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>
43 #include "kasan/kasan.h"
48 * 1. slab_mutex (Global Mutex)
50 * 3. slab_lock(page) (Only on some arches and for debugging)
54 * The role of the slab_mutex is to protect the list of all the slabs
55 * and to synchronize major metadata changes to slab cache structures.
57 * The slab_lock is only used for debugging and on arches that do not
58 * have the ability to do a cmpxchg_double. It only protects the second
59 * double word in the page struct. Meaning
60 * A. page->freelist -> List of object free in a page
61 * B. page->counters -> Counters of objects
62 * C. page->frozen -> frozen state
64 * If a slab is frozen then it is exempt from list management. It is not
65 * on any list. The processor that froze the slab is the one who can
66 * perform list operations on the page. Other processors may put objects
67 * onto the freelist but the processor that froze the slab is the only
68 * one that can retrieve the objects from the page's freelist.
70 * The list_lock protects the partial and full list on each node and
71 * the partial slab counter. If taken then no new slabs may be added or
72 * removed from the lists nor make the number of partial slabs be modified.
73 * (Note that the total number of slabs is an atomic value that may be
74 * modified without taking the list lock).
76 * The list_lock is a centralized lock and thus we avoid taking it as
77 * much as possible. As long as SLUB does not have to handle partial
78 * slabs, operations can continue without any centralized lock. F.e.
79 * allocating a long series of objects that fill up slabs does not require
81 * Interrupts are disabled during allocation and deallocation in order to
82 * make the slab allocator safe to use in the context of an irq. In addition
83 * interrupts are disabled to ensure that the processor does not change
84 * while handling per_cpu slabs, due to kernel preemption.
86 * SLUB assigns one slab for allocation to each processor.
87 * Allocations only occur from these slabs called cpu slabs.
89 * Slabs with free elements are kept on a partial list and during regular
90 * operations no list for full slabs is used. If an object in a full slab is
91 * freed then the slab will show up again on the partial lists.
92 * We track full slabs for debugging purposes though because otherwise we
93 * cannot scan all objects.
95 * Slabs are freed when they become empty. Teardown and setup is
96 * minimal so we rely on the page allocators per cpu caches for
97 * fast frees and allocs.
99 * Overloading of page flags that are otherwise used for LRU management.
101 * PageActive The slab is frozen and exempt from list processing.
102 * This means that the slab is dedicated to a purpose
103 * such as satisfying allocations for a specific
104 * processor. Objects may be freed in the slab while
105 * it is frozen but slab_free will then skip the usual
106 * list operations. It is up to the processor holding
107 * the slab to integrate the slab into the slab lists
108 * when the slab is no longer needed.
110 * One use of this flag is to mark slabs that are
111 * used for allocations. Then such a slab becomes a cpu
112 * slab. The cpu slab may be equipped with an additional
113 * freelist that allows lockless access to
114 * free objects in addition to the regular freelist
115 * that requires the slab lock.
117 * PageError Slab requires special handling due to debug
118 * options set. This moves slab handling out of
119 * the fast path and disables lockless freelists.
122 static inline int kmem_cache_debug(struct kmem_cache *s)
124 #ifdef CONFIG_SLUB_DEBUG
125 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
131 static inline void *fixup_red_left(struct kmem_cache *s, void *p)
133 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
134 p += s->red_left_pad;
139 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
141 #ifdef CONFIG_SLUB_CPU_PARTIAL
142 return !kmem_cache_debug(s);
149 * Issues still to be resolved:
151 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
153 * - Variable sizing of the per node arrays
156 /* Enable to test recovery from slab corruption on boot */
157 #undef SLUB_RESILIENCY_TEST
159 /* Enable to log cmpxchg failures */
160 #undef SLUB_DEBUG_CMPXCHG
163 * Mininum number of partial slabs. These will be left on the partial
164 * lists even if they are empty. kmem_cache_shrink may reclaim them.
166 #define MIN_PARTIAL 5
169 * Maximum number of desirable partial slabs.
170 * The existence of more partial slabs makes kmem_cache_shrink
171 * sort the partial list by the number of objects in use.
173 #define MAX_PARTIAL 10
175 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_STORE_USER)
179 * Debugging flags that require metadata to be stored in the slab. These get
180 * disabled when slub_debug=O is used and a cache's min order increases with
183 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
186 #define OO_MASK ((1 << OO_SHIFT) - 1)
187 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
189 /* Internal SLUB flags */
190 #define __OBJECT_POISON 0x80000000UL /* Poison object */
191 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
194 static struct notifier_block slab_notifier;
198 * Tracking user of a slab.
200 #define TRACK_ADDRS_COUNT 16
202 unsigned long addr; /* Called from address */
203 #ifdef CONFIG_STACKTRACE
204 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
206 int cpu; /* Was running on cpu */
207 int pid; /* Pid context */
208 unsigned long when; /* When did the operation occur */
211 enum track_item { TRACK_ALLOC, TRACK_FREE };
214 static int sysfs_slab_add(struct kmem_cache *);
215 static int sysfs_slab_alias(struct kmem_cache *, const char *);
216 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
218 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
221 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void stat(const struct kmem_cache *s, enum stat_item si)
226 #ifdef CONFIG_SLUB_STATS
228 * The rmw is racy on a preemptible kernel but this is acceptable, so
229 * avoid this_cpu_add()'s irq-disable overhead.
231 raw_cpu_inc(s->cpu_slab->stat[si]);
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 static inline void *get_freepointer(struct kmem_cache *s, void *object)
241 return *(void **)(object + s->offset);
244 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
246 prefetch(object + s->offset);
249 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
253 if (!debug_pagealloc_enabled())
254 return get_freepointer(s, object);
255 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
259 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
261 *(void **)(object + s->offset) = fp;
264 /* Loop over all objects in a slab */
265 #define for_each_object(__p, __s, __addr, __objects) \
266 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
269 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
270 for (__p = (__addr), __idx = 1; __idx <= __objects;\
271 __p += (__s)->size, __idx++)
273 /* Determine object index from a given position */
274 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
276 return (p - addr) / s->size;
279 static inline int order_objects(int order, unsigned long size, int reserved)
281 return ((PAGE_SIZE << order) - reserved) / size;
284 static inline struct kmem_cache_order_objects oo_make(int order,
285 unsigned long size, int reserved)
287 struct kmem_cache_order_objects x = {
288 (order << OO_SHIFT) + order_objects(order, size, reserved)
294 static inline int oo_order(struct kmem_cache_order_objects x)
296 return x.x >> OO_SHIFT;
299 static inline int oo_objects(struct kmem_cache_order_objects x)
301 return x.x & OO_MASK;
305 * Per slab locking using the pagelock
307 static __always_inline void slab_lock(struct page *page)
309 VM_BUG_ON_PAGE(PageTail(page), page);
310 bit_spin_lock(PG_locked, &page->flags);
313 static __always_inline void slab_unlock(struct page *page)
315 VM_BUG_ON_PAGE(PageTail(page), page);
316 __bit_spin_unlock(PG_locked, &page->flags);
319 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
322 tmp.counters = counters_new;
324 * page->counters can cover frozen/inuse/objects as well
325 * as page->_count. If we assign to ->counters directly
326 * we run the risk of losing updates to page->_count, so
327 * be careful and only assign to the fields we need.
329 page->frozen = tmp.frozen;
330 page->inuse = tmp.inuse;
331 page->objects = tmp.objects;
334 /* Interrupts must be disabled (for the fallback code to work right) */
335 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
336 void *freelist_old, unsigned long counters_old,
337 void *freelist_new, unsigned long counters_new,
340 VM_BUG_ON(!irqs_disabled());
341 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
342 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
343 if (s->flags & __CMPXCHG_DOUBLE) {
344 if (cmpxchg_double(&page->freelist, &page->counters,
345 freelist_old, counters_old,
346 freelist_new, counters_new))
352 if (page->freelist == freelist_old &&
353 page->counters == counters_old) {
354 page->freelist = freelist_new;
355 set_page_slub_counters(page, counters_new);
363 stat(s, CMPXCHG_DOUBLE_FAIL);
365 #ifdef SLUB_DEBUG_CMPXCHG
366 pr_info("%s %s: cmpxchg double redo ", n, s->name);
372 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
373 void *freelist_old, unsigned long counters_old,
374 void *freelist_new, unsigned long counters_new,
377 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
378 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
379 if (s->flags & __CMPXCHG_DOUBLE) {
380 if (cmpxchg_double(&page->freelist, &page->counters,
381 freelist_old, counters_old,
382 freelist_new, counters_new))
389 local_irq_save(flags);
391 if (page->freelist == freelist_old &&
392 page->counters == counters_old) {
393 page->freelist = freelist_new;
394 set_page_slub_counters(page, counters_new);
396 local_irq_restore(flags);
400 local_irq_restore(flags);
404 stat(s, CMPXCHG_DOUBLE_FAIL);
406 #ifdef SLUB_DEBUG_CMPXCHG
407 pr_info("%s %s: cmpxchg double redo ", n, s->name);
413 #ifdef CONFIG_SLUB_DEBUG
415 * Determine a map of object in use on a page.
417 * Node listlock must be held to guarantee that the page does
418 * not vanish from under us.
420 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
423 void *addr = page_address(page);
425 for (p = page->freelist; p; p = get_freepointer(s, p))
426 set_bit(slab_index(p, s, addr), map);
429 static inline int size_from_object(struct kmem_cache *s)
431 if (s->flags & SLAB_RED_ZONE)
432 return s->size - s->red_left_pad;
437 static inline void *restore_red_left(struct kmem_cache *s, void *p)
439 if (s->flags & SLAB_RED_ZONE)
440 p -= s->red_left_pad;
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();
479 /* Verify that a pointer has an address that is valid within a slab page */
480 static inline int check_valid_pointer(struct kmem_cache *s,
481 struct page *page, void *object)
488 base = page_address(page);
489 object = restore_red_left(s, object);
490 if (object < base || object >= base + page->objects * s->size ||
491 (object - base) % s->size) {
498 static void print_section(char *text, u8 *addr, unsigned int length)
500 metadata_access_enable();
501 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
503 metadata_access_disable();
506 static struct track *get_track(struct kmem_cache *s, void *object,
507 enum track_item alloc)
512 p = object + s->offset + sizeof(void *);
514 p = object + s->inuse;
519 static void set_track(struct kmem_cache *s, void *object,
520 enum track_item alloc, unsigned long addr)
522 struct track *p = get_track(s, object, alloc);
525 #ifdef CONFIG_STACKTRACE
526 struct stack_trace trace;
529 trace.nr_entries = 0;
530 trace.max_entries = TRACK_ADDRS_COUNT;
531 trace.entries = p->addrs;
533 metadata_access_enable();
534 save_stack_trace(&trace);
535 metadata_access_disable();
537 /* See rant in lockdep.c */
538 if (trace.nr_entries != 0 &&
539 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
542 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
546 p->cpu = smp_processor_id();
547 p->pid = current->pid;
550 memset(p, 0, sizeof(struct track));
553 static void init_tracking(struct kmem_cache *s, void *object)
555 if (!(s->flags & SLAB_STORE_USER))
558 set_track(s, object, TRACK_FREE, 0UL);
559 set_track(s, object, TRACK_ALLOC, 0UL);
562 static void print_track(const char *s, struct track *t)
567 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
568 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
569 #ifdef CONFIG_STACKTRACE
572 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
574 pr_err("\t%pS\n", (void *)t->addrs[i]);
581 static void print_tracking(struct kmem_cache *s, void *object)
583 if (!(s->flags & SLAB_STORE_USER))
586 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
587 print_track("Freed", get_track(s, object, TRACK_FREE));
590 static void print_page_info(struct page *page)
592 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
593 page, page->objects, page->inuse, page->freelist, page->flags);
597 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
599 struct va_format vaf;
605 pr_err("=============================================================================\n");
606 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
607 pr_err("-----------------------------------------------------------------------------\n\n");
609 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
613 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
615 struct va_format vaf;
621 pr_err("FIX %s: %pV\n", s->name, &vaf);
625 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
627 unsigned int off; /* Offset of last byte */
628 u8 *addr = page_address(page);
630 print_tracking(s, p);
632 print_page_info(page);
634 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
635 p, p - addr, get_freepointer(s, p));
637 if (s->flags & SLAB_RED_ZONE)
638 print_section("Redzone ", p - s->red_left_pad, s->red_left_pad);
639 else if (p > addr + 16)
640 print_section("Bytes b4 ", p - 16, 16);
642 print_section("Object ", p, min_t(unsigned long, s->object_size,
644 if (s->flags & SLAB_RED_ZONE)
645 print_section("Redzone ", p + s->object_size,
646 s->inuse - s->object_size);
649 off = s->offset + sizeof(void *);
653 if (s->flags & SLAB_STORE_USER)
654 off += 2 * sizeof(struct track);
656 if (off != size_from_object(s))
657 /* Beginning of the filler is the free pointer */
658 print_section("Padding ", p + off, size_from_object(s) - off);
663 void object_err(struct kmem_cache *s, struct page *page,
664 u8 *object, char *reason)
666 slab_bug(s, "%s", reason);
667 print_trailer(s, page, object);
670 static void slab_err(struct kmem_cache *s, struct page *page,
671 const char *fmt, ...)
677 vsnprintf(buf, sizeof(buf), fmt, args);
679 slab_bug(s, "%s", buf);
680 print_page_info(page);
684 static void init_object(struct kmem_cache *s, void *object, u8 val)
688 if (s->flags & SLAB_RED_ZONE)
689 memset(p - s->red_left_pad, val, s->red_left_pad);
691 if (s->flags & __OBJECT_POISON) {
692 memset(p, POISON_FREE, s->object_size - 1);
693 p[s->object_size - 1] = POISON_END;
696 if (s->flags & SLAB_RED_ZONE)
697 memset(p + s->object_size, val, s->inuse - s->object_size);
700 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
701 void *from, void *to)
703 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
704 memset(from, data, to - from);
707 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
708 u8 *object, char *what,
709 u8 *start, unsigned int value, unsigned int bytes)
714 metadata_access_enable();
715 fault = memchr_inv(start, value, bytes);
716 metadata_access_disable();
721 while (end > fault && end[-1] == value)
724 slab_bug(s, "%s overwritten", what);
725 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
726 fault, end - 1, fault[0], value);
727 print_trailer(s, page, object);
729 restore_bytes(s, what, value, fault, end);
737 * Bytes of the object to be managed.
738 * If the freepointer may overlay the object then the free
739 * pointer is the first word of the object.
741 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
744 * object + s->object_size
745 * Padding to reach word boundary. This is also used for Redzoning.
746 * Padding is extended by another word if Redzoning is enabled and
747 * object_size == inuse.
749 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
750 * 0xcc (RED_ACTIVE) for objects in use.
753 * Meta data starts here.
755 * A. Free pointer (if we cannot overwrite object on free)
756 * B. Tracking data for SLAB_STORE_USER
757 * C. Padding to reach required alignment boundary or at mininum
758 * one word if debugging is on to be able to detect writes
759 * before the word boundary.
761 * Padding is done using 0x5a (POISON_INUSE)
764 * Nothing is used beyond s->size.
766 * If slabcaches are merged then the object_size and inuse boundaries are mostly
767 * ignored. And therefore no slab options that rely on these boundaries
768 * may be used with merged slabcaches.
771 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
773 unsigned long off = s->inuse; /* The end of info */
776 /* Freepointer is placed after the object. */
777 off += sizeof(void *);
779 if (s->flags & SLAB_STORE_USER)
780 /* We also have user information there */
781 off += 2 * sizeof(struct track);
783 if (size_from_object(s) == off)
786 return check_bytes_and_report(s, page, p, "Object padding",
787 p + off, POISON_INUSE, size_from_object(s) - off);
790 /* Check the pad bytes at the end of a slab page */
791 static int slab_pad_check(struct kmem_cache *s, struct page *page)
799 if (!(s->flags & SLAB_POISON))
802 start = page_address(page);
803 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
804 end = start + length;
805 remainder = length % s->size;
809 metadata_access_enable();
810 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
811 metadata_access_disable();
814 while (end > fault && end[-1] == POISON_INUSE)
817 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
818 print_section("Padding ", end - remainder, remainder);
820 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
824 static int check_object(struct kmem_cache *s, struct page *page,
825 void *object, u8 val)
828 u8 *endobject = object + s->object_size;
830 if (s->flags & SLAB_RED_ZONE) {
831 if (!check_bytes_and_report(s, page, object, "Redzone",
832 object - s->red_left_pad, val, s->red_left_pad))
835 if (!check_bytes_and_report(s, page, object, "Redzone",
836 endobject, val, s->inuse - s->object_size))
839 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
840 check_bytes_and_report(s, page, p, "Alignment padding",
841 endobject, POISON_INUSE,
842 s->inuse - s->object_size);
846 if (s->flags & SLAB_POISON) {
847 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
848 (!check_bytes_and_report(s, page, p, "Poison", p,
849 POISON_FREE, s->object_size - 1) ||
850 !check_bytes_and_report(s, page, p, "Poison",
851 p + s->object_size - 1, POISON_END, 1)))
854 * check_pad_bytes cleans up on its own.
856 check_pad_bytes(s, page, p);
859 if (!s->offset && val == SLUB_RED_ACTIVE)
861 * Object and freepointer overlap. Cannot check
862 * freepointer while object is allocated.
866 /* Check free pointer validity */
867 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
868 object_err(s, page, p, "Freepointer corrupt");
870 * No choice but to zap it and thus lose the remainder
871 * of the free objects in this slab. May cause
872 * another error because the object count is now wrong.
874 set_freepointer(s, p, NULL);
880 static int check_slab(struct kmem_cache *s, struct page *page)
884 VM_BUG_ON(!irqs_disabled());
886 if (!PageSlab(page)) {
887 slab_err(s, page, "Not a valid slab page");
891 maxobj = order_objects(compound_order(page), s->size, s->reserved);
892 if (page->objects > maxobj) {
893 slab_err(s, page, "objects %u > max %u",
894 page->objects, maxobj);
897 if (page->inuse > page->objects) {
898 slab_err(s, page, "inuse %u > max %u",
899 page->inuse, page->objects);
902 /* Slab_pad_check fixes things up after itself */
903 slab_pad_check(s, page);
908 * Determine if a certain object on a page is on the freelist. Must hold the
909 * slab lock to guarantee that the chains are in a consistent state.
911 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
919 while (fp && nr <= page->objects) {
922 if (!check_valid_pointer(s, page, fp)) {
924 object_err(s, page, object,
925 "Freechain corrupt");
926 set_freepointer(s, object, NULL);
928 slab_err(s, page, "Freepointer corrupt");
929 page->freelist = NULL;
930 page->inuse = page->objects;
931 slab_fix(s, "Freelist cleared");
937 fp = get_freepointer(s, object);
941 max_objects = order_objects(compound_order(page), s->size, s->reserved);
942 if (max_objects > MAX_OBJS_PER_PAGE)
943 max_objects = MAX_OBJS_PER_PAGE;
945 if (page->objects != max_objects) {
946 slab_err(s, page, "Wrong number of objects. Found %d but "
947 "should be %d", page->objects, max_objects);
948 page->objects = max_objects;
949 slab_fix(s, "Number of objects adjusted.");
951 if (page->inuse != page->objects - nr) {
952 slab_err(s, page, "Wrong object count. Counter is %d but "
953 "counted were %d", page->inuse, page->objects - nr);
954 page->inuse = page->objects - nr;
955 slab_fix(s, "Object count adjusted.");
957 return search == NULL;
960 static void trace(struct kmem_cache *s, struct page *page, void *object,
963 if (s->flags & SLAB_TRACE) {
964 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
966 alloc ? "alloc" : "free",
971 print_section("Object ", (void *)object,
979 * Tracking of fully allocated slabs for debugging purposes.
981 static void add_full(struct kmem_cache *s,
982 struct kmem_cache_node *n, struct page *page)
984 if (!(s->flags & SLAB_STORE_USER))
987 lockdep_assert_held(&n->list_lock);
988 list_add(&page->lru, &n->full);
991 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
993 if (!(s->flags & SLAB_STORE_USER))
996 lockdep_assert_held(&n->list_lock);
997 list_del(&page->lru);
1000 /* Tracking of the number of slabs for debugging purposes */
1001 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1003 struct kmem_cache_node *n = get_node(s, node);
1005 return atomic_long_read(&n->nr_slabs);
1008 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1010 return atomic_long_read(&n->nr_slabs);
1013 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1015 struct kmem_cache_node *n = get_node(s, node);
1018 * May be called early in order to allocate a slab for the
1019 * kmem_cache_node structure. Solve the chicken-egg
1020 * dilemma by deferring the increment of the count during
1021 * bootstrap (see early_kmem_cache_node_alloc).
1024 atomic_long_inc(&n->nr_slabs);
1025 atomic_long_add(objects, &n->total_objects);
1028 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1030 struct kmem_cache_node *n = get_node(s, node);
1032 atomic_long_dec(&n->nr_slabs);
1033 atomic_long_sub(objects, &n->total_objects);
1036 /* Object debug checks for alloc/free paths */
1037 static void *setup_object_debug(struct kmem_cache *s, struct page *page,
1040 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1043 object = fixup_red_left(s, object);
1044 init_object(s, object, SLUB_RED_INACTIVE);
1045 init_tracking(s, object);
1050 static noinline int alloc_debug_processing(struct kmem_cache *s,
1052 void *object, unsigned long addr)
1054 if (!check_slab(s, page))
1057 if (!check_valid_pointer(s, page, object)) {
1058 object_err(s, page, object, "Freelist Pointer check fails");
1062 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1065 /* Success perform special debug activities for allocs */
1066 if (s->flags & SLAB_STORE_USER)
1067 set_track(s, object, TRACK_ALLOC, addr);
1068 trace(s, page, object, 1);
1069 init_object(s, object, SLUB_RED_ACTIVE);
1073 if (PageSlab(page)) {
1075 * If this is a slab page then lets do the best we can
1076 * to avoid issues in the future. Marking all objects
1077 * as used avoids touching the remaining objects.
1079 slab_fix(s, "Marking all objects used");
1080 page->inuse = page->objects;
1081 page->freelist = NULL;
1086 /* Supports checking bulk free of a constructed freelist */
1087 static noinline struct kmem_cache_node *free_debug_processing(
1088 struct kmem_cache *s, struct page *page,
1089 void *head, void *tail, int bulk_cnt,
1090 unsigned long addr, unsigned long *flags)
1092 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1093 void *object = head;
1096 spin_lock_irqsave(&n->list_lock, *flags);
1099 if (!check_slab(s, page))
1105 if (!check_valid_pointer(s, page, object)) {
1106 slab_err(s, page, "Invalid object pointer 0x%p", object);
1110 if (on_freelist(s, page, object)) {
1111 object_err(s, page, object, "Object already free");
1115 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1118 if (unlikely(s != page->slab_cache)) {
1119 if (!PageSlab(page)) {
1120 slab_err(s, page, "Attempt to free object(0x%p) "
1121 "outside of slab", object);
1122 } else if (!page->slab_cache) {
1123 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1127 object_err(s, page, object,
1128 "page slab pointer corrupt.");
1132 if (s->flags & SLAB_STORE_USER)
1133 set_track(s, object, TRACK_FREE, addr);
1134 trace(s, page, object, 0);
1135 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1136 init_object(s, object, SLUB_RED_INACTIVE);
1138 /* Reached end of constructed freelist yet? */
1139 if (object != tail) {
1140 object = get_freepointer(s, object);
1144 if (cnt != bulk_cnt)
1145 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1150 * Keep node_lock to preserve integrity
1151 * until the object is actually freed
1157 spin_unlock_irqrestore(&n->list_lock, *flags);
1158 slab_fix(s, "Object at 0x%p not freed", object);
1162 static int __init setup_slub_debug(char *str)
1164 slub_debug = DEBUG_DEFAULT_FLAGS;
1165 if (*str++ != '=' || !*str)
1167 * No options specified. Switch on full debugging.
1173 * No options but restriction on slabs. This means full
1174 * debugging for slabs matching a pattern.
1181 * Switch off all debugging measures.
1186 * Determine which debug features should be switched on
1188 for (; *str && *str != ','; str++) {
1189 switch (tolower(*str)) {
1191 slub_debug |= SLAB_DEBUG_FREE;
1194 slub_debug |= SLAB_RED_ZONE;
1197 slub_debug |= SLAB_POISON;
1200 slub_debug |= SLAB_STORE_USER;
1203 slub_debug |= SLAB_TRACE;
1206 slub_debug |= SLAB_FAILSLAB;
1210 * Avoid enabling debugging on caches if its minimum
1211 * order would increase as a result.
1213 disable_higher_order_debug = 1;
1216 pr_err("slub_debug option '%c' unknown. skipped\n",
1223 slub_debug_slabs = str + 1;
1228 __setup("slub_debug", setup_slub_debug);
1230 unsigned long kmem_cache_flags(unsigned long object_size,
1231 unsigned long flags, const char *name,
1232 void (*ctor)(void *))
1235 * Enable debugging if selected on the kernel commandline.
1237 if (slub_debug && (!slub_debug_slabs || (name &&
1238 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1239 flags |= slub_debug;
1243 #else /* !CONFIG_SLUB_DEBUG */
1244 static inline void *setup_object_debug(struct kmem_cache *s,
1245 struct page *page, void *object) { return object; }
1247 static inline int alloc_debug_processing(struct kmem_cache *s,
1248 struct page *page, void *object, unsigned long addr) { return 0; }
1250 static inline struct kmem_cache_node *free_debug_processing(
1251 struct kmem_cache *s, struct page *page,
1252 void *head, void *tail, int bulk_cnt,
1253 unsigned long addr, unsigned long *flags) { return NULL; }
1255 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1257 static inline int check_object(struct kmem_cache *s, struct page *page,
1258 void *object, u8 val) { return 1; }
1259 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1260 struct page *page) {}
1261 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1262 struct page *page) {}
1263 unsigned long kmem_cache_flags(unsigned long object_size,
1264 unsigned long flags, const char *name,
1265 void (*ctor)(void *))
1269 #define slub_debug 0
1271 #define disable_higher_order_debug 0
1273 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1275 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1277 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1279 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1282 #endif /* CONFIG_SLUB_DEBUG */
1285 * Hooks for other subsystems that check memory allocations. In a typical
1286 * production configuration these hooks all should produce no code at all.
1288 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1290 kmemleak_alloc(ptr, size, 1, flags);
1291 kasan_kmalloc_large(ptr, size);
1294 static inline void kfree_hook(const void *x)
1297 kasan_kfree_large(x);
1300 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1302 kmemleak_free_recursive(x, s->flags);
1305 * Trouble is that we may no longer disable interrupts in the fast path
1306 * So in order to make the debug calls that expect irqs to be
1307 * disabled we need to disable interrupts temporarily.
1309 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1311 unsigned long flags;
1313 local_irq_save(flags);
1314 kmemcheck_slab_free(s, x, s->object_size);
1315 debug_check_no_locks_freed(x, s->object_size);
1316 local_irq_restore(flags);
1319 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1320 debug_check_no_obj_freed(x, s->object_size);
1322 kasan_slab_free(s, x);
1325 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1326 void *head, void *tail)
1329 * Compiler cannot detect this function can be removed if slab_free_hook()
1330 * evaluates to nothing. Thus, catch all relevant config debug options here.
1332 #if defined(CONFIG_KMEMCHECK) || \
1333 defined(CONFIG_LOCKDEP) || \
1334 defined(CONFIG_DEBUG_KMEMLEAK) || \
1335 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1336 defined(CONFIG_KASAN)
1338 void *object = head;
1339 void *tail_obj = tail ? : head;
1342 slab_free_hook(s, object);
1343 } while ((object != tail_obj) &&
1344 (object = get_freepointer(s, object)));
1348 static void *setup_object(struct kmem_cache *s, struct page *page,
1351 object = setup_object_debug(s, page, object);
1352 if (unlikely(s->ctor)) {
1353 kasan_unpoison_object_data(s, object);
1355 kasan_poison_object_data(s, object);
1362 * Slab allocation and freeing
1364 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1365 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1368 int order = oo_order(oo);
1370 flags |= __GFP_NOTRACK;
1372 if (node == NUMA_NO_NODE)
1373 page = alloc_pages(flags, order);
1375 page = __alloc_pages_node(node, flags, order);
1377 if (page && memcg_charge_slab(page, flags, order, s)) {
1378 __free_pages(page, order);
1385 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1388 struct kmem_cache_order_objects oo = s->oo;
1393 flags &= gfp_allowed_mask;
1395 if (gfpflags_allow_blocking(flags))
1398 flags |= s->allocflags;
1401 * Let the initial higher-order allocation fail under memory pressure
1402 * so we fall-back to the minimum order allocation.
1404 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1405 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1406 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_DIRECT_RECLAIM;
1408 page = alloc_slab_page(s, alloc_gfp, node, oo);
1409 if (unlikely(!page)) {
1413 * Allocation may have failed due to fragmentation.
1414 * Try a lower order alloc if possible
1416 page = alloc_slab_page(s, alloc_gfp, node, oo);
1417 if (unlikely(!page))
1419 stat(s, ORDER_FALLBACK);
1422 if (kmemcheck_enabled &&
1423 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1424 int pages = 1 << oo_order(oo);
1426 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1429 * Objects from caches that have a constructor don't get
1430 * cleared when they're allocated, so we need to do it here.
1433 kmemcheck_mark_uninitialized_pages(page, pages);
1435 kmemcheck_mark_unallocated_pages(page, pages);
1438 page->objects = oo_objects(oo);
1440 order = compound_order(page);
1441 page->slab_cache = s;
1442 __SetPageSlab(page);
1443 if (page_is_pfmemalloc(page))
1444 SetPageSlabPfmemalloc(page);
1446 start = page_address(page);
1448 if (unlikely(s->flags & SLAB_POISON))
1449 memset(start, POISON_INUSE, PAGE_SIZE << order);
1451 kasan_poison_slab(page);
1453 for_each_object_idx(p, idx, s, start, page->objects) {
1454 void *object = setup_object(s, page, p);
1456 if (likely(idx < page->objects)) {
1457 set_freepointer(s, object,
1458 fixup_red_left(s, p + s->size));
1460 set_freepointer(s, object, NULL);
1463 page->freelist = fixup_red_left(s, start);
1464 page->inuse = page->objects;
1468 if (gfpflags_allow_blocking(flags))
1469 local_irq_disable();
1473 mod_zone_page_state(page_zone(page),
1474 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1475 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1478 inc_slabs_node(s, page_to_nid(page), page->objects);
1483 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1485 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1486 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1490 return allocate_slab(s,
1491 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1494 static void __free_slab(struct kmem_cache *s, struct page *page)
1496 int order = compound_order(page);
1497 int pages = 1 << order;
1499 if (kmem_cache_debug(s)) {
1502 slab_pad_check(s, page);
1503 for_each_object(p, s, page_address(page),
1505 void *object = fixup_red_left(s, p);
1507 check_object(s, page, object, SLUB_RED_INACTIVE);
1511 kmemcheck_free_shadow(page, compound_order(page));
1513 mod_zone_page_state(page_zone(page),
1514 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1515 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1518 __ClearPageSlabPfmemalloc(page);
1519 __ClearPageSlab(page);
1521 page_mapcount_reset(page);
1522 if (current->reclaim_state)
1523 current->reclaim_state->reclaimed_slab += pages;
1524 memcg_uncharge_slab(page, order, s);
1525 __free_pages(page, order);
1528 #define need_reserve_slab_rcu \
1529 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1531 static void rcu_free_slab(struct rcu_head *h)
1535 if (need_reserve_slab_rcu)
1536 page = virt_to_head_page(h);
1538 page = container_of((struct list_head *)h, struct page, lru);
1540 __free_slab(page->slab_cache, page);
1543 static void free_slab(struct kmem_cache *s, struct page *page)
1545 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1546 struct rcu_head *head;
1548 if (need_reserve_slab_rcu) {
1549 int order = compound_order(page);
1550 int offset = (PAGE_SIZE << order) - s->reserved;
1552 VM_BUG_ON(s->reserved != sizeof(*head));
1553 head = page_address(page) + offset;
1555 head = &page->rcu_head;
1558 call_rcu(head, rcu_free_slab);
1560 __free_slab(s, page);
1563 static void discard_slab(struct kmem_cache *s, struct page *page)
1565 dec_slabs_node(s, page_to_nid(page), page->objects);
1570 * Management of partially allocated slabs.
1573 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1576 if (tail == DEACTIVATE_TO_TAIL)
1577 list_add_tail(&page->lru, &n->partial);
1579 list_add(&page->lru, &n->partial);
1582 static inline void add_partial(struct kmem_cache_node *n,
1583 struct page *page, int tail)
1585 lockdep_assert_held(&n->list_lock);
1586 __add_partial(n, page, tail);
1589 static inline void remove_partial(struct kmem_cache_node *n,
1592 lockdep_assert_held(&n->list_lock);
1593 list_del(&page->lru);
1598 * Remove slab from the partial list, freeze it and
1599 * return the pointer to the freelist.
1601 * Returns a list of objects or NULL if it fails.
1603 static inline void *acquire_slab(struct kmem_cache *s,
1604 struct kmem_cache_node *n, struct page *page,
1605 int mode, int *objects)
1608 unsigned long counters;
1611 lockdep_assert_held(&n->list_lock);
1614 * Zap the freelist and set the frozen bit.
1615 * The old freelist is the list of objects for the
1616 * per cpu allocation list.
1618 freelist = page->freelist;
1619 counters = page->counters;
1620 new.counters = counters;
1621 *objects = new.objects - new.inuse;
1623 new.inuse = page->objects;
1624 new.freelist = NULL;
1626 new.freelist = freelist;
1629 VM_BUG_ON(new.frozen);
1632 if (!__cmpxchg_double_slab(s, page,
1634 new.freelist, new.counters,
1638 remove_partial(n, page);
1643 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1644 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1647 * Try to allocate a partial slab from a specific node.
1649 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1650 struct kmem_cache_cpu *c, gfp_t flags)
1652 struct page *page, *page2;
1653 void *object = NULL;
1658 * Racy check. If we mistakenly see no partial slabs then we
1659 * just allocate an empty slab. If we mistakenly try to get a
1660 * partial slab and there is none available then get_partials()
1663 if (!n || !n->nr_partial)
1666 spin_lock(&n->list_lock);
1667 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1670 if (!pfmemalloc_match(page, flags))
1673 t = acquire_slab(s, n, page, object == NULL, &objects);
1677 available += objects;
1680 stat(s, ALLOC_FROM_PARTIAL);
1683 put_cpu_partial(s, page, 0);
1684 stat(s, CPU_PARTIAL_NODE);
1686 if (!kmem_cache_has_cpu_partial(s)
1687 || available > s->cpu_partial / 2)
1691 spin_unlock(&n->list_lock);
1696 * Get a page from somewhere. Search in increasing NUMA distances.
1698 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1699 struct kmem_cache_cpu *c)
1702 struct zonelist *zonelist;
1705 enum zone_type high_zoneidx = gfp_zone(flags);
1707 unsigned int cpuset_mems_cookie;
1710 * The defrag ratio allows a configuration of the tradeoffs between
1711 * inter node defragmentation and node local allocations. A lower
1712 * defrag_ratio increases the tendency to do local allocations
1713 * instead of attempting to obtain partial slabs from other nodes.
1715 * If the defrag_ratio is set to 0 then kmalloc() always
1716 * returns node local objects. If the ratio is higher then kmalloc()
1717 * may return off node objects because partial slabs are obtained
1718 * from other nodes and filled up.
1720 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1721 * defrag_ratio = 1000) then every (well almost) allocation will
1722 * first attempt to defrag slab caches on other nodes. This means
1723 * scanning over all nodes to look for partial slabs which may be
1724 * expensive if we do it every time we are trying to find a slab
1725 * with available objects.
1727 if (!s->remote_node_defrag_ratio ||
1728 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1732 cpuset_mems_cookie = read_mems_allowed_begin();
1733 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1734 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1735 struct kmem_cache_node *n;
1737 n = get_node(s, zone_to_nid(zone));
1739 if (n && cpuset_zone_allowed(zone, flags) &&
1740 n->nr_partial > s->min_partial) {
1741 object = get_partial_node(s, n, c, flags);
1744 * Don't check read_mems_allowed_retry()
1745 * here - if mems_allowed was updated in
1746 * parallel, that was a harmless race
1747 * between allocation and the cpuset
1754 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1760 * Get a partial page, lock it and return it.
1762 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1763 struct kmem_cache_cpu *c)
1766 int searchnode = node;
1768 if (node == NUMA_NO_NODE)
1769 searchnode = numa_mem_id();
1770 else if (!node_present_pages(node))
1771 searchnode = node_to_mem_node(node);
1773 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1774 if (object || node != NUMA_NO_NODE)
1777 return get_any_partial(s, flags, c);
1780 #ifdef CONFIG_PREEMPT
1782 * Calculate the next globally unique transaction for disambiguiation
1783 * during cmpxchg. The transactions start with the cpu number and are then
1784 * incremented by CONFIG_NR_CPUS.
1786 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1789 * No preemption supported therefore also no need to check for
1795 static inline unsigned long next_tid(unsigned long tid)
1797 return tid + TID_STEP;
1800 static inline unsigned int tid_to_cpu(unsigned long tid)
1802 return tid % TID_STEP;
1805 static inline unsigned long tid_to_event(unsigned long tid)
1807 return tid / TID_STEP;
1810 static inline unsigned int init_tid(int cpu)
1815 static inline void note_cmpxchg_failure(const char *n,
1816 const struct kmem_cache *s, unsigned long tid)
1818 #ifdef SLUB_DEBUG_CMPXCHG
1819 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1821 pr_info("%s %s: cmpxchg redo ", n, s->name);
1823 #ifdef CONFIG_PREEMPT
1824 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1825 pr_warn("due to cpu change %d -> %d\n",
1826 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1829 if (tid_to_event(tid) != tid_to_event(actual_tid))
1830 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1831 tid_to_event(tid), tid_to_event(actual_tid));
1833 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1834 actual_tid, tid, next_tid(tid));
1836 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1839 static void init_kmem_cache_cpus(struct kmem_cache *s)
1843 for_each_possible_cpu(cpu)
1844 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1848 * Remove the cpu slab
1850 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1853 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1854 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1856 enum slab_modes l = M_NONE, m = M_NONE;
1858 int tail = DEACTIVATE_TO_HEAD;
1862 if (page->freelist) {
1863 stat(s, DEACTIVATE_REMOTE_FREES);
1864 tail = DEACTIVATE_TO_TAIL;
1868 * Stage one: Free all available per cpu objects back
1869 * to the page freelist while it is still frozen. Leave the
1872 * There is no need to take the list->lock because the page
1875 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1877 unsigned long counters;
1880 prior = page->freelist;
1881 counters = page->counters;
1882 set_freepointer(s, freelist, prior);
1883 new.counters = counters;
1885 VM_BUG_ON(!new.frozen);
1887 } while (!__cmpxchg_double_slab(s, page,
1889 freelist, new.counters,
1890 "drain percpu freelist"));
1892 freelist = nextfree;
1896 * Stage two: Ensure that the page is unfrozen while the
1897 * list presence reflects the actual number of objects
1900 * We setup the list membership and then perform a cmpxchg
1901 * with the count. If there is a mismatch then the page
1902 * is not unfrozen but the page is on the wrong list.
1904 * Then we restart the process which may have to remove
1905 * the page from the list that we just put it on again
1906 * because the number of objects in the slab may have
1911 old.freelist = page->freelist;
1912 old.counters = page->counters;
1913 VM_BUG_ON(!old.frozen);
1915 /* Determine target state of the slab */
1916 new.counters = old.counters;
1919 set_freepointer(s, freelist, old.freelist);
1920 new.freelist = freelist;
1922 new.freelist = old.freelist;
1926 if (!new.inuse && n->nr_partial >= s->min_partial)
1928 else if (new.freelist) {
1933 * Taking the spinlock removes the possiblity
1934 * that acquire_slab() will see a slab page that
1937 spin_lock(&n->list_lock);
1941 if (kmem_cache_debug(s) && !lock) {
1944 * This also ensures that the scanning of full
1945 * slabs from diagnostic functions will not see
1948 spin_lock(&n->list_lock);
1956 remove_partial(n, page);
1958 else if (l == M_FULL)
1960 remove_full(s, n, page);
1962 if (m == M_PARTIAL) {
1964 add_partial(n, page, tail);
1967 } else if (m == M_FULL) {
1969 stat(s, DEACTIVATE_FULL);
1970 add_full(s, n, page);
1976 if (!__cmpxchg_double_slab(s, page,
1977 old.freelist, old.counters,
1978 new.freelist, new.counters,
1983 spin_unlock(&n->list_lock);
1986 stat(s, DEACTIVATE_EMPTY);
1987 discard_slab(s, page);
1993 * Unfreeze all the cpu partial slabs.
1995 * This function must be called with interrupts disabled
1996 * for the cpu using c (or some other guarantee must be there
1997 * to guarantee no concurrent accesses).
1999 static void unfreeze_partials(struct kmem_cache *s,
2000 struct kmem_cache_cpu *c)
2002 #ifdef CONFIG_SLUB_CPU_PARTIAL
2003 struct kmem_cache_node *n = NULL, *n2 = NULL;
2004 struct page *page, *discard_page = NULL;
2006 while ((page = c->partial)) {
2010 c->partial = page->next;
2012 n2 = get_node(s, page_to_nid(page));
2015 spin_unlock(&n->list_lock);
2018 spin_lock(&n->list_lock);
2023 old.freelist = page->freelist;
2024 old.counters = page->counters;
2025 VM_BUG_ON(!old.frozen);
2027 new.counters = old.counters;
2028 new.freelist = old.freelist;
2032 } while (!__cmpxchg_double_slab(s, page,
2033 old.freelist, old.counters,
2034 new.freelist, new.counters,
2035 "unfreezing slab"));
2037 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2038 page->next = discard_page;
2039 discard_page = page;
2041 add_partial(n, page, DEACTIVATE_TO_TAIL);
2042 stat(s, FREE_ADD_PARTIAL);
2047 spin_unlock(&n->list_lock);
2049 while (discard_page) {
2050 page = discard_page;
2051 discard_page = discard_page->next;
2053 stat(s, DEACTIVATE_EMPTY);
2054 discard_slab(s, page);
2061 * Put a page that was just frozen (in __slab_free) into a partial page
2062 * slot if available. This is done without interrupts disabled and without
2063 * preemption disabled. The cmpxchg is racy and may put the partial page
2064 * onto a random cpus partial slot.
2066 * If we did not find a slot then simply move all the partials to the
2067 * per node partial list.
2069 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2071 #ifdef CONFIG_SLUB_CPU_PARTIAL
2072 struct page *oldpage;
2080 oldpage = this_cpu_read(s->cpu_slab->partial);
2083 pobjects = oldpage->pobjects;
2084 pages = oldpage->pages;
2085 if (drain && pobjects > s->cpu_partial) {
2086 unsigned long flags;
2088 * partial array is full. Move the existing
2089 * set to the per node partial list.
2091 local_irq_save(flags);
2092 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2093 local_irq_restore(flags);
2097 stat(s, CPU_PARTIAL_DRAIN);
2102 pobjects += page->objects - page->inuse;
2104 page->pages = pages;
2105 page->pobjects = pobjects;
2106 page->next = oldpage;
2108 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2110 if (unlikely(!s->cpu_partial)) {
2111 unsigned long flags;
2113 local_irq_save(flags);
2114 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2115 local_irq_restore(flags);
2121 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2123 stat(s, CPUSLAB_FLUSH);
2124 deactivate_slab(s, c->page, c->freelist);
2126 c->tid = next_tid(c->tid);
2134 * Called from IPI handler with interrupts disabled.
2136 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2138 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2144 unfreeze_partials(s, c);
2148 static void flush_cpu_slab(void *d)
2150 struct kmem_cache *s = d;
2152 __flush_cpu_slab(s, smp_processor_id());
2155 static bool has_cpu_slab(int cpu, void *info)
2157 struct kmem_cache *s = info;
2158 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2160 return c->page || c->partial;
2163 static void flush_all(struct kmem_cache *s)
2165 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2169 * Check if the objects in a per cpu structure fit numa
2170 * locality expectations.
2172 static inline int node_match(struct page *page, int node)
2175 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2181 #ifdef CONFIG_SLUB_DEBUG
2182 static int count_free(struct page *page)
2184 return page->objects - page->inuse;
2187 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2189 return atomic_long_read(&n->total_objects);
2191 #endif /* CONFIG_SLUB_DEBUG */
2193 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2194 static unsigned long count_partial(struct kmem_cache_node *n,
2195 int (*get_count)(struct page *))
2197 unsigned long flags;
2198 unsigned long x = 0;
2201 spin_lock_irqsave(&n->list_lock, flags);
2202 list_for_each_entry(page, &n->partial, lru)
2203 x += get_count(page);
2204 spin_unlock_irqrestore(&n->list_lock, flags);
2207 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2209 static noinline void
2210 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2212 #ifdef CONFIG_SLUB_DEBUG
2213 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2214 DEFAULT_RATELIMIT_BURST);
2216 struct kmem_cache_node *n;
2218 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2221 pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2223 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2224 s->name, s->object_size, s->size, oo_order(s->oo),
2227 if (oo_order(s->min) > get_order(s->object_size))
2228 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2231 for_each_kmem_cache_node(s, node, n) {
2232 unsigned long nr_slabs;
2233 unsigned long nr_objs;
2234 unsigned long nr_free;
2236 nr_free = count_partial(n, count_free);
2237 nr_slabs = node_nr_slabs(n);
2238 nr_objs = node_nr_objs(n);
2240 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2241 node, nr_slabs, nr_objs, nr_free);
2246 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2247 int node, struct kmem_cache_cpu **pc)
2250 struct kmem_cache_cpu *c = *pc;
2253 freelist = get_partial(s, flags, node, c);
2258 page = new_slab(s, flags, node);
2260 c = raw_cpu_ptr(s->cpu_slab);
2265 * No other reference to the page yet so we can
2266 * muck around with it freely without cmpxchg
2268 freelist = page->freelist;
2269 page->freelist = NULL;
2271 stat(s, ALLOC_SLAB);
2280 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2282 if (unlikely(PageSlabPfmemalloc(page)))
2283 return gfp_pfmemalloc_allowed(gfpflags);
2289 * Check the page->freelist of a page and either transfer the freelist to the
2290 * per cpu freelist or deactivate the page.
2292 * The page is still frozen if the return value is not NULL.
2294 * If this function returns NULL then the page has been unfrozen.
2296 * This function must be called with interrupt disabled.
2298 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2301 unsigned long counters;
2305 freelist = page->freelist;
2306 counters = page->counters;
2308 new.counters = counters;
2309 VM_BUG_ON(!new.frozen);
2311 new.inuse = page->objects;
2312 new.frozen = freelist != NULL;
2314 } while (!__cmpxchg_double_slab(s, page,
2323 * Slow path. The lockless freelist is empty or we need to perform
2326 * Processing is still very fast if new objects have been freed to the
2327 * regular freelist. In that case we simply take over the regular freelist
2328 * as the lockless freelist and zap the regular freelist.
2330 * If that is not working then we fall back to the partial lists. We take the
2331 * first element of the freelist as the object to allocate now and move the
2332 * rest of the freelist to the lockless freelist.
2334 * And if we were unable to get a new slab from the partial slab lists then
2335 * we need to allocate a new slab. This is the slowest path since it involves
2336 * a call to the page allocator and the setup of a new slab.
2338 * Version of __slab_alloc to use when we know that interrupts are
2339 * already disabled (which is the case for bulk allocation).
2341 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2342 unsigned long addr, struct kmem_cache_cpu *c)
2352 if (unlikely(!node_match(page, node))) {
2353 int searchnode = node;
2355 if (node != NUMA_NO_NODE && !node_present_pages(node))
2356 searchnode = node_to_mem_node(node);
2358 if (unlikely(!node_match(page, searchnode))) {
2359 stat(s, ALLOC_NODE_MISMATCH);
2360 deactivate_slab(s, page, c->freelist);
2368 * By rights, we should be searching for a slab page that was
2369 * PFMEMALLOC but right now, we are losing the pfmemalloc
2370 * information when the page leaves the per-cpu allocator
2372 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2373 deactivate_slab(s, page, c->freelist);
2379 /* must check again c->freelist in case of cpu migration or IRQ */
2380 freelist = c->freelist;
2384 freelist = get_freelist(s, page);
2388 stat(s, DEACTIVATE_BYPASS);
2392 stat(s, ALLOC_REFILL);
2396 * freelist is pointing to the list of objects to be used.
2397 * page is pointing to the page from which the objects are obtained.
2398 * That page must be frozen for per cpu allocations to work.
2400 VM_BUG_ON(!c->page->frozen);
2401 c->freelist = get_freepointer(s, freelist);
2402 c->tid = next_tid(c->tid);
2408 page = c->page = c->partial;
2409 c->partial = page->next;
2410 stat(s, CPU_PARTIAL_ALLOC);
2415 freelist = new_slab_objects(s, gfpflags, node, &c);
2417 if (unlikely(!freelist)) {
2418 slab_out_of_memory(s, gfpflags, node);
2423 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2426 /* Only entered in the debug case */
2427 if (kmem_cache_debug(s) &&
2428 !alloc_debug_processing(s, page, freelist, addr))
2429 goto new_slab; /* Slab failed checks. Next slab needed */
2431 deactivate_slab(s, page, get_freepointer(s, freelist));
2438 * Another one that disabled interrupt and compensates for possible
2439 * cpu changes by refetching the per cpu area pointer.
2441 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2442 unsigned long addr, struct kmem_cache_cpu *c)
2445 unsigned long flags;
2447 local_irq_save(flags);
2448 #ifdef CONFIG_PREEMPT
2450 * We may have been preempted and rescheduled on a different
2451 * cpu before disabling interrupts. Need to reload cpu area
2454 c = this_cpu_ptr(s->cpu_slab);
2457 p = ___slab_alloc(s, gfpflags, node, addr, c);
2458 local_irq_restore(flags);
2463 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2464 * have the fastpath folded into their functions. So no function call
2465 * overhead for requests that can be satisfied on the fastpath.
2467 * The fastpath works by first checking if the lockless freelist can be used.
2468 * If not then __slab_alloc is called for slow processing.
2470 * Otherwise we can simply pick the next object from the lockless free list.
2472 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2473 gfp_t gfpflags, int node, unsigned long addr)
2476 struct kmem_cache_cpu *c;
2480 s = slab_pre_alloc_hook(s, gfpflags);
2485 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2486 * enabled. We may switch back and forth between cpus while
2487 * reading from one cpu area. That does not matter as long
2488 * as we end up on the original cpu again when doing the cmpxchg.
2490 * We should guarantee that tid and kmem_cache are retrieved on
2491 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2492 * to check if it is matched or not.
2495 tid = this_cpu_read(s->cpu_slab->tid);
2496 c = raw_cpu_ptr(s->cpu_slab);
2497 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2498 unlikely(tid != READ_ONCE(c->tid)));
2501 * Irqless object alloc/free algorithm used here depends on sequence
2502 * of fetching cpu_slab's data. tid should be fetched before anything
2503 * on c to guarantee that object and page associated with previous tid
2504 * won't be used with current tid. If we fetch tid first, object and
2505 * page could be one associated with next tid and our alloc/free
2506 * request will be failed. In this case, we will retry. So, no problem.
2511 * The transaction ids are globally unique per cpu and per operation on
2512 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2513 * occurs on the right processor and that there was no operation on the
2514 * linked list in between.
2517 object = c->freelist;
2519 if (unlikely(!object || !node_match(page, node))) {
2520 object = __slab_alloc(s, gfpflags, node, addr, c);
2521 stat(s, ALLOC_SLOWPATH);
2523 void *next_object = get_freepointer_safe(s, object);
2526 * The cmpxchg will only match if there was no additional
2527 * operation and if we are on the right processor.
2529 * The cmpxchg does the following atomically (without lock
2531 * 1. Relocate first pointer to the current per cpu area.
2532 * 2. Verify that tid and freelist have not been changed
2533 * 3. If they were not changed replace tid and freelist
2535 * Since this is without lock semantics the protection is only
2536 * against code executing on this cpu *not* from access by
2539 if (unlikely(!this_cpu_cmpxchg_double(
2540 s->cpu_slab->freelist, s->cpu_slab->tid,
2542 next_object, next_tid(tid)))) {
2544 note_cmpxchg_failure("slab_alloc", s, tid);
2547 prefetch_freepointer(s, next_object);
2548 stat(s, ALLOC_FASTPATH);
2551 if (unlikely(gfpflags & __GFP_ZERO) && object)
2552 memset(object, 0, s->object_size);
2554 slab_post_alloc_hook(s, gfpflags, 1, &object);
2559 static __always_inline void *slab_alloc(struct kmem_cache *s,
2560 gfp_t gfpflags, unsigned long addr)
2562 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2565 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2567 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2569 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2574 EXPORT_SYMBOL(kmem_cache_alloc);
2576 #ifdef CONFIG_TRACING
2577 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2579 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2580 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2581 kasan_kmalloc(s, ret, size);
2584 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2588 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2590 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2592 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2593 s->object_size, s->size, gfpflags, node);
2597 EXPORT_SYMBOL(kmem_cache_alloc_node);
2599 #ifdef CONFIG_TRACING
2600 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2602 int node, size_t size)
2604 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2606 trace_kmalloc_node(_RET_IP_, ret,
2607 size, s->size, gfpflags, node);
2609 kasan_kmalloc(s, ret, size);
2612 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2617 * Slow path handling. This may still be called frequently since objects
2618 * have a longer lifetime than the cpu slabs in most processing loads.
2620 * So we still attempt to reduce cache line usage. Just take the slab
2621 * lock and free the item. If there is no additional partial page
2622 * handling required then we can return immediately.
2624 static void __slab_free(struct kmem_cache *s, struct page *page,
2625 void *head, void *tail, int cnt,
2632 unsigned long counters;
2633 struct kmem_cache_node *n = NULL;
2634 unsigned long uninitialized_var(flags);
2636 stat(s, FREE_SLOWPATH);
2638 if (kmem_cache_debug(s) &&
2639 !(n = free_debug_processing(s, page, head, tail, cnt,
2645 spin_unlock_irqrestore(&n->list_lock, flags);
2648 prior = page->freelist;
2649 counters = page->counters;
2650 set_freepointer(s, tail, prior);
2651 new.counters = counters;
2652 was_frozen = new.frozen;
2654 if ((!new.inuse || !prior) && !was_frozen) {
2656 if (kmem_cache_has_cpu_partial(s) && !prior) {
2659 * Slab was on no list before and will be
2661 * We can defer the list move and instead
2666 } else { /* Needs to be taken off a list */
2668 n = get_node(s, page_to_nid(page));
2670 * Speculatively acquire the list_lock.
2671 * If the cmpxchg does not succeed then we may
2672 * drop the list_lock without any processing.
2674 * Otherwise the list_lock will synchronize with
2675 * other processors updating the list of slabs.
2677 spin_lock_irqsave(&n->list_lock, flags);
2682 } while (!cmpxchg_double_slab(s, page,
2690 * If we just froze the page then put it onto the
2691 * per cpu partial list.
2693 if (new.frozen && !was_frozen) {
2694 put_cpu_partial(s, page, 1);
2695 stat(s, CPU_PARTIAL_FREE);
2698 * The list lock was not taken therefore no list
2699 * activity can be necessary.
2702 stat(s, FREE_FROZEN);
2706 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2710 * Objects left in the slab. If it was not on the partial list before
2713 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2714 if (kmem_cache_debug(s))
2715 remove_full(s, n, page);
2716 add_partial(n, page, DEACTIVATE_TO_TAIL);
2717 stat(s, FREE_ADD_PARTIAL);
2719 spin_unlock_irqrestore(&n->list_lock, flags);
2725 * Slab on the partial list.
2727 remove_partial(n, page);
2728 stat(s, FREE_REMOVE_PARTIAL);
2730 /* Slab must be on the full list */
2731 remove_full(s, n, page);
2734 spin_unlock_irqrestore(&n->list_lock, flags);
2736 discard_slab(s, page);
2740 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2741 * can perform fastpath freeing without additional function calls.
2743 * The fastpath is only possible if we are freeing to the current cpu slab
2744 * of this processor. This typically the case if we have just allocated
2747 * If fastpath is not possible then fall back to __slab_free where we deal
2748 * with all sorts of special processing.
2750 * Bulk free of a freelist with several objects (all pointing to the
2751 * same page) possible by specifying head and tail ptr, plus objects
2752 * count (cnt). Bulk free indicated by tail pointer being set.
2754 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2755 void *head, void *tail, int cnt,
2758 void *tail_obj = tail ? : head;
2759 struct kmem_cache_cpu *c;
2762 slab_free_freelist_hook(s, head, tail);
2766 * Determine the currently cpus per cpu slab.
2767 * The cpu may change afterward. However that does not matter since
2768 * data is retrieved via this pointer. If we are on the same cpu
2769 * during the cmpxchg then the free will succeed.
2772 tid = this_cpu_read(s->cpu_slab->tid);
2773 c = raw_cpu_ptr(s->cpu_slab);
2774 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2775 unlikely(tid != READ_ONCE(c->tid)));
2777 /* Same with comment on barrier() in slab_alloc_node() */
2780 if (likely(page == c->page)) {
2781 set_freepointer(s, tail_obj, c->freelist);
2783 if (unlikely(!this_cpu_cmpxchg_double(
2784 s->cpu_slab->freelist, s->cpu_slab->tid,
2786 head, next_tid(tid)))) {
2788 note_cmpxchg_failure("slab_free", s, tid);
2791 stat(s, FREE_FASTPATH);
2793 __slab_free(s, page, head, tail_obj, cnt, addr);
2797 void kmem_cache_free(struct kmem_cache *s, void *x)
2799 s = cache_from_obj(s, x);
2802 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2803 trace_kmem_cache_free(_RET_IP_, x);
2805 EXPORT_SYMBOL(kmem_cache_free);
2807 struct detached_freelist {
2812 struct kmem_cache *s;
2816 * This function progressively scans the array with free objects (with
2817 * a limited look ahead) and extract objects belonging to the same
2818 * page. It builds a detached freelist directly within the given
2819 * page/objects. This can happen without any need for
2820 * synchronization, because the objects are owned by running process.
2821 * The freelist is build up as a single linked list in the objects.
2822 * The idea is, that this detached freelist can then be bulk
2823 * transferred to the real freelist(s), but only requiring a single
2824 * synchronization primitive. Look ahead in the array is limited due
2825 * to performance reasons.
2828 int build_detached_freelist(struct kmem_cache *s, size_t size,
2829 void **p, struct detached_freelist *df)
2831 size_t first_skipped_index = 0;
2836 /* Always re-init detached_freelist */
2841 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2842 } while (!object && size);
2847 page = virt_to_head_page(object);
2849 /* Handle kalloc'ed objects */
2850 if (unlikely(!PageSlab(page))) {
2851 BUG_ON(!PageCompound(page));
2853 __free_kmem_pages(page, compound_order(page));
2854 p[size] = NULL; /* mark object processed */
2857 /* Derive kmem_cache from object */
2858 df->s = page->slab_cache;
2860 df->s = cache_from_obj(s, object); /* Support for memcg */
2863 /* Start new detached freelist */
2865 set_freepointer(df->s, object, NULL);
2867 df->freelist = object;
2868 p[size] = NULL; /* mark object processed */
2874 continue; /* Skip processed objects */
2876 /* df->page is always set at this point */
2877 if (df->page == virt_to_head_page(object)) {
2878 /* Opportunity build freelist */
2879 set_freepointer(df->s, object, df->freelist);
2880 df->freelist = object;
2882 p[size] = NULL; /* mark object processed */
2887 /* Limit look ahead search */
2891 if (!first_skipped_index)
2892 first_skipped_index = size + 1;
2895 return first_skipped_index;
2898 /* Note that interrupts must be enabled when calling this function. */
2899 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2905 struct detached_freelist df;
2907 size = build_detached_freelist(s, size, p, &df);
2908 if (unlikely(!df.page))
2911 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2912 } while (likely(size));
2914 EXPORT_SYMBOL(kmem_cache_free_bulk);
2916 /* Note that interrupts must be enabled when calling this function. */
2917 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2920 struct kmem_cache_cpu *c;
2923 /* memcg and kmem_cache debug support */
2924 s = slab_pre_alloc_hook(s, flags);
2928 * Drain objects in the per cpu slab, while disabling local
2929 * IRQs, which protects against PREEMPT and interrupts
2930 * handlers invoking normal fastpath.
2932 local_irq_disable();
2933 c = this_cpu_ptr(s->cpu_slab);
2935 for (i = 0; i < size; i++) {
2936 void *object = c->freelist;
2938 if (unlikely(!object)) {
2940 * Invoking slow path likely have side-effect
2941 * of re-populating per CPU c->freelist
2943 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2945 if (unlikely(!p[i]))
2948 c = this_cpu_ptr(s->cpu_slab);
2949 continue; /* goto for-loop */
2951 c->freelist = get_freepointer(s, object);
2954 c->tid = next_tid(c->tid);
2957 /* Clear memory outside IRQ disabled fastpath loop */
2958 if (unlikely(flags & __GFP_ZERO)) {
2961 for (j = 0; j < i; j++)
2962 memset(p[j], 0, s->object_size);
2965 /* memcg and kmem_cache debug support */
2966 slab_post_alloc_hook(s, flags, size, p);
2970 slab_post_alloc_hook(s, flags, i, p);
2971 __kmem_cache_free_bulk(s, i, p);
2974 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2978 * Object placement in a slab is made very easy because we always start at
2979 * offset 0. If we tune the size of the object to the alignment then we can
2980 * get the required alignment by putting one properly sized object after
2983 * Notice that the allocation order determines the sizes of the per cpu
2984 * caches. Each processor has always one slab available for allocations.
2985 * Increasing the allocation order reduces the number of times that slabs
2986 * must be moved on and off the partial lists and is therefore a factor in
2991 * Mininum / Maximum order of slab pages. This influences locking overhead
2992 * and slab fragmentation. A higher order reduces the number of partial slabs
2993 * and increases the number of allocations possible without having to
2994 * take the list_lock.
2996 static int slub_min_order;
2997 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2998 static int slub_min_objects;
3001 * Calculate the order of allocation given an slab object size.
3003 * The order of allocation has significant impact on performance and other
3004 * system components. Generally order 0 allocations should be preferred since
3005 * order 0 does not cause fragmentation in the page allocator. Larger objects
3006 * be problematic to put into order 0 slabs because there may be too much
3007 * unused space left. We go to a higher order if more than 1/16th of the slab
3010 * In order to reach satisfactory performance we must ensure that a minimum
3011 * number of objects is in one slab. Otherwise we may generate too much
3012 * activity on the partial lists which requires taking the list_lock. This is
3013 * less a concern for large slabs though which are rarely used.
3015 * slub_max_order specifies the order where we begin to stop considering the
3016 * number of objects in a slab as critical. If we reach slub_max_order then
3017 * we try to keep the page order as low as possible. So we accept more waste
3018 * of space in favor of a small page order.
3020 * Higher order allocations also allow the placement of more objects in a
3021 * slab and thereby reduce object handling overhead. If the user has
3022 * requested a higher mininum order then we start with that one instead of
3023 * the smallest order which will fit the object.
3025 static inline int slab_order(int size, int min_objects,
3026 int max_order, int fract_leftover, int reserved)
3030 int min_order = slub_min_order;
3032 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3033 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3035 for (order = max(min_order, get_order(min_objects * size + reserved));
3036 order <= max_order; order++) {
3038 unsigned long slab_size = PAGE_SIZE << order;
3040 rem = (slab_size - reserved) % size;
3042 if (rem <= slab_size / fract_leftover)
3049 static inline int calculate_order(int size, int reserved)
3057 * Attempt to find best configuration for a slab. This
3058 * works by first attempting to generate a layout with
3059 * the best configuration and backing off gradually.
3061 * First we increase the acceptable waste in a slab. Then
3062 * we reduce the minimum objects required in a slab.
3064 min_objects = slub_min_objects;
3066 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3067 max_objects = order_objects(slub_max_order, size, reserved);
3068 min_objects = min(min_objects, max_objects);
3070 while (min_objects > 1) {
3072 while (fraction >= 4) {
3073 order = slab_order(size, min_objects,
3074 slub_max_order, fraction, reserved);
3075 if (order <= slub_max_order)
3083 * We were unable to place multiple objects in a slab. Now
3084 * lets see if we can place a single object there.
3086 order = slab_order(size, 1, slub_max_order, 1, reserved);
3087 if (order <= slub_max_order)
3091 * Doh this slab cannot be placed using slub_max_order.
3093 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3094 if (order < MAX_ORDER)
3100 init_kmem_cache_node(struct kmem_cache_node *n)
3103 spin_lock_init(&n->list_lock);
3104 INIT_LIST_HEAD(&n->partial);
3105 #ifdef CONFIG_SLUB_DEBUG
3106 atomic_long_set(&n->nr_slabs, 0);
3107 atomic_long_set(&n->total_objects, 0);
3108 INIT_LIST_HEAD(&n->full);
3112 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3114 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3115 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3118 * Must align to double word boundary for the double cmpxchg
3119 * instructions to work; see __pcpu_double_call_return_bool().
3121 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3122 2 * sizeof(void *));
3127 init_kmem_cache_cpus(s);
3132 static struct kmem_cache *kmem_cache_node;
3135 * No kmalloc_node yet so do it by hand. We know that this is the first
3136 * slab on the node for this slabcache. There are no concurrent accesses
3139 * Note that this function only works on the kmem_cache_node
3140 * when allocating for the kmem_cache_node. This is used for bootstrapping
3141 * memory on a fresh node that has no slab structures yet.
3143 static void early_kmem_cache_node_alloc(int node)
3146 struct kmem_cache_node *n;
3148 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3150 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3153 if (page_to_nid(page) != node) {
3154 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3155 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3160 page->freelist = get_freepointer(kmem_cache_node, n);
3163 kmem_cache_node->node[node] = n;
3164 #ifdef CONFIG_SLUB_DEBUG
3165 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3166 init_tracking(kmem_cache_node, n);
3168 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node));
3169 init_kmem_cache_node(n);
3170 inc_slabs_node(kmem_cache_node, node, page->objects);
3173 * No locks need to be taken here as it has just been
3174 * initialized and there is no concurrent access.
3176 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3179 static void free_kmem_cache_nodes(struct kmem_cache *s)
3182 struct kmem_cache_node *n;
3184 for_each_kmem_cache_node(s, node, n) {
3185 kmem_cache_free(kmem_cache_node, n);
3186 s->node[node] = NULL;
3190 void __kmem_cache_release(struct kmem_cache *s)
3192 free_percpu(s->cpu_slab);
3193 free_kmem_cache_nodes(s);
3196 static int init_kmem_cache_nodes(struct kmem_cache *s)
3200 for_each_node_state(node, N_NORMAL_MEMORY) {
3201 struct kmem_cache_node *n;
3203 if (slab_state == DOWN) {
3204 early_kmem_cache_node_alloc(node);
3207 n = kmem_cache_alloc_node(kmem_cache_node,
3211 free_kmem_cache_nodes(s);
3216 init_kmem_cache_node(n);
3221 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3223 if (min < MIN_PARTIAL)
3225 else if (min > MAX_PARTIAL)
3227 s->min_partial = min;
3231 * calculate_sizes() determines the order and the distribution of data within
3234 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3236 unsigned long flags = s->flags;
3237 unsigned long size = s->object_size;
3241 * Round up object size to the next word boundary. We can only
3242 * place the free pointer at word boundaries and this determines
3243 * the possible location of the free pointer.
3245 size = ALIGN(size, sizeof(void *));
3247 #ifdef CONFIG_SLUB_DEBUG
3249 * Determine if we can poison the object itself. If the user of
3250 * the slab may touch the object after free or before allocation
3251 * then we should never poison the object itself.
3253 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3255 s->flags |= __OBJECT_POISON;
3257 s->flags &= ~__OBJECT_POISON;
3261 * If we are Redzoning then check if there is some space between the
3262 * end of the object and the free pointer. If not then add an
3263 * additional word to have some bytes to store Redzone information.
3265 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3266 size += sizeof(void *);
3270 * With that we have determined the number of bytes in actual use
3271 * by the object. This is the potential offset to the free pointer.
3275 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3278 * Relocate free pointer after the object if it is not
3279 * permitted to overwrite the first word of the object on
3282 * This is the case if we do RCU, have a constructor or
3283 * destructor or are poisoning the objects.
3286 size += sizeof(void *);
3289 #ifdef CONFIG_SLUB_DEBUG
3290 if (flags & SLAB_STORE_USER)
3292 * Need to store information about allocs and frees after
3295 size += 2 * sizeof(struct track);
3297 if (flags & SLAB_RED_ZONE)
3299 * Add some empty padding so that we can catch
3300 * overwrites from earlier objects rather than let
3301 * tracking information or the free pointer be
3302 * corrupted if a user writes before the start
3305 size += sizeof(void *);
3307 if (flags & SLAB_RED_ZONE) {
3308 s->red_left_pad = sizeof(void *);
3310 s->red_left_pad = min_t(int, s->red_left_pad,
3311 KASAN_SHADOW_SCALE_SIZE);
3313 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3314 size += s->red_left_pad;
3319 * SLUB stores one object immediately after another beginning from
3320 * offset 0. In order to align the objects we have to simply size
3321 * each object to conform to the alignment.
3323 size = ALIGN(size, s->align);
3325 if (forced_order >= 0)
3326 order = forced_order;
3328 order = calculate_order(size, s->reserved);
3335 s->allocflags |= __GFP_COMP;
3337 if (s->flags & SLAB_CACHE_DMA)
3338 s->allocflags |= GFP_DMA;
3340 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3341 s->allocflags |= __GFP_RECLAIMABLE;
3344 * Determine the number of objects per slab
3346 s->oo = oo_make(order, size, s->reserved);
3347 s->min = oo_make(get_order(size), size, s->reserved);
3348 if (oo_objects(s->oo) > oo_objects(s->max))
3351 return !!oo_objects(s->oo);
3354 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3356 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3359 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3360 s->reserved = sizeof(struct rcu_head);
3362 if (!calculate_sizes(s, -1))
3364 if (disable_higher_order_debug) {
3366 * Disable debugging flags that store metadata if the min slab
3369 if (get_order(s->size) > get_order(s->object_size)) {
3370 s->flags &= ~DEBUG_METADATA_FLAGS;
3372 if (!calculate_sizes(s, -1))
3377 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3378 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3379 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3380 /* Enable fast mode */
3381 s->flags |= __CMPXCHG_DOUBLE;
3385 * The larger the object size is, the more pages we want on the partial
3386 * list to avoid pounding the page allocator excessively.
3388 set_min_partial(s, ilog2(s->size) / 2);
3391 * cpu_partial determined the maximum number of objects kept in the
3392 * per cpu partial lists of a processor.
3394 * Per cpu partial lists mainly contain slabs that just have one
3395 * object freed. If they are used for allocation then they can be
3396 * filled up again with minimal effort. The slab will never hit the
3397 * per node partial lists and therefore no locking will be required.
3399 * This setting also determines
3401 * A) The number of objects from per cpu partial slabs dumped to the
3402 * per node list when we reach the limit.
3403 * B) The number of objects in cpu partial slabs to extract from the
3404 * per node list when we run out of per cpu objects. We only fetch
3405 * 50% to keep some capacity around for frees.
3407 if (!kmem_cache_has_cpu_partial(s))
3409 else if (s->size >= PAGE_SIZE)
3411 else if (s->size >= 1024)
3413 else if (s->size >= 256)
3414 s->cpu_partial = 13;
3416 s->cpu_partial = 30;
3419 s->remote_node_defrag_ratio = 1000;
3421 if (!init_kmem_cache_nodes(s))
3424 if (alloc_kmem_cache_cpus(s))
3427 free_kmem_cache_nodes(s);
3429 if (flags & SLAB_PANIC)
3430 panic("Cannot create slab %s size=%lu realsize=%u "
3431 "order=%u offset=%u flags=%lx\n",
3432 s->name, (unsigned long)s->size, s->size,
3433 oo_order(s->oo), s->offset, flags);
3437 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3440 #ifdef CONFIG_SLUB_DEBUG
3441 void *addr = page_address(page);
3443 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3444 sizeof(long), GFP_ATOMIC);
3447 slab_err(s, page, text, s->name);
3450 get_map(s, page, map);
3451 for_each_object(p, s, addr, page->objects) {
3452 void *object = fixup_red_left(s, p);
3454 if (!test_bit(slab_index(p, s, addr), map)) {
3455 pr_err("INFO: Object 0x%p @offset=%tu\n",
3456 object, object - addr);
3457 print_tracking(s, object);
3466 * Attempt to free all partial slabs on a node.
3467 * This is called from __kmem_cache_shutdown(). We must take list_lock
3468 * because sysfs file might still access partial list after the shutdowning.
3470 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3472 struct page *page, *h;
3474 BUG_ON(irqs_disabled());
3475 spin_lock_irq(&n->list_lock);
3476 list_for_each_entry_safe(page, h, &n->partial, lru) {
3478 remove_partial(n, page);
3479 discard_slab(s, page);
3481 list_slab_objects(s, page,
3482 "Objects remaining in %s on __kmem_cache_shutdown()");
3485 spin_unlock_irq(&n->list_lock);
3489 * Release all resources used by a slab cache.
3491 int __kmem_cache_shutdown(struct kmem_cache *s)
3494 struct kmem_cache_node *n;
3497 /* Attempt to free all objects */
3498 for_each_kmem_cache_node(s, node, n) {
3500 if (n->nr_partial || slabs_node(s, node))
3506 /********************************************************************
3508 *******************************************************************/
3510 static int __init setup_slub_min_order(char *str)
3512 get_option(&str, &slub_min_order);
3517 __setup("slub_min_order=", setup_slub_min_order);
3519 static int __init setup_slub_max_order(char *str)
3521 get_option(&str, &slub_max_order);
3522 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3527 __setup("slub_max_order=", setup_slub_max_order);
3529 static int __init setup_slub_min_objects(char *str)
3531 get_option(&str, &slub_min_objects);
3536 __setup("slub_min_objects=", setup_slub_min_objects);
3538 void *__kmalloc(size_t size, gfp_t flags)
3540 struct kmem_cache *s;
3543 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3544 return kmalloc_large(size, flags);
3546 s = kmalloc_slab(size, flags);
3548 if (unlikely(ZERO_OR_NULL_PTR(s)))
3551 ret = slab_alloc(s, flags, _RET_IP_);
3553 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3555 kasan_kmalloc(s, ret, size);
3559 EXPORT_SYMBOL(__kmalloc);
3562 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3567 flags |= __GFP_COMP | __GFP_NOTRACK;
3568 page = alloc_kmem_pages_node(node, flags, get_order(size));
3570 ptr = page_address(page);
3572 kmalloc_large_node_hook(ptr, size, flags);
3576 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3578 struct kmem_cache *s;
3581 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3582 ret = kmalloc_large_node(size, flags, node);
3584 trace_kmalloc_node(_RET_IP_, ret,
3585 size, PAGE_SIZE << get_order(size),
3591 s = kmalloc_slab(size, flags);
3593 if (unlikely(ZERO_OR_NULL_PTR(s)))
3596 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3598 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3600 kasan_kmalloc(s, ret, size);
3604 EXPORT_SYMBOL(__kmalloc_node);
3607 static size_t __ksize(const void *object)
3611 if (unlikely(object == ZERO_SIZE_PTR))
3614 page = virt_to_head_page(object);
3616 if (unlikely(!PageSlab(page))) {
3617 WARN_ON(!PageCompound(page));
3618 return PAGE_SIZE << compound_order(page);
3621 return slab_ksize(page->slab_cache);
3624 size_t ksize(const void *object)
3626 size_t size = __ksize(object);
3627 /* We assume that ksize callers could use whole allocated area,
3628 so we need unpoison this area. */
3629 kasan_krealloc(object, size);
3632 EXPORT_SYMBOL(ksize);
3634 void kfree(const void *x)
3637 void *object = (void *)x;
3639 trace_kfree(_RET_IP_, x);
3641 if (unlikely(ZERO_OR_NULL_PTR(x)))
3644 page = virt_to_head_page(x);
3645 if (unlikely(!PageSlab(page))) {
3646 BUG_ON(!PageCompound(page));
3648 __free_kmem_pages(page, compound_order(page));
3651 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3653 EXPORT_SYMBOL(kfree);
3655 #define SHRINK_PROMOTE_MAX 32
3658 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3659 * up most to the head of the partial lists. New allocations will then
3660 * fill those up and thus they can be removed from the partial lists.
3662 * The slabs with the least items are placed last. This results in them
3663 * being allocated from last increasing the chance that the last objects
3664 * are freed in them.
3666 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3670 struct kmem_cache_node *n;
3673 struct list_head discard;
3674 struct list_head promote[SHRINK_PROMOTE_MAX];
3675 unsigned long flags;
3680 * Disable empty slabs caching. Used to avoid pinning offline
3681 * memory cgroups by kmem pages that can be freed.
3687 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3688 * so we have to make sure the change is visible.
3690 kick_all_cpus_sync();
3694 for_each_kmem_cache_node(s, node, n) {
3695 INIT_LIST_HEAD(&discard);
3696 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3697 INIT_LIST_HEAD(promote + i);
3699 spin_lock_irqsave(&n->list_lock, flags);
3702 * Build lists of slabs to discard or promote.
3704 * Note that concurrent frees may occur while we hold the
3705 * list_lock. page->inuse here is the upper limit.
3707 list_for_each_entry_safe(page, t, &n->partial, lru) {
3708 int free = page->objects - page->inuse;
3710 /* Do not reread page->inuse */
3713 /* We do not keep full slabs on the list */
3716 if (free == page->objects) {
3717 list_move(&page->lru, &discard);
3719 } else if (free <= SHRINK_PROMOTE_MAX)
3720 list_move(&page->lru, promote + free - 1);
3724 * Promote the slabs filled up most to the head of the
3727 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3728 list_splice(promote + i, &n->partial);
3730 spin_unlock_irqrestore(&n->list_lock, flags);
3732 /* Release empty slabs */
3733 list_for_each_entry_safe(page, t, &discard, lru)
3734 discard_slab(s, page);
3736 if (slabs_node(s, node))
3743 static int slab_mem_going_offline_callback(void *arg)
3745 struct kmem_cache *s;
3747 mutex_lock(&slab_mutex);
3748 list_for_each_entry(s, &slab_caches, list)
3749 __kmem_cache_shrink(s, false);
3750 mutex_unlock(&slab_mutex);
3755 static void slab_mem_offline_callback(void *arg)
3757 struct kmem_cache_node *n;
3758 struct kmem_cache *s;
3759 struct memory_notify *marg = arg;
3762 offline_node = marg->status_change_nid_normal;
3765 * If the node still has available memory. we need kmem_cache_node
3768 if (offline_node < 0)
3771 mutex_lock(&slab_mutex);
3772 list_for_each_entry(s, &slab_caches, list) {
3773 n = get_node(s, offline_node);
3776 * if n->nr_slabs > 0, slabs still exist on the node
3777 * that is going down. We were unable to free them,
3778 * and offline_pages() function shouldn't call this
3779 * callback. So, we must fail.
3781 BUG_ON(slabs_node(s, offline_node));
3783 s->node[offline_node] = NULL;
3784 kmem_cache_free(kmem_cache_node, n);
3787 mutex_unlock(&slab_mutex);
3790 static int slab_mem_going_online_callback(void *arg)
3792 struct kmem_cache_node *n;
3793 struct kmem_cache *s;
3794 struct memory_notify *marg = arg;
3795 int nid = marg->status_change_nid_normal;
3799 * If the node's memory is already available, then kmem_cache_node is
3800 * already created. Nothing to do.
3806 * We are bringing a node online. No memory is available yet. We must
3807 * allocate a kmem_cache_node structure in order to bring the node
3810 mutex_lock(&slab_mutex);
3811 list_for_each_entry(s, &slab_caches, list) {
3813 * XXX: kmem_cache_alloc_node will fallback to other nodes
3814 * since memory is not yet available from the node that
3817 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3822 init_kmem_cache_node(n);
3826 mutex_unlock(&slab_mutex);
3830 static int slab_memory_callback(struct notifier_block *self,
3831 unsigned long action, void *arg)
3836 case MEM_GOING_ONLINE:
3837 ret = slab_mem_going_online_callback(arg);
3839 case MEM_GOING_OFFLINE:
3840 ret = slab_mem_going_offline_callback(arg);
3843 case MEM_CANCEL_ONLINE:
3844 slab_mem_offline_callback(arg);
3847 case MEM_CANCEL_OFFLINE:
3851 ret = notifier_from_errno(ret);
3857 static struct notifier_block slab_memory_callback_nb = {
3858 .notifier_call = slab_memory_callback,
3859 .priority = SLAB_CALLBACK_PRI,
3862 /********************************************************************
3863 * Basic setup of slabs
3864 *******************************************************************/
3867 * Used for early kmem_cache structures that were allocated using
3868 * the page allocator. Allocate them properly then fix up the pointers
3869 * that may be pointing to the wrong kmem_cache structure.
3872 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3875 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3876 struct kmem_cache_node *n;
3878 memcpy(s, static_cache, kmem_cache->object_size);
3881 * This runs very early, and only the boot processor is supposed to be
3882 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3885 __flush_cpu_slab(s, smp_processor_id());
3886 for_each_kmem_cache_node(s, node, n) {
3889 list_for_each_entry(p, &n->partial, lru)
3892 #ifdef CONFIG_SLUB_DEBUG
3893 list_for_each_entry(p, &n->full, lru)
3897 slab_init_memcg_params(s);
3898 list_add(&s->list, &slab_caches);
3902 void __init kmem_cache_init(void)
3904 static __initdata struct kmem_cache boot_kmem_cache,
3905 boot_kmem_cache_node;
3907 if (debug_guardpage_minorder())
3910 kmem_cache_node = &boot_kmem_cache_node;
3911 kmem_cache = &boot_kmem_cache;
3913 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3914 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3916 register_hotmemory_notifier(&slab_memory_callback_nb);
3918 /* Able to allocate the per node structures */
3919 slab_state = PARTIAL;
3921 create_boot_cache(kmem_cache, "kmem_cache",
3922 offsetof(struct kmem_cache, node) +
3923 nr_node_ids * sizeof(struct kmem_cache_node *),
3924 SLAB_HWCACHE_ALIGN);
3926 kmem_cache = bootstrap(&boot_kmem_cache);
3929 * Allocate kmem_cache_node properly from the kmem_cache slab.
3930 * kmem_cache_node is separately allocated so no need to
3931 * update any list pointers.
3933 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3935 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3936 setup_kmalloc_cache_index_table();
3937 create_kmalloc_caches(0);
3940 register_cpu_notifier(&slab_notifier);
3943 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3945 slub_min_order, slub_max_order, slub_min_objects,
3946 nr_cpu_ids, nr_node_ids);
3949 void __init kmem_cache_init_late(void)
3954 __kmem_cache_alias(const char *name, size_t size, size_t align,
3955 unsigned long flags, void (*ctor)(void *))
3957 struct kmem_cache *s, *c;
3959 s = find_mergeable(size, align, flags, name, ctor);
3964 * Adjust the object sizes so that we clear
3965 * the complete object on kzalloc.
3967 s->object_size = max(s->object_size, (int)size);
3968 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3970 for_each_memcg_cache(c, s) {
3971 c->object_size = s->object_size;
3972 c->inuse = max_t(int, c->inuse,
3973 ALIGN(size, sizeof(void *)));
3976 if (sysfs_slab_alias(s, name)) {
3985 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3989 err = kmem_cache_open(s, flags);
3993 /* Mutex is not taken during early boot */
3994 if (slab_state <= UP)
3997 memcg_propagate_slab_attrs(s);
3998 err = sysfs_slab_add(s);
4000 __kmem_cache_release(s);
4007 * Use the cpu notifier to insure that the cpu slabs are flushed when
4010 static int slab_cpuup_callback(struct notifier_block *nfb,
4011 unsigned long action, void *hcpu)
4013 long cpu = (long)hcpu;
4014 struct kmem_cache *s;
4015 unsigned long flags;
4018 case CPU_UP_CANCELED:
4019 case CPU_UP_CANCELED_FROZEN:
4021 case CPU_DEAD_FROZEN:
4022 mutex_lock(&slab_mutex);
4023 list_for_each_entry(s, &slab_caches, list) {
4024 local_irq_save(flags);
4025 __flush_cpu_slab(s, cpu);
4026 local_irq_restore(flags);
4028 mutex_unlock(&slab_mutex);
4036 static struct notifier_block slab_notifier = {
4037 .notifier_call = slab_cpuup_callback
4042 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4044 struct kmem_cache *s;
4047 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4048 return kmalloc_large(size, gfpflags);
4050 s = kmalloc_slab(size, gfpflags);
4052 if (unlikely(ZERO_OR_NULL_PTR(s)))
4055 ret = slab_alloc(s, gfpflags, caller);
4057 /* Honor the call site pointer we received. */
4058 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4064 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4065 int node, unsigned long caller)
4067 struct kmem_cache *s;
4070 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4071 ret = kmalloc_large_node(size, gfpflags, node);
4073 trace_kmalloc_node(caller, ret,
4074 size, PAGE_SIZE << get_order(size),
4080 s = kmalloc_slab(size, gfpflags);
4082 if (unlikely(ZERO_OR_NULL_PTR(s)))
4085 ret = slab_alloc_node(s, gfpflags, node, caller);
4087 /* Honor the call site pointer we received. */
4088 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4095 static int count_inuse(struct page *page)
4100 static int count_total(struct page *page)
4102 return page->objects;
4106 #ifdef CONFIG_SLUB_DEBUG
4107 static int validate_slab(struct kmem_cache *s, struct page *page,
4111 void *addr = page_address(page);
4113 if (!check_slab(s, page) ||
4114 !on_freelist(s, page, NULL))
4117 /* Now we know that a valid freelist exists */
4118 bitmap_zero(map, page->objects);
4120 get_map(s, page, map);
4121 for_each_object(p, s, addr, page->objects) {
4122 void *object = fixup_red_left(s, p);
4124 if (test_bit(slab_index(p, s, addr), map))
4125 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
4129 for_each_object(p, s, addr, page->objects) {
4130 void *object = fixup_red_left(s, p);
4132 if (!test_bit(slab_index(p, s, addr), map))
4133 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
4140 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4144 validate_slab(s, page, map);
4148 static int validate_slab_node(struct kmem_cache *s,
4149 struct kmem_cache_node *n, unsigned long *map)
4151 unsigned long count = 0;
4153 unsigned long flags;
4155 spin_lock_irqsave(&n->list_lock, flags);
4157 list_for_each_entry(page, &n->partial, lru) {
4158 validate_slab_slab(s, page, map);
4161 if (count != n->nr_partial)
4162 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4163 s->name, count, n->nr_partial);
4165 if (!(s->flags & SLAB_STORE_USER))
4168 list_for_each_entry(page, &n->full, lru) {
4169 validate_slab_slab(s, page, map);
4172 if (count != atomic_long_read(&n->nr_slabs))
4173 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4174 s->name, count, atomic_long_read(&n->nr_slabs));
4177 spin_unlock_irqrestore(&n->list_lock, flags);
4181 static long validate_slab_cache(struct kmem_cache *s)
4184 unsigned long count = 0;
4185 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4186 sizeof(unsigned long), GFP_KERNEL);
4187 struct kmem_cache_node *n;
4193 for_each_kmem_cache_node(s, node, n)
4194 count += validate_slab_node(s, n, map);
4199 * Generate lists of code addresses where slabcache objects are allocated
4204 unsigned long count;
4211 DECLARE_BITMAP(cpus, NR_CPUS);
4217 unsigned long count;
4218 struct location *loc;
4221 static void free_loc_track(struct loc_track *t)
4224 free_pages((unsigned long)t->loc,
4225 get_order(sizeof(struct location) * t->max));
4228 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4233 order = get_order(sizeof(struct location) * max);
4235 l = (void *)__get_free_pages(flags, order);
4240 memcpy(l, t->loc, sizeof(struct location) * t->count);
4248 static int add_location(struct loc_track *t, struct kmem_cache *s,
4249 const struct track *track)
4251 long start, end, pos;
4253 unsigned long caddr;
4254 unsigned long age = jiffies - track->when;
4260 pos = start + (end - start + 1) / 2;
4263 * There is nothing at "end". If we end up there
4264 * we need to add something to before end.
4269 caddr = t->loc[pos].addr;
4270 if (track->addr == caddr) {
4276 if (age < l->min_time)
4278 if (age > l->max_time)
4281 if (track->pid < l->min_pid)
4282 l->min_pid = track->pid;
4283 if (track->pid > l->max_pid)
4284 l->max_pid = track->pid;
4286 cpumask_set_cpu(track->cpu,
4287 to_cpumask(l->cpus));
4289 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4293 if (track->addr < caddr)
4300 * Not found. Insert new tracking element.
4302 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4308 (t->count - pos) * sizeof(struct location));
4311 l->addr = track->addr;
4315 l->min_pid = track->pid;
4316 l->max_pid = track->pid;
4317 cpumask_clear(to_cpumask(l->cpus));
4318 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4319 nodes_clear(l->nodes);
4320 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4324 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4325 struct page *page, enum track_item alloc,
4328 void *addr = page_address(page);
4331 bitmap_zero(map, page->objects);
4332 get_map(s, page, map);
4334 for_each_object(p, s, addr, page->objects) {
4335 void *object = fixup_red_left(s, p);
4337 if (!test_bit(slab_index(p, s, addr), map))
4338 add_location(t, s, get_track(s, object, alloc));
4342 static int list_locations(struct kmem_cache *s, char *buf,
4343 enum track_item alloc)
4347 struct loc_track t = { 0, 0, NULL };
4349 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4350 sizeof(unsigned long), GFP_KERNEL);
4351 struct kmem_cache_node *n;
4353 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4356 return sprintf(buf, "Out of memory\n");
4358 /* Push back cpu slabs */
4361 for_each_kmem_cache_node(s, node, n) {
4362 unsigned long flags;
4365 if (!atomic_long_read(&n->nr_slabs))
4368 spin_lock_irqsave(&n->list_lock, flags);
4369 list_for_each_entry(page, &n->partial, lru)
4370 process_slab(&t, s, page, alloc, map);
4371 list_for_each_entry(page, &n->full, lru)
4372 process_slab(&t, s, page, alloc, map);
4373 spin_unlock_irqrestore(&n->list_lock, flags);
4376 for (i = 0; i < t.count; i++) {
4377 struct location *l = &t.loc[i];
4379 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4381 len += sprintf(buf + len, "%7ld ", l->count);
4384 len += sprintf(buf + len, "%pS", (void *)l->addr);
4386 len += sprintf(buf + len, "<not-available>");
4388 if (l->sum_time != l->min_time) {
4389 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4391 (long)div_u64(l->sum_time, l->count),
4394 len += sprintf(buf + len, " age=%ld",
4397 if (l->min_pid != l->max_pid)
4398 len += sprintf(buf + len, " pid=%ld-%ld",
4399 l->min_pid, l->max_pid);
4401 len += sprintf(buf + len, " pid=%ld",
4404 if (num_online_cpus() > 1 &&
4405 !cpumask_empty(to_cpumask(l->cpus)) &&
4406 len < PAGE_SIZE - 60)
4407 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4409 cpumask_pr_args(to_cpumask(l->cpus)));
4411 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4412 len < PAGE_SIZE - 60)
4413 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4415 nodemask_pr_args(&l->nodes));
4417 len += sprintf(buf + len, "\n");
4423 len += sprintf(buf, "No data\n");
4428 #ifdef SLUB_RESILIENCY_TEST
4429 static void __init resiliency_test(void)
4433 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4435 pr_err("SLUB resiliency testing\n");
4436 pr_err("-----------------------\n");
4437 pr_err("A. Corruption after allocation\n");
4439 p = kzalloc(16, GFP_KERNEL);
4441 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4444 validate_slab_cache(kmalloc_caches[4]);
4446 /* Hmmm... The next two are dangerous */
4447 p = kzalloc(32, GFP_KERNEL);
4448 p[32 + sizeof(void *)] = 0x34;
4449 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4451 pr_err("If allocated object is overwritten then not detectable\n\n");
4453 validate_slab_cache(kmalloc_caches[5]);
4454 p = kzalloc(64, GFP_KERNEL);
4455 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4457 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4459 pr_err("If allocated object is overwritten then not detectable\n\n");
4460 validate_slab_cache(kmalloc_caches[6]);
4462 pr_err("\nB. Corruption after free\n");
4463 p = kzalloc(128, GFP_KERNEL);
4466 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4467 validate_slab_cache(kmalloc_caches[7]);
4469 p = kzalloc(256, GFP_KERNEL);
4472 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4473 validate_slab_cache(kmalloc_caches[8]);
4475 p = kzalloc(512, GFP_KERNEL);
4478 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4479 validate_slab_cache(kmalloc_caches[9]);
4483 static void resiliency_test(void) {};
4488 enum slab_stat_type {
4489 SL_ALL, /* All slabs */
4490 SL_PARTIAL, /* Only partially allocated slabs */
4491 SL_CPU, /* Only slabs used for cpu caches */
4492 SL_OBJECTS, /* Determine allocated objects not slabs */
4493 SL_TOTAL /* Determine object capacity not slabs */
4496 #define SO_ALL (1 << SL_ALL)
4497 #define SO_PARTIAL (1 << SL_PARTIAL)
4498 #define SO_CPU (1 << SL_CPU)
4499 #define SO_OBJECTS (1 << SL_OBJECTS)
4500 #define SO_TOTAL (1 << SL_TOTAL)
4502 static ssize_t show_slab_objects(struct kmem_cache *s,
4503 char *buf, unsigned long flags)
4505 unsigned long total = 0;
4508 unsigned long *nodes;
4510 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4514 if (flags & SO_CPU) {
4517 for_each_possible_cpu(cpu) {
4518 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4523 page = READ_ONCE(c->page);
4527 node = page_to_nid(page);
4528 if (flags & SO_TOTAL)
4530 else if (flags & SO_OBJECTS)
4538 page = READ_ONCE(c->partial);
4540 node = page_to_nid(page);
4541 if (flags & SO_TOTAL)
4543 else if (flags & SO_OBJECTS)
4554 #ifdef CONFIG_SLUB_DEBUG
4555 if (flags & SO_ALL) {
4556 struct kmem_cache_node *n;
4558 for_each_kmem_cache_node(s, node, n) {
4560 if (flags & SO_TOTAL)
4561 x = atomic_long_read(&n->total_objects);
4562 else if (flags & SO_OBJECTS)
4563 x = atomic_long_read(&n->total_objects) -
4564 count_partial(n, count_free);
4566 x = atomic_long_read(&n->nr_slabs);
4573 if (flags & SO_PARTIAL) {
4574 struct kmem_cache_node *n;
4576 for_each_kmem_cache_node(s, node, n) {
4577 if (flags & SO_TOTAL)
4578 x = count_partial(n, count_total);
4579 else if (flags & SO_OBJECTS)
4580 x = count_partial(n, count_inuse);
4587 x = sprintf(buf, "%lu", total);
4589 for (node = 0; node < nr_node_ids; node++)
4591 x += sprintf(buf + x, " N%d=%lu",
4596 return x + sprintf(buf + x, "\n");
4599 #ifdef CONFIG_SLUB_DEBUG
4600 static int any_slab_objects(struct kmem_cache *s)
4603 struct kmem_cache_node *n;
4605 for_each_kmem_cache_node(s, node, n)
4606 if (atomic_long_read(&n->total_objects))
4613 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4614 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4616 struct slab_attribute {
4617 struct attribute attr;
4618 ssize_t (*show)(struct kmem_cache *s, char *buf);
4619 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4622 #define SLAB_ATTR_RO(_name) \
4623 static struct slab_attribute _name##_attr = \
4624 __ATTR(_name, 0400, _name##_show, NULL)
4626 #define SLAB_ATTR(_name) \
4627 static struct slab_attribute _name##_attr = \
4628 __ATTR(_name, 0600, _name##_show, _name##_store)
4630 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4632 return sprintf(buf, "%d\n", s->size);
4634 SLAB_ATTR_RO(slab_size);
4636 static ssize_t align_show(struct kmem_cache *s, char *buf)
4638 return sprintf(buf, "%d\n", s->align);
4640 SLAB_ATTR_RO(align);
4642 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4644 return sprintf(buf, "%d\n", s->object_size);
4646 SLAB_ATTR_RO(object_size);
4648 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4650 return sprintf(buf, "%d\n", oo_objects(s->oo));
4652 SLAB_ATTR_RO(objs_per_slab);
4654 static ssize_t order_store(struct kmem_cache *s,
4655 const char *buf, size_t length)
4657 unsigned long order;
4660 err = kstrtoul(buf, 10, &order);
4664 if (order > slub_max_order || order < slub_min_order)
4667 calculate_sizes(s, order);
4671 static ssize_t order_show(struct kmem_cache *s, char *buf)
4673 return sprintf(buf, "%d\n", oo_order(s->oo));
4677 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4679 return sprintf(buf, "%lu\n", s->min_partial);
4682 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4688 err = kstrtoul(buf, 10, &min);
4692 set_min_partial(s, min);
4695 SLAB_ATTR(min_partial);
4697 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4699 return sprintf(buf, "%u\n", s->cpu_partial);
4702 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4705 unsigned long objects;
4708 err = kstrtoul(buf, 10, &objects);
4711 if (objects && !kmem_cache_has_cpu_partial(s))
4714 s->cpu_partial = objects;
4718 SLAB_ATTR(cpu_partial);
4720 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4724 return sprintf(buf, "%pS\n", s->ctor);
4728 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4730 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4732 SLAB_ATTR_RO(aliases);
4734 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4736 return show_slab_objects(s, buf, SO_PARTIAL);
4738 SLAB_ATTR_RO(partial);
4740 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4742 return show_slab_objects(s, buf, SO_CPU);
4744 SLAB_ATTR_RO(cpu_slabs);
4746 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4748 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4750 SLAB_ATTR_RO(objects);
4752 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4754 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4756 SLAB_ATTR_RO(objects_partial);
4758 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4765 for_each_online_cpu(cpu) {
4766 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4769 pages += page->pages;
4770 objects += page->pobjects;
4774 len = sprintf(buf, "%d(%d)", objects, pages);
4777 for_each_online_cpu(cpu) {
4778 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4780 if (page && len < PAGE_SIZE - 20)
4781 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4782 page->pobjects, page->pages);
4785 return len + sprintf(buf + len, "\n");
4787 SLAB_ATTR_RO(slabs_cpu_partial);
4789 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4791 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4794 static ssize_t reclaim_account_store(struct kmem_cache *s,
4795 const char *buf, size_t length)
4797 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4799 s->flags |= SLAB_RECLAIM_ACCOUNT;
4802 SLAB_ATTR(reclaim_account);
4804 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4806 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4808 SLAB_ATTR_RO(hwcache_align);
4810 #ifdef CONFIG_ZONE_DMA
4811 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4813 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4815 SLAB_ATTR_RO(cache_dma);
4818 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4820 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4822 SLAB_ATTR_RO(destroy_by_rcu);
4824 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4826 return sprintf(buf, "%d\n", s->reserved);
4828 SLAB_ATTR_RO(reserved);
4830 #ifdef CONFIG_SLUB_DEBUG
4831 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4833 return show_slab_objects(s, buf, SO_ALL);
4835 SLAB_ATTR_RO(slabs);
4837 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4839 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4841 SLAB_ATTR_RO(total_objects);
4843 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4845 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4848 static ssize_t sanity_checks_store(struct kmem_cache *s,
4849 const char *buf, size_t length)
4851 s->flags &= ~SLAB_DEBUG_FREE;
4852 if (buf[0] == '1') {
4853 s->flags &= ~__CMPXCHG_DOUBLE;
4854 s->flags |= SLAB_DEBUG_FREE;
4858 SLAB_ATTR(sanity_checks);
4860 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4862 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4865 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4869 * Tracing a merged cache is going to give confusing results
4870 * as well as cause other issues like converting a mergeable
4871 * cache into an umergeable one.
4873 if (s->refcount > 1)
4876 s->flags &= ~SLAB_TRACE;
4877 if (buf[0] == '1') {
4878 s->flags &= ~__CMPXCHG_DOUBLE;
4879 s->flags |= SLAB_TRACE;
4885 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4887 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4890 static ssize_t red_zone_store(struct kmem_cache *s,
4891 const char *buf, size_t length)
4893 if (any_slab_objects(s))
4896 s->flags &= ~SLAB_RED_ZONE;
4897 if (buf[0] == '1') {
4898 s->flags &= ~__CMPXCHG_DOUBLE;
4899 s->flags |= SLAB_RED_ZONE;
4901 calculate_sizes(s, -1);
4904 SLAB_ATTR(red_zone);
4906 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4908 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4911 static ssize_t poison_store(struct kmem_cache *s,
4912 const char *buf, size_t length)
4914 if (any_slab_objects(s))
4917 s->flags &= ~SLAB_POISON;
4918 if (buf[0] == '1') {
4919 s->flags &= ~__CMPXCHG_DOUBLE;
4920 s->flags |= SLAB_POISON;
4922 calculate_sizes(s, -1);
4927 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4929 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4932 static ssize_t store_user_store(struct kmem_cache *s,
4933 const char *buf, size_t length)
4935 if (any_slab_objects(s))
4938 s->flags &= ~SLAB_STORE_USER;
4939 if (buf[0] == '1') {
4940 s->flags &= ~__CMPXCHG_DOUBLE;
4941 s->flags |= SLAB_STORE_USER;
4943 calculate_sizes(s, -1);
4946 SLAB_ATTR(store_user);
4948 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4953 static ssize_t validate_store(struct kmem_cache *s,
4954 const char *buf, size_t length)
4958 if (buf[0] == '1') {
4959 ret = validate_slab_cache(s);
4965 SLAB_ATTR(validate);
4967 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4969 if (!(s->flags & SLAB_STORE_USER))
4971 return list_locations(s, buf, TRACK_ALLOC);
4973 SLAB_ATTR_RO(alloc_calls);
4975 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4977 if (!(s->flags & SLAB_STORE_USER))
4979 return list_locations(s, buf, TRACK_FREE);
4981 SLAB_ATTR_RO(free_calls);
4982 #endif /* CONFIG_SLUB_DEBUG */
4984 #ifdef CONFIG_FAILSLAB
4985 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4987 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4990 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4993 if (s->refcount > 1)
4996 s->flags &= ~SLAB_FAILSLAB;
4998 s->flags |= SLAB_FAILSLAB;
5001 SLAB_ATTR(failslab);
5004 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5009 static ssize_t shrink_store(struct kmem_cache *s,
5010 const char *buf, size_t length)
5013 kmem_cache_shrink(s);
5021 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5023 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5026 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5027 const char *buf, size_t length)
5029 unsigned long ratio;
5032 err = kstrtoul(buf, 10, &ratio);
5037 s->remote_node_defrag_ratio = ratio * 10;
5041 SLAB_ATTR(remote_node_defrag_ratio);
5044 #ifdef CONFIG_SLUB_STATS
5045 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5047 unsigned long sum = 0;
5050 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5055 for_each_online_cpu(cpu) {
5056 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5062 len = sprintf(buf, "%lu", sum);
5065 for_each_online_cpu(cpu) {
5066 if (data[cpu] && len < PAGE_SIZE - 20)
5067 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5071 return len + sprintf(buf + len, "\n");
5074 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5078 for_each_online_cpu(cpu)
5079 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5082 #define STAT_ATTR(si, text) \
5083 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5085 return show_stat(s, buf, si); \
5087 static ssize_t text##_store(struct kmem_cache *s, \
5088 const char *buf, size_t length) \
5090 if (buf[0] != '0') \
5092 clear_stat(s, si); \
5097 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5098 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5099 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5100 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5101 STAT_ATTR(FREE_FROZEN, free_frozen);
5102 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5103 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5104 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5105 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5106 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5107 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5108 STAT_ATTR(FREE_SLAB, free_slab);
5109 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5110 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5111 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5112 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5113 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5114 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5115 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5116 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5117 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5118 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5119 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5120 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5121 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5122 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5125 static struct attribute *slab_attrs[] = {
5126 &slab_size_attr.attr,
5127 &object_size_attr.attr,
5128 &objs_per_slab_attr.attr,
5130 &min_partial_attr.attr,
5131 &cpu_partial_attr.attr,
5133 &objects_partial_attr.attr,
5135 &cpu_slabs_attr.attr,
5139 &hwcache_align_attr.attr,
5140 &reclaim_account_attr.attr,
5141 &destroy_by_rcu_attr.attr,
5143 &reserved_attr.attr,
5144 &slabs_cpu_partial_attr.attr,
5145 #ifdef CONFIG_SLUB_DEBUG
5146 &total_objects_attr.attr,
5148 &sanity_checks_attr.attr,
5150 &red_zone_attr.attr,
5152 &store_user_attr.attr,
5153 &validate_attr.attr,
5154 &alloc_calls_attr.attr,
5155 &free_calls_attr.attr,
5157 #ifdef CONFIG_ZONE_DMA
5158 &cache_dma_attr.attr,
5161 &remote_node_defrag_ratio_attr.attr,
5163 #ifdef CONFIG_SLUB_STATS
5164 &alloc_fastpath_attr.attr,
5165 &alloc_slowpath_attr.attr,
5166 &free_fastpath_attr.attr,
5167 &free_slowpath_attr.attr,
5168 &free_frozen_attr.attr,
5169 &free_add_partial_attr.attr,
5170 &free_remove_partial_attr.attr,
5171 &alloc_from_partial_attr.attr,
5172 &alloc_slab_attr.attr,
5173 &alloc_refill_attr.attr,
5174 &alloc_node_mismatch_attr.attr,
5175 &free_slab_attr.attr,
5176 &cpuslab_flush_attr.attr,
5177 &deactivate_full_attr.attr,
5178 &deactivate_empty_attr.attr,
5179 &deactivate_to_head_attr.attr,
5180 &deactivate_to_tail_attr.attr,
5181 &deactivate_remote_frees_attr.attr,
5182 &deactivate_bypass_attr.attr,
5183 &order_fallback_attr.attr,
5184 &cmpxchg_double_fail_attr.attr,
5185 &cmpxchg_double_cpu_fail_attr.attr,
5186 &cpu_partial_alloc_attr.attr,
5187 &cpu_partial_free_attr.attr,
5188 &cpu_partial_node_attr.attr,
5189 &cpu_partial_drain_attr.attr,
5191 #ifdef CONFIG_FAILSLAB
5192 &failslab_attr.attr,
5198 static struct attribute_group slab_attr_group = {
5199 .attrs = slab_attrs,
5202 static ssize_t slab_attr_show(struct kobject *kobj,
5203 struct attribute *attr,
5206 struct slab_attribute *attribute;
5207 struct kmem_cache *s;
5210 attribute = to_slab_attr(attr);
5213 if (!attribute->show)
5216 err = attribute->show(s, buf);
5221 static ssize_t slab_attr_store(struct kobject *kobj,
5222 struct attribute *attr,
5223 const char *buf, size_t len)
5225 struct slab_attribute *attribute;
5226 struct kmem_cache *s;
5229 attribute = to_slab_attr(attr);
5232 if (!attribute->store)
5235 err = attribute->store(s, buf, len);
5237 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5238 struct kmem_cache *c;
5240 mutex_lock(&slab_mutex);
5241 if (s->max_attr_size < len)
5242 s->max_attr_size = len;
5245 * This is a best effort propagation, so this function's return
5246 * value will be determined by the parent cache only. This is
5247 * basically because not all attributes will have a well
5248 * defined semantics for rollbacks - most of the actions will
5249 * have permanent effects.
5251 * Returning the error value of any of the children that fail
5252 * is not 100 % defined, in the sense that users seeing the
5253 * error code won't be able to know anything about the state of
5256 * Only returning the error code for the parent cache at least
5257 * has well defined semantics. The cache being written to
5258 * directly either failed or succeeded, in which case we loop
5259 * through the descendants with best-effort propagation.
5261 for_each_memcg_cache(c, s)
5262 attribute->store(c, buf, len);
5263 mutex_unlock(&slab_mutex);
5269 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5273 char *buffer = NULL;
5274 struct kmem_cache *root_cache;
5276 if (is_root_cache(s))
5279 root_cache = s->memcg_params.root_cache;
5282 * This mean this cache had no attribute written. Therefore, no point
5283 * in copying default values around
5285 if (!root_cache->max_attr_size)
5288 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5291 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5293 if (!attr || !attr->store || !attr->show)
5297 * It is really bad that we have to allocate here, so we will
5298 * do it only as a fallback. If we actually allocate, though,
5299 * we can just use the allocated buffer until the end.
5301 * Most of the slub attributes will tend to be very small in
5302 * size, but sysfs allows buffers up to a page, so they can
5303 * theoretically happen.
5307 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5310 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5311 if (WARN_ON(!buffer))
5316 attr->show(root_cache, buf);
5317 attr->store(s, buf, strlen(buf));
5321 free_page((unsigned long)buffer);
5325 static void kmem_cache_release(struct kobject *k)
5327 slab_kmem_cache_release(to_slab(k));
5330 static const struct sysfs_ops slab_sysfs_ops = {
5331 .show = slab_attr_show,
5332 .store = slab_attr_store,
5335 static struct kobj_type slab_ktype = {
5336 .sysfs_ops = &slab_sysfs_ops,
5337 .release = kmem_cache_release,
5340 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5342 struct kobj_type *ktype = get_ktype(kobj);
5344 if (ktype == &slab_ktype)
5349 static const struct kset_uevent_ops slab_uevent_ops = {
5350 .filter = uevent_filter,
5353 static struct kset *slab_kset;
5355 static inline struct kset *cache_kset(struct kmem_cache *s)
5358 if (!is_root_cache(s))
5359 return s->memcg_params.root_cache->memcg_kset;
5364 #define ID_STR_LENGTH 64
5366 /* Create a unique string id for a slab cache:
5368 * Format :[flags-]size
5370 static char *create_unique_id(struct kmem_cache *s)
5372 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5379 * First flags affecting slabcache operations. We will only
5380 * get here for aliasable slabs so we do not need to support
5381 * too many flags. The flags here must cover all flags that
5382 * are matched during merging to guarantee that the id is
5385 if (s->flags & SLAB_CACHE_DMA)
5387 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5389 if (s->flags & SLAB_DEBUG_FREE)
5391 if (!(s->flags & SLAB_NOTRACK))
5393 if (s->flags & SLAB_ACCOUNT)
5397 p += sprintf(p, "%07d", s->size);
5399 BUG_ON(p > name + ID_STR_LENGTH - 1);
5403 static int sysfs_slab_add(struct kmem_cache *s)
5407 int unmergeable = slab_unmergeable(s);
5411 * Slabcache can never be merged so we can use the name proper.
5412 * This is typically the case for debug situations. In that
5413 * case we can catch duplicate names easily.
5415 sysfs_remove_link(&slab_kset->kobj, s->name);
5419 * Create a unique name for the slab as a target
5422 name = create_unique_id(s);
5425 s->kobj.kset = cache_kset(s);
5426 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5430 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5435 if (is_root_cache(s)) {
5436 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5437 if (!s->memcg_kset) {
5444 kobject_uevent(&s->kobj, KOBJ_ADD);
5446 /* Setup first alias */
5447 sysfs_slab_alias(s, s->name);
5454 kobject_del(&s->kobj);
5458 void sysfs_slab_remove(struct kmem_cache *s)
5460 if (slab_state < FULL)
5462 * Sysfs has not been setup yet so no need to remove the
5468 kset_unregister(s->memcg_kset);
5470 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5471 kobject_del(&s->kobj);
5472 kobject_put(&s->kobj);
5476 * Need to buffer aliases during bootup until sysfs becomes
5477 * available lest we lose that information.
5479 struct saved_alias {
5480 struct kmem_cache *s;
5482 struct saved_alias *next;
5485 static struct saved_alias *alias_list;
5487 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5489 struct saved_alias *al;
5491 if (slab_state == FULL) {
5493 * If we have a leftover link then remove it.
5495 sysfs_remove_link(&slab_kset->kobj, name);
5496 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5499 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5505 al->next = alias_list;
5510 static int __init slab_sysfs_init(void)
5512 struct kmem_cache *s;
5515 mutex_lock(&slab_mutex);
5517 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5519 mutex_unlock(&slab_mutex);
5520 pr_err("Cannot register slab subsystem.\n");
5526 list_for_each_entry(s, &slab_caches, list) {
5527 err = sysfs_slab_add(s);
5529 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5533 while (alias_list) {
5534 struct saved_alias *al = alias_list;
5536 alias_list = alias_list->next;
5537 err = sysfs_slab_alias(al->s, al->name);
5539 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5544 mutex_unlock(&slab_mutex);
5549 __initcall(slab_sysfs_init);
5550 #endif /* CONFIG_SYSFS */
5553 * The /proc/slabinfo ABI
5555 #ifdef CONFIG_SLABINFO
5556 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5558 unsigned long nr_slabs = 0;
5559 unsigned long nr_objs = 0;
5560 unsigned long nr_free = 0;
5562 struct kmem_cache_node *n;
5564 for_each_kmem_cache_node(s, node, n) {
5565 nr_slabs += node_nr_slabs(n);
5566 nr_objs += node_nr_objs(n);
5567 nr_free += count_partial(n, count_free);
5570 sinfo->active_objs = nr_objs - nr_free;
5571 sinfo->num_objs = nr_objs;
5572 sinfo->active_slabs = nr_slabs;
5573 sinfo->num_slabs = nr_slabs;
5574 sinfo->objects_per_slab = oo_objects(s->oo);
5575 sinfo->cache_order = oo_order(s->oo);
5578 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5582 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5583 size_t count, loff_t *ppos)
5587 #endif /* CONFIG_SLABINFO */