2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 void *fixup_red_left(struct kmem_cache *s, void *p)
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
145 * Issues still to be resolved:
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 * - Variable sizing of the per node arrays
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 #define MIN_PARTIAL 5
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
169 #define MAX_PARTIAL 10
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193 /* Internal SLUB flags */
194 #define __OBJECT_POISON 0x80000000UL /* Poison object */
195 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
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);
256 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
260 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
262 *(void **)(object + s->offset) = fp;
265 /* Loop over all objects in a slab */
266 #define for_each_object(__p, __s, __addr, __objects) \
267 for (__p = fixup_red_left(__s, __addr); \
268 __p < (__addr) + (__objects) * (__s)->size; \
271 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
272 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
273 __idx <= __objects; \
274 __p += (__s)->size, __idx++)
276 /* Determine object index from a given position */
277 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
279 return (p - addr) / s->size;
282 static inline int order_objects(int order, unsigned long size, int reserved)
284 return ((PAGE_SIZE << order) - reserved) / size;
287 static inline struct kmem_cache_order_objects oo_make(int order,
288 unsigned long size, int reserved)
290 struct kmem_cache_order_objects x = {
291 (order << OO_SHIFT) + order_objects(order, size, reserved)
297 static inline int oo_order(struct kmem_cache_order_objects x)
299 return x.x >> OO_SHIFT;
302 static inline int oo_objects(struct kmem_cache_order_objects x)
304 return x.x & OO_MASK;
308 * Per slab locking using the pagelock
310 static __always_inline void slab_lock(struct page *page)
312 VM_BUG_ON_PAGE(PageTail(page), page);
313 bit_spin_lock(PG_locked, &page->flags);
316 static __always_inline void slab_unlock(struct page *page)
318 VM_BUG_ON_PAGE(PageTail(page), page);
319 __bit_spin_unlock(PG_locked, &page->flags);
322 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
325 tmp.counters = counters_new;
327 * page->counters can cover frozen/inuse/objects as well
328 * as page->_refcount. If we assign to ->counters directly
329 * we run the risk of losing updates to page->_refcount, so
330 * be careful and only assign to the fields we need.
332 page->frozen = tmp.frozen;
333 page->inuse = tmp.inuse;
334 page->objects = tmp.objects;
337 /* Interrupts must be disabled (for the fallback code to work right) */
338 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
339 void *freelist_old, unsigned long counters_old,
340 void *freelist_new, unsigned long counters_new,
343 VM_BUG_ON(!irqs_disabled());
344 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
345 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
346 if (s->flags & __CMPXCHG_DOUBLE) {
347 if (cmpxchg_double(&page->freelist, &page->counters,
348 freelist_old, counters_old,
349 freelist_new, counters_new))
355 if (page->freelist == freelist_old &&
356 page->counters == counters_old) {
357 page->freelist = freelist_new;
358 set_page_slub_counters(page, counters_new);
366 stat(s, CMPXCHG_DOUBLE_FAIL);
368 #ifdef SLUB_DEBUG_CMPXCHG
369 pr_info("%s %s: cmpxchg double redo ", n, s->name);
375 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
380 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
381 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
382 if (s->flags & __CMPXCHG_DOUBLE) {
383 if (cmpxchg_double(&page->freelist, &page->counters,
384 freelist_old, counters_old,
385 freelist_new, counters_new))
392 local_irq_save(flags);
394 if (page->freelist == freelist_old &&
395 page->counters == counters_old) {
396 page->freelist = freelist_new;
397 set_page_slub_counters(page, counters_new);
399 local_irq_restore(flags);
403 local_irq_restore(flags);
407 stat(s, CMPXCHG_DOUBLE_FAIL);
409 #ifdef SLUB_DEBUG_CMPXCHG
410 pr_info("%s %s: cmpxchg double redo ", n, s->name);
416 #ifdef CONFIG_SLUB_DEBUG
418 * Determine a map of object in use on a page.
420 * Node listlock must be held to guarantee that the page does
421 * not vanish from under us.
423 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
426 void *addr = page_address(page);
428 for (p = page->freelist; p; p = get_freepointer(s, p))
429 set_bit(slab_index(p, s, addr), map);
432 static inline int size_from_object(struct kmem_cache *s)
434 if (s->flags & SLAB_RED_ZONE)
435 return s->size - s->red_left_pad;
440 static inline void *restore_red_left(struct kmem_cache *s, void *p)
442 if (s->flags & SLAB_RED_ZONE)
443 p -= s->red_left_pad;
451 #if defined(CONFIG_SLUB_DEBUG_ON)
452 static int slub_debug = DEBUG_DEFAULT_FLAGS;
454 static int slub_debug;
457 static char *slub_debug_slabs;
458 static int disable_higher_order_debug;
461 * slub is about to manipulate internal object metadata. This memory lies
462 * outside the range of the allocated object, so accessing it would normally
463 * be reported by kasan as a bounds error. metadata_access_enable() is used
464 * to tell kasan that these accesses are OK.
466 static inline void metadata_access_enable(void)
468 kasan_disable_current();
471 static inline void metadata_access_disable(void)
473 kasan_enable_current();
480 /* Verify that a pointer has an address that is valid within a slab page */
481 static inline int check_valid_pointer(struct kmem_cache *s,
482 struct page *page, void *object)
489 base = page_address(page);
490 object = restore_red_left(s, object);
491 if (object < base || object >= base + page->objects * s->size ||
492 (object - base) % s->size) {
499 static void print_section(char *text, u8 *addr, unsigned int length)
501 metadata_access_enable();
502 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
504 metadata_access_disable();
507 static struct track *get_track(struct kmem_cache *s, void *object,
508 enum track_item alloc)
513 p = object + s->offset + sizeof(void *);
515 p = object + s->inuse;
520 static void set_track(struct kmem_cache *s, void *object,
521 enum track_item alloc, unsigned long addr)
523 struct track *p = get_track(s, object, alloc);
526 #ifdef CONFIG_STACKTRACE
527 struct stack_trace trace;
530 trace.nr_entries = 0;
531 trace.max_entries = TRACK_ADDRS_COUNT;
532 trace.entries = p->addrs;
534 metadata_access_enable();
535 save_stack_trace(&trace);
536 metadata_access_disable();
538 /* See rant in lockdep.c */
539 if (trace.nr_entries != 0 &&
540 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
543 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
547 p->cpu = smp_processor_id();
548 p->pid = current->pid;
551 memset(p, 0, sizeof(struct track));
554 static void init_tracking(struct kmem_cache *s, void *object)
556 if (!(s->flags & SLAB_STORE_USER))
559 set_track(s, object, TRACK_FREE, 0UL);
560 set_track(s, object, TRACK_ALLOC, 0UL);
563 static void print_track(const char *s, struct track *t)
568 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
569 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
570 #ifdef CONFIG_STACKTRACE
573 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
575 pr_err("\t%pS\n", (void *)t->addrs[i]);
582 static void print_tracking(struct kmem_cache *s, void *object)
584 if (!(s->flags & SLAB_STORE_USER))
587 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
588 print_track("Freed", get_track(s, object, TRACK_FREE));
591 static void print_page_info(struct page *page)
593 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
594 page, page->objects, page->inuse, page->freelist, page->flags);
598 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
600 struct va_format vaf;
606 pr_err("=============================================================================\n");
607 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
608 pr_err("-----------------------------------------------------------------------------\n\n");
610 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
614 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
616 struct va_format vaf;
622 pr_err("FIX %s: %pV\n", s->name, &vaf);
626 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
628 unsigned int off; /* Offset of last byte */
629 u8 *addr = page_address(page);
631 print_tracking(s, p);
633 print_page_info(page);
635 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
636 p, p - addr, get_freepointer(s, p));
638 if (s->flags & SLAB_RED_ZONE)
639 print_section("Redzone ", p - s->red_left_pad, s->red_left_pad);
640 else if (p > addr + 16)
641 print_section("Bytes b4 ", p - 16, 16);
643 print_section("Object ", p, min_t(unsigned long, s->object_size,
645 if (s->flags & SLAB_RED_ZONE)
646 print_section("Redzone ", p + s->object_size,
647 s->inuse - s->object_size);
650 off = s->offset + sizeof(void *);
654 if (s->flags & SLAB_STORE_USER)
655 off += 2 * sizeof(struct track);
657 off += kasan_metadata_size(s);
659 if (off != size_from_object(s))
660 /* Beginning of the filler is the free pointer */
661 print_section("Padding ", p + off, size_from_object(s) - off);
666 void object_err(struct kmem_cache *s, struct page *page,
667 u8 *object, char *reason)
669 slab_bug(s, "%s", reason);
670 print_trailer(s, page, object);
673 static void slab_err(struct kmem_cache *s, struct page *page,
674 const char *fmt, ...)
680 vsnprintf(buf, sizeof(buf), fmt, args);
682 slab_bug(s, "%s", buf);
683 print_page_info(page);
687 static void init_object(struct kmem_cache *s, void *object, u8 val)
691 if (s->flags & SLAB_RED_ZONE)
692 memset(p - s->red_left_pad, val, s->red_left_pad);
694 if (s->flags & __OBJECT_POISON) {
695 memset(p, POISON_FREE, s->object_size - 1);
696 p[s->object_size - 1] = POISON_END;
699 if (s->flags & SLAB_RED_ZONE)
700 memset(p + s->object_size, val, s->inuse - s->object_size);
703 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
704 void *from, void *to)
706 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
707 memset(from, data, to - from);
710 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
711 u8 *object, char *what,
712 u8 *start, unsigned int value, unsigned int bytes)
717 metadata_access_enable();
718 fault = memchr_inv(start, value, bytes);
719 metadata_access_disable();
724 while (end > fault && end[-1] == value)
727 slab_bug(s, "%s overwritten", what);
728 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
729 fault, end - 1, fault[0], value);
730 print_trailer(s, page, object);
732 restore_bytes(s, what, value, fault, end);
740 * Bytes of the object to be managed.
741 * If the freepointer may overlay the object then the free
742 * pointer is the first word of the object.
744 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
747 * object + s->object_size
748 * Padding to reach word boundary. This is also used for Redzoning.
749 * Padding is extended by another word if Redzoning is enabled and
750 * object_size == inuse.
752 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
753 * 0xcc (RED_ACTIVE) for objects in use.
756 * Meta data starts here.
758 * A. Free pointer (if we cannot overwrite object on free)
759 * B. Tracking data for SLAB_STORE_USER
760 * C. Padding to reach required alignment boundary or at mininum
761 * one word if debugging is on to be able to detect writes
762 * before the word boundary.
764 * Padding is done using 0x5a (POISON_INUSE)
767 * Nothing is used beyond s->size.
769 * If slabcaches are merged then the object_size and inuse boundaries are mostly
770 * ignored. And therefore no slab options that rely on these boundaries
771 * may be used with merged slabcaches.
774 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
776 unsigned long off = s->inuse; /* The end of info */
779 /* Freepointer is placed after the object. */
780 off += sizeof(void *);
782 if (s->flags & SLAB_STORE_USER)
783 /* We also have user information there */
784 off += 2 * sizeof(struct track);
786 off += kasan_metadata_size(s);
788 if (size_from_object(s) == off)
791 return check_bytes_and_report(s, page, p, "Object padding",
792 p + off, POISON_INUSE, size_from_object(s) - off);
795 /* Check the pad bytes at the end of a slab page */
796 static int slab_pad_check(struct kmem_cache *s, struct page *page)
804 if (!(s->flags & SLAB_POISON))
807 start = page_address(page);
808 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
809 end = start + length;
810 remainder = length % s->size;
814 metadata_access_enable();
815 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
816 metadata_access_disable();
819 while (end > fault && end[-1] == POISON_INUSE)
822 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
823 print_section("Padding ", end - remainder, remainder);
825 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
829 static int check_object(struct kmem_cache *s, struct page *page,
830 void *object, u8 val)
833 u8 *endobject = object + s->object_size;
835 if (s->flags & SLAB_RED_ZONE) {
836 if (!check_bytes_and_report(s, page, object, "Redzone",
837 object - s->red_left_pad, val, s->red_left_pad))
840 if (!check_bytes_and_report(s, page, object, "Redzone",
841 endobject, val, s->inuse - s->object_size))
844 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
845 check_bytes_and_report(s, page, p, "Alignment padding",
846 endobject, POISON_INUSE,
847 s->inuse - s->object_size);
851 if (s->flags & SLAB_POISON) {
852 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
853 (!check_bytes_and_report(s, page, p, "Poison", p,
854 POISON_FREE, s->object_size - 1) ||
855 !check_bytes_and_report(s, page, p, "Poison",
856 p + s->object_size - 1, POISON_END, 1)))
859 * check_pad_bytes cleans up on its own.
861 check_pad_bytes(s, page, p);
864 if (!s->offset && val == SLUB_RED_ACTIVE)
866 * Object and freepointer overlap. Cannot check
867 * freepointer while object is allocated.
871 /* Check free pointer validity */
872 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
873 object_err(s, page, p, "Freepointer corrupt");
875 * No choice but to zap it and thus lose the remainder
876 * of the free objects in this slab. May cause
877 * another error because the object count is now wrong.
879 set_freepointer(s, p, NULL);
885 static int check_slab(struct kmem_cache *s, struct page *page)
889 VM_BUG_ON(!irqs_disabled());
891 if (!PageSlab(page)) {
892 slab_err(s, page, "Not a valid slab page");
896 maxobj = order_objects(compound_order(page), s->size, s->reserved);
897 if (page->objects > maxobj) {
898 slab_err(s, page, "objects %u > max %u",
899 page->objects, maxobj);
902 if (page->inuse > page->objects) {
903 slab_err(s, page, "inuse %u > max %u",
904 page->inuse, page->objects);
907 /* Slab_pad_check fixes things up after itself */
908 slab_pad_check(s, page);
913 * Determine if a certain object on a page is on the freelist. Must hold the
914 * slab lock to guarantee that the chains are in a consistent state.
916 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
924 while (fp && nr <= page->objects) {
927 if (!check_valid_pointer(s, page, fp)) {
929 object_err(s, page, object,
930 "Freechain corrupt");
931 set_freepointer(s, object, NULL);
933 slab_err(s, page, "Freepointer corrupt");
934 page->freelist = NULL;
935 page->inuse = page->objects;
936 slab_fix(s, "Freelist cleared");
942 fp = get_freepointer(s, object);
946 max_objects = order_objects(compound_order(page), s->size, s->reserved);
947 if (max_objects > MAX_OBJS_PER_PAGE)
948 max_objects = MAX_OBJS_PER_PAGE;
950 if (page->objects != max_objects) {
951 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
952 page->objects, max_objects);
953 page->objects = max_objects;
954 slab_fix(s, "Number of objects adjusted.");
956 if (page->inuse != page->objects - nr) {
957 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
958 page->inuse, page->objects - nr);
959 page->inuse = page->objects - nr;
960 slab_fix(s, "Object count adjusted.");
962 return search == NULL;
965 static void trace(struct kmem_cache *s, struct page *page, void *object,
968 if (s->flags & SLAB_TRACE) {
969 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
971 alloc ? "alloc" : "free",
976 print_section("Object ", (void *)object,
984 * Tracking of fully allocated slabs for debugging purposes.
986 static void add_full(struct kmem_cache *s,
987 struct kmem_cache_node *n, struct page *page)
989 if (!(s->flags & SLAB_STORE_USER))
992 lockdep_assert_held(&n->list_lock);
993 list_add(&page->lru, &n->full);
996 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
998 if (!(s->flags & SLAB_STORE_USER))
1001 lockdep_assert_held(&n->list_lock);
1002 list_del(&page->lru);
1005 /* Tracking of the number of slabs for debugging purposes */
1006 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1008 struct kmem_cache_node *n = get_node(s, node);
1010 return atomic_long_read(&n->nr_slabs);
1013 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1015 return atomic_long_read(&n->nr_slabs);
1018 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1020 struct kmem_cache_node *n = get_node(s, node);
1023 * May be called early in order to allocate a slab for the
1024 * kmem_cache_node structure. Solve the chicken-egg
1025 * dilemma by deferring the increment of the count during
1026 * bootstrap (see early_kmem_cache_node_alloc).
1029 atomic_long_inc(&n->nr_slabs);
1030 atomic_long_add(objects, &n->total_objects);
1033 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1035 struct kmem_cache_node *n = get_node(s, node);
1037 atomic_long_dec(&n->nr_slabs);
1038 atomic_long_sub(objects, &n->total_objects);
1041 /* Object debug checks for alloc/free paths */
1042 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1045 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1048 init_object(s, object, SLUB_RED_INACTIVE);
1049 init_tracking(s, object);
1052 static inline int alloc_consistency_checks(struct kmem_cache *s,
1054 void *object, unsigned long addr)
1056 if (!check_slab(s, page))
1059 if (!check_valid_pointer(s, page, object)) {
1060 object_err(s, page, object, "Freelist Pointer check fails");
1064 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1070 static noinline int alloc_debug_processing(struct kmem_cache *s,
1072 void *object, unsigned long addr)
1074 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1075 if (!alloc_consistency_checks(s, page, object, addr))
1079 /* Success perform special debug activities for allocs */
1080 if (s->flags & SLAB_STORE_USER)
1081 set_track(s, object, TRACK_ALLOC, addr);
1082 trace(s, page, object, 1);
1083 init_object(s, object, SLUB_RED_ACTIVE);
1087 if (PageSlab(page)) {
1089 * If this is a slab page then lets do the best we can
1090 * to avoid issues in the future. Marking all objects
1091 * as used avoids touching the remaining objects.
1093 slab_fix(s, "Marking all objects used");
1094 page->inuse = page->objects;
1095 page->freelist = NULL;
1100 static inline int free_consistency_checks(struct kmem_cache *s,
1101 struct page *page, void *object, unsigned long addr)
1103 if (!check_valid_pointer(s, page, object)) {
1104 slab_err(s, page, "Invalid object pointer 0x%p", object);
1108 if (on_freelist(s, page, object)) {
1109 object_err(s, page, object, "Object already free");
1113 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1116 if (unlikely(s != page->slab_cache)) {
1117 if (!PageSlab(page)) {
1118 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1120 } else if (!page->slab_cache) {
1121 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1125 object_err(s, page, object,
1126 "page slab pointer corrupt.");
1132 /* Supports checking bulk free of a constructed freelist */
1133 static noinline int free_debug_processing(
1134 struct kmem_cache *s, struct page *page,
1135 void *head, void *tail, int bulk_cnt,
1138 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1139 void *object = head;
1141 unsigned long uninitialized_var(flags);
1144 raw_spin_lock_irqsave(&n->list_lock, flags);
1147 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1148 if (!check_slab(s, page))
1155 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1156 if (!free_consistency_checks(s, page, object, addr))
1160 if (s->flags & SLAB_STORE_USER)
1161 set_track(s, object, TRACK_FREE, addr);
1162 trace(s, page, object, 0);
1163 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1164 init_object(s, object, SLUB_RED_INACTIVE);
1166 /* Reached end of constructed freelist yet? */
1167 if (object != tail) {
1168 object = get_freepointer(s, object);
1174 if (cnt != bulk_cnt)
1175 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1179 raw_spin_unlock_irqrestore(&n->list_lock, flags);
1181 slab_fix(s, "Object at 0x%p not freed", object);
1185 static int __init setup_slub_debug(char *str)
1187 slub_debug = DEBUG_DEFAULT_FLAGS;
1188 if (*str++ != '=' || !*str)
1190 * No options specified. Switch on full debugging.
1196 * No options but restriction on slabs. This means full
1197 * debugging for slabs matching a pattern.
1204 * Switch off all debugging measures.
1209 * Determine which debug features should be switched on
1211 for (; *str && *str != ','; str++) {
1212 switch (tolower(*str)) {
1214 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1217 slub_debug |= SLAB_RED_ZONE;
1220 slub_debug |= SLAB_POISON;
1223 slub_debug |= SLAB_STORE_USER;
1226 slub_debug |= SLAB_TRACE;
1229 slub_debug |= SLAB_FAILSLAB;
1233 * Avoid enabling debugging on caches if its minimum
1234 * order would increase as a result.
1236 disable_higher_order_debug = 1;
1239 pr_err("slub_debug option '%c' unknown. skipped\n",
1246 slub_debug_slabs = str + 1;
1251 __setup("slub_debug", setup_slub_debug);
1253 unsigned long kmem_cache_flags(unsigned long object_size,
1254 unsigned long flags, const char *name,
1255 void (*ctor)(void *))
1258 * Enable debugging if selected on the kernel commandline.
1260 if (slub_debug && (!slub_debug_slabs || (name &&
1261 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1262 flags |= slub_debug;
1266 #else /* !CONFIG_SLUB_DEBUG */
1267 static inline void setup_object_debug(struct kmem_cache *s,
1268 struct page *page, void *object) {}
1270 static inline int alloc_debug_processing(struct kmem_cache *s,
1271 struct page *page, void *object, unsigned long addr) { return 0; }
1273 static inline int free_debug_processing(
1274 struct kmem_cache *s, struct page *page,
1275 void *head, void *tail, int bulk_cnt,
1276 unsigned long addr) { return 0; }
1278 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1280 static inline int check_object(struct kmem_cache *s, struct page *page,
1281 void *object, u8 val) { return 1; }
1282 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1283 struct page *page) {}
1284 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1285 struct page *page) {}
1286 unsigned long kmem_cache_flags(unsigned long object_size,
1287 unsigned long flags, const char *name,
1288 void (*ctor)(void *))
1292 #define slub_debug 0
1294 #define disable_higher_order_debug 0
1296 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1298 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1300 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1302 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1305 #endif /* CONFIG_SLUB_DEBUG */
1307 struct slub_free_list {
1308 raw_spinlock_t lock;
1309 struct list_head list;
1311 static DEFINE_PER_CPU(struct slub_free_list, slub_free_list);
1314 * Hooks for other subsystems that check memory allocations. In a typical
1315 * production configuration these hooks all should produce no code at all.
1317 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1319 kmemleak_alloc(ptr, size, 1, flags);
1320 kasan_kmalloc_large(ptr, size, flags);
1323 static inline void kfree_hook(const void *x)
1326 kasan_kfree_large(x);
1329 static inline void *slab_free_hook(struct kmem_cache *s, void *x)
1333 kmemleak_free_recursive(x, s->flags);
1336 * Trouble is that we may no longer disable interrupts in the fast path
1337 * So in order to make the debug calls that expect irqs to be
1338 * disabled we need to disable interrupts temporarily.
1340 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1342 unsigned long flags;
1344 local_irq_save(flags);
1345 kmemcheck_slab_free(s, x, s->object_size);
1346 debug_check_no_locks_freed(x, s->object_size);
1347 local_irq_restore(flags);
1350 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1351 debug_check_no_obj_freed(x, s->object_size);
1353 freeptr = get_freepointer(s, x);
1355 * kasan_slab_free() may put x into memory quarantine, delaying its
1356 * reuse. In this case the object's freelist pointer is changed.
1358 kasan_slab_free(s, x);
1362 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1363 void *head, void *tail)
1366 * Compiler cannot detect this function can be removed if slab_free_hook()
1367 * evaluates to nothing. Thus, catch all relevant config debug options here.
1369 #if defined(CONFIG_KMEMCHECK) || \
1370 defined(CONFIG_LOCKDEP) || \
1371 defined(CONFIG_DEBUG_KMEMLEAK) || \
1372 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1373 defined(CONFIG_KASAN)
1375 void *object = head;
1376 void *tail_obj = tail ? : head;
1380 freeptr = slab_free_hook(s, object);
1381 } while ((object != tail_obj) && (object = freeptr));
1385 static void setup_object(struct kmem_cache *s, struct page *page,
1388 setup_object_debug(s, page, object);
1389 kasan_init_slab_obj(s, object);
1390 if (unlikely(s->ctor)) {
1391 kasan_unpoison_object_data(s, object);
1393 kasan_poison_object_data(s, object);
1398 * Slab allocation and freeing
1400 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1401 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1404 int order = oo_order(oo);
1406 flags |= __GFP_NOTRACK;
1408 if (node == NUMA_NO_NODE)
1409 page = alloc_pages(flags, order);
1411 page = __alloc_pages_node(node, flags, order);
1413 if (page && memcg_charge_slab(page, flags, order, s)) {
1414 __free_pages(page, order);
1421 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1422 /* Pre-initialize the random sequence cache */
1423 static int init_cache_random_seq(struct kmem_cache *s)
1426 unsigned long i, count = oo_objects(s->oo);
1428 err = cache_random_seq_create(s, count, GFP_KERNEL);
1430 pr_err("SLUB: Unable to initialize free list for %s\n",
1435 /* Transform to an offset on the set of pages */
1436 if (s->random_seq) {
1437 for (i = 0; i < count; i++)
1438 s->random_seq[i] *= s->size;
1443 /* Initialize each random sequence freelist per cache */
1444 static void __init init_freelist_randomization(void)
1446 struct kmem_cache *s;
1448 mutex_lock(&slab_mutex);
1450 list_for_each_entry(s, &slab_caches, list)
1451 init_cache_random_seq(s);
1453 mutex_unlock(&slab_mutex);
1456 /* Get the next entry on the pre-computed freelist randomized */
1457 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1458 unsigned long *pos, void *start,
1459 unsigned long page_limit,
1460 unsigned long freelist_count)
1465 * If the target page allocation failed, the number of objects on the
1466 * page might be smaller than the usual size defined by the cache.
1469 idx = s->random_seq[*pos];
1471 if (*pos >= freelist_count)
1473 } while (unlikely(idx >= page_limit));
1475 return (char *)start + idx;
1478 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1479 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1484 unsigned long idx, pos, page_limit, freelist_count;
1486 if (page->objects < 2 || !s->random_seq)
1489 freelist_count = oo_objects(s->oo);
1490 pos = get_random_int() % freelist_count;
1492 page_limit = page->objects * s->size;
1493 start = fixup_red_left(s, page_address(page));
1495 /* First entry is used as the base of the freelist */
1496 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1498 page->freelist = cur;
1500 for (idx = 1; idx < page->objects; idx++) {
1501 setup_object(s, page, cur);
1502 next = next_freelist_entry(s, page, &pos, start, page_limit,
1504 set_freepointer(s, cur, next);
1507 setup_object(s, page, cur);
1508 set_freepointer(s, cur, NULL);
1513 static inline int init_cache_random_seq(struct kmem_cache *s)
1517 static inline void init_freelist_randomization(void) { }
1518 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1522 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1524 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1527 struct kmem_cache_order_objects oo = s->oo;
1532 bool enableirqs = false;
1534 flags &= gfp_allowed_mask;
1536 if (gfpflags_allow_blocking(flags))
1538 #ifdef CONFIG_PREEMPT_RT_FULL
1539 if (system_state == SYSTEM_RUNNING)
1545 flags |= s->allocflags;
1548 * Let the initial higher-order allocation fail under memory pressure
1549 * so we fall-back to the minimum order allocation.
1551 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1552 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1553 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1555 page = alloc_slab_page(s, alloc_gfp, node, oo);
1556 if (unlikely(!page)) {
1560 * Allocation may have failed due to fragmentation.
1561 * Try a lower order alloc if possible
1563 page = alloc_slab_page(s, alloc_gfp, node, oo);
1564 if (unlikely(!page))
1566 stat(s, ORDER_FALLBACK);
1569 if (kmemcheck_enabled &&
1570 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1571 int pages = 1 << oo_order(oo);
1573 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1576 * Objects from caches that have a constructor don't get
1577 * cleared when they're allocated, so we need to do it here.
1580 kmemcheck_mark_uninitialized_pages(page, pages);
1582 kmemcheck_mark_unallocated_pages(page, pages);
1585 page->objects = oo_objects(oo);
1587 order = compound_order(page);
1588 page->slab_cache = s;
1589 __SetPageSlab(page);
1590 if (page_is_pfmemalloc(page))
1591 SetPageSlabPfmemalloc(page);
1593 start = page_address(page);
1595 if (unlikely(s->flags & SLAB_POISON))
1596 memset(start, POISON_INUSE, PAGE_SIZE << order);
1598 kasan_poison_slab(page);
1600 shuffle = shuffle_freelist(s, page);
1603 for_each_object_idx(p, idx, s, start, page->objects) {
1604 setup_object(s, page, p);
1605 if (likely(idx < page->objects))
1606 set_freepointer(s, p, p + s->size);
1608 set_freepointer(s, p, NULL);
1610 page->freelist = fixup_red_left(s, start);
1613 page->inuse = page->objects;
1618 local_irq_disable();
1622 mod_zone_page_state(page_zone(page),
1623 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1624 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1627 inc_slabs_node(s, page_to_nid(page), page->objects);
1632 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1634 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1635 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1636 flags &= ~GFP_SLAB_BUG_MASK;
1637 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1638 invalid_mask, &invalid_mask, flags, &flags);
1641 return allocate_slab(s,
1642 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1645 static void __free_slab(struct kmem_cache *s, struct page *page)
1647 int order = compound_order(page);
1648 int pages = 1 << order;
1650 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1653 slab_pad_check(s, page);
1654 for_each_object(p, s, page_address(page),
1656 check_object(s, page, p, SLUB_RED_INACTIVE);
1659 kmemcheck_free_shadow(page, compound_order(page));
1661 mod_zone_page_state(page_zone(page),
1662 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1663 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1666 __ClearPageSlabPfmemalloc(page);
1667 __ClearPageSlab(page);
1669 page_mapcount_reset(page);
1670 if (current->reclaim_state)
1671 current->reclaim_state->reclaimed_slab += pages;
1672 memcg_uncharge_slab(page, order, s);
1673 __free_pages(page, order);
1676 static void free_delayed(struct list_head *h)
1678 while(!list_empty(h)) {
1679 struct page *page = list_first_entry(h, struct page, lru);
1681 list_del(&page->lru);
1682 __free_slab(page->slab_cache, page);
1686 #define need_reserve_slab_rcu \
1687 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1689 static void rcu_free_slab(struct rcu_head *h)
1693 if (need_reserve_slab_rcu)
1694 page = virt_to_head_page(h);
1696 page = container_of((struct list_head *)h, struct page, lru);
1698 __free_slab(page->slab_cache, page);
1701 static void free_slab(struct kmem_cache *s, struct page *page)
1703 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1704 struct rcu_head *head;
1706 if (need_reserve_slab_rcu) {
1707 int order = compound_order(page);
1708 int offset = (PAGE_SIZE << order) - s->reserved;
1710 VM_BUG_ON(s->reserved != sizeof(*head));
1711 head = page_address(page) + offset;
1713 head = &page->rcu_head;
1716 call_rcu(head, rcu_free_slab);
1717 } else if (irqs_disabled()) {
1718 struct slub_free_list *f = this_cpu_ptr(&slub_free_list);
1720 raw_spin_lock(&f->lock);
1721 list_add(&page->lru, &f->list);
1722 raw_spin_unlock(&f->lock);
1724 __free_slab(s, page);
1727 static void discard_slab(struct kmem_cache *s, struct page *page)
1729 dec_slabs_node(s, page_to_nid(page), page->objects);
1734 * Management of partially allocated slabs.
1737 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1740 if (tail == DEACTIVATE_TO_TAIL)
1741 list_add_tail(&page->lru, &n->partial);
1743 list_add(&page->lru, &n->partial);
1746 static inline void add_partial(struct kmem_cache_node *n,
1747 struct page *page, int tail)
1749 lockdep_assert_held(&n->list_lock);
1750 __add_partial(n, page, tail);
1753 static inline void remove_partial(struct kmem_cache_node *n,
1756 lockdep_assert_held(&n->list_lock);
1757 list_del(&page->lru);
1762 * Remove slab from the partial list, freeze it and
1763 * return the pointer to the freelist.
1765 * Returns a list of objects or NULL if it fails.
1767 static inline void *acquire_slab(struct kmem_cache *s,
1768 struct kmem_cache_node *n, struct page *page,
1769 int mode, int *objects)
1772 unsigned long counters;
1775 lockdep_assert_held(&n->list_lock);
1778 * Zap the freelist and set the frozen bit.
1779 * The old freelist is the list of objects for the
1780 * per cpu allocation list.
1782 freelist = page->freelist;
1783 counters = page->counters;
1784 new.counters = counters;
1785 *objects = new.objects - new.inuse;
1787 new.inuse = page->objects;
1788 new.freelist = NULL;
1790 new.freelist = freelist;
1793 VM_BUG_ON(new.frozen);
1796 if (!__cmpxchg_double_slab(s, page,
1798 new.freelist, new.counters,
1802 remove_partial(n, page);
1807 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1808 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1811 * Try to allocate a partial slab from a specific node.
1813 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1814 struct kmem_cache_cpu *c, gfp_t flags)
1816 struct page *page, *page2;
1817 void *object = NULL;
1822 * Racy check. If we mistakenly see no partial slabs then we
1823 * just allocate an empty slab. If we mistakenly try to get a
1824 * partial slab and there is none available then get_partials()
1827 if (!n || !n->nr_partial)
1830 raw_spin_lock(&n->list_lock);
1831 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1834 if (!pfmemalloc_match(page, flags))
1837 t = acquire_slab(s, n, page, object == NULL, &objects);
1841 available += objects;
1844 stat(s, ALLOC_FROM_PARTIAL);
1847 put_cpu_partial(s, page, 0);
1848 stat(s, CPU_PARTIAL_NODE);
1850 if (!kmem_cache_has_cpu_partial(s)
1851 || available > s->cpu_partial / 2)
1855 raw_spin_unlock(&n->list_lock);
1860 * Get a page from somewhere. Search in increasing NUMA distances.
1862 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1863 struct kmem_cache_cpu *c)
1866 struct zonelist *zonelist;
1869 enum zone_type high_zoneidx = gfp_zone(flags);
1871 unsigned int cpuset_mems_cookie;
1874 * The defrag ratio allows a configuration of the tradeoffs between
1875 * inter node defragmentation and node local allocations. A lower
1876 * defrag_ratio increases the tendency to do local allocations
1877 * instead of attempting to obtain partial slabs from other nodes.
1879 * If the defrag_ratio is set to 0 then kmalloc() always
1880 * returns node local objects. If the ratio is higher then kmalloc()
1881 * may return off node objects because partial slabs are obtained
1882 * from other nodes and filled up.
1884 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1885 * (which makes defrag_ratio = 1000) then every (well almost)
1886 * allocation will first attempt to defrag slab caches on other nodes.
1887 * This means scanning over all nodes to look for partial slabs which
1888 * may be expensive if we do it every time we are trying to find a slab
1889 * with available objects.
1891 if (!s->remote_node_defrag_ratio ||
1892 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1896 cpuset_mems_cookie = read_mems_allowed_begin();
1897 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1898 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1899 struct kmem_cache_node *n;
1901 n = get_node(s, zone_to_nid(zone));
1903 if (n && cpuset_zone_allowed(zone, flags) &&
1904 n->nr_partial > s->min_partial) {
1905 object = get_partial_node(s, n, c, flags);
1908 * Don't check read_mems_allowed_retry()
1909 * here - if mems_allowed was updated in
1910 * parallel, that was a harmless race
1911 * between allocation and the cpuset
1918 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1924 * Get a partial page, lock it and return it.
1926 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1927 struct kmem_cache_cpu *c)
1930 int searchnode = node;
1932 if (node == NUMA_NO_NODE)
1933 searchnode = numa_mem_id();
1934 else if (!node_present_pages(node))
1935 searchnode = node_to_mem_node(node);
1937 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1938 if (object || node != NUMA_NO_NODE)
1941 return get_any_partial(s, flags, c);
1944 #ifdef CONFIG_PREEMPT
1946 * Calculate the next globally unique transaction for disambiguiation
1947 * during cmpxchg. The transactions start with the cpu number and are then
1948 * incremented by CONFIG_NR_CPUS.
1950 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1953 * No preemption supported therefore also no need to check for
1959 static inline unsigned long next_tid(unsigned long tid)
1961 return tid + TID_STEP;
1964 static inline unsigned int tid_to_cpu(unsigned long tid)
1966 return tid % TID_STEP;
1969 static inline unsigned long tid_to_event(unsigned long tid)
1971 return tid / TID_STEP;
1974 static inline unsigned int init_tid(int cpu)
1979 static inline void note_cmpxchg_failure(const char *n,
1980 const struct kmem_cache *s, unsigned long tid)
1982 #ifdef SLUB_DEBUG_CMPXCHG
1983 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1985 pr_info("%s %s: cmpxchg redo ", n, s->name);
1987 #ifdef CONFIG_PREEMPT
1988 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1989 pr_warn("due to cpu change %d -> %d\n",
1990 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1993 if (tid_to_event(tid) != tid_to_event(actual_tid))
1994 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1995 tid_to_event(tid), tid_to_event(actual_tid));
1997 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1998 actual_tid, tid, next_tid(tid));
2000 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2003 static void init_kmem_cache_cpus(struct kmem_cache *s)
2007 for_each_possible_cpu(cpu)
2008 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2012 * Remove the cpu slab
2014 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2017 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2018 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2020 enum slab_modes l = M_NONE, m = M_NONE;
2022 int tail = DEACTIVATE_TO_HEAD;
2026 if (page->freelist) {
2027 stat(s, DEACTIVATE_REMOTE_FREES);
2028 tail = DEACTIVATE_TO_TAIL;
2032 * Stage one: Free all available per cpu objects back
2033 * to the page freelist while it is still frozen. Leave the
2036 * There is no need to take the list->lock because the page
2039 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2041 unsigned long counters;
2044 prior = page->freelist;
2045 counters = page->counters;
2046 set_freepointer(s, freelist, prior);
2047 new.counters = counters;
2049 VM_BUG_ON(!new.frozen);
2051 } while (!__cmpxchg_double_slab(s, page,
2053 freelist, new.counters,
2054 "drain percpu freelist"));
2056 freelist = nextfree;
2060 * Stage two: Ensure that the page is unfrozen while the
2061 * list presence reflects the actual number of objects
2064 * We setup the list membership and then perform a cmpxchg
2065 * with the count. If there is a mismatch then the page
2066 * is not unfrozen but the page is on the wrong list.
2068 * Then we restart the process which may have to remove
2069 * the page from the list that we just put it on again
2070 * because the number of objects in the slab may have
2075 old.freelist = page->freelist;
2076 old.counters = page->counters;
2077 VM_BUG_ON(!old.frozen);
2079 /* Determine target state of the slab */
2080 new.counters = old.counters;
2083 set_freepointer(s, freelist, old.freelist);
2084 new.freelist = freelist;
2086 new.freelist = old.freelist;
2090 if (!new.inuse && n->nr_partial >= s->min_partial)
2092 else if (new.freelist) {
2097 * Taking the spinlock removes the possiblity
2098 * that acquire_slab() will see a slab page that
2101 raw_spin_lock(&n->list_lock);
2105 if (kmem_cache_debug(s) && !lock) {
2108 * This also ensures that the scanning of full
2109 * slabs from diagnostic functions will not see
2112 raw_spin_lock(&n->list_lock);
2120 remove_partial(n, page);
2122 else if (l == M_FULL)
2124 remove_full(s, n, page);
2126 if (m == M_PARTIAL) {
2128 add_partial(n, page, tail);
2131 } else if (m == M_FULL) {
2133 stat(s, DEACTIVATE_FULL);
2134 add_full(s, n, page);
2140 if (!__cmpxchg_double_slab(s, page,
2141 old.freelist, old.counters,
2142 new.freelist, new.counters,
2147 raw_spin_unlock(&n->list_lock);
2150 stat(s, DEACTIVATE_EMPTY);
2151 discard_slab(s, page);
2157 * Unfreeze all the cpu partial slabs.
2159 * This function must be called with interrupts disabled
2160 * for the cpu using c (or some other guarantee must be there
2161 * to guarantee no concurrent accesses).
2163 static void unfreeze_partials(struct kmem_cache *s,
2164 struct kmem_cache_cpu *c)
2166 #ifdef CONFIG_SLUB_CPU_PARTIAL
2167 struct kmem_cache_node *n = NULL, *n2 = NULL;
2168 struct page *page, *discard_page = NULL;
2170 while ((page = c->partial)) {
2174 c->partial = page->next;
2176 n2 = get_node(s, page_to_nid(page));
2179 raw_spin_unlock(&n->list_lock);
2182 raw_spin_lock(&n->list_lock);
2187 old.freelist = page->freelist;
2188 old.counters = page->counters;
2189 VM_BUG_ON(!old.frozen);
2191 new.counters = old.counters;
2192 new.freelist = old.freelist;
2196 } while (!__cmpxchg_double_slab(s, page,
2197 old.freelist, old.counters,
2198 new.freelist, new.counters,
2199 "unfreezing slab"));
2201 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2202 page->next = discard_page;
2203 discard_page = page;
2205 add_partial(n, page, DEACTIVATE_TO_TAIL);
2206 stat(s, FREE_ADD_PARTIAL);
2211 raw_spin_unlock(&n->list_lock);
2213 while (discard_page) {
2214 page = discard_page;
2215 discard_page = discard_page->next;
2217 stat(s, DEACTIVATE_EMPTY);
2218 discard_slab(s, page);
2225 * Put a page that was just frozen (in __slab_free) into a partial page
2226 * slot if available. This is done without interrupts disabled and without
2227 * preemption disabled. The cmpxchg is racy and may put the partial page
2228 * onto a random cpus partial slot.
2230 * If we did not find a slot then simply move all the partials to the
2231 * per node partial list.
2233 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2235 #ifdef CONFIG_SLUB_CPU_PARTIAL
2236 struct page *oldpage;
2244 oldpage = this_cpu_read(s->cpu_slab->partial);
2247 pobjects = oldpage->pobjects;
2248 pages = oldpage->pages;
2249 if (drain && pobjects > s->cpu_partial) {
2250 struct slub_free_list *f;
2251 unsigned long flags;
2254 * partial array is full. Move the existing
2255 * set to the per node partial list.
2257 local_irq_save(flags);
2258 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2259 f = this_cpu_ptr(&slub_free_list);
2260 raw_spin_lock(&f->lock);
2261 list_splice_init(&f->list, &tofree);
2262 raw_spin_unlock(&f->lock);
2263 local_irq_restore(flags);
2264 free_delayed(&tofree);
2268 stat(s, CPU_PARTIAL_DRAIN);
2273 pobjects += page->objects - page->inuse;
2275 page->pages = pages;
2276 page->pobjects = pobjects;
2277 page->next = oldpage;
2279 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2281 if (unlikely(!s->cpu_partial)) {
2282 unsigned long flags;
2284 local_irq_save(flags);
2285 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2286 local_irq_restore(flags);
2292 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2294 stat(s, CPUSLAB_FLUSH);
2295 deactivate_slab(s, c->page, c->freelist);
2297 c->tid = next_tid(c->tid);
2305 * Called from IPI handler with interrupts disabled.
2307 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2309 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2315 unfreeze_partials(s, c);
2319 static void flush_cpu_slab(void *d)
2321 struct kmem_cache *s = d;
2323 __flush_cpu_slab(s, smp_processor_id());
2326 static bool has_cpu_slab(int cpu, void *info)
2328 struct kmem_cache *s = info;
2329 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2331 return c->page || c->partial;
2334 static void flush_all(struct kmem_cache *s)
2339 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2340 for_each_online_cpu(cpu) {
2341 struct slub_free_list *f;
2343 if (!has_cpu_slab(cpu, s))
2346 f = &per_cpu(slub_free_list, cpu);
2347 raw_spin_lock_irq(&f->lock);
2348 list_splice_init(&f->list, &tofree);
2349 raw_spin_unlock_irq(&f->lock);
2350 free_delayed(&tofree);
2355 * Use the cpu notifier to insure that the cpu slabs are flushed when
2358 static int slub_cpu_dead(unsigned int cpu)
2360 struct kmem_cache *s;
2361 unsigned long flags;
2363 mutex_lock(&slab_mutex);
2364 list_for_each_entry(s, &slab_caches, list) {
2365 local_irq_save(flags);
2366 __flush_cpu_slab(s, cpu);
2367 local_irq_restore(flags);
2369 mutex_unlock(&slab_mutex);
2374 * Check if the objects in a per cpu structure fit numa
2375 * locality expectations.
2377 static inline int node_match(struct page *page, int node)
2380 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2386 #ifdef CONFIG_SLUB_DEBUG
2387 static int count_free(struct page *page)
2389 return page->objects - page->inuse;
2392 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2394 return atomic_long_read(&n->total_objects);
2396 #endif /* CONFIG_SLUB_DEBUG */
2398 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2399 static unsigned long count_partial(struct kmem_cache_node *n,
2400 int (*get_count)(struct page *))
2402 unsigned long flags;
2403 unsigned long x = 0;
2406 raw_spin_lock_irqsave(&n->list_lock, flags);
2407 list_for_each_entry(page, &n->partial, lru)
2408 x += get_count(page);
2409 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2412 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2414 static noinline void
2415 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2417 #ifdef CONFIG_SLUB_DEBUG
2418 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2419 DEFAULT_RATELIMIT_BURST);
2421 struct kmem_cache_node *n;
2423 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2426 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2427 nid, gfpflags, &gfpflags);
2428 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2429 s->name, s->object_size, s->size, oo_order(s->oo),
2432 if (oo_order(s->min) > get_order(s->object_size))
2433 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2436 for_each_kmem_cache_node(s, node, n) {
2437 unsigned long nr_slabs;
2438 unsigned long nr_objs;
2439 unsigned long nr_free;
2441 nr_free = count_partial(n, count_free);
2442 nr_slabs = node_nr_slabs(n);
2443 nr_objs = node_nr_objs(n);
2445 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2446 node, nr_slabs, nr_objs, nr_free);
2451 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2452 int node, struct kmem_cache_cpu **pc)
2455 struct kmem_cache_cpu *c = *pc;
2458 freelist = get_partial(s, flags, node, c);
2463 page = new_slab(s, flags, node);
2465 c = raw_cpu_ptr(s->cpu_slab);
2470 * No other reference to the page yet so we can
2471 * muck around with it freely without cmpxchg
2473 freelist = page->freelist;
2474 page->freelist = NULL;
2476 stat(s, ALLOC_SLAB);
2485 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2487 if (unlikely(PageSlabPfmemalloc(page)))
2488 return gfp_pfmemalloc_allowed(gfpflags);
2494 * Check the page->freelist of a page and either transfer the freelist to the
2495 * per cpu freelist or deactivate the page.
2497 * The page is still frozen if the return value is not NULL.
2499 * If this function returns NULL then the page has been unfrozen.
2501 * This function must be called with interrupt disabled.
2503 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2506 unsigned long counters;
2510 freelist = page->freelist;
2511 counters = page->counters;
2513 new.counters = counters;
2514 VM_BUG_ON(!new.frozen);
2516 new.inuse = page->objects;
2517 new.frozen = freelist != NULL;
2519 } while (!__cmpxchg_double_slab(s, page,
2528 * Slow path. The lockless freelist is empty or we need to perform
2531 * Processing is still very fast if new objects have been freed to the
2532 * regular freelist. In that case we simply take over the regular freelist
2533 * as the lockless freelist and zap the regular freelist.
2535 * If that is not working then we fall back to the partial lists. We take the
2536 * first element of the freelist as the object to allocate now and move the
2537 * rest of the freelist to the lockless freelist.
2539 * And if we were unable to get a new slab from the partial slab lists then
2540 * we need to allocate a new slab. This is the slowest path since it involves
2541 * a call to the page allocator and the setup of a new slab.
2543 * Version of __slab_alloc to use when we know that interrupts are
2544 * already disabled (which is the case for bulk allocation).
2546 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2547 unsigned long addr, struct kmem_cache_cpu *c,
2548 struct list_head *to_free)
2550 struct slub_free_list *f;
2559 if (unlikely(!node_match(page, node))) {
2560 int searchnode = node;
2562 if (node != NUMA_NO_NODE && !node_present_pages(node))
2563 searchnode = node_to_mem_node(node);
2565 if (unlikely(!node_match(page, searchnode))) {
2566 stat(s, ALLOC_NODE_MISMATCH);
2567 deactivate_slab(s, page, c->freelist);
2575 * By rights, we should be searching for a slab page that was
2576 * PFMEMALLOC but right now, we are losing the pfmemalloc
2577 * information when the page leaves the per-cpu allocator
2579 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2580 deactivate_slab(s, page, c->freelist);
2586 /* must check again c->freelist in case of cpu migration or IRQ */
2587 freelist = c->freelist;
2591 freelist = get_freelist(s, page);
2595 stat(s, DEACTIVATE_BYPASS);
2599 stat(s, ALLOC_REFILL);
2603 * freelist is pointing to the list of objects to be used.
2604 * page is pointing to the page from which the objects are obtained.
2605 * That page must be frozen for per cpu allocations to work.
2607 VM_BUG_ON(!c->page->frozen);
2608 c->freelist = get_freepointer(s, freelist);
2609 c->tid = next_tid(c->tid);
2612 f = this_cpu_ptr(&slub_free_list);
2613 raw_spin_lock(&f->lock);
2614 list_splice_init(&f->list, to_free);
2615 raw_spin_unlock(&f->lock);
2622 page = c->page = c->partial;
2623 c->partial = page->next;
2624 stat(s, CPU_PARTIAL_ALLOC);
2629 freelist = new_slab_objects(s, gfpflags, node, &c);
2631 if (unlikely(!freelist)) {
2632 slab_out_of_memory(s, gfpflags, node);
2637 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2640 /* Only entered in the debug case */
2641 if (kmem_cache_debug(s) &&
2642 !alloc_debug_processing(s, page, freelist, addr))
2643 goto new_slab; /* Slab failed checks. Next slab needed */
2645 deactivate_slab(s, page, get_freepointer(s, freelist));
2652 * Another one that disabled interrupt and compensates for possible
2653 * cpu changes by refetching the per cpu area pointer.
2655 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2656 unsigned long addr, struct kmem_cache_cpu *c)
2659 unsigned long flags;
2662 local_irq_save(flags);
2663 #ifdef CONFIG_PREEMPT
2665 * We may have been preempted and rescheduled on a different
2666 * cpu before disabling interrupts. Need to reload cpu area
2669 c = this_cpu_ptr(s->cpu_slab);
2672 p = ___slab_alloc(s, gfpflags, node, addr, c, &tofree);
2673 local_irq_restore(flags);
2674 free_delayed(&tofree);
2679 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2680 * have the fastpath folded into their functions. So no function call
2681 * overhead for requests that can be satisfied on the fastpath.
2683 * The fastpath works by first checking if the lockless freelist can be used.
2684 * If not then __slab_alloc is called for slow processing.
2686 * Otherwise we can simply pick the next object from the lockless free list.
2688 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2689 gfp_t gfpflags, int node, unsigned long addr)
2692 struct kmem_cache_cpu *c;
2696 s = slab_pre_alloc_hook(s, gfpflags);
2701 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2702 * enabled. We may switch back and forth between cpus while
2703 * reading from one cpu area. That does not matter as long
2704 * as we end up on the original cpu again when doing the cmpxchg.
2706 * We should guarantee that tid and kmem_cache are retrieved on
2707 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2708 * to check if it is matched or not.
2711 tid = this_cpu_read(s->cpu_slab->tid);
2712 c = raw_cpu_ptr(s->cpu_slab);
2713 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2714 unlikely(tid != READ_ONCE(c->tid)));
2717 * Irqless object alloc/free algorithm used here depends on sequence
2718 * of fetching cpu_slab's data. tid should be fetched before anything
2719 * on c to guarantee that object and page associated with previous tid
2720 * won't be used with current tid. If we fetch tid first, object and
2721 * page could be one associated with next tid and our alloc/free
2722 * request will be failed. In this case, we will retry. So, no problem.
2727 * The transaction ids are globally unique per cpu and per operation on
2728 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2729 * occurs on the right processor and that there was no operation on the
2730 * linked list in between.
2733 object = c->freelist;
2735 if (unlikely(!object || !node_match(page, node))) {
2736 object = __slab_alloc(s, gfpflags, node, addr, c);
2737 stat(s, ALLOC_SLOWPATH);
2739 void *next_object = get_freepointer_safe(s, object);
2742 * The cmpxchg will only match if there was no additional
2743 * operation and if we are on the right processor.
2745 * The cmpxchg does the following atomically (without lock
2747 * 1. Relocate first pointer to the current per cpu area.
2748 * 2. Verify that tid and freelist have not been changed
2749 * 3. If they were not changed replace tid and freelist
2751 * Since this is without lock semantics the protection is only
2752 * against code executing on this cpu *not* from access by
2755 if (unlikely(!this_cpu_cmpxchg_double(
2756 s->cpu_slab->freelist, s->cpu_slab->tid,
2758 next_object, next_tid(tid)))) {
2760 note_cmpxchg_failure("slab_alloc", s, tid);
2763 prefetch_freepointer(s, next_object);
2764 stat(s, ALLOC_FASTPATH);
2767 if (unlikely(gfpflags & __GFP_ZERO) && object)
2768 memset(object, 0, s->object_size);
2770 slab_post_alloc_hook(s, gfpflags, 1, &object);
2775 static __always_inline void *slab_alloc(struct kmem_cache *s,
2776 gfp_t gfpflags, unsigned long addr)
2778 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2781 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2783 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2785 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2790 EXPORT_SYMBOL(kmem_cache_alloc);
2792 #ifdef CONFIG_TRACING
2793 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2795 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2796 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2797 kasan_kmalloc(s, ret, size, gfpflags);
2800 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2804 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2806 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2808 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2809 s->object_size, s->size, gfpflags, node);
2813 EXPORT_SYMBOL(kmem_cache_alloc_node);
2815 #ifdef CONFIG_TRACING
2816 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2818 int node, size_t size)
2820 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2822 trace_kmalloc_node(_RET_IP_, ret,
2823 size, s->size, gfpflags, node);
2825 kasan_kmalloc(s, ret, size, gfpflags);
2828 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2833 * Slow path handling. This may still be called frequently since objects
2834 * have a longer lifetime than the cpu slabs in most processing loads.
2836 * So we still attempt to reduce cache line usage. Just take the slab
2837 * lock and free the item. If there is no additional partial page
2838 * handling required then we can return immediately.
2840 static void __slab_free(struct kmem_cache *s, struct page *page,
2841 void *head, void *tail, int cnt,
2848 unsigned long counters;
2849 struct kmem_cache_node *n = NULL;
2850 unsigned long uninitialized_var(flags);
2852 stat(s, FREE_SLOWPATH);
2854 if (kmem_cache_debug(s) &&
2855 !free_debug_processing(s, page, head, tail, cnt, addr))
2860 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2863 prior = page->freelist;
2864 counters = page->counters;
2865 set_freepointer(s, tail, prior);
2866 new.counters = counters;
2867 was_frozen = new.frozen;
2869 if ((!new.inuse || !prior) && !was_frozen) {
2871 if (kmem_cache_has_cpu_partial(s) && !prior) {
2874 * Slab was on no list before and will be
2876 * We can defer the list move and instead
2881 } else { /* Needs to be taken off a list */
2883 n = get_node(s, page_to_nid(page));
2885 * Speculatively acquire the list_lock.
2886 * If the cmpxchg does not succeed then we may
2887 * drop the list_lock without any processing.
2889 * Otherwise the list_lock will synchronize with
2890 * other processors updating the list of slabs.
2892 raw_spin_lock_irqsave(&n->list_lock, flags);
2897 } while (!cmpxchg_double_slab(s, page,
2905 * If we just froze the page then put it onto the
2906 * per cpu partial list.
2908 if (new.frozen && !was_frozen) {
2909 put_cpu_partial(s, page, 1);
2910 stat(s, CPU_PARTIAL_FREE);
2913 * The list lock was not taken therefore no list
2914 * activity can be necessary.
2917 stat(s, FREE_FROZEN);
2921 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2925 * Objects left in the slab. If it was not on the partial list before
2928 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2929 if (kmem_cache_debug(s))
2930 remove_full(s, n, page);
2931 add_partial(n, page, DEACTIVATE_TO_TAIL);
2932 stat(s, FREE_ADD_PARTIAL);
2934 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2940 * Slab on the partial list.
2942 remove_partial(n, page);
2943 stat(s, FREE_REMOVE_PARTIAL);
2945 /* Slab must be on the full list */
2946 remove_full(s, n, page);
2949 raw_spin_unlock_irqrestore(&n->list_lock, flags);
2951 discard_slab(s, page);
2955 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2956 * can perform fastpath freeing without additional function calls.
2958 * The fastpath is only possible if we are freeing to the current cpu slab
2959 * of this processor. This typically the case if we have just allocated
2962 * If fastpath is not possible then fall back to __slab_free where we deal
2963 * with all sorts of special processing.
2965 * Bulk free of a freelist with several objects (all pointing to the
2966 * same page) possible by specifying head and tail ptr, plus objects
2967 * count (cnt). Bulk free indicated by tail pointer being set.
2969 static __always_inline void do_slab_free(struct kmem_cache *s,
2970 struct page *page, void *head, void *tail,
2971 int cnt, unsigned long addr)
2973 void *tail_obj = tail ? : head;
2974 struct kmem_cache_cpu *c;
2978 * Determine the currently cpus per cpu slab.
2979 * The cpu may change afterward. However that does not matter since
2980 * data is retrieved via this pointer. If we are on the same cpu
2981 * during the cmpxchg then the free will succeed.
2984 tid = this_cpu_read(s->cpu_slab->tid);
2985 c = raw_cpu_ptr(s->cpu_slab);
2986 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2987 unlikely(tid != READ_ONCE(c->tid)));
2989 /* Same with comment on barrier() in slab_alloc_node() */
2992 if (likely(page == c->page)) {
2993 set_freepointer(s, tail_obj, c->freelist);
2995 if (unlikely(!this_cpu_cmpxchg_double(
2996 s->cpu_slab->freelist, s->cpu_slab->tid,
2998 head, next_tid(tid)))) {
3000 note_cmpxchg_failure("slab_free", s, tid);
3003 stat(s, FREE_FASTPATH);
3005 __slab_free(s, page, head, tail_obj, cnt, addr);
3009 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3010 void *head, void *tail, int cnt,
3013 slab_free_freelist_hook(s, head, tail);
3015 * slab_free_freelist_hook() could have put the items into quarantine.
3016 * If so, no need to free them.
3018 if (s->flags & SLAB_KASAN && !(s->flags & SLAB_DESTROY_BY_RCU))
3020 do_slab_free(s, page, head, tail, cnt, addr);
3024 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3026 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3030 void kmem_cache_free(struct kmem_cache *s, void *x)
3032 s = cache_from_obj(s, x);
3035 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3036 trace_kmem_cache_free(_RET_IP_, x);
3038 EXPORT_SYMBOL(kmem_cache_free);
3040 struct detached_freelist {
3045 struct kmem_cache *s;
3049 * This function progressively scans the array with free objects (with
3050 * a limited look ahead) and extract objects belonging to the same
3051 * page. It builds a detached freelist directly within the given
3052 * page/objects. This can happen without any need for
3053 * synchronization, because the objects are owned by running process.
3054 * The freelist is build up as a single linked list in the objects.
3055 * The idea is, that this detached freelist can then be bulk
3056 * transferred to the real freelist(s), but only requiring a single
3057 * synchronization primitive. Look ahead in the array is limited due
3058 * to performance reasons.
3061 int build_detached_freelist(struct kmem_cache *s, size_t size,
3062 void **p, struct detached_freelist *df)
3064 size_t first_skipped_index = 0;
3069 /* Always re-init detached_freelist */
3074 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3075 } while (!object && size);
3080 page = virt_to_head_page(object);
3082 /* Handle kalloc'ed objects */
3083 if (unlikely(!PageSlab(page))) {
3084 BUG_ON(!PageCompound(page));
3086 __free_pages(page, compound_order(page));
3087 p[size] = NULL; /* mark object processed */
3090 /* Derive kmem_cache from object */
3091 df->s = page->slab_cache;
3093 df->s = cache_from_obj(s, object); /* Support for memcg */
3096 /* Start new detached freelist */
3098 set_freepointer(df->s, object, NULL);
3100 df->freelist = object;
3101 p[size] = NULL; /* mark object processed */
3107 continue; /* Skip processed objects */
3109 /* df->page is always set at this point */
3110 if (df->page == virt_to_head_page(object)) {
3111 /* Opportunity build freelist */
3112 set_freepointer(df->s, object, df->freelist);
3113 df->freelist = object;
3115 p[size] = NULL; /* mark object processed */
3120 /* Limit look ahead search */
3124 if (!first_skipped_index)
3125 first_skipped_index = size + 1;
3128 return first_skipped_index;
3131 /* Note that interrupts must be enabled when calling this function. */
3132 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3138 struct detached_freelist df;
3140 size = build_detached_freelist(s, size, p, &df);
3141 if (unlikely(!df.page))
3144 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3145 } while (likely(size));
3147 EXPORT_SYMBOL(kmem_cache_free_bulk);
3149 /* Note that interrupts must be enabled when calling this function. */
3150 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3153 struct kmem_cache_cpu *c;
3157 /* memcg and kmem_cache debug support */
3158 s = slab_pre_alloc_hook(s, flags);
3162 * Drain objects in the per cpu slab, while disabling local
3163 * IRQs, which protects against PREEMPT and interrupts
3164 * handlers invoking normal fastpath.
3166 local_irq_disable();
3167 c = this_cpu_ptr(s->cpu_slab);
3169 for (i = 0; i < size; i++) {
3170 void *object = c->freelist;
3172 if (unlikely(!object)) {
3174 * Invoking slow path likely have side-effect
3175 * of re-populating per CPU c->freelist
3177 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3178 _RET_IP_, c, &to_free);
3179 if (unlikely(!p[i]))
3182 c = this_cpu_ptr(s->cpu_slab);
3183 continue; /* goto for-loop */
3185 c->freelist = get_freepointer(s, object);
3188 c->tid = next_tid(c->tid);
3190 free_delayed(&to_free);
3192 /* Clear memory outside IRQ disabled fastpath loop */
3193 if (unlikely(flags & __GFP_ZERO)) {
3196 for (j = 0; j < i; j++)
3197 memset(p[j], 0, s->object_size);
3200 /* memcg and kmem_cache debug support */
3201 slab_post_alloc_hook(s, flags, size, p);
3205 slab_post_alloc_hook(s, flags, i, p);
3206 __kmem_cache_free_bulk(s, i, p);
3209 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3213 * Object placement in a slab is made very easy because we always start at
3214 * offset 0. If we tune the size of the object to the alignment then we can
3215 * get the required alignment by putting one properly sized object after
3218 * Notice that the allocation order determines the sizes of the per cpu
3219 * caches. Each processor has always one slab available for allocations.
3220 * Increasing the allocation order reduces the number of times that slabs
3221 * must be moved on and off the partial lists and is therefore a factor in
3226 * Mininum / Maximum order of slab pages. This influences locking overhead
3227 * and slab fragmentation. A higher order reduces the number of partial slabs
3228 * and increases the number of allocations possible without having to
3229 * take the list_lock.
3231 static int slub_min_order;
3232 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3233 static int slub_min_objects;
3236 * Calculate the order of allocation given an slab object size.
3238 * The order of allocation has significant impact on performance and other
3239 * system components. Generally order 0 allocations should be preferred since
3240 * order 0 does not cause fragmentation in the page allocator. Larger objects
3241 * be problematic to put into order 0 slabs because there may be too much
3242 * unused space left. We go to a higher order if more than 1/16th of the slab
3245 * In order to reach satisfactory performance we must ensure that a minimum
3246 * number of objects is in one slab. Otherwise we may generate too much
3247 * activity on the partial lists which requires taking the list_lock. This is
3248 * less a concern for large slabs though which are rarely used.
3250 * slub_max_order specifies the order where we begin to stop considering the
3251 * number of objects in a slab as critical. If we reach slub_max_order then
3252 * we try to keep the page order as low as possible. So we accept more waste
3253 * of space in favor of a small page order.
3255 * Higher order allocations also allow the placement of more objects in a
3256 * slab and thereby reduce object handling overhead. If the user has
3257 * requested a higher mininum order then we start with that one instead of
3258 * the smallest order which will fit the object.
3260 static inline int slab_order(int size, int min_objects,
3261 int max_order, int fract_leftover, int reserved)
3265 int min_order = slub_min_order;
3267 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3268 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3270 for (order = max(min_order, get_order(min_objects * size + reserved));
3271 order <= max_order; order++) {
3273 unsigned long slab_size = PAGE_SIZE << order;
3275 rem = (slab_size - reserved) % size;
3277 if (rem <= slab_size / fract_leftover)
3284 static inline int calculate_order(int size, int reserved)
3292 * Attempt to find best configuration for a slab. This
3293 * works by first attempting to generate a layout with
3294 * the best configuration and backing off gradually.
3296 * First we increase the acceptable waste in a slab. Then
3297 * we reduce the minimum objects required in a slab.
3299 min_objects = slub_min_objects;
3301 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3302 max_objects = order_objects(slub_max_order, size, reserved);
3303 min_objects = min(min_objects, max_objects);
3305 while (min_objects > 1) {
3307 while (fraction >= 4) {
3308 order = slab_order(size, min_objects,
3309 slub_max_order, fraction, reserved);
3310 if (order <= slub_max_order)
3318 * We were unable to place multiple objects in a slab. Now
3319 * lets see if we can place a single object there.
3321 order = slab_order(size, 1, slub_max_order, 1, reserved);
3322 if (order <= slub_max_order)
3326 * Doh this slab cannot be placed using slub_max_order.
3328 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3329 if (order < MAX_ORDER)
3335 init_kmem_cache_node(struct kmem_cache_node *n)
3338 raw_spin_lock_init(&n->list_lock);
3339 INIT_LIST_HEAD(&n->partial);
3340 #ifdef CONFIG_SLUB_DEBUG
3341 atomic_long_set(&n->nr_slabs, 0);
3342 atomic_long_set(&n->total_objects, 0);
3343 INIT_LIST_HEAD(&n->full);
3347 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3349 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3350 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3353 * Must align to double word boundary for the double cmpxchg
3354 * instructions to work; see __pcpu_double_call_return_bool().
3356 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3357 2 * sizeof(void *));
3362 init_kmem_cache_cpus(s);
3367 static struct kmem_cache *kmem_cache_node;
3370 * No kmalloc_node yet so do it by hand. We know that this is the first
3371 * slab on the node for this slabcache. There are no concurrent accesses
3374 * Note that this function only works on the kmem_cache_node
3375 * when allocating for the kmem_cache_node. This is used for bootstrapping
3376 * memory on a fresh node that has no slab structures yet.
3378 static void early_kmem_cache_node_alloc(int node)
3381 struct kmem_cache_node *n;
3383 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3385 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3388 if (page_to_nid(page) != node) {
3389 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3390 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3395 page->freelist = get_freepointer(kmem_cache_node, n);
3398 kmem_cache_node->node[node] = n;
3399 #ifdef CONFIG_SLUB_DEBUG
3400 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3401 init_tracking(kmem_cache_node, n);
3403 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3405 init_kmem_cache_node(n);
3406 inc_slabs_node(kmem_cache_node, node, page->objects);
3409 * No locks need to be taken here as it has just been
3410 * initialized and there is no concurrent access.
3412 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3415 static void free_kmem_cache_nodes(struct kmem_cache *s)
3418 struct kmem_cache_node *n;
3420 for_each_kmem_cache_node(s, node, n) {
3421 kmem_cache_free(kmem_cache_node, n);
3422 s->node[node] = NULL;
3426 void __kmem_cache_release(struct kmem_cache *s)
3428 cache_random_seq_destroy(s);
3429 free_percpu(s->cpu_slab);
3430 free_kmem_cache_nodes(s);
3433 static int init_kmem_cache_nodes(struct kmem_cache *s)
3437 for_each_node_state(node, N_NORMAL_MEMORY) {
3438 struct kmem_cache_node *n;
3440 if (slab_state == DOWN) {
3441 early_kmem_cache_node_alloc(node);
3444 n = kmem_cache_alloc_node(kmem_cache_node,
3448 free_kmem_cache_nodes(s);
3453 init_kmem_cache_node(n);
3458 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3460 if (min < MIN_PARTIAL)
3462 else if (min > MAX_PARTIAL)
3464 s->min_partial = min;
3468 * calculate_sizes() determines the order and the distribution of data within
3471 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3473 unsigned long flags = s->flags;
3474 size_t size = s->object_size;
3478 * Round up object size to the next word boundary. We can only
3479 * place the free pointer at word boundaries and this determines
3480 * the possible location of the free pointer.
3482 size = ALIGN(size, sizeof(void *));
3484 #ifdef CONFIG_SLUB_DEBUG
3486 * Determine if we can poison the object itself. If the user of
3487 * the slab may touch the object after free or before allocation
3488 * then we should never poison the object itself.
3490 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3492 s->flags |= __OBJECT_POISON;
3494 s->flags &= ~__OBJECT_POISON;
3498 * If we are Redzoning then check if there is some space between the
3499 * end of the object and the free pointer. If not then add an
3500 * additional word to have some bytes to store Redzone information.
3502 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3503 size += sizeof(void *);
3507 * With that we have determined the number of bytes in actual use
3508 * by the object. This is the potential offset to the free pointer.
3512 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3515 * Relocate free pointer after the object if it is not
3516 * permitted to overwrite the first word of the object on
3519 * This is the case if we do RCU, have a constructor or
3520 * destructor or are poisoning the objects.
3523 size += sizeof(void *);
3526 #ifdef CONFIG_SLUB_DEBUG
3527 if (flags & SLAB_STORE_USER)
3529 * Need to store information about allocs and frees after
3532 size += 2 * sizeof(struct track);
3535 kasan_cache_create(s, &size, &s->flags);
3536 #ifdef CONFIG_SLUB_DEBUG
3537 if (flags & SLAB_RED_ZONE) {
3539 * Add some empty padding so that we can catch
3540 * overwrites from earlier objects rather than let
3541 * tracking information or the free pointer be
3542 * corrupted if a user writes before the start
3545 size += sizeof(void *);
3547 s->red_left_pad = sizeof(void *);
3548 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3549 size += s->red_left_pad;
3554 * SLUB stores one object immediately after another beginning from
3555 * offset 0. In order to align the objects we have to simply size
3556 * each object to conform to the alignment.
3558 size = ALIGN(size, s->align);
3560 if (forced_order >= 0)
3561 order = forced_order;
3563 order = calculate_order(size, s->reserved);
3570 s->allocflags |= __GFP_COMP;
3572 if (s->flags & SLAB_CACHE_DMA)
3573 s->allocflags |= GFP_DMA;
3575 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3576 s->allocflags |= __GFP_RECLAIMABLE;
3579 * Determine the number of objects per slab
3581 s->oo = oo_make(order, size, s->reserved);
3582 s->min = oo_make(get_order(size), size, s->reserved);
3583 if (oo_objects(s->oo) > oo_objects(s->max))
3586 return !!oo_objects(s->oo);
3589 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3591 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3594 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3595 s->reserved = sizeof(struct rcu_head);
3597 if (!calculate_sizes(s, -1))
3599 if (disable_higher_order_debug) {
3601 * Disable debugging flags that store metadata if the min slab
3604 if (get_order(s->size) > get_order(s->object_size)) {
3605 s->flags &= ~DEBUG_METADATA_FLAGS;
3607 if (!calculate_sizes(s, -1))
3612 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3613 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3614 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3615 /* Enable fast mode */
3616 s->flags |= __CMPXCHG_DOUBLE;
3620 * The larger the object size is, the more pages we want on the partial
3621 * list to avoid pounding the page allocator excessively.
3623 set_min_partial(s, ilog2(s->size) / 2);
3626 * cpu_partial determined the maximum number of objects kept in the
3627 * per cpu partial lists of a processor.
3629 * Per cpu partial lists mainly contain slabs that just have one
3630 * object freed. If they are used for allocation then they can be
3631 * filled up again with minimal effort. The slab will never hit the
3632 * per node partial lists and therefore no locking will be required.
3634 * This setting also determines
3636 * A) The number of objects from per cpu partial slabs dumped to the
3637 * per node list when we reach the limit.
3638 * B) The number of objects in cpu partial slabs to extract from the
3639 * per node list when we run out of per cpu objects. We only fetch
3640 * 50% to keep some capacity around for frees.
3642 if (!kmem_cache_has_cpu_partial(s))
3644 else if (s->size >= PAGE_SIZE)
3646 else if (s->size >= 1024)
3648 else if (s->size >= 256)
3649 s->cpu_partial = 13;
3651 s->cpu_partial = 30;
3654 s->remote_node_defrag_ratio = 1000;
3657 /* Initialize the pre-computed randomized freelist if slab is up */
3658 if (slab_state >= UP) {
3659 if (init_cache_random_seq(s))
3663 if (!init_kmem_cache_nodes(s))
3666 if (alloc_kmem_cache_cpus(s))
3669 free_kmem_cache_nodes(s);
3671 if (flags & SLAB_PANIC)
3672 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3673 s->name, (unsigned long)s->size, s->size,
3674 oo_order(s->oo), s->offset, flags);
3678 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3681 #ifdef CONFIG_SLUB_DEBUG
3682 #ifdef CONFIG_PREEMPT_RT_BASE
3683 /* XXX move out of irq-off section */
3684 slab_err(s, page, text, s->name);
3686 void *addr = page_address(page);
3688 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3689 sizeof(long), GFP_ATOMIC);
3692 slab_err(s, page, text, s->name);
3695 get_map(s, page, map);
3696 for_each_object(p, s, addr, page->objects) {
3698 if (!test_bit(slab_index(p, s, addr), map)) {
3699 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3700 print_tracking(s, p);
3710 * Attempt to free all partial slabs on a node.
3711 * This is called from __kmem_cache_shutdown(). We must take list_lock
3712 * because sysfs file might still access partial list after the shutdowning.
3714 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3717 struct page *page, *h;
3719 BUG_ON(irqs_disabled());
3720 raw_spin_lock_irq(&n->list_lock);
3721 list_for_each_entry_safe(page, h, &n->partial, lru) {
3723 remove_partial(n, page);
3724 list_add(&page->lru, &discard);
3726 list_slab_objects(s, page,
3727 "Objects remaining in %s on __kmem_cache_shutdown()");
3730 raw_spin_unlock_irq(&n->list_lock);
3732 list_for_each_entry_safe(page, h, &discard, lru)
3733 discard_slab(s, page);
3737 * Release all resources used by a slab cache.
3739 int __kmem_cache_shutdown(struct kmem_cache *s)
3742 struct kmem_cache_node *n;
3745 /* Attempt to free all objects */
3746 for_each_kmem_cache_node(s, node, n) {
3748 if (n->nr_partial || slabs_node(s, node))
3754 /********************************************************************
3756 *******************************************************************/
3758 static int __init setup_slub_min_order(char *str)
3760 get_option(&str, &slub_min_order);
3765 __setup("slub_min_order=", setup_slub_min_order);
3767 static int __init setup_slub_max_order(char *str)
3769 get_option(&str, &slub_max_order);
3770 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3775 __setup("slub_max_order=", setup_slub_max_order);
3777 static int __init setup_slub_min_objects(char *str)
3779 get_option(&str, &slub_min_objects);
3784 __setup("slub_min_objects=", setup_slub_min_objects);
3786 void *__kmalloc(size_t size, gfp_t flags)
3788 struct kmem_cache *s;
3791 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3792 return kmalloc_large(size, flags);
3794 s = kmalloc_slab(size, flags);
3796 if (unlikely(ZERO_OR_NULL_PTR(s)))
3799 ret = slab_alloc(s, flags, _RET_IP_);
3801 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3803 kasan_kmalloc(s, ret, size, flags);
3807 EXPORT_SYMBOL(__kmalloc);
3810 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3815 flags |= __GFP_COMP | __GFP_NOTRACK;
3816 page = alloc_pages_node(node, flags, get_order(size));
3818 ptr = page_address(page);
3820 kmalloc_large_node_hook(ptr, size, flags);
3824 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3826 struct kmem_cache *s;
3829 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3830 ret = kmalloc_large_node(size, flags, node);
3832 trace_kmalloc_node(_RET_IP_, ret,
3833 size, PAGE_SIZE << get_order(size),
3839 s = kmalloc_slab(size, flags);
3841 if (unlikely(ZERO_OR_NULL_PTR(s)))
3844 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3846 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3848 kasan_kmalloc(s, ret, size, flags);
3852 EXPORT_SYMBOL(__kmalloc_node);
3855 #ifdef CONFIG_HARDENED_USERCOPY
3857 * Rejects objects that are incorrectly sized.
3859 * Returns NULL if check passes, otherwise const char * to name of cache
3860 * to indicate an error.
3862 const char *__check_heap_object(const void *ptr, unsigned long n,
3865 struct kmem_cache *s;
3866 unsigned long offset;
3869 /* Find object and usable object size. */
3870 s = page->slab_cache;
3871 object_size = slab_ksize(s);
3873 /* Reject impossible pointers. */
3874 if (ptr < page_address(page))
3877 /* Find offset within object. */
3878 offset = (ptr - page_address(page)) % s->size;
3880 /* Adjust for redzone and reject if within the redzone. */
3881 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3882 if (offset < s->red_left_pad)
3884 offset -= s->red_left_pad;
3887 /* Allow address range falling entirely within object size. */
3888 if (offset <= object_size && n <= object_size - offset)
3893 #endif /* CONFIG_HARDENED_USERCOPY */
3895 static size_t __ksize(const void *object)
3899 if (unlikely(object == ZERO_SIZE_PTR))
3902 page = virt_to_head_page(object);
3904 if (unlikely(!PageSlab(page))) {
3905 WARN_ON(!PageCompound(page));
3906 return PAGE_SIZE << compound_order(page);
3909 return slab_ksize(page->slab_cache);
3912 size_t ksize(const void *object)
3914 size_t size = __ksize(object);
3915 /* We assume that ksize callers could use whole allocated area,
3916 * so we need to unpoison this area.
3918 kasan_unpoison_shadow(object, size);
3921 EXPORT_SYMBOL(ksize);
3923 void kfree(const void *x)
3926 void *object = (void *)x;
3928 trace_kfree(_RET_IP_, x);
3930 if (unlikely(ZERO_OR_NULL_PTR(x)))
3933 page = virt_to_head_page(x);
3934 if (unlikely(!PageSlab(page))) {
3935 BUG_ON(!PageCompound(page));
3937 __free_pages(page, compound_order(page));
3940 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3942 EXPORT_SYMBOL(kfree);
3944 #define SHRINK_PROMOTE_MAX 32
3947 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3948 * up most to the head of the partial lists. New allocations will then
3949 * fill those up and thus they can be removed from the partial lists.
3951 * The slabs with the least items are placed last. This results in them
3952 * being allocated from last increasing the chance that the last objects
3953 * are freed in them.
3955 int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3959 struct kmem_cache_node *n;
3962 struct list_head discard;
3963 struct list_head promote[SHRINK_PROMOTE_MAX];
3964 unsigned long flags;
3969 * Disable empty slabs caching. Used to avoid pinning offline
3970 * memory cgroups by kmem pages that can be freed.
3976 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3977 * so we have to make sure the change is visible.
3979 synchronize_sched();
3983 for_each_kmem_cache_node(s, node, n) {
3984 INIT_LIST_HEAD(&discard);
3985 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3986 INIT_LIST_HEAD(promote + i);
3988 raw_spin_lock_irqsave(&n->list_lock, flags);
3991 * Build lists of slabs to discard or promote.
3993 * Note that concurrent frees may occur while we hold the
3994 * list_lock. page->inuse here is the upper limit.
3996 list_for_each_entry_safe(page, t, &n->partial, lru) {
3997 int free = page->objects - page->inuse;
3999 /* Do not reread page->inuse */
4002 /* We do not keep full slabs on the list */
4005 if (free == page->objects) {
4006 list_move(&page->lru, &discard);
4008 } else if (free <= SHRINK_PROMOTE_MAX)
4009 list_move(&page->lru, promote + free - 1);
4013 * Promote the slabs filled up most to the head of the
4016 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4017 list_splice(promote + i, &n->partial);
4019 raw_spin_unlock_irqrestore(&n->list_lock, flags);
4021 /* Release empty slabs */
4022 list_for_each_entry_safe(page, t, &discard, lru)
4023 discard_slab(s, page);
4025 if (slabs_node(s, node))
4032 static int slab_mem_going_offline_callback(void *arg)
4034 struct kmem_cache *s;
4036 mutex_lock(&slab_mutex);
4037 list_for_each_entry(s, &slab_caches, list)
4038 __kmem_cache_shrink(s, false);
4039 mutex_unlock(&slab_mutex);
4044 static void slab_mem_offline_callback(void *arg)
4046 struct kmem_cache_node *n;
4047 struct kmem_cache *s;
4048 struct memory_notify *marg = arg;
4051 offline_node = marg->status_change_nid_normal;
4054 * If the node still has available memory. we need kmem_cache_node
4057 if (offline_node < 0)
4060 mutex_lock(&slab_mutex);
4061 list_for_each_entry(s, &slab_caches, list) {
4062 n = get_node(s, offline_node);
4065 * if n->nr_slabs > 0, slabs still exist on the node
4066 * that is going down. We were unable to free them,
4067 * and offline_pages() function shouldn't call this
4068 * callback. So, we must fail.
4070 BUG_ON(slabs_node(s, offline_node));
4072 s->node[offline_node] = NULL;
4073 kmem_cache_free(kmem_cache_node, n);
4076 mutex_unlock(&slab_mutex);
4079 static int slab_mem_going_online_callback(void *arg)
4081 struct kmem_cache_node *n;
4082 struct kmem_cache *s;
4083 struct memory_notify *marg = arg;
4084 int nid = marg->status_change_nid_normal;
4088 * If the node's memory is already available, then kmem_cache_node is
4089 * already created. Nothing to do.
4095 * We are bringing a node online. No memory is available yet. We must
4096 * allocate a kmem_cache_node structure in order to bring the node
4099 mutex_lock(&slab_mutex);
4100 list_for_each_entry(s, &slab_caches, list) {
4102 * XXX: kmem_cache_alloc_node will fallback to other nodes
4103 * since memory is not yet available from the node that
4106 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4111 init_kmem_cache_node(n);
4115 mutex_unlock(&slab_mutex);
4119 static int slab_memory_callback(struct notifier_block *self,
4120 unsigned long action, void *arg)
4125 case MEM_GOING_ONLINE:
4126 ret = slab_mem_going_online_callback(arg);
4128 case MEM_GOING_OFFLINE:
4129 ret = slab_mem_going_offline_callback(arg);
4132 case MEM_CANCEL_ONLINE:
4133 slab_mem_offline_callback(arg);
4136 case MEM_CANCEL_OFFLINE:
4140 ret = notifier_from_errno(ret);
4146 static struct notifier_block slab_memory_callback_nb = {
4147 .notifier_call = slab_memory_callback,
4148 .priority = SLAB_CALLBACK_PRI,
4151 /********************************************************************
4152 * Basic setup of slabs
4153 *******************************************************************/
4156 * Used for early kmem_cache structures that were allocated using
4157 * the page allocator. Allocate them properly then fix up the pointers
4158 * that may be pointing to the wrong kmem_cache structure.
4161 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4164 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4165 struct kmem_cache_node *n;
4167 memcpy(s, static_cache, kmem_cache->object_size);
4170 * This runs very early, and only the boot processor is supposed to be
4171 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4174 __flush_cpu_slab(s, smp_processor_id());
4175 for_each_kmem_cache_node(s, node, n) {
4178 list_for_each_entry(p, &n->partial, lru)
4181 #ifdef CONFIG_SLUB_DEBUG
4182 list_for_each_entry(p, &n->full, lru)
4186 slab_init_memcg_params(s);
4187 list_add(&s->list, &slab_caches);
4191 void __init kmem_cache_init(void)
4193 static __initdata struct kmem_cache boot_kmem_cache,
4194 boot_kmem_cache_node;
4197 for_each_possible_cpu(cpu) {
4198 raw_spin_lock_init(&per_cpu(slub_free_list, cpu).lock);
4199 INIT_LIST_HEAD(&per_cpu(slub_free_list, cpu).list);
4202 if (debug_guardpage_minorder())
4205 kmem_cache_node = &boot_kmem_cache_node;
4206 kmem_cache = &boot_kmem_cache;
4208 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4209 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
4211 register_hotmemory_notifier(&slab_memory_callback_nb);
4213 /* Able to allocate the per node structures */
4214 slab_state = PARTIAL;
4216 create_boot_cache(kmem_cache, "kmem_cache",
4217 offsetof(struct kmem_cache, node) +
4218 nr_node_ids * sizeof(struct kmem_cache_node *),
4219 SLAB_HWCACHE_ALIGN);
4221 kmem_cache = bootstrap(&boot_kmem_cache);
4224 * Allocate kmem_cache_node properly from the kmem_cache slab.
4225 * kmem_cache_node is separately allocated so no need to
4226 * update any list pointers.
4228 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4230 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4231 setup_kmalloc_cache_index_table();
4232 create_kmalloc_caches(0);
4234 /* Setup random freelists for each cache */
4235 init_freelist_randomization();
4237 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4240 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4242 slub_min_order, slub_max_order, slub_min_objects,
4243 nr_cpu_ids, nr_node_ids);
4246 void __init kmem_cache_init_late(void)
4251 __kmem_cache_alias(const char *name, size_t size, size_t align,
4252 unsigned long flags, void (*ctor)(void *))
4254 struct kmem_cache *s, *c;
4256 s = find_mergeable(size, align, flags, name, ctor);
4261 * Adjust the object sizes so that we clear
4262 * the complete object on kzalloc.
4264 s->object_size = max(s->object_size, (int)size);
4265 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
4267 for_each_memcg_cache(c, s) {
4268 c->object_size = s->object_size;
4269 c->inuse = max_t(int, c->inuse,
4270 ALIGN(size, sizeof(void *)));
4273 if (sysfs_slab_alias(s, name)) {
4282 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4286 err = kmem_cache_open(s, flags);
4290 /* Mutex is not taken during early boot */
4291 if (slab_state <= UP)
4294 memcg_propagate_slab_attrs(s);
4295 err = sysfs_slab_add(s);
4297 __kmem_cache_release(s);
4302 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4304 struct kmem_cache *s;
4307 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4308 return kmalloc_large(size, gfpflags);
4310 s = kmalloc_slab(size, gfpflags);
4312 if (unlikely(ZERO_OR_NULL_PTR(s)))
4315 ret = slab_alloc(s, gfpflags, caller);
4317 /* Honor the call site pointer we received. */
4318 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4324 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4325 int node, unsigned long caller)
4327 struct kmem_cache *s;
4330 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4331 ret = kmalloc_large_node(size, gfpflags, node);
4333 trace_kmalloc_node(caller, ret,
4334 size, PAGE_SIZE << get_order(size),
4340 s = kmalloc_slab(size, gfpflags);
4342 if (unlikely(ZERO_OR_NULL_PTR(s)))
4345 ret = slab_alloc_node(s, gfpflags, node, caller);
4347 /* Honor the call site pointer we received. */
4348 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4355 static int count_inuse(struct page *page)
4360 static int count_total(struct page *page)
4362 return page->objects;
4366 #ifdef CONFIG_SLUB_DEBUG
4367 static int validate_slab(struct kmem_cache *s, struct page *page,
4371 void *addr = page_address(page);
4373 if (!check_slab(s, page) ||
4374 !on_freelist(s, page, NULL))
4377 /* Now we know that a valid freelist exists */
4378 bitmap_zero(map, page->objects);
4380 get_map(s, page, map);
4381 for_each_object(p, s, addr, page->objects) {
4382 if (test_bit(slab_index(p, s, addr), map))
4383 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4387 for_each_object(p, s, addr, page->objects)
4388 if (!test_bit(slab_index(p, s, addr), map))
4389 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4394 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4398 validate_slab(s, page, map);
4402 static int validate_slab_node(struct kmem_cache *s,
4403 struct kmem_cache_node *n, unsigned long *map)
4405 unsigned long count = 0;
4407 unsigned long flags;
4409 raw_spin_lock_irqsave(&n->list_lock, flags);
4411 list_for_each_entry(page, &n->partial, lru) {
4412 validate_slab_slab(s, page, map);
4415 if (count != n->nr_partial)
4416 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4417 s->name, count, n->nr_partial);
4419 if (!(s->flags & SLAB_STORE_USER))
4422 list_for_each_entry(page, &n->full, lru) {
4423 validate_slab_slab(s, page, map);
4426 if (count != atomic_long_read(&n->nr_slabs))
4427 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4428 s->name, count, atomic_long_read(&n->nr_slabs));
4431 raw_spin_unlock_irqrestore(&n->list_lock, flags);
4435 static long validate_slab_cache(struct kmem_cache *s)
4438 unsigned long count = 0;
4439 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4440 sizeof(unsigned long), GFP_KERNEL);
4441 struct kmem_cache_node *n;
4447 for_each_kmem_cache_node(s, node, n)
4448 count += validate_slab_node(s, n, map);
4453 * Generate lists of code addresses where slabcache objects are allocated
4458 unsigned long count;
4465 DECLARE_BITMAP(cpus, NR_CPUS);
4471 unsigned long count;
4472 struct location *loc;
4475 static void free_loc_track(struct loc_track *t)
4478 free_pages((unsigned long)t->loc,
4479 get_order(sizeof(struct location) * t->max));
4482 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4487 order = get_order(sizeof(struct location) * max);
4489 l = (void *)__get_free_pages(flags, order);
4494 memcpy(l, t->loc, sizeof(struct location) * t->count);
4502 static int add_location(struct loc_track *t, struct kmem_cache *s,
4503 const struct track *track)
4505 long start, end, pos;
4507 unsigned long caddr;
4508 unsigned long age = jiffies - track->when;
4514 pos = start + (end - start + 1) / 2;
4517 * There is nothing at "end". If we end up there
4518 * we need to add something to before end.
4523 caddr = t->loc[pos].addr;
4524 if (track->addr == caddr) {
4530 if (age < l->min_time)
4532 if (age > l->max_time)
4535 if (track->pid < l->min_pid)
4536 l->min_pid = track->pid;
4537 if (track->pid > l->max_pid)
4538 l->max_pid = track->pid;
4540 cpumask_set_cpu(track->cpu,
4541 to_cpumask(l->cpus));
4543 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4547 if (track->addr < caddr)
4554 * Not found. Insert new tracking element.
4556 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4562 (t->count - pos) * sizeof(struct location));
4565 l->addr = track->addr;
4569 l->min_pid = track->pid;
4570 l->max_pid = track->pid;
4571 cpumask_clear(to_cpumask(l->cpus));
4572 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4573 nodes_clear(l->nodes);
4574 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4578 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4579 struct page *page, enum track_item alloc,
4582 void *addr = page_address(page);
4585 bitmap_zero(map, page->objects);
4586 get_map(s, page, map);
4588 for_each_object(p, s, addr, page->objects)
4589 if (!test_bit(slab_index(p, s, addr), map))
4590 add_location(t, s, get_track(s, p, alloc));
4593 static int list_locations(struct kmem_cache *s, char *buf,
4594 enum track_item alloc)
4598 struct loc_track t = { 0, 0, NULL };
4600 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4601 sizeof(unsigned long), GFP_KERNEL);
4602 struct kmem_cache_node *n;
4604 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4607 return sprintf(buf, "Out of memory\n");
4609 /* Push back cpu slabs */
4612 for_each_kmem_cache_node(s, node, n) {
4613 unsigned long flags;
4616 if (!atomic_long_read(&n->nr_slabs))
4619 raw_spin_lock_irqsave(&n->list_lock, flags);
4620 list_for_each_entry(page, &n->partial, lru)
4621 process_slab(&t, s, page, alloc, map);
4622 list_for_each_entry(page, &n->full, lru)
4623 process_slab(&t, s, page, alloc, map);
4624 raw_spin_unlock_irqrestore(&n->list_lock, flags);
4627 for (i = 0; i < t.count; i++) {
4628 struct location *l = &t.loc[i];
4630 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4632 len += sprintf(buf + len, "%7ld ", l->count);
4635 len += sprintf(buf + len, "%pS", (void *)l->addr);
4637 len += sprintf(buf + len, "<not-available>");
4639 if (l->sum_time != l->min_time) {
4640 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4642 (long)div_u64(l->sum_time, l->count),
4645 len += sprintf(buf + len, " age=%ld",
4648 if (l->min_pid != l->max_pid)
4649 len += sprintf(buf + len, " pid=%ld-%ld",
4650 l->min_pid, l->max_pid);
4652 len += sprintf(buf + len, " pid=%ld",
4655 if (num_online_cpus() > 1 &&
4656 !cpumask_empty(to_cpumask(l->cpus)) &&
4657 len < PAGE_SIZE - 60)
4658 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4660 cpumask_pr_args(to_cpumask(l->cpus)));
4662 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4663 len < PAGE_SIZE - 60)
4664 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4666 nodemask_pr_args(&l->nodes));
4668 len += sprintf(buf + len, "\n");
4674 len += sprintf(buf, "No data\n");
4679 #ifdef SLUB_RESILIENCY_TEST
4680 static void __init resiliency_test(void)
4684 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4686 pr_err("SLUB resiliency testing\n");
4687 pr_err("-----------------------\n");
4688 pr_err("A. Corruption after allocation\n");
4690 p = kzalloc(16, GFP_KERNEL);
4692 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4695 validate_slab_cache(kmalloc_caches[4]);
4697 /* Hmmm... The next two are dangerous */
4698 p = kzalloc(32, GFP_KERNEL);
4699 p[32 + sizeof(void *)] = 0x34;
4700 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4702 pr_err("If allocated object is overwritten then not detectable\n\n");
4704 validate_slab_cache(kmalloc_caches[5]);
4705 p = kzalloc(64, GFP_KERNEL);
4706 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4708 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4710 pr_err("If allocated object is overwritten then not detectable\n\n");
4711 validate_slab_cache(kmalloc_caches[6]);
4713 pr_err("\nB. Corruption after free\n");
4714 p = kzalloc(128, GFP_KERNEL);
4717 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4718 validate_slab_cache(kmalloc_caches[7]);
4720 p = kzalloc(256, GFP_KERNEL);
4723 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4724 validate_slab_cache(kmalloc_caches[8]);
4726 p = kzalloc(512, GFP_KERNEL);
4729 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4730 validate_slab_cache(kmalloc_caches[9]);
4734 static void resiliency_test(void) {};
4739 enum slab_stat_type {
4740 SL_ALL, /* All slabs */
4741 SL_PARTIAL, /* Only partially allocated slabs */
4742 SL_CPU, /* Only slabs used for cpu caches */
4743 SL_OBJECTS, /* Determine allocated objects not slabs */
4744 SL_TOTAL /* Determine object capacity not slabs */
4747 #define SO_ALL (1 << SL_ALL)
4748 #define SO_PARTIAL (1 << SL_PARTIAL)
4749 #define SO_CPU (1 << SL_CPU)
4750 #define SO_OBJECTS (1 << SL_OBJECTS)
4751 #define SO_TOTAL (1 << SL_TOTAL)
4753 static ssize_t show_slab_objects(struct kmem_cache *s,
4754 char *buf, unsigned long flags)
4756 unsigned long total = 0;
4759 unsigned long *nodes;
4761 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4765 if (flags & SO_CPU) {
4768 for_each_possible_cpu(cpu) {
4769 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4774 page = READ_ONCE(c->page);
4778 node = page_to_nid(page);
4779 if (flags & SO_TOTAL)
4781 else if (flags & SO_OBJECTS)
4789 page = READ_ONCE(c->partial);
4791 node = page_to_nid(page);
4792 if (flags & SO_TOTAL)
4794 else if (flags & SO_OBJECTS)
4805 #ifdef CONFIG_SLUB_DEBUG
4806 if (flags & SO_ALL) {
4807 struct kmem_cache_node *n;
4809 for_each_kmem_cache_node(s, node, n) {
4811 if (flags & SO_TOTAL)
4812 x = atomic_long_read(&n->total_objects);
4813 else if (flags & SO_OBJECTS)
4814 x = atomic_long_read(&n->total_objects) -
4815 count_partial(n, count_free);
4817 x = atomic_long_read(&n->nr_slabs);
4824 if (flags & SO_PARTIAL) {
4825 struct kmem_cache_node *n;
4827 for_each_kmem_cache_node(s, node, n) {
4828 if (flags & SO_TOTAL)
4829 x = count_partial(n, count_total);
4830 else if (flags & SO_OBJECTS)
4831 x = count_partial(n, count_inuse);
4838 x = sprintf(buf, "%lu", total);
4840 for (node = 0; node < nr_node_ids; node++)
4842 x += sprintf(buf + x, " N%d=%lu",
4847 return x + sprintf(buf + x, "\n");
4850 #ifdef CONFIG_SLUB_DEBUG
4851 static int any_slab_objects(struct kmem_cache *s)
4854 struct kmem_cache_node *n;
4856 for_each_kmem_cache_node(s, node, n)
4857 if (atomic_long_read(&n->total_objects))
4864 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4865 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4867 struct slab_attribute {
4868 struct attribute attr;
4869 ssize_t (*show)(struct kmem_cache *s, char *buf);
4870 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4873 #define SLAB_ATTR_RO(_name) \
4874 static struct slab_attribute _name##_attr = \
4875 __ATTR(_name, 0400, _name##_show, NULL)
4877 #define SLAB_ATTR(_name) \
4878 static struct slab_attribute _name##_attr = \
4879 __ATTR(_name, 0600, _name##_show, _name##_store)
4881 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4883 return sprintf(buf, "%d\n", s->size);
4885 SLAB_ATTR_RO(slab_size);
4887 static ssize_t align_show(struct kmem_cache *s, char *buf)
4889 return sprintf(buf, "%d\n", s->align);
4891 SLAB_ATTR_RO(align);
4893 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4895 return sprintf(buf, "%d\n", s->object_size);
4897 SLAB_ATTR_RO(object_size);
4899 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4901 return sprintf(buf, "%d\n", oo_objects(s->oo));
4903 SLAB_ATTR_RO(objs_per_slab);
4905 static ssize_t order_store(struct kmem_cache *s,
4906 const char *buf, size_t length)
4908 unsigned long order;
4911 err = kstrtoul(buf, 10, &order);
4915 if (order > slub_max_order || order < slub_min_order)
4918 calculate_sizes(s, order);
4922 static ssize_t order_show(struct kmem_cache *s, char *buf)
4924 return sprintf(buf, "%d\n", oo_order(s->oo));
4928 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4930 return sprintf(buf, "%lu\n", s->min_partial);
4933 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4939 err = kstrtoul(buf, 10, &min);
4943 set_min_partial(s, min);
4946 SLAB_ATTR(min_partial);
4948 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4950 return sprintf(buf, "%u\n", s->cpu_partial);
4953 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4956 unsigned long objects;
4959 err = kstrtoul(buf, 10, &objects);
4962 if (objects && !kmem_cache_has_cpu_partial(s))
4965 s->cpu_partial = objects;
4969 SLAB_ATTR(cpu_partial);
4971 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4975 return sprintf(buf, "%pS\n", s->ctor);
4979 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4981 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4983 SLAB_ATTR_RO(aliases);
4985 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4987 return show_slab_objects(s, buf, SO_PARTIAL);
4989 SLAB_ATTR_RO(partial);
4991 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4993 return show_slab_objects(s, buf, SO_CPU);
4995 SLAB_ATTR_RO(cpu_slabs);
4997 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4999 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5001 SLAB_ATTR_RO(objects);
5003 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5005 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5007 SLAB_ATTR_RO(objects_partial);
5009 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5016 for_each_online_cpu(cpu) {
5017 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
5020 pages += page->pages;
5021 objects += page->pobjects;
5025 len = sprintf(buf, "%d(%d)", objects, pages);
5028 for_each_online_cpu(cpu) {
5029 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
5031 if (page && len < PAGE_SIZE - 20)
5032 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5033 page->pobjects, page->pages);
5036 return len + sprintf(buf + len, "\n");
5038 SLAB_ATTR_RO(slabs_cpu_partial);
5040 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5042 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5045 static ssize_t reclaim_account_store(struct kmem_cache *s,
5046 const char *buf, size_t length)
5048 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5050 s->flags |= SLAB_RECLAIM_ACCOUNT;
5053 SLAB_ATTR(reclaim_account);
5055 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5057 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5059 SLAB_ATTR_RO(hwcache_align);
5061 #ifdef CONFIG_ZONE_DMA
5062 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5064 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5066 SLAB_ATTR_RO(cache_dma);
5069 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5071 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
5073 SLAB_ATTR_RO(destroy_by_rcu);
5075 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
5077 return sprintf(buf, "%d\n", s->reserved);
5079 SLAB_ATTR_RO(reserved);
5081 #ifdef CONFIG_SLUB_DEBUG
5082 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5084 return show_slab_objects(s, buf, SO_ALL);
5086 SLAB_ATTR_RO(slabs);
5088 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5090 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5092 SLAB_ATTR_RO(total_objects);
5094 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5096 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5099 static ssize_t sanity_checks_store(struct kmem_cache *s,
5100 const char *buf, size_t length)
5102 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5103 if (buf[0] == '1') {
5104 s->flags &= ~__CMPXCHG_DOUBLE;
5105 s->flags |= SLAB_CONSISTENCY_CHECKS;
5109 SLAB_ATTR(sanity_checks);
5111 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5113 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5116 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5120 * Tracing a merged cache is going to give confusing results
5121 * as well as cause other issues like converting a mergeable
5122 * cache into an umergeable one.
5124 if (s->refcount > 1)
5127 s->flags &= ~SLAB_TRACE;
5128 if (buf[0] == '1') {
5129 s->flags &= ~__CMPXCHG_DOUBLE;
5130 s->flags |= SLAB_TRACE;
5136 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5138 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5141 static ssize_t red_zone_store(struct kmem_cache *s,
5142 const char *buf, size_t length)
5144 if (any_slab_objects(s))
5147 s->flags &= ~SLAB_RED_ZONE;
5148 if (buf[0] == '1') {
5149 s->flags |= SLAB_RED_ZONE;
5151 calculate_sizes(s, -1);
5154 SLAB_ATTR(red_zone);
5156 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5158 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5161 static ssize_t poison_store(struct kmem_cache *s,
5162 const char *buf, size_t length)
5164 if (any_slab_objects(s))
5167 s->flags &= ~SLAB_POISON;
5168 if (buf[0] == '1') {
5169 s->flags |= SLAB_POISON;
5171 calculate_sizes(s, -1);
5176 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5178 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5181 static ssize_t store_user_store(struct kmem_cache *s,
5182 const char *buf, size_t length)
5184 if (any_slab_objects(s))
5187 s->flags &= ~SLAB_STORE_USER;
5188 if (buf[0] == '1') {
5189 s->flags &= ~__CMPXCHG_DOUBLE;
5190 s->flags |= SLAB_STORE_USER;
5192 calculate_sizes(s, -1);
5195 SLAB_ATTR(store_user);
5197 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5202 static ssize_t validate_store(struct kmem_cache *s,
5203 const char *buf, size_t length)
5207 if (buf[0] == '1') {
5208 ret = validate_slab_cache(s);
5214 SLAB_ATTR(validate);
5216 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5218 if (!(s->flags & SLAB_STORE_USER))
5220 return list_locations(s, buf, TRACK_ALLOC);
5222 SLAB_ATTR_RO(alloc_calls);
5224 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5226 if (!(s->flags & SLAB_STORE_USER))
5228 return list_locations(s, buf, TRACK_FREE);
5230 SLAB_ATTR_RO(free_calls);
5231 #endif /* CONFIG_SLUB_DEBUG */
5233 #ifdef CONFIG_FAILSLAB
5234 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5236 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5239 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5242 if (s->refcount > 1)
5245 s->flags &= ~SLAB_FAILSLAB;
5247 s->flags |= SLAB_FAILSLAB;
5250 SLAB_ATTR(failslab);
5253 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5258 static ssize_t shrink_store(struct kmem_cache *s,
5259 const char *buf, size_t length)
5262 kmem_cache_shrink(s);
5270 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5272 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5275 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5276 const char *buf, size_t length)
5278 unsigned long ratio;
5281 err = kstrtoul(buf, 10, &ratio);
5286 s->remote_node_defrag_ratio = ratio * 10;
5290 SLAB_ATTR(remote_node_defrag_ratio);
5293 #ifdef CONFIG_SLUB_STATS
5294 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5296 unsigned long sum = 0;
5299 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5304 for_each_online_cpu(cpu) {
5305 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5311 len = sprintf(buf, "%lu", sum);
5314 for_each_online_cpu(cpu) {
5315 if (data[cpu] && len < PAGE_SIZE - 20)
5316 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5320 return len + sprintf(buf + len, "\n");
5323 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5327 for_each_online_cpu(cpu)
5328 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5331 #define STAT_ATTR(si, text) \
5332 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5334 return show_stat(s, buf, si); \
5336 static ssize_t text##_store(struct kmem_cache *s, \
5337 const char *buf, size_t length) \
5339 if (buf[0] != '0') \
5341 clear_stat(s, si); \
5346 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5347 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5348 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5349 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5350 STAT_ATTR(FREE_FROZEN, free_frozen);
5351 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5352 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5353 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5354 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5355 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5356 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5357 STAT_ATTR(FREE_SLAB, free_slab);
5358 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5359 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5360 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5361 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5362 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5363 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5364 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5365 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5366 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5367 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5368 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5369 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5370 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5371 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5374 static struct attribute *slab_attrs[] = {
5375 &slab_size_attr.attr,
5376 &object_size_attr.attr,
5377 &objs_per_slab_attr.attr,
5379 &min_partial_attr.attr,
5380 &cpu_partial_attr.attr,
5382 &objects_partial_attr.attr,
5384 &cpu_slabs_attr.attr,
5388 &hwcache_align_attr.attr,
5389 &reclaim_account_attr.attr,
5390 &destroy_by_rcu_attr.attr,
5392 &reserved_attr.attr,
5393 &slabs_cpu_partial_attr.attr,
5394 #ifdef CONFIG_SLUB_DEBUG
5395 &total_objects_attr.attr,
5397 &sanity_checks_attr.attr,
5399 &red_zone_attr.attr,
5401 &store_user_attr.attr,
5402 &validate_attr.attr,
5403 &alloc_calls_attr.attr,
5404 &free_calls_attr.attr,
5406 #ifdef CONFIG_ZONE_DMA
5407 &cache_dma_attr.attr,
5410 &remote_node_defrag_ratio_attr.attr,
5412 #ifdef CONFIG_SLUB_STATS
5413 &alloc_fastpath_attr.attr,
5414 &alloc_slowpath_attr.attr,
5415 &free_fastpath_attr.attr,
5416 &free_slowpath_attr.attr,
5417 &free_frozen_attr.attr,
5418 &free_add_partial_attr.attr,
5419 &free_remove_partial_attr.attr,
5420 &alloc_from_partial_attr.attr,
5421 &alloc_slab_attr.attr,
5422 &alloc_refill_attr.attr,
5423 &alloc_node_mismatch_attr.attr,
5424 &free_slab_attr.attr,
5425 &cpuslab_flush_attr.attr,
5426 &deactivate_full_attr.attr,
5427 &deactivate_empty_attr.attr,
5428 &deactivate_to_head_attr.attr,
5429 &deactivate_to_tail_attr.attr,
5430 &deactivate_remote_frees_attr.attr,
5431 &deactivate_bypass_attr.attr,
5432 &order_fallback_attr.attr,
5433 &cmpxchg_double_fail_attr.attr,
5434 &cmpxchg_double_cpu_fail_attr.attr,
5435 &cpu_partial_alloc_attr.attr,
5436 &cpu_partial_free_attr.attr,
5437 &cpu_partial_node_attr.attr,
5438 &cpu_partial_drain_attr.attr,
5440 #ifdef CONFIG_FAILSLAB
5441 &failslab_attr.attr,
5447 static struct attribute_group slab_attr_group = {
5448 .attrs = slab_attrs,
5451 static ssize_t slab_attr_show(struct kobject *kobj,
5452 struct attribute *attr,
5455 struct slab_attribute *attribute;
5456 struct kmem_cache *s;
5459 attribute = to_slab_attr(attr);
5462 if (!attribute->show)
5465 err = attribute->show(s, buf);
5470 static ssize_t slab_attr_store(struct kobject *kobj,
5471 struct attribute *attr,
5472 const char *buf, size_t len)
5474 struct slab_attribute *attribute;
5475 struct kmem_cache *s;
5478 attribute = to_slab_attr(attr);
5481 if (!attribute->store)
5484 err = attribute->store(s, buf, len);
5486 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5487 struct kmem_cache *c;
5489 mutex_lock(&slab_mutex);
5490 if (s->max_attr_size < len)
5491 s->max_attr_size = len;
5494 * This is a best effort propagation, so this function's return
5495 * value will be determined by the parent cache only. This is
5496 * basically because not all attributes will have a well
5497 * defined semantics for rollbacks - most of the actions will
5498 * have permanent effects.
5500 * Returning the error value of any of the children that fail
5501 * is not 100 % defined, in the sense that users seeing the
5502 * error code won't be able to know anything about the state of
5505 * Only returning the error code for the parent cache at least
5506 * has well defined semantics. The cache being written to
5507 * directly either failed or succeeded, in which case we loop
5508 * through the descendants with best-effort propagation.
5510 for_each_memcg_cache(c, s)
5511 attribute->store(c, buf, len);
5512 mutex_unlock(&slab_mutex);
5518 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5522 char *buffer = NULL;
5523 struct kmem_cache *root_cache;
5525 if (is_root_cache(s))
5528 root_cache = s->memcg_params.root_cache;
5531 * This mean this cache had no attribute written. Therefore, no point
5532 * in copying default values around
5534 if (!root_cache->max_attr_size)
5537 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5540 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5542 if (!attr || !attr->store || !attr->show)
5546 * It is really bad that we have to allocate here, so we will
5547 * do it only as a fallback. If we actually allocate, though,
5548 * we can just use the allocated buffer until the end.
5550 * Most of the slub attributes will tend to be very small in
5551 * size, but sysfs allows buffers up to a page, so they can
5552 * theoretically happen.
5556 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5559 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5560 if (WARN_ON(!buffer))
5565 attr->show(root_cache, buf);
5566 attr->store(s, buf, strlen(buf));
5570 free_page((unsigned long)buffer);
5574 static void kmem_cache_release(struct kobject *k)
5576 slab_kmem_cache_release(to_slab(k));
5579 static const struct sysfs_ops slab_sysfs_ops = {
5580 .show = slab_attr_show,
5581 .store = slab_attr_store,
5584 static struct kobj_type slab_ktype = {
5585 .sysfs_ops = &slab_sysfs_ops,
5586 .release = kmem_cache_release,
5589 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5591 struct kobj_type *ktype = get_ktype(kobj);
5593 if (ktype == &slab_ktype)
5598 static const struct kset_uevent_ops slab_uevent_ops = {
5599 .filter = uevent_filter,
5602 static struct kset *slab_kset;
5604 static inline struct kset *cache_kset(struct kmem_cache *s)
5607 if (!is_root_cache(s))
5608 return s->memcg_params.root_cache->memcg_kset;
5613 #define ID_STR_LENGTH 64
5615 /* Create a unique string id for a slab cache:
5617 * Format :[flags-]size
5619 static char *create_unique_id(struct kmem_cache *s)
5621 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5628 * First flags affecting slabcache operations. We will only
5629 * get here for aliasable slabs so we do not need to support
5630 * too many flags. The flags here must cover all flags that
5631 * are matched during merging to guarantee that the id is
5634 if (s->flags & SLAB_CACHE_DMA)
5636 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5638 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5640 if (!(s->flags & SLAB_NOTRACK))
5642 if (s->flags & SLAB_ACCOUNT)
5646 p += sprintf(p, "%07d", s->size);
5648 BUG_ON(p > name + ID_STR_LENGTH - 1);
5652 static int sysfs_slab_add(struct kmem_cache *s)
5656 int unmergeable = slab_unmergeable(s);
5660 * Slabcache can never be merged so we can use the name proper.
5661 * This is typically the case for debug situations. In that
5662 * case we can catch duplicate names easily.
5664 sysfs_remove_link(&slab_kset->kobj, s->name);
5668 * Create a unique name for the slab as a target
5671 name = create_unique_id(s);
5674 s->kobj.kset = cache_kset(s);
5675 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5679 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5684 if (is_root_cache(s)) {
5685 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5686 if (!s->memcg_kset) {
5693 kobject_uevent(&s->kobj, KOBJ_ADD);
5695 /* Setup first alias */
5696 sysfs_slab_alias(s, s->name);
5703 kobject_del(&s->kobj);
5707 void sysfs_slab_remove(struct kmem_cache *s)
5709 if (slab_state < FULL)
5711 * Sysfs has not been setup yet so no need to remove the
5717 kset_unregister(s->memcg_kset);
5719 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5720 kobject_del(&s->kobj);
5721 kobject_put(&s->kobj);
5725 * Need to buffer aliases during bootup until sysfs becomes
5726 * available lest we lose that information.
5728 struct saved_alias {
5729 struct kmem_cache *s;
5731 struct saved_alias *next;
5734 static struct saved_alias *alias_list;
5736 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5738 struct saved_alias *al;
5740 if (slab_state == FULL) {
5742 * If we have a leftover link then remove it.
5744 sysfs_remove_link(&slab_kset->kobj, name);
5745 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5748 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5754 al->next = alias_list;
5759 static int __init slab_sysfs_init(void)
5761 struct kmem_cache *s;
5764 mutex_lock(&slab_mutex);
5766 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5768 mutex_unlock(&slab_mutex);
5769 pr_err("Cannot register slab subsystem.\n");
5775 list_for_each_entry(s, &slab_caches, list) {
5776 err = sysfs_slab_add(s);
5778 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5782 while (alias_list) {
5783 struct saved_alias *al = alias_list;
5785 alias_list = alias_list->next;
5786 err = sysfs_slab_alias(al->s, al->name);
5788 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5793 mutex_unlock(&slab_mutex);
5798 __initcall(slab_sysfs_init);
5799 #endif /* CONFIG_SYSFS */
5802 * The /proc/slabinfo ABI
5804 #ifdef CONFIG_SLABINFO
5805 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5807 unsigned long nr_slabs = 0;
5808 unsigned long nr_objs = 0;
5809 unsigned long nr_free = 0;
5811 struct kmem_cache_node *n;
5813 for_each_kmem_cache_node(s, node, n) {
5814 nr_slabs += node_nr_slabs(n);
5815 nr_objs += node_nr_objs(n);
5816 nr_free += count_partial(n, count_free);
5819 sinfo->active_objs = nr_objs - nr_free;
5820 sinfo->num_objs = nr_objs;
5821 sinfo->active_slabs = nr_slabs;
5822 sinfo->num_slabs = nr_slabs;
5823 sinfo->objects_per_slab = oo_objects(s->oo);
5824 sinfo->cache_order = oo_order(s->oo);
5827 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5831 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5832 size_t count, loff_t *ppos)
5836 #endif /* CONFIG_SLABINFO */