "%d\n", s->align);
4599 SLAB_ATTR_RO(align);
4601 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4603 return sprintf(buf, "%d\n", s->object_size);
4605 SLAB_ATTR_RO(object_size);
4607 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4609 return sprintf(buf, "%d\n", oo_objects(s->oo));
4611 SLAB_ATTR_RO(objs_per_slab);
4613 static ssize_t order_store(struct kmem_cache *s,
4614 const char *buf, size_t length)
4616 unsigned long order;
4619 err = kstrtoul(buf, 10, &order);
4623 if (order > slub_max_order || order < slub_min_order)
4626 calculate_sizes(s, order);
4630 static ssize_t order_show(struct kmem_cache *s, char *buf)
4632 return sprintf(buf, "%d\n", oo_order(s->oo));
4636 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4638 return sprintf(buf, "%lu\n", s->min_partial);
4641 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4647 err = kstrtoul(buf, 10, &min);
4651 set_min_partial(s, min);
4654 SLAB_ATTR(min_partial);
4656 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4658 return sprintf(buf, "%u\n", s->cpu_partial);
4661 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4664 unsigned long objects;
4667 err = kstrtoul(buf, 10, &objects);
4670 if (objects && !kmem_cache_has_cpu_partial(s))
4673 s->cpu_partial = objects;
4677 SLAB_ATTR(cpu_partial);
4679 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4683 return sprintf(buf, "%pS\n", s->ctor);
4687 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4689 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4691 SLAB_ATTR_RO(aliases);
4693 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4695 return show_slab_objects(s, buf, SO_PARTIAL);
4697 SLAB_ATTR_RO(partial);
4699 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4701 return show_slab_objects(s, buf, SO_CPU);
4703 SLAB_ATTR_RO(cpu_slabs);
4705 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4707 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4709 SLAB_ATTR_RO(objects);
4711 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4713 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4715 SLAB_ATTR_RO(objects_partial);
4717 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4724 for_each_online_cpu(cpu) {
4725 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4728 pages += page->pages;
4729 objects += page->pobjects;
4733 len = sprintf(buf, "%d(%d)", objects, pages);
4736 for_each_online_cpu(cpu) {
4737 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4739 if (page && len < PAGE_SIZE - 20)
4740 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4741 page->pobjects, page->pages);
4744 return len + sprintf(buf + len, "\n");
4746 SLAB_ATTR_RO(slabs_cpu_partial);
4748 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4750 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4753 static ssize_t reclaim_account_store(struct kmem_cache *s,
4754 const char *buf, size_t length)
4756 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4758 s->flags |= SLAB_RECLAIM_ACCOUNT;
4761 SLAB_ATTR(reclaim_account);
4763 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4765 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4767 SLAB_ATTR_RO(hwcache_align);
4769 #ifdef CONFIG_ZONE_DMA
4770 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4772 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4774 SLAB_ATTR_RO(cache_dma);
4777 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4779 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4781 SLAB_ATTR_RO(destroy_by_rcu);
4783 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4785 return sprintf(buf, "%d\n", s->reserved);
4787 SLAB_ATTR_RO(reserved);
4789 #ifdef CONFIG_SLUB_DEBUG
4790 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4792 return show_slab_objects(s, buf, SO_ALL);
4794 SLAB_ATTR_RO(slabs);
4796 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4798 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4800 SLAB_ATTR_RO(total_objects);
4802 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4804 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
4807 static ssize_t sanity_checks_store(struct kmem_cache *s,
4808 const char *buf, size_t length)
4810 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
4811 if (buf[0] == '1') {
4812 s->flags &= ~__CMPXCHG_DOUBLE;
4813 s->flags |= SLAB_CONSISTENCY_CHECKS;
4817 SLAB_ATTR(sanity_checks);
4819 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4821 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4824 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4828 * Tracing a merged cache is going to give confusing results
4829 * as well as cause other issues like converting a mergeable
4830 * cache into an umergeable one.
4832 if (s->refcount > 1)
4835 s->flags &= ~SLAB_TRACE;
4836 if (buf[0] == '1') {
4837 s->flags &= ~__CMPXCHG_DOUBLE;
4838 s->flags |= SLAB_TRACE;
4844 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4846 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4849 static ssize_t red_zone_store(struct kmem_cache *s,
4850 const char *buf, size_t length)
4852 if (any_slab_objects(s))
4855 s->flags &= ~SLAB_RED_ZONE;
4856 if (buf[0] == '1') {
4857 s->flags |= SLAB_RED_ZONE;
4859 calculate_sizes(s, -1);
4862 SLAB_ATTR(red_zone);
4864 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4866 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4869 static ssize_t poison_store(struct kmem_cache *s,
4870 const char *buf, size_t length)
4872 if (any_slab_objects(s))
4875 s->flags &= ~SLAB_POISON;
4876 if (buf[0] == '1') {
4877 s->flags |= SLAB_POISON;
4879 calculate_sizes(s, -1);
4884 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4886 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4889 static ssize_t store_user_store(struct kmem_cache *s,
4890 const char *buf, size_t length)
4892 if (any_slab_objects(s))
4895 s->flags &= ~SLAB_STORE_USER;
4896 if (buf[0] == '1') {
4897 s->flags &= ~__CMPXCHG_DOUBLE;
4898 s->flags |= SLAB_STORE_USER;
4900 calculate_sizes(s, -1);
4903 SLAB_ATTR(store_user);
4905 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4910 static ssize_t validate_store(struct kmem_cache *s,
4911 const char *buf, size_t length)
4915 if (buf[0] == '1') {
4916 ret = validate_slab_cache(s);
4922 SLAB_ATTR(validate);
4924 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4926 if (!(s->flags & SLAB_STORE_USER))
4928 return list_locations(s, buf, TRACK_ALLOC);
4930 SLAB_ATTR_RO(alloc_calls);
4932 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4934 if (!(s->flags & SLAB_STORE_USER))
4936 return list_locations(s, buf, TRACK_FREE);
4938 SLAB_ATTR_RO(free_calls);
4939 #endif /* CONFIG_SLUB_DEBUG */
4941 #ifdef CONFIG_FAILSLAB
4942 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4944 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4947 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4950 if (s->refcount > 1)
4953 s->flags &= ~SLAB_FAILSLAB;
4955 s->flags |= SLAB_FAILSLAB;
4958 SLAB_ATTR(failslab);
4961 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4966 static ssize_t shrink_store(struct kmem_cache *s,
4967 const char *buf, size_t length)
4970 kmem_cache_shrink(s);
4978 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4980 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4983 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4984 const char *buf, size_t length)
4986 unsigned long ratio;
4989 err = kstrtoul(buf, 10, &ratio);
4994 s->remote_node_defrag_ratio = ratio * 10;
4998 SLAB_ATTR(remote_node_defrag_ratio);
5001 #ifdef CONFIG_SLUB_STATS
5002 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5004 unsigned long sum = 0;
5007 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5012 for_each_online_cpu(cpu) {
5013 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5019 len = sprintf(buf, "%lu", sum);
5022 for_each_online_cpu(cpu) {
5023 if (data[cpu] && len < PAGE_SIZE - 20)
5024 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5028 return len + sprintf(buf + len, "\n");
5031 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5035 for_each_online_cpu(cpu)
5036 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5039 #define STAT_ATTR(si, text) \
5040 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5042 return show_stat(s, buf, si); \
5044 static ssize_t text##_store(struct kmem_cache *s, \
5045 const char *buf, size_t length) \
5047 if (buf[0] != '0') \
5049 clear_stat(s, si); \
5054 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5055 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5056 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5057 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5058 STAT_ATTR(FREE_FROZEN, free_frozen);
5059 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5060 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5061 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5062 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5063 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5064 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5065 STAT_ATTR(FREE_SLAB, free_slab);
5066 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5067 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5068 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5069 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5070 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5071 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5072 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5073 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5074 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5075 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5076 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5077 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5078 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5079 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5082 static struct attribute *slab_attrs[] = {
5083 &slab_size_attr.attr,
5084 &object_size_attr.attr,
5085 &objs_per_slab_attr.attr,
5087 &min_partial_attr.attr,
5088 &cpu_partial_attr.attr,
5090 &objects_partial_attr.attr,
5092 &cpu_slabs_attr.attr,
5096 &hwcache_align_attr.attr,
5097 &reclaim_account_attr.attr,
5098 &destroy_by_rcu_attr.attr,
5100 &reserved_attr.attr,
5101 &slabs_cpu_partial_attr.attr,
5102 #ifdef CONFIG_SLUB_DEBUG
5103 &total_objects_attr.attr,
5105 &sanity_checks_attr.attr,
5107 &red_zone_attr.attr,
5109 &store_user_attr.attr,
5110 &validate_attr.attr,
5111 &alloc_calls_attr.attr,
5112 &free_calls_attr.attr,
5114 #ifdef CONFIG_ZONE_DMA
5115 &cache_dma_attr.attr,
5118 &remote_node_defrag_ratio_attr.attr,
5120 #ifdef CONFIG_SLUB_STATS
5121 &alloc_fastpath_attr.attr,
5122 &alloc_slowpath_attr.attr,
5123 &free_fastpath_attr.attr,
5124 &free_slowpath_attr.attr,
5125 &free_frozen_attr.attr,
5126 &free_add_partial_attr.attr,
5127 &free_remove_partial_attr.attr,
5128 &alloc_from_partial_attr.attr,
5129 &alloc_slab_attr.attr,
5130 &alloc_refill_attr.attr,
5131 &alloc_node_mismatch_attr.attr,
5132 &free_slab_attr.attr,
5133 &cpuslab_flush_attr.attr,
5134 &deactivate_full_attr.attr,
5135 &deactivate_empty_attr.attr,
5136 &deactivate_to_head_attr.attr,
5137 &deactivate_to_tail_attr.attr,
5138 &deactivate_remote_frees_attr.attr,
5139 &deactivate_bypass_attr.attr,
5140 &order_fallback_attr.attr,
5141 &cmpxchg_double_fail_attr.attr,
5142 &cmpxchg_double_cpu_fail_attr.attr,
5143 &cpu_partial_alloc_attr.attr,
5144 &cpu_partial_free_attr.attr,
5145 &cpu_partial_node_attr.attr,
5146 &cpu_partial_drain_attr.attr,
5148 #ifdef CONFIG_FAILSLAB
5149 &failslab_attr.attr,
5155 static struct attribute_group slab_attr_group = {
5156 .attrs = slab_attrs,
5159 static ssize_t slab_attr_show(struct kobject *kobj,
5160 struct attribute *attr,
5163 struct slab_attribute *attribute;
5164 struct kmem_cache *s;
5167 attribute = to_slab_attr(attr);
5170 if (!attribute->show)
5173 err = attribute->show(s, buf);
5178 static ssize_t slab_attr_store(struct kobject *kobj,
5179 struct attribute *attr,
5180 const char *buf, size_t len)
5182 struct slab_attribute *attribute;
5183 struct kmem_cache *s;
5186 attribute = to_slab_attr(attr);
5189 if (!attribute->store)
5192 err = attribute->store(s, buf, len);
5194 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5195 struct kmem_cache *c;
5197 mutex_lock(&slab_mutex);
5198 if (s->max_attr_size < len)
5199 s->max_attr_size = len;
5202 * This is a best effort propagation, so this function's return
5203 * value will be determined by the parent cache only. This is
5204 * basically because not all attributes will have a well
5205 * defined semantics for rollbacks - most of the actions will
5206 * have permanent effects.
5208 * Returning the error value of any of the children that fail
5209 * is not 100 % defined, in the sense that users seeing the
5210 * error code won't be able to know anything about the state of
5213 * Only returning the error code for the parent cache at least
5214 * has well defined semantics. The cache being written to
5215 * directly either failed or succeeded, in which case we loop
5216 * through the descendants with best-effort propagation.
5218 for_each_memcg_cache(c, s)
5219 attribute->store(c, buf, len);
5220 mutex_unlock(&slab_mutex);
5226 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5230 char *buffer = NULL;
5231 struct kmem_cache *root_cache;
5233 if (is_root_cache(s))
5236 root_cache = s->memcg_params.root_cache;
5239 * This mean this cache had no attribute written. Therefore, no point
5240 * in copying default values around
5242 if (!root_cache->max_attr_size)
5245 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5248 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5250 if (!attr || !attr->store || !attr->show)
5254 * It is really bad that we have to allocate here, so we will
5255 * do it only as a fallback. If we actually allocate, though,
5256 * we can just use the allocated buffer until the end.
5258 * Most of the slub attributes will tend to be very small in
5259 * size, but sysfs allows buffers up to a page, so they can
5260 * theoretically happen.
5264 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5267 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5268 if (WARN_ON(!buffer))
5273 attr->show(root_cache, buf);
5274 attr->store(s, buf, strlen(buf));
5278 free_page((unsigned long)buffer);
5282 static void kmem_cache_release(struct kobject *k)
5284 slab_kmem_cache_release(to_slab(k));
5287 static const struct sysfs_ops slab_sysfs_ops = {
5288 .show = slab_attr_show,
5289 .store = slab_attr_store,
5292 static struct kobj_type slab_ktype = {
5293 .sysfs_ops = &slab_sysfs_ops,
5294 .release = kmem_cache_release,
5297 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5299 struct kobj_type *ktype = get_ktype(kobj);
5301 if (ktype == &slab_ktype)
5306 static const struct kset_uevent_ops slab_uevent_ops = {
5307 .filter = uevent_filter,
5310 static struct kset *slab_kset;
5312 static inline struct kset *cache_kset(struct kmem_cache *s)
5315 if (!is_root_cache(s))
5316 return s->memcg_params.root_cache->memcg_kset;
5321 #define ID_STR_LENGTH 64
5323 /* Create a unique string id for a slab cache:
5325 * Format :[flags-]size
5327 static char *create_unique_id(struct kmem_cache *s)
5329 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5336 * First flags affecting slabcache operations. We will only
5337 * get here for aliasable slabs so we do not need to support
5338 * too many flags. The flags here must cover all flags that
5339 * are matched during merging to guarantee that the id is
5342 if (s->flags & SLAB_CACHE_DMA)
5344 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5346 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5348 if (!(s->flags & SLAB_NOTRACK))
5350 if (s->flags & SLAB_ACCOUNT)
5354 p += sprintf(p, "%07d", s->size);
5356 BUG_ON(p > name + ID_STR_LENGTH - 1);
5360 static int sysfs_slab_add(struct kmem_cache *s)
5364 int unmergeable = slab_unmergeable(s);
5368 * Slabcache can never be merged so we can use the name proper.
5369 * This is typically the case for debug situations. In that
5370 * case we can catch duplicate names easily.
5372 sysfs_remove_link(&slab_kset->kobj, s->name);
5376 * Create a unique name for the slab as a target
5379 name = create_unique_id(s);
5382 s->kobj.kset = cache_kset(s);
5383 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5387 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5392 if (is_root_cache(s)) {
5393 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5394 if (!s->memcg_kset) {
5401 kobject_uevent(&s->kobj, KOBJ_ADD);
5403 /* Setup first alias */
5404 sysfs_slab_alias(s, s->name);
5411 kobject_del(&s->kobj);
5415 void sysfs_slab_remove(struct kmem_cache *s)
5417 if (slab_state < FULL)
5419 * Sysfs has not been setup yet so no need to remove the
5425 kset_unregister(s->memcg_kset);
5427 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5428 kobject_del(&s->kobj);
5429 kobject_put(&s->kobj);
5433 * Need to buffer aliases during bootup until sysfs becomes
5434 * available lest we lose that information.
5436 struct saved_alias {
5437 struct kmem_cache *s;
5439 struct saved_alias *next;
5442 static struct saved_alias *alias_list;
5444 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5446 struct saved_alias *al;
5448 if (slab_state == FULL) {
5450 * If we have a leftover link then remove it.
5452 sysfs_remove_link(&slab_kset->kobj, name);
5453 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5456 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5462 al->next = alias_list;
5467 static int __init slab_sysfs_init(void)
5469 struct kmem_cache *s;
5472 mutex_lock(&slab_mutex);
5474 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5476 mutex_unlock(&slab_mutex);
5477 pr_err("Cannot register slab subsystem.\n");
5483 list_for_each_entry(s, &slab_caches, list) {
5484 err = sysfs_slab_add(s);
5486 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5490 while (alias_list) {
5491 struct saved_alias *al = alias_list;
5493 alias_list = alias_list->next;
5494 err = sysfs_slab_alias(al->s, al->name);
5496 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5501 mutex_unlock(&slab_mutex);
5506 __initcall(slab_sysfs_init);
5507 #endif /* CONFIG_SYSFS */
5510 * The /proc/slabinfo ABI
5512 #ifdef CONFIG_SLABINFO
5513 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5515 unsigned long nr_slabs = 0;
5516 unsigned long nr_objs = 0;
5517 unsigned long nr_free = 0;
5519 struct kmem_cache_node *n;
5521 for_each_kmem_cache_node(s, node, n) {
5522 nr_slabs += node_nr_slabs(n);
5523 nr_objs += node_nr_objs(n);
5524 nr_free += count_partial(n, count_free);
5527 sinfo->active_objs = nr_objs - nr_free;
5528 sinfo->num_objs = nr_objs;
5529 sinfo->active_slabs = nr_slabs;
5530 sinfo->num_slabs = nr_slabs;
5531 sinfo->objects_per_slab = oo_objects(s->oo);
5532 sinfo->cache_order = oo_order(s->oo);
5535 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5539 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5540 size_t count, loff_t *ppos)
5544 #endif /* CONFIG_SLABINFO */