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);
217 static void sysfs_slab_remove(struct kmem_cache *s);
219 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
220 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
222 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
223 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
226 static inline void stat(const struct kmem_cache *s, enum stat_item si)
228 #ifdef CONFIG_SLUB_STATS
230 * The rmw is racy on a preemptible kernel but this is acceptable, so
231 * avoid this_cpu_add()'s irq-disable overhead.
233 raw_cpu_inc(s->cpu_slab->stat[si]);
237 /********************************************************************
238 * Core slab cache functions
239 *******************************************************************/
241 static inline void *get_freepointer(struct kmem_cache *s, void *object)
243 return *(void **)(object + s->offset);
246 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
248 prefetch(object + s->offset);
251 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
255 if (!debug_pagealloc_enabled())
256 return get_freepointer(s, object);
258 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
262 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
264 *(void **)(object + s->offset) = fp;
267 /* Loop over all objects in a slab */
268 #define for_each_object(__p, __s, __addr, __objects) \
269 for (__p = fixup_red_left(__s, __addr); \
270 __p < (__addr) + (__objects) * (__s)->size; \
273 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
274 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
275 __idx <= __objects; \
276 __p += (__s)->size, __idx++)
278 /* Determine object index from a given position */
279 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
281 return (p - addr) / s->size;
284 static inline int order_objects(int order, unsigned long size, int reserved)
286 return ((PAGE_SIZE << order) - reserved) / size;
289 static inline struct kmem_cache_order_objects oo_make(int order,
290 unsigned long size, int reserved)
292 struct kmem_cache_order_objects x = {
293 (order << OO_SHIFT) + order_objects(order, size, reserved)
299 static inline int oo_order(struct kmem_cache_order_objects x)
301 return x.x >> OO_SHIFT;
304 static inline int oo_objects(struct kmem_cache_order_objects x)
306 return x.x & OO_MASK;
310 * Per slab locking using the pagelock
312 static __always_inline void slab_lock(struct page *page)
314 VM_BUG_ON_PAGE(PageTail(page), page);
315 bit_spin_lock(PG_locked, &page->flags);
318 static __always_inline void slab_unlock(struct page *page)
320 VM_BUG_ON_PAGE(PageTail(page), page);
321 __bit_spin_unlock(PG_locked, &page->flags);
324 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
327 tmp.counters = counters_new;
329 * page->counters can cover frozen/inuse/objects as well
330 * as page->_refcount. If we assign to ->counters directly
331 * we run the risk of losing updates to page->_refcount, so
332 * be careful and only assign to the fields we need.
334 page->frozen = tmp.frozen;
335 page->inuse = tmp.inuse;
336 page->objects = tmp.objects;
339 /* Interrupts must be disabled (for the fallback code to work right) */
340 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
341 void *freelist_old, unsigned long counters_old,
342 void *freelist_new, unsigned long counters_new,
345 VM_BUG_ON(!irqs_disabled());
346 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
347 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
348 if (s->flags & __CMPXCHG_DOUBLE) {
349 if (cmpxchg_double(&page->freelist, &page->counters,
350 freelist_old, counters_old,
351 freelist_new, counters_new))
357 if (page->freelist == freelist_old &&
358 page->counters == counters_old) {
359 page->freelist = freelist_new;
360 set_page_slub_counters(page, counters_new);
368 stat(s, CMPXCHG_DOUBLE_FAIL);
370 #ifdef SLUB_DEBUG_CMPXCHG
371 pr_info("%s %s: cmpxchg double redo ", n, s->name);
377 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
378 void *freelist_old, unsigned long counters_old,
379 void *freelist_new, unsigned long counters_new,
382 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
383 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
384 if (s->flags & __CMPXCHG_DOUBLE) {
385 if (cmpxchg_double(&page->freelist, &page->counters,
386 freelist_old, counters_old,
387 freelist_new, counters_new))
394 local_irq_save(flags);
396 if (page->freelist == freelist_old &&
397 page->counters == counters_old) {
398 page->freelist = freelist_new;
399 set_page_slub_counters(page, counters_new);
401 local_irq_restore(flags);
405 local_irq_restore(flags);
409 stat(s, CMPXCHG_DOUBLE_FAIL);
411 #ifdef SLUB_DEBUG_CMPXCHG
412 pr_info("%s %s: cmpxchg double redo ", n, s->name);
418 #ifdef CONFIG_SLUB_DEBUG
420 * Determine a map of object in use on a page.
422 * Node listlock must be held to guarantee that the page does
423 * not vanish from under us.
425 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
428 void *addr = page_address(page);
430 for (p = page->freelist; p; p = get_freepointer(s, p))
431 set_bit(slab_index(p, s, addr), map);
434 static inline int size_from_object(struct kmem_cache *s)
436 if (s->flags & SLAB_RED_ZONE)
437 return s->size - s->red_left_pad;
442 static inline void *restore_red_left(struct kmem_cache *s, void *p)
444 if (s->flags & SLAB_RED_ZONE)
445 p -= s->red_left_pad;
453 #if defined(CONFIG_SLUB_DEBUG_ON)
454 static int slub_debug = DEBUG_DEFAULT_FLAGS;
456 static int slub_debug;
459 static char *slub_debug_slabs;
460 static int disable_higher_order_debug;
463 * slub is about to manipulate internal object metadata. This memory lies
464 * outside the range of the allocated object, so accessing it would normally
465 * be reported by kasan as a bounds error. metadata_access_enable() is used
466 * to tell kasan that these accesses are OK.
468 static inline void metadata_access_enable(void)
470 kasan_disable_current();
473 static inline void metadata_access_disable(void)
475 kasan_enable_current();
482 /* Verify that a pointer has an address that is valid within a slab page */
483 static inline int check_valid_pointer(struct kmem_cache *s,
484 struct page *page, void *object)
491 base = page_address(page);
492 object = restore_red_left(s, object);
493 if (object < base || object >= base + page->objects * s->size ||
494 (object - base) % s->size) {
501 static void print_section(char *level, char *text, u8 *addr,
504 metadata_access_enable();
505 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
507 metadata_access_disable();
510 static struct track *get_track(struct kmem_cache *s, void *object,
511 enum track_item alloc)
516 p = object + s->offset + sizeof(void *);
518 p = object + s->inuse;
523 static void set_track(struct kmem_cache *s, void *object,
524 enum track_item alloc, unsigned long addr)
526 struct track *p = get_track(s, object, alloc);
529 #ifdef CONFIG_STACKTRACE
530 struct stack_trace trace;
533 trace.nr_entries = 0;
534 trace.max_entries = TRACK_ADDRS_COUNT;
535 trace.entries = p->addrs;
537 metadata_access_enable();
538 save_stack_trace(&trace);
539 metadata_access_disable();
541 /* See rant in lockdep.c */
542 if (trace.nr_entries != 0 &&
543 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
546 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
550 p->cpu = smp_processor_id();
551 p->pid = current->pid;
554 memset(p, 0, sizeof(struct track));
557 static void init_tracking(struct kmem_cache *s, void *object)
559 if (!(s->flags & SLAB_STORE_USER))
562 set_track(s, object, TRACK_FREE, 0UL);
563 set_track(s, object, TRACK_ALLOC, 0UL);
566 static void print_track(const char *s, struct track *t)
571 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
572 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
573 #ifdef CONFIG_STACKTRACE
576 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
578 pr_err("\t%pS\n", (void *)t->addrs[i]);
585 static void print_tracking(struct kmem_cache *s, void *object)
587 if (!(s->flags & SLAB_STORE_USER))
590 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
591 print_track("Freed", get_track(s, object, TRACK_FREE));
594 static void print_page_info(struct page *page)
596 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
597 page, page->objects, page->inuse, page->freelist, page->flags);
601 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
603 struct va_format vaf;
609 pr_err("=============================================================================\n");
610 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
611 pr_err("-----------------------------------------------------------------------------\n\n");
613 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
617 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
619 struct va_format vaf;
625 pr_err("FIX %s: %pV\n", s->name, &vaf);
629 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
631 unsigned int off; /* Offset of last byte */
632 u8 *addr = page_address(page);
634 print_tracking(s, p);
636 print_page_info(page);
638 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
639 p, p - addr, get_freepointer(s, p));
641 if (s->flags & SLAB_RED_ZONE)
642 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
644 else if (p > addr + 16)
645 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
647 print_section(KERN_ERR, "Object ", p,
648 min_t(unsigned long, s->object_size, PAGE_SIZE));
649 if (s->flags & SLAB_RED_ZONE)
650 print_section(KERN_ERR, "Redzone ", p + s->object_size,
651 s->inuse - s->object_size);
654 off = s->offset + sizeof(void *);
658 if (s->flags & SLAB_STORE_USER)
659 off += 2 * sizeof(struct track);
661 off += kasan_metadata_size(s);
663 if (off != size_from_object(s))
664 /* Beginning of the filler is the free pointer */
665 print_section(KERN_ERR, "Padding ", p + off,
666 size_from_object(s) - off);
671 void object_err(struct kmem_cache *s, struct page *page,
672 u8 *object, char *reason)
674 slab_bug(s, "%s", reason);
675 print_trailer(s, page, object);
678 static void slab_err(struct kmem_cache *s, struct page *page,
679 const char *fmt, ...)
685 vsnprintf(buf, sizeof(buf), fmt, args);
687 slab_bug(s, "%s", buf);
688 print_page_info(page);
692 static void init_object(struct kmem_cache *s, void *object, u8 val)
696 if (s->flags & SLAB_RED_ZONE)
697 memset(p - s->red_left_pad, val, s->red_left_pad);
699 if (s->flags & __OBJECT_POISON) {
700 memset(p, POISON_FREE, s->object_size - 1);
701 p[s->object_size - 1] = POISON_END;
704 if (s->flags & SLAB_RED_ZONE)
705 memset(p + s->object_size, val, s->inuse - s->object_size);
708 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
709 void *from, void *to)
711 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
712 memset(from, data, to - from);
715 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
716 u8 *object, char *what,
717 u8 *start, unsigned int value, unsigned int bytes)
722 metadata_access_enable();
723 fault = memchr_inv(start, value, bytes);
724 metadata_access_disable();
729 while (end > fault && end[-1] == value)
732 slab_bug(s, "%s overwritten", what);
733 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
734 fault, end - 1, fault[0], value);
735 print_trailer(s, page, object);
737 restore_bytes(s, what, value, fault, end);
745 * Bytes of the object to be managed.
746 * If the freepointer may overlay the object then the free
747 * pointer is the first word of the object.
749 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
752 * object + s->object_size
753 * Padding to reach word boundary. This is also used for Redzoning.
754 * Padding is extended by another word if Redzoning is enabled and
755 * object_size == inuse.
757 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
758 * 0xcc (RED_ACTIVE) for objects in use.
761 * Meta data starts here.
763 * A. Free pointer (if we cannot overwrite object on free)
764 * B. Tracking data for SLAB_STORE_USER
765 * C. Padding to reach required alignment boundary or at mininum
766 * one word if debugging is on to be able to detect writes
767 * before the word boundary.
769 * Padding is done using 0x5a (POISON_INUSE)
772 * Nothing is used beyond s->size.
774 * If slabcaches are merged then the object_size and inuse boundaries are mostly
775 * ignored. And therefore no slab options that rely on these boundaries
776 * may be used with merged slabcaches.
779 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
781 unsigned long off = s->inuse; /* The end of info */
784 /* Freepointer is placed after the object. */
785 off += sizeof(void *);
787 if (s->flags & SLAB_STORE_USER)
788 /* We also have user information there */
789 off += 2 * sizeof(struct track);
791 off += kasan_metadata_size(s);
793 if (size_from_object(s) == off)
796 return check_bytes_and_report(s, page, p, "Object padding",
797 p + off, POISON_INUSE, size_from_object(s) - off);
800 /* Check the pad bytes at the end of a slab page */
801 static int slab_pad_check(struct kmem_cache *s, struct page *page)
809 if (!(s->flags & SLAB_POISON))
812 start = page_address(page);
813 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
814 end = start + length;
815 remainder = length % s->size;
819 metadata_access_enable();
820 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
821 metadata_access_disable();
824 while (end > fault && end[-1] == POISON_INUSE)
827 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
828 print_section(KERN_ERR, "Padding ", end - remainder, remainder);
830 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
834 static int check_object(struct kmem_cache *s, struct page *page,
835 void *object, u8 val)
838 u8 *endobject = object + s->object_size;
840 if (s->flags & SLAB_RED_ZONE) {
841 if (!check_bytes_and_report(s, page, object, "Redzone",
842 object - s->red_left_pad, val, s->red_left_pad))
845 if (!check_bytes_and_report(s, page, object, "Redzone",
846 endobject, val, s->inuse - s->object_size))
849 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
850 check_bytes_and_report(s, page, p, "Alignment padding",
851 endobject, POISON_INUSE,
852 s->inuse - s->object_size);
856 if (s->flags & SLAB_POISON) {
857 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
858 (!check_bytes_and_report(s, page, p, "Poison", p,
859 POISON_FREE, s->object_size - 1) ||
860 !check_bytes_and_report(s, page, p, "Poison",
861 p + s->object_size - 1, POISON_END, 1)))
864 * check_pad_bytes cleans up on its own.
866 check_pad_bytes(s, page, p);
869 if (!s->offset && val == SLUB_RED_ACTIVE)
871 * Object and freepointer overlap. Cannot check
872 * freepointer while object is allocated.
876 /* Check free pointer validity */
877 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
878 object_err(s, page, p, "Freepointer corrupt");
880 * No choice but to zap it and thus lose the remainder
881 * of the free objects in this slab. May cause
882 * another error because the object count is now wrong.
884 set_freepointer(s, p, NULL);
890 static int check_slab(struct kmem_cache *s, struct page *page)
894 VM_BUG_ON(!irqs_disabled());
896 if (!PageSlab(page)) {
897 slab_err(s, page, "Not a valid slab page");
901 maxobj = order_objects(compound_order(page), s->size, s->reserved);
902 if (page->objects > maxobj) {
903 slab_err(s, page, "objects %u > max %u",
904 page->objects, maxobj);
907 if (page->inuse > page->objects) {
908 slab_err(s, page, "inuse %u > max %u",
909 page->inuse, page->objects);
912 /* Slab_pad_check fixes things up after itself */
913 slab_pad_check(s, page);
918 * Determine if a certain object on a page is on the freelist. Must hold the
919 * slab lock to guarantee that the chains are in a consistent state.
921 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
929 while (fp && nr <= page->objects) {
932 if (!check_valid_pointer(s, page, fp)) {
934 object_err(s, page, object,
935 "Freechain corrupt");
936 set_freepointer(s, object, NULL);
938 slab_err(s, page, "Freepointer corrupt");
939 page->freelist = NULL;
940 page->inuse = page->objects;
941 slab_fix(s, "Freelist cleared");
947 fp = get_freepointer(s, object);
951 max_objects = order_objects(compound_order(page), s->size, s->reserved);
952 if (max_objects > MAX_OBJS_PER_PAGE)
953 max_objects = MAX_OBJS_PER_PAGE;
955 if (page->objects != max_objects) {
956 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
957 page->objects, max_objects);
958 page->objects = max_objects;
959 slab_fix(s, "Number of objects adjusted.");
961 if (page->inuse != page->objects - nr) {
962 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
963 page->inuse, page->objects - nr);
964 page->inuse = page->objects - nr;
965 slab_fix(s, "Object count adjusted.");
967 return search == NULL;
970 static void trace(struct kmem_cache *s, struct page *page, void *object,
973 if (s->flags & SLAB_TRACE) {
974 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
976 alloc ? "alloc" : "free",
981 print_section(KERN_INFO, "Object ", (void *)object,
989 * Tracking of fully allocated slabs for debugging purposes.
991 static void add_full(struct kmem_cache *s,
992 struct kmem_cache_node *n, struct page *page)
994 if (!(s->flags & SLAB_STORE_USER))
997 lockdep_assert_held(&n->list_lock);
998 list_add(&page->lru, &n->full);
1001 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1003 if (!(s->flags & SLAB_STORE_USER))
1006 lockdep_assert_held(&n->list_lock);
1007 list_del(&page->lru);
1010 /* Tracking of the number of slabs for debugging purposes */
1011 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1013 struct kmem_cache_node *n = get_node(s, node);
1015 return atomic_long_read(&n->nr_slabs);
1018 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1020 return atomic_long_read(&n->nr_slabs);
1023 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1025 struct kmem_cache_node *n = get_node(s, node);
1028 * May be called early in order to allocate a slab for the
1029 * kmem_cache_node structure. Solve the chicken-egg
1030 * dilemma by deferring the increment of the count during
1031 * bootstrap (see early_kmem_cache_node_alloc).
1034 atomic_long_inc(&n->nr_slabs);
1035 atomic_long_add(objects, &n->total_objects);
1038 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1040 struct kmem_cache_node *n = get_node(s, node);
1042 atomic_long_dec(&n->nr_slabs);
1043 atomic_long_sub(objects, &n->total_objects);
1046 /* Object debug checks for alloc/free paths */
1047 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1050 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1053 init_object(s, object, SLUB_RED_INACTIVE);
1054 init_tracking(s, object);
1057 static inline int alloc_consistency_checks(struct kmem_cache *s,
1059 void *object, unsigned long addr)
1061 if (!check_slab(s, page))
1064 if (!check_valid_pointer(s, page, object)) {
1065 object_err(s, page, object, "Freelist Pointer check fails");
1069 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1075 static noinline int alloc_debug_processing(struct kmem_cache *s,
1077 void *object, unsigned long addr)
1079 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1080 if (!alloc_consistency_checks(s, page, object, addr))
1084 /* Success perform special debug activities for allocs */
1085 if (s->flags & SLAB_STORE_USER)
1086 set_track(s, object, TRACK_ALLOC, addr);
1087 trace(s, page, object, 1);
1088 init_object(s, object, SLUB_RED_ACTIVE);
1092 if (PageSlab(page)) {
1094 * If this is a slab page then lets do the best we can
1095 * to avoid issues in the future. Marking all objects
1096 * as used avoids touching the remaining objects.
1098 slab_fix(s, "Marking all objects used");
1099 page->inuse = page->objects;
1100 page->freelist = NULL;
1105 static inline int free_consistency_checks(struct kmem_cache *s,
1106 struct page *page, void *object, unsigned long addr)
1108 if (!check_valid_pointer(s, page, object)) {
1109 slab_err(s, page, "Invalid object pointer 0x%p", object);
1113 if (on_freelist(s, page, object)) {
1114 object_err(s, page, object, "Object already free");
1118 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1121 if (unlikely(s != page->slab_cache)) {
1122 if (!PageSlab(page)) {
1123 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1125 } else if (!page->slab_cache) {
1126 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1130 object_err(s, page, object,
1131 "page slab pointer corrupt.");
1137 /* Supports checking bulk free of a constructed freelist */
1138 static noinline int free_debug_processing(
1139 struct kmem_cache *s, struct page *page,
1140 void *head, void *tail, int bulk_cnt,
1143 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1144 void *object = head;
1146 unsigned long uninitialized_var(flags);
1149 spin_lock_irqsave(&n->list_lock, flags);
1152 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1153 if (!check_slab(s, page))
1160 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1161 if (!free_consistency_checks(s, page, object, addr))
1165 if (s->flags & SLAB_STORE_USER)
1166 set_track(s, object, TRACK_FREE, addr);
1167 trace(s, page, object, 0);
1168 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1169 init_object(s, object, SLUB_RED_INACTIVE);
1171 /* Reached end of constructed freelist yet? */
1172 if (object != tail) {
1173 object = get_freepointer(s, object);
1179 if (cnt != bulk_cnt)
1180 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1184 spin_unlock_irqrestore(&n->list_lock, flags);
1186 slab_fix(s, "Object at 0x%p not freed", object);
1190 static int __init setup_slub_debug(char *str)
1192 slub_debug = DEBUG_DEFAULT_FLAGS;
1193 if (*str++ != '=' || !*str)
1195 * No options specified. Switch on full debugging.
1201 * No options but restriction on slabs. This means full
1202 * debugging for slabs matching a pattern.
1209 * Switch off all debugging measures.
1214 * Determine which debug features should be switched on
1216 for (; *str && *str != ','; str++) {
1217 switch (tolower(*str)) {
1219 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1222 slub_debug |= SLAB_RED_ZONE;
1225 slub_debug |= SLAB_POISON;
1228 slub_debug |= SLAB_STORE_USER;
1231 slub_debug |= SLAB_TRACE;
1234 slub_debug |= SLAB_FAILSLAB;
1238 * Avoid enabling debugging on caches if its minimum
1239 * order would increase as a result.
1241 disable_higher_order_debug = 1;
1244 pr_err("slub_debug option '%c' unknown. skipped\n",
1251 slub_debug_slabs = str + 1;
1256 __setup("slub_debug", setup_slub_debug);
1258 unsigned long kmem_cache_flags(unsigned long object_size,
1259 unsigned long flags, const char *name,
1260 void (*ctor)(void *))
1263 * Enable debugging if selected on the kernel commandline.
1265 if (slub_debug && (!slub_debug_slabs || (name &&
1266 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1267 flags |= slub_debug;
1271 #else /* !CONFIG_SLUB_DEBUG */
1272 static inline void setup_object_debug(struct kmem_cache *s,
1273 struct page *page, void *object) {}
1275 static inline int alloc_debug_processing(struct kmem_cache *s,
1276 struct page *page, void *object, unsigned long addr) { return 0; }
1278 static inline int free_debug_processing(
1279 struct kmem_cache *s, struct page *page,
1280 void *head, void *tail, int bulk_cnt,
1281 unsigned long addr) { return 0; }
1283 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1285 static inline int check_object(struct kmem_cache *s, struct page *page,
1286 void *object, u8 val) { return 1; }
1287 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1288 struct page *page) {}
1289 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1290 struct page *page) {}
1291 unsigned long kmem_cache_flags(unsigned long object_size,
1292 unsigned long flags, const char *name,
1293 void (*ctor)(void *))
1297 #define slub_debug 0
1299 #define disable_higher_order_debug 0
1301 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1303 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1305 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1307 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1310 #endif /* CONFIG_SLUB_DEBUG */
1313 * Hooks for other subsystems that check memory allocations. In a typical
1314 * production configuration these hooks all should produce no code at all.
1316 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1318 kmemleak_alloc(ptr, size, 1, flags);
1319 kasan_kmalloc_large(ptr, size, flags);
1322 static inline void kfree_hook(const void *x)
1325 kasan_kfree_large(x);
1328 static inline void *slab_free_hook(struct kmem_cache *s, void *x)
1332 kmemleak_free_recursive(x, s->flags);
1335 * Trouble is that we may no longer disable interrupts in the fast path
1336 * So in order to make the debug calls that expect irqs to be
1337 * disabled we need to disable interrupts temporarily.
1339 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1341 unsigned long flags;
1343 local_irq_save(flags);
1344 kmemcheck_slab_free(s, x, s->object_size);
1345 debug_check_no_locks_freed(x, s->object_size);
1346 local_irq_restore(flags);
1349 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1350 debug_check_no_obj_freed(x, s->object_size);
1352 freeptr = get_freepointer(s, x);
1354 * kasan_slab_free() may put x into memory quarantine, delaying its
1355 * reuse. In this case the object's freelist pointer is changed.
1357 kasan_slab_free(s, x);
1361 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1362 void *head, void *tail)
1365 * Compiler cannot detect this function can be removed if slab_free_hook()
1366 * evaluates to nothing. Thus, catch all relevant config debug options here.
1368 #if defined(CONFIG_KMEMCHECK) || \
1369 defined(CONFIG_LOCKDEP) || \
1370 defined(CONFIG_DEBUG_KMEMLEAK) || \
1371 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1372 defined(CONFIG_KASAN)
1374 void *object = head;
1375 void *tail_obj = tail ? : head;
1379 freeptr = slab_free_hook(s, object);
1380 } while ((object != tail_obj) && (object = freeptr));
1384 static void setup_object(struct kmem_cache *s, struct page *page,
1387 setup_object_debug(s, page, object);
1388 kasan_init_slab_obj(s, object);
1389 if (unlikely(s->ctor)) {
1390 kasan_unpoison_object_data(s, object);
1392 kasan_poison_object_data(s, object);
1397 * Slab allocation and freeing
1399 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1400 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1403 int order = oo_order(oo);
1405 flags |= __GFP_NOTRACK;
1407 if (node == NUMA_NO_NODE)
1408 page = alloc_pages(flags, order);
1410 page = __alloc_pages_node(node, flags, order);
1412 if (page && memcg_charge_slab(page, flags, order, s)) {
1413 __free_pages(page, order);
1420 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1421 /* Pre-initialize the random sequence cache */
1422 static int init_cache_random_seq(struct kmem_cache *s)
1425 unsigned long i, count = oo_objects(s->oo);
1427 /* Bailout if already initialised */
1431 err = cache_random_seq_create(s, count, GFP_KERNEL);
1433 pr_err("SLUB: Unable to initialize free list for %s\n",
1438 /* Transform to an offset on the set of pages */
1439 if (s->random_seq) {
1440 for (i = 0; i < count; i++)
1441 s->random_seq[i] *= s->size;
1446 /* Initialize each random sequence freelist per cache */
1447 static void __init init_freelist_randomization(void)
1449 struct kmem_cache *s;
1451 mutex_lock(&slab_mutex);
1453 list_for_each_entry(s, &slab_caches, list)
1454 init_cache_random_seq(s);
1456 mutex_unlock(&slab_mutex);
1459 /* Get the next entry on the pre-computed freelist randomized */
1460 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1461 unsigned long *pos, void *start,
1462 unsigned long page_limit,
1463 unsigned long freelist_count)
1468 * If the target page allocation failed, the number of objects on the
1469 * page might be smaller than the usual size defined by the cache.
1472 idx = s->random_seq[*pos];
1474 if (*pos >= freelist_count)
1476 } while (unlikely(idx >= page_limit));
1478 return (char *)start + idx;
1481 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1482 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1487 unsigned long idx, pos, page_limit, freelist_count;
1489 if (page->objects < 2 || !s->random_seq)
1492 freelist_count = oo_objects(s->oo);
1493 pos = get_random_int() % freelist_count;
1495 page_limit = page->objects * s->size;
1496 start = fixup_red_left(s, page_address(page));
1498 /* First entry is used as the base of the freelist */
1499 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1501 page->freelist = cur;
1503 for (idx = 1; idx < page->objects; idx++) {
1504 setup_object(s, page, cur);
1505 next = next_freelist_entry(s, page, &pos, start, page_limit,
1507 set_freepointer(s, cur, next);
1510 setup_object(s, page, cur);
1511 set_freepointer(s, cur, NULL);
1516 static inline int init_cache_random_seq(struct kmem_cache *s)
1520 static inline void init_freelist_randomization(void) { }
1521 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1525 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1527 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1530 struct kmem_cache_order_objects oo = s->oo;
1536 flags &= gfp_allowed_mask;
1538 if (gfpflags_allow_blocking(flags))
1541 flags |= s->allocflags;
1544 * Let the initial higher-order allocation fail under memory pressure
1545 * so we fall-back to the minimum order allocation.
1547 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1548 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1549 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1551 page = alloc_slab_page(s, alloc_gfp, node, oo);
1552 if (unlikely(!page)) {
1556 * Allocation may have failed due to fragmentation.
1557 * Try a lower order alloc if possible
1559 page = alloc_slab_page(s, alloc_gfp, node, oo);
1560 if (unlikely(!page))
1562 stat(s, ORDER_FALLBACK);
1565 if (kmemcheck_enabled &&
1566 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1567 int pages = 1 << oo_order(oo);
1569 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1572 * Objects from caches that have a constructor don't get
1573 * cleared when they're allocated, so we need to do it here.
1576 kmemcheck_mark_uninitialized_pages(page, pages);
1578 kmemcheck_mark_unallocated_pages(page, pages);
1581 page->objects = oo_objects(oo);
1583 order = compound_order(page);
1584 page->slab_cache = s;
1585 __SetPageSlab(page);
1586 if (page_is_pfmemalloc(page))
1587 SetPageSlabPfmemalloc(page);
1589 start = page_address(page);
1591 if (unlikely(s->flags & SLAB_POISON))
1592 memset(start, POISON_INUSE, PAGE_SIZE << order);
1594 kasan_poison_slab(page);
1596 shuffle = shuffle_freelist(s, page);
1599 for_each_object_idx(p, idx, s, start, page->objects) {
1600 setup_object(s, page, p);
1601 if (likely(idx < page->objects))
1602 set_freepointer(s, p, p + s->size);
1604 set_freepointer(s, p, NULL);
1606 page->freelist = fixup_red_left(s, start);
1609 page->inuse = page->objects;
1613 if (gfpflags_allow_blocking(flags))
1614 local_irq_disable();
1618 mod_zone_page_state(page_zone(page),
1619 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1620 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1623 inc_slabs_node(s, page_to_nid(page), page->objects);
1628 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1630 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1631 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1632 flags &= ~GFP_SLAB_BUG_MASK;
1633 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1634 invalid_mask, &invalid_mask, flags, &flags);
1638 return allocate_slab(s,
1639 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1642 static void __free_slab(struct kmem_cache *s, struct page *page)
1644 int order = compound_order(page);
1645 int pages = 1 << order;
1647 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1650 slab_pad_check(s, page);
1651 for_each_object(p, s, page_address(page),
1653 check_object(s, page, p, SLUB_RED_INACTIVE);
1656 kmemcheck_free_shadow(page, compound_order(page));
1658 mod_zone_page_state(page_zone(page),
1659 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1660 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1663 __ClearPageSlabPfmemalloc(page);
1664 __ClearPageSlab(page);
1666 page_mapcount_reset(page);
1667 if (current->reclaim_state)
1668 current->reclaim_state->reclaimed_slab += pages;
1669 memcg_uncharge_slab(page, order, s);
1670 __free_pages(page, order);
1673 #define need_reserve_slab_rcu \
1674 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1676 static void rcu_free_slab(struct rcu_head *h)
1680 if (need_reserve_slab_rcu)
1681 page = virt_to_head_page(h);
1683 page = container_of((struct list_head *)h, struct page, lru);
1685 __free_slab(page->slab_cache, page);
1688 static void free_slab(struct kmem_cache *s, struct page *page)
1690 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1691 struct rcu_head *head;
1693 if (need_reserve_slab_rcu) {
1694 int order = compound_order(page);
1695 int offset = (PAGE_SIZE << order) - s->reserved;
1697 VM_BUG_ON(s->reserved != sizeof(*head));
1698 head = page_address(page) + offset;
1700 head = &page->rcu_head;
1703 call_rcu(head, rcu_free_slab);
1705 __free_slab(s, page);
1708 static void discard_slab(struct kmem_cache *s, struct page *page)
1710 dec_slabs_node(s, page_to_nid(page), page->objects);
1715 * Management of partially allocated slabs.
1718 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1721 if (tail == DEACTIVATE_TO_TAIL)
1722 list_add_tail(&page->lru, &n->partial);
1724 list_add(&page->lru, &n->partial);
1727 static inline void add_partial(struct kmem_cache_node *n,
1728 struct page *page, int tail)
1730 lockdep_assert_held(&n->list_lock);
1731 __add_partial(n, page, tail);
1734 static inline void remove_partial(struct kmem_cache_node *n,
1737 lockdep_assert_held(&n->list_lock);
1738 list_del(&page->lru);
1743 * Remove slab from the partial list, freeze it and
1744 * return the pointer to the freelist.
1746 * Returns a list of objects or NULL if it fails.
1748 static inline void *acquire_slab(struct kmem_cache *s,
1749 struct kmem_cache_node *n, struct page *page,
1750 int mode, int *objects)
1753 unsigned long counters;
1756 lockdep_assert_held(&n->list_lock);
1759 * Zap the freelist and set the frozen bit.
1760 * The old freelist is the list of objects for the
1761 * per cpu allocation list.
1763 freelist = page->freelist;
1764 counters = page->counters;
1765 new.counters = counters;
1766 *objects = new.objects - new.inuse;
1768 new.inuse = page->objects;
1769 new.freelist = NULL;
1771 new.freelist = freelist;
1774 VM_BUG_ON(new.frozen);
1777 if (!__cmpxchg_double_slab(s, page,
1779 new.freelist, new.counters,
1783 remove_partial(n, page);
1788 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1789 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1792 * Try to allocate a partial slab from a specific node.
1794 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1795 struct kmem_cache_cpu *c, gfp_t flags)
1797 struct page *page, *page2;
1798 void *object = NULL;
1803 * Racy check. If we mistakenly see no partial slabs then we
1804 * just allocate an empty slab. If we mistakenly try to get a
1805 * partial slab and there is none available then get_partials()
1808 if (!n || !n->nr_partial)
1811 spin_lock(&n->list_lock);
1812 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1815 if (!pfmemalloc_match(page, flags))
1818 t = acquire_slab(s, n, page, object == NULL, &objects);
1822 available += objects;
1825 stat(s, ALLOC_FROM_PARTIAL);
1828 put_cpu_partial(s, page, 0);
1829 stat(s, CPU_PARTIAL_NODE);
1831 if (!kmem_cache_has_cpu_partial(s)
1832 || available > s->cpu_partial / 2)
1836 spin_unlock(&n->list_lock);
1841 * Get a page from somewhere. Search in increasing NUMA distances.
1843 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1844 struct kmem_cache_cpu *c)
1847 struct zonelist *zonelist;
1850 enum zone_type high_zoneidx = gfp_zone(flags);
1852 unsigned int cpuset_mems_cookie;
1855 * The defrag ratio allows a configuration of the tradeoffs between
1856 * inter node defragmentation and node local allocations. A lower
1857 * defrag_ratio increases the tendency to do local allocations
1858 * instead of attempting to obtain partial slabs from other nodes.
1860 * If the defrag_ratio is set to 0 then kmalloc() always
1861 * returns node local objects. If the ratio is higher then kmalloc()
1862 * may return off node objects because partial slabs are obtained
1863 * from other nodes and filled up.
1865 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1866 * (which makes defrag_ratio = 1000) then every (well almost)
1867 * allocation will first attempt to defrag slab caches on other nodes.
1868 * This means scanning over all nodes to look for partial slabs which
1869 * may be expensive if we do it every time we are trying to find a slab
1870 * with available objects.
1872 if (!s->remote_node_defrag_ratio ||
1873 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1877 cpuset_mems_cookie = read_mems_allowed_begin();
1878 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1879 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1880 struct kmem_cache_node *n;
1882 n = get_node(s, zone_to_nid(zone));
1884 if (n && cpuset_zone_allowed(zone, flags) &&
1885 n->nr_partial > s->min_partial) {
1886 object = get_partial_node(s, n, c, flags);
1889 * Don't check read_mems_allowed_retry()
1890 * here - if mems_allowed was updated in
1891 * parallel, that was a harmless race
1892 * between allocation and the cpuset
1899 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1905 * Get a partial page, lock it and return it.
1907 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1908 struct kmem_cache_cpu *c)
1911 int searchnode = node;
1913 if (node == NUMA_NO_NODE)
1914 searchnode = numa_mem_id();
1915 else if (!node_present_pages(node))
1916 searchnode = node_to_mem_node(node);
1918 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1919 if (object || node != NUMA_NO_NODE)
1922 return get_any_partial(s, flags, c);
1925 #ifdef CONFIG_PREEMPT
1927 * Calculate the next globally unique transaction for disambiguiation
1928 * during cmpxchg. The transactions start with the cpu number and are then
1929 * incremented by CONFIG_NR_CPUS.
1931 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1934 * No preemption supported therefore also no need to check for
1940 static inline unsigned long next_tid(unsigned long tid)
1942 return tid + TID_STEP;
1945 static inline unsigned int tid_to_cpu(unsigned long tid)
1947 return tid % TID_STEP;
1950 static inline unsigned long tid_to_event(unsigned long tid)
1952 return tid / TID_STEP;
1955 static inline unsigned int init_tid(int cpu)
1960 static inline void note_cmpxchg_failure(const char *n,
1961 const struct kmem_cache *s, unsigned long tid)
1963 #ifdef SLUB_DEBUG_CMPXCHG
1964 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1966 pr_info("%s %s: cmpxchg redo ", n, s->name);
1968 #ifdef CONFIG_PREEMPT
1969 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1970 pr_warn("due to cpu change %d -> %d\n",
1971 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1974 if (tid_to_event(tid) != tid_to_event(actual_tid))
1975 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1976 tid_to_event(tid), tid_to_event(actual_tid));
1978 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1979 actual_tid, tid, next_tid(tid));
1981 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1984 static void init_kmem_cache_cpus(struct kmem_cache *s)
1988 for_each_possible_cpu(cpu)
1989 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1993 * Remove the cpu slab
1995 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1998 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1999 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2001 enum slab_modes l = M_NONE, m = M_NONE;
2003 int tail = DEACTIVATE_TO_HEAD;
2007 if (page->freelist) {
2008 stat(s, DEACTIVATE_REMOTE_FREES);
2009 tail = DEACTIVATE_TO_TAIL;
2013 * Stage one: Free all available per cpu objects back
2014 * to the page freelist while it is still frozen. Leave the
2017 * There is no need to take the list->lock because the page
2020 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2022 unsigned long counters;
2025 prior = page->freelist;
2026 counters = page->counters;
2027 set_freepointer(s, freelist, prior);
2028 new.counters = counters;
2030 VM_BUG_ON(!new.frozen);
2032 } while (!__cmpxchg_double_slab(s, page,
2034 freelist, new.counters,
2035 "drain percpu freelist"));
2037 freelist = nextfree;
2041 * Stage two: Ensure that the page is unfrozen while the
2042 * list presence reflects the actual number of objects
2045 * We setup the list membership and then perform a cmpxchg
2046 * with the count. If there is a mismatch then the page
2047 * is not unfrozen but the page is on the wrong list.
2049 * Then we restart the process which may have to remove
2050 * the page from the list that we just put it on again
2051 * because the number of objects in the slab may have
2056 old.freelist = page->freelist;
2057 old.counters = page->counters;
2058 VM_BUG_ON(!old.frozen);
2060 /* Determine target state of the slab */
2061 new.counters = old.counters;
2064 set_freepointer(s, freelist, old.freelist);
2065 new.freelist = freelist;
2067 new.freelist = old.freelist;
2071 if (!new.inuse && n->nr_partial >= s->min_partial)
2073 else if (new.freelist) {
2078 * Taking the spinlock removes the possiblity
2079 * that acquire_slab() will see a slab page that
2082 spin_lock(&n->list_lock);
2086 if (kmem_cache_debug(s) && !lock) {
2089 * This also ensures that the scanning of full
2090 * slabs from diagnostic functions will not see
2093 spin_lock(&n->list_lock);
2101 remove_partial(n, page);
2103 else if (l == M_FULL)
2105 remove_full(s, n, page);
2107 if (m == M_PARTIAL) {
2109 add_partial(n, page, tail);
2112 } else if (m == M_FULL) {
2114 stat(s, DEACTIVATE_FULL);
2115 add_full(s, n, page);
2121 if (!__cmpxchg_double_slab(s, page,
2122 old.freelist, old.counters,
2123 new.freelist, new.counters,
2128 spin_unlock(&n->list_lock);
2131 stat(s, DEACTIVATE_EMPTY);
2132 discard_slab(s, page);
2138 * Unfreeze all the cpu partial slabs.
2140 * This function must be called with interrupts disabled
2141 * for the cpu using c (or some other guarantee must be there
2142 * to guarantee no concurrent accesses).
2144 static void unfreeze_partials(struct kmem_cache *s,
2145 struct kmem_cache_cpu *c)
2147 #ifdef CONFIG_SLUB_CPU_PARTIAL
2148 struct kmem_cache_node *n = NULL, *n2 = NULL;
2149 struct page *page, *discard_page = NULL;
2151 while ((page = c->partial)) {
2155 c->partial = page->next;
2157 n2 = get_node(s, page_to_nid(page));
2160 spin_unlock(&n->list_lock);
2163 spin_lock(&n->list_lock);
2168 old.freelist = page->freelist;
2169 old.counters = page->counters;
2170 VM_BUG_ON(!old.frozen);
2172 new.counters = old.counters;
2173 new.freelist = old.freelist;
2177 } while (!__cmpxchg_double_slab(s, page,
2178 old.freelist, old.counters,
2179 new.freelist, new.counters,
2180 "unfreezing slab"));
2182 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2183 page->next = discard_page;
2184 discard_page = page;
2186 add_partial(n, page, DEACTIVATE_TO_TAIL);
2187 stat(s, FREE_ADD_PARTIAL);
2192 spin_unlock(&n->list_lock);
2194 while (discard_page) {
2195 page = discard_page;
2196 discard_page = discard_page->next;
2198 stat(s, DEACTIVATE_EMPTY);
2199 discard_slab(s, page);
2206 * Put a page that was just frozen (in __slab_free) into a partial page
2207 * slot if available. This is done without interrupts disabled and without
2208 * preemption disabled. The cmpxchg is racy and may put the partial page
2209 * onto a random cpus partial slot.
2211 * If we did not find a slot then simply move all the partials to the
2212 * per node partial list.
2214 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2216 #ifdef CONFIG_SLUB_CPU_PARTIAL
2217 struct page *oldpage;
2225 oldpage = this_cpu_read(s->cpu_slab->partial);
2228 pobjects = oldpage->pobjects;
2229 pages = oldpage->pages;
2230 if (drain && pobjects > s->cpu_partial) {
2231 unsigned long flags;
2233 * partial array is full. Move the existing
2234 * set to the per node partial list.
2236 local_irq_save(flags);
2237 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2238 local_irq_restore(flags);
2242 stat(s, CPU_PARTIAL_DRAIN);
2247 pobjects += page->objects - page->inuse;
2249 page->pages = pages;
2250 page->pobjects = pobjects;
2251 page->next = oldpage;
2253 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2255 if (unlikely(!s->cpu_partial)) {
2256 unsigned long flags;
2258 local_irq_save(flags);
2259 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2260 local_irq_restore(flags);
2266 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2268 stat(s, CPUSLAB_FLUSH);
2269 deactivate_slab(s, c->page, c->freelist);
2271 c->tid = next_tid(c->tid);
2279 * Called from IPI handler with interrupts disabled.
2281 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2283 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2289 unfreeze_partials(s, c);
2293 static void flush_cpu_slab(void *d)
2295 struct kmem_cache *s = d;
2297 __flush_cpu_slab(s, smp_processor_id());
2300 static bool has_cpu_slab(int cpu, void *info)
2302 struct kmem_cache *s = info;
2303 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2305 return c->page || c->partial;
2308 static void flush_all(struct kmem_cache *s)
2310 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2314 * Use the cpu notifier to insure that the cpu slabs are flushed when
2317 static int slub_cpu_dead(unsigned int cpu)
2319 struct kmem_cache *s;
2320 unsigned long flags;
2322 mutex_lock(&slab_mutex);
2323 list_for_each_entry(s, &slab_caches, list) {
2324 local_irq_save(flags);
2325 __flush_cpu_slab(s, cpu);
2326 local_irq_restore(flags);
2328 mutex_unlock(&slab_mutex);
2333 * Check if the objects in a per cpu structure fit numa
2334 * locality expectations.
2336 static inline int node_match(struct page *page, int node)
2339 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2345 #ifdef CONFIG_SLUB_DEBUG
2346 static int count_free(struct page *page)
2348 return page->objects - page->inuse;
2351 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2353 return atomic_long_read(&n->total_objects);
2355 #endif /* CONFIG_SLUB_DEBUG */
2357 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2358 static unsigned long count_partial(struct kmem_cache_node *n,
2359 int (*get_count)(struct page *))
2361 unsigned long flags;
2362 unsigned long x = 0;
2365 spin_lock_irqsave(&n->list_lock, flags);
2366 list_for_each_entry(page, &n->partial, lru)
2367 x += get_count(page);
2368 spin_unlock_irqrestore(&n->list_lock, flags);
2371 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2373 static noinline void
2374 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2376 #ifdef CONFIG_SLUB_DEBUG
2377 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2378 DEFAULT_RATELIMIT_BURST);
2380 struct kmem_cache_node *n;
2382 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2385 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2386 nid, gfpflags, &gfpflags);
2387 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2388 s->name, s->object_size, s->size, oo_order(s->oo),
2391 if (oo_order(s->min) > get_order(s->object_size))
2392 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2395 for_each_kmem_cache_node(s, node, n) {
2396 unsigned long nr_slabs;
2397 unsigned long nr_objs;
2398 unsigned long nr_free;
2400 nr_free = count_partial(n, count_free);
2401 nr_slabs = node_nr_slabs(n);
2402 nr_objs = node_nr_objs(n);
2404 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2405 node, nr_slabs, nr_objs, nr_free);
2410 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2411 int node, struct kmem_cache_cpu **pc)
2414 struct kmem_cache_cpu *c = *pc;
2417 freelist = get_partial(s, flags, node, c);
2422 page = new_slab(s, flags, node);
2424 c = raw_cpu_ptr(s->cpu_slab);
2429 * No other reference to the page yet so we can
2430 * muck around with it freely without cmpxchg
2432 freelist = page->freelist;
2433 page->freelist = NULL;
2435 stat(s, ALLOC_SLAB);
2444 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2446 if (unlikely(PageSlabPfmemalloc(page)))
2447 return gfp_pfmemalloc_allowed(gfpflags);
2453 * Check the page->freelist of a page and either transfer the freelist to the
2454 * per cpu freelist or deactivate the page.
2456 * The page is still frozen if the return value is not NULL.
2458 * If this function returns NULL then the page has been unfrozen.
2460 * This function must be called with interrupt disabled.
2462 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2465 unsigned long counters;
2469 freelist = page->freelist;
2470 counters = page->counters;
2472 new.counters = counters;
2473 VM_BUG_ON(!new.frozen);
2475 new.inuse = page->objects;
2476 new.frozen = freelist != NULL;
2478 } while (!__cmpxchg_double_slab(s, page,
2487 * Slow path. The lockless freelist is empty or we need to perform
2490 * Processing is still very fast if new objects have been freed to the
2491 * regular freelist. In that case we simply take over the regular freelist
2492 * as the lockless freelist and zap the regular freelist.
2494 * If that is not working then we fall back to the partial lists. We take the
2495 * first element of the freelist as the object to allocate now and move the
2496 * rest of the freelist to the lockless freelist.
2498 * And if we were unable to get a new slab from the partial slab lists then
2499 * we need to allocate a new slab. This is the slowest path since it involves
2500 * a call to the page allocator and the setup of a new slab.
2502 * Version of __slab_alloc to use when we know that interrupts are
2503 * already disabled (which is the case for bulk allocation).
2505 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2506 unsigned long addr, struct kmem_cache_cpu *c)
2516 if (unlikely(!node_match(page, node))) {
2517 int searchnode = node;
2519 if (node != NUMA_NO_NODE && !node_present_pages(node))
2520 searchnode = node_to_mem_node(node);
2522 if (unlikely(!node_match(page, searchnode))) {
2523 stat(s, ALLOC_NODE_MISMATCH);
2524 deactivate_slab(s, page, c->freelist);
2532 * By rights, we should be searching for a slab page that was
2533 * PFMEMALLOC but right now, we are losing the pfmemalloc
2534 * information when the page leaves the per-cpu allocator
2536 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2537 deactivate_slab(s, page, c->freelist);
2543 /* must check again c->freelist in case of cpu migration or IRQ */
2544 freelist = c->freelist;
2548 freelist = get_freelist(s, page);
2552 stat(s, DEACTIVATE_BYPASS);
2556 stat(s, ALLOC_REFILL);
2560 * freelist is pointing to the list of objects to be used.
2561 * page is pointing to the page from which the objects are obtained.
2562 * That page must be frozen for per cpu allocations to work.
2564 VM_BUG_ON(!c->page->frozen);
2565 c->freelist = get_freepointer(s, freelist);
2566 c->tid = next_tid(c->tid);
2572 page = c->page = c->partial;
2573 c->partial = page->next;
2574 stat(s, CPU_PARTIAL_ALLOC);
2578 freelist = new_slab_objects(s, gfpflags, node, &c);
2580 if (unlikely(!freelist)) {
2581 slab_out_of_memory(s, gfpflags, node);
2586 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2589 /* Only entered in the debug case */
2590 if (kmem_cache_debug(s) &&
2591 !alloc_debug_processing(s, page, freelist, addr))
2592 goto new_slab; /* Slab failed checks. Next slab needed */
2594 deactivate_slab(s, page, get_freepointer(s, freelist));
2601 * Another one that disabled interrupt and compensates for possible
2602 * cpu changes by refetching the per cpu area pointer.
2604 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2605 unsigned long addr, struct kmem_cache_cpu *c)
2608 unsigned long flags;
2610 local_irq_save(flags);
2611 #ifdef CONFIG_PREEMPT
2613 * We may have been preempted and rescheduled on a different
2614 * cpu before disabling interrupts. Need to reload cpu area
2617 c = this_cpu_ptr(s->cpu_slab);
2620 p = ___slab_alloc(s, gfpflags, node, addr, c);
2621 local_irq_restore(flags);
2626 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2627 * have the fastpath folded into their functions. So no function call
2628 * overhead for requests that can be satisfied on the fastpath.
2630 * The fastpath works by first checking if the lockless freelist can be used.
2631 * If not then __slab_alloc is called for slow processing.
2633 * Otherwise we can simply pick the next object from the lockless free list.
2635 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2636 gfp_t gfpflags, int node, unsigned long addr)
2639 struct kmem_cache_cpu *c;
2643 s = slab_pre_alloc_hook(s, gfpflags);
2648 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2649 * enabled. We may switch back and forth between cpus while
2650 * reading from one cpu area. That does not matter as long
2651 * as we end up on the original cpu again when doing the cmpxchg.
2653 * We should guarantee that tid and kmem_cache are retrieved on
2654 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2655 * to check if it is matched or not.
2658 tid = this_cpu_read(s->cpu_slab->tid);
2659 c = raw_cpu_ptr(s->cpu_slab);
2660 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2661 unlikely(tid != READ_ONCE(c->tid)));
2664 * Irqless object alloc/free algorithm used here depends on sequence
2665 * of fetching cpu_slab's data. tid should be fetched before anything
2666 * on c to guarantee that object and page associated with previous tid
2667 * won't be used with current tid. If we fetch tid first, object and
2668 * page could be one associated with next tid and our alloc/free
2669 * request will be failed. In this case, we will retry. So, no problem.
2674 * The transaction ids are globally unique per cpu and per operation on
2675 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2676 * occurs on the right processor and that there was no operation on the
2677 * linked list in between.
2680 object = c->freelist;
2682 if (unlikely(!object || !node_match(page, node))) {
2683 object = __slab_alloc(s, gfpflags, node, addr, c);
2684 stat(s, ALLOC_SLOWPATH);
2686 void *next_object = get_freepointer_safe(s, object);
2689 * The cmpxchg will only match if there was no additional
2690 * operation and if we are on the right processor.
2692 * The cmpxchg does the following atomically (without lock
2694 * 1. Relocate first pointer to the current per cpu area.
2695 * 2. Verify that tid and freelist have not been changed
2696 * 3. If they were not changed replace tid and freelist
2698 * Since this is without lock semantics the protection is only
2699 * against code executing on this cpu *not* from access by
2702 if (unlikely(!this_cpu_cmpxchg_double(
2703 s->cpu_slab->freelist, s->cpu_slab->tid,
2705 next_object, next_tid(tid)))) {
2707 note_cmpxchg_failure("slab_alloc", s, tid);
2710 prefetch_freepointer(s, next_object);
2711 stat(s, ALLOC_FASTPATH);
2714 if (unlikely(gfpflags & __GFP_ZERO) && object)
2715 memset(object, 0, s->object_size);
2717 slab_post_alloc_hook(s, gfpflags, 1, &object);
2722 static __always_inline void *slab_alloc(struct kmem_cache *s,
2723 gfp_t gfpflags, unsigned long addr)
2725 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2728 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2730 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2732 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2737 EXPORT_SYMBOL(kmem_cache_alloc);
2739 #ifdef CONFIG_TRACING
2740 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2742 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2743 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2744 kasan_kmalloc(s, ret, size, gfpflags);
2747 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2751 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2753 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2755 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2756 s->object_size, s->size, gfpflags, node);
2760 EXPORT_SYMBOL(kmem_cache_alloc_node);
2762 #ifdef CONFIG_TRACING
2763 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2765 int node, size_t size)
2767 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2769 trace_kmalloc_node(_RET_IP_, ret,
2770 size, s->size, gfpflags, node);
2772 kasan_kmalloc(s, ret, size, gfpflags);
2775 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2780 * Slow path handling. This may still be called frequently since objects
2781 * have a longer lifetime than the cpu slabs in most processing loads.
2783 * So we still attempt to reduce cache line usage. Just take the slab
2784 * lock and free the item. If there is no additional partial page
2785 * handling required then we can return immediately.
2787 static void __slab_free(struct kmem_cache *s, struct page *page,
2788 void *head, void *tail, int cnt,
2795 unsigned long counters;
2796 struct kmem_cache_node *n = NULL;
2797 unsigned long uninitialized_var(flags);
2799 stat(s, FREE_SLOWPATH);
2801 if (kmem_cache_debug(s) &&
2802 !free_debug_processing(s, page, head, tail, cnt, addr))
2807 spin_unlock_irqrestore(&n->list_lock, flags);
2810 prior = page->freelist;
2811 counters = page->counters;
2812 set_freepointer(s, tail, prior);
2813 new.counters = counters;
2814 was_frozen = new.frozen;
2816 if ((!new.inuse || !prior) && !was_frozen) {
2818 if (kmem_cache_has_cpu_partial(s) && !prior) {
2821 * Slab was on no list before and will be
2823 * We can defer the list move and instead
2828 } else { /* Needs to be taken off a list */
2830 n = get_node(s, page_to_nid(page));
2832 * Speculatively acquire the list_lock.
2833 * If the cmpxchg does not succeed then we may
2834 * drop the list_lock without any processing.
2836 * Otherwise the list_lock will synchronize with
2837 * other processors updating the list of slabs.
2839 spin_lock_irqsave(&n->list_lock, flags);
2844 } while (!cmpxchg_double_slab(s, page,
2852 * If we just froze the page then put it onto the
2853 * per cpu partial list.
2855 if (new.frozen && !was_frozen) {
2856 put_cpu_partial(s, page, 1);
2857 stat(s, CPU_PARTIAL_FREE);
2860 * The list lock was not taken therefore no list
2861 * activity can be necessary.
2864 stat(s, FREE_FROZEN);
2868 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2872 * Objects left in the slab. If it was not on the partial list before
2875 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2876 if (kmem_cache_debug(s))
2877 remove_full(s, n, page);
2878 add_partial(n, page, DEACTIVATE_TO_TAIL);
2879 stat(s, FREE_ADD_PARTIAL);
2881 spin_unlock_irqrestore(&n->list_lock, flags);
2887 * Slab on the partial list.
2889 remove_partial(n, page);
2890 stat(s, FREE_REMOVE_PARTIAL);
2892 /* Slab must be on the full list */
2893 remove_full(s, n, page);
2896 spin_unlock_irqrestore(&n->list_lock, flags);
2898 discard_slab(s, page);
2902 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2903 * can perform fastpath freeing without additional function calls.
2905 * The fastpath is only possible if we are freeing to the current cpu slab
2906 * of this processor. This typically the case if we have just allocated
2909 * If fastpath is not possible then fall back to __slab_free where we deal
2910 * with all sorts of special processing.
2912 * Bulk free of a freelist with several objects (all pointing to the
2913 * same page) possible by specifying head and tail ptr, plus objects
2914 * count (cnt). Bulk free indicated by tail pointer being set.
2916 static __always_inline void do_slab_free(struct kmem_cache *s,
2917 struct page *page, void *head, void *tail,
2918 int cnt, unsigned long addr)
2920 void *tail_obj = tail ? : head;
2921 struct kmem_cache_cpu *c;
2925 * Determine the currently cpus per cpu slab.
2926 * The cpu may change afterward. However that does not matter since
2927 * data is retrieved via this pointer. If we are on the same cpu
2928 * during the cmpxchg then the free will succeed.
2931 tid = this_cpu_read(s->cpu_slab->tid);
2932 c = raw_cpu_ptr(s->cpu_slab);
2933 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2934 unlikely(tid != READ_ONCE(c->tid)));
2936 /* Same with comment on barrier() in slab_alloc_node() */
2939 if (likely(page == c->page)) {
2940 set_freepointer(s, tail_obj, c->freelist);
2942 if (unlikely(!this_cpu_cmpxchg_double(
2943 s->cpu_slab->freelist, s->cpu_slab->tid,
2945 head, next_tid(tid)))) {
2947 note_cmpxchg_failure("slab_free", s, tid);
2950 stat(s, FREE_FASTPATH);
2952 __slab_free(s, page, head, tail_obj, cnt, addr);
2956 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2957 void *head, void *tail, int cnt,
2960 slab_free_freelist_hook(s, head, tail);
2962 * slab_free_freelist_hook() could have put the items into quarantine.
2963 * If so, no need to free them.
2965 if (s->flags & SLAB_KASAN && !(s->flags & SLAB_TYPESAFE_BY_RCU))
2967 do_slab_free(s, page, head, tail, cnt, addr);
2971 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
2973 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
2977 void kmem_cache_free(struct kmem_cache *s, void *x)
2979 s = cache_from_obj(s, x);
2982 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2983 trace_kmem_cache_free(_RET_IP_, x);
2985 EXPORT_SYMBOL(kmem_cache_free);
2987 struct detached_freelist {
2992 struct kmem_cache *s;
2996 * This function progressively scans the array with free objects (with
2997 * a limited look ahead) and extract objects belonging to the same
2998 * page. It builds a detached freelist directly within the given
2999 * page/objects. This can happen without any need for
3000 * synchronization, because the objects are owned by running process.
3001 * The freelist is build up as a single linked list in the objects.
3002 * The idea is, that this detached freelist can then be bulk
3003 * transferred to the real freelist(s), but only requiring a single
3004 * synchronization primitive. Look ahead in the array is limited due
3005 * to performance reasons.
3008 int build_detached_freelist(struct kmem_cache *s, size_t size,
3009 void **p, struct detached_freelist *df)
3011 size_t first_skipped_index = 0;
3016 /* Always re-init detached_freelist */
3021 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3022 } while (!object && size);
3027 page = virt_to_head_page(object);
3029 /* Handle kalloc'ed objects */
3030 if (unlikely(!PageSlab(page))) {
3031 BUG_ON(!PageCompound(page));
3033 __free_pages(page, compound_order(page));
3034 p[size] = NULL; /* mark object processed */
3037 /* Derive kmem_cache from object */
3038 df->s = page->slab_cache;
3040 df->s = cache_from_obj(s, object); /* Support for memcg */
3043 /* Start new detached freelist */
3045 set_freepointer(df->s, object, NULL);
3047 df->freelist = object;
3048 p[size] = NULL; /* mark object processed */
3054 continue; /* Skip processed objects */
3056 /* df->page is always set at this point */
3057 if (df->page == virt_to_head_page(object)) {
3058 /* Opportunity build freelist */
3059 set_freepointer(df->s, object, df->freelist);
3060 df->freelist = object;
3062 p[size] = NULL; /* mark object processed */
3067 /* Limit look ahead search */
3071 if (!first_skipped_index)
3072 first_skipped_index = size + 1;
3075 return first_skipped_index;
3078 /* Note that interrupts must be enabled when calling this function. */
3079 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3085 struct detached_freelist df;
3087 size = build_detached_freelist(s, size, p, &df);
3091 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3092 } while (likely(size));
3094 EXPORT_SYMBOL(kmem_cache_free_bulk);
3096 /* Note that interrupts must be enabled when calling this function. */
3097 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3100 struct kmem_cache_cpu *c;
3103 /* memcg and kmem_cache debug support */
3104 s = slab_pre_alloc_hook(s, flags);
3108 * Drain objects in the per cpu slab, while disabling local
3109 * IRQs, which protects against PREEMPT and interrupts
3110 * handlers invoking normal fastpath.
3112 local_irq_disable();
3113 c = this_cpu_ptr(s->cpu_slab);
3115 for (i = 0; i < size; i++) {
3116 void *object = c->freelist;
3118 if (unlikely(!object)) {
3120 * Invoking slow path likely have side-effect
3121 * of re-populating per CPU c->freelist
3123 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3125 if (unlikely(!p[i]))
3128 c = this_cpu_ptr(s->cpu_slab);
3129 continue; /* goto for-loop */
3131 c->freelist = get_freepointer(s, object);
3134 c->tid = next_tid(c->tid);
3137 /* Clear memory outside IRQ disabled fastpath loop */
3138 if (unlikely(flags & __GFP_ZERO)) {
3141 for (j = 0; j < i; j++)
3142 memset(p[j], 0, s->object_size);
3145 /* memcg and kmem_cache debug support */
3146 slab_post_alloc_hook(s, flags, size, p);
3150 slab_post_alloc_hook(s, flags, i, p);
3151 __kmem_cache_free_bulk(s, i, p);
3154 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3158 * Object placement in a slab is made very easy because we always start at
3159 * offset 0. If we tune the size of the object to the alignment then we can
3160 * get the required alignment by putting one properly sized object after
3163 * Notice that the allocation order determines the sizes of the per cpu
3164 * caches. Each processor has always one slab available for allocations.
3165 * Increasing the allocation order reduces the number of times that slabs
3166 * must be moved on and off the partial lists and is therefore a factor in
3171 * Mininum / Maximum order of slab pages. This influences locking overhead
3172 * and slab fragmentation. A higher order reduces the number of partial slabs
3173 * and increases the number of allocations possible without having to
3174 * take the list_lock.
3176 static int slub_min_order;
3177 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3178 static int slub_min_objects;
3181 * Calculate the order of allocation given an slab object size.
3183 * The order of allocation has significant impact on performance and other
3184 * system components. Generally order 0 allocations should be preferred since
3185 * order 0 does not cause fragmentation in the page allocator. Larger objects
3186 * be problematic to put into order 0 slabs because there may be too much
3187 * unused space left. We go to a higher order if more than 1/16th of the slab
3190 * In order to reach satisfactory performance we must ensure that a minimum
3191 * number of objects is in one slab. Otherwise we may generate too much
3192 * activity on the partial lists which requires taking the list_lock. This is
3193 * less a concern for large slabs though which are rarely used.
3195 * slub_max_order specifies the order where we begin to stop considering the
3196 * number of objects in a slab as critical. If we reach slub_max_order then
3197 * we try to keep the page order as low as possible. So we accept more waste
3198 * of space in favor of a small page order.
3200 * Higher order allocations also allow the placement of more objects in a
3201 * slab and thereby reduce object handling overhead. If the user has
3202 * requested a higher mininum order then we start with that one instead of
3203 * the smallest order which will fit the object.
3205 static inline int slab_order(int size, int min_objects,
3206 int max_order, int fract_leftover, int reserved)
3210 int min_order = slub_min_order;
3212 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3213 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3215 for (order = max(min_order, get_order(min_objects * size + reserved));
3216 order <= max_order; order++) {
3218 unsigned long slab_size = PAGE_SIZE << order;
3220 rem = (slab_size - reserved) % size;
3222 if (rem <= slab_size / fract_leftover)
3229 static inline int calculate_order(int size, int reserved)
3237 * Attempt to find best configuration for a slab. This
3238 * works by first attempting to generate a layout with
3239 * the best configuration and backing off gradually.
3241 * First we increase the acceptable waste in a slab. Then
3242 * we reduce the minimum objects required in a slab.
3244 min_objects = slub_min_objects;
3246 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3247 max_objects = order_objects(slub_max_order, size, reserved);
3248 min_objects = min(min_objects, max_objects);
3250 while (min_objects > 1) {
3252 while (fraction >= 4) {
3253 order = slab_order(size, min_objects,
3254 slub_max_order, fraction, reserved);
3255 if (order <= slub_max_order)
3263 * We were unable to place multiple objects in a slab. Now
3264 * lets see if we can place a single object there.
3266 order = slab_order(size, 1, slub_max_order, 1, reserved);
3267 if (order <= slub_max_order)
3271 * Doh this slab cannot be placed using slub_max_order.
3273 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3274 if (order < MAX_ORDER)
3280 init_kmem_cache_node(struct kmem_cache_node *n)
3283 spin_lock_init(&n->list_lock);
3284 INIT_LIST_HEAD(&n->partial);
3285 #ifdef CONFIG_SLUB_DEBUG
3286 atomic_long_set(&n->nr_slabs, 0);
3287 atomic_long_set(&n->total_objects, 0);
3288 INIT_LIST_HEAD(&n->full);
3292 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3294 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3295 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3298 * Must align to double word boundary for the double cmpxchg
3299 * instructions to work; see __pcpu_double_call_return_bool().
3301 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3302 2 * sizeof(void *));
3307 init_kmem_cache_cpus(s);
3312 static struct kmem_cache *kmem_cache_node;
3315 * No kmalloc_node yet so do it by hand. We know that this is the first
3316 * slab on the node for this slabcache. There are no concurrent accesses
3319 * Note that this function only works on the kmem_cache_node
3320 * when allocating for the kmem_cache_node. This is used for bootstrapping
3321 * memory on a fresh node that has no slab structures yet.
3323 static void early_kmem_cache_node_alloc(int node)
3326 struct kmem_cache_node *n;
3328 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3330 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3333 if (page_to_nid(page) != node) {
3334 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3335 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3340 page->freelist = get_freepointer(kmem_cache_node, n);
3343 kmem_cache_node->node[node] = n;
3344 #ifdef CONFIG_SLUB_DEBUG
3345 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3346 init_tracking(kmem_cache_node, n);
3348 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3350 init_kmem_cache_node(n);
3351 inc_slabs_node(kmem_cache_node, node, page->objects);
3354 * No locks need to be taken here as it has just been
3355 * initialized and there is no concurrent access.
3357 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3360 static void free_kmem_cache_nodes(struct kmem_cache *s)
3363 struct kmem_cache_node *n;
3365 for_each_kmem_cache_node(s, node, n) {
3366 kmem_cache_free(kmem_cache_node, n);
3367 s->node[node] = NULL;
3371 void __kmem_cache_release(struct kmem_cache *s)
3373 cache_random_seq_destroy(s);
3374 free_percpu(s->cpu_slab);
3375 free_kmem_cache_nodes(s);
3378 static int init_kmem_cache_nodes(struct kmem_cache *s)
3382 for_each_node_state(node, N_NORMAL_MEMORY) {
3383 struct kmem_cache_node *n;
3385 if (slab_state == DOWN) {
3386 early_kmem_cache_node_alloc(node);
3389 n = kmem_cache_alloc_node(kmem_cache_node,
3393 free_kmem_cache_nodes(s);
3398 init_kmem_cache_node(n);
3403 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3405 if (min < MIN_PARTIAL)
3407 else if (min > MAX_PARTIAL)
3409 s->min_partial = min;
3413 * calculate_sizes() determines the order and the distribution of data within
3416 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3418 unsigned long flags = s->flags;
3419 size_t size = s->object_size;
3423 * Round up object size to the next word boundary. We can only
3424 * place the free pointer at word boundaries and this determines
3425 * the possible location of the free pointer.
3427 size = ALIGN(size, sizeof(void *));
3429 #ifdef CONFIG_SLUB_DEBUG
3431 * Determine if we can poison the object itself. If the user of
3432 * the slab may touch the object after free or before allocation
3433 * then we should never poison the object itself.
3435 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3437 s->flags |= __OBJECT_POISON;
3439 s->flags &= ~__OBJECT_POISON;
3443 * If we are Redzoning then check if there is some space between the
3444 * end of the object and the free pointer. If not then add an
3445 * additional word to have some bytes to store Redzone information.
3447 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3448 size += sizeof(void *);
3452 * With that we have determined the number of bytes in actual use
3453 * by the object. This is the potential offset to the free pointer.
3457 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3460 * Relocate free pointer after the object if it is not
3461 * permitted to overwrite the first word of the object on
3464 * This is the case if we do RCU, have a constructor or
3465 * destructor or are poisoning the objects.
3468 size += sizeof(void *);
3471 #ifdef CONFIG_SLUB_DEBUG
3472 if (flags & SLAB_STORE_USER)
3474 * Need to store information about allocs and frees after
3477 size += 2 * sizeof(struct track);
3480 kasan_cache_create(s, &size, &s->flags);
3481 #ifdef CONFIG_SLUB_DEBUG
3482 if (flags & SLAB_RED_ZONE) {
3484 * Add some empty padding so that we can catch
3485 * overwrites from earlier objects rather than let
3486 * tracking information or the free pointer be
3487 * corrupted if a user writes before the start
3490 size += sizeof(void *);
3492 s->red_left_pad = sizeof(void *);
3493 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3494 size += s->red_left_pad;
3499 * SLUB stores one object immediately after another beginning from
3500 * offset 0. In order to align the objects we have to simply size
3501 * each object to conform to the alignment.
3503 size = ALIGN(size, s->align);
3505 if (forced_order >= 0)
3506 order = forced_order;
3508 order = calculate_order(size, s->reserved);
3515 s->allocflags |= __GFP_COMP;
3517 if (s->flags & SLAB_CACHE_DMA)
3518 s->allocflags |= GFP_DMA;
3520 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3521 s->allocflags |= __GFP_RECLAIMABLE;
3524 * Determine the number of objects per slab
3526 s->oo = oo_make(order, size, s->reserved);
3527 s->min = oo_make(get_order(size), size, s->reserved);
3528 if (oo_objects(s->oo) > oo_objects(s->max))
3531 return !!oo_objects(s->oo);
3534 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3536 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3539 if (need_reserve_slab_rcu && (s->flags & SLAB_TYPESAFE_BY_RCU))
3540 s->reserved = sizeof(struct rcu_head);
3542 if (!calculate_sizes(s, -1))
3544 if (disable_higher_order_debug) {
3546 * Disable debugging flags that store metadata if the min slab
3549 if (get_order(s->size) > get_order(s->object_size)) {
3550 s->flags &= ~DEBUG_METADATA_FLAGS;
3552 if (!calculate_sizes(s, -1))
3557 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3558 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3559 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3560 /* Enable fast mode */
3561 s->flags |= __CMPXCHG_DOUBLE;
3565 * The larger the object size is, the more pages we want on the partial
3566 * list to avoid pounding the page allocator excessively.
3568 set_min_partial(s, ilog2(s->size) / 2);
3571 * cpu_partial determined the maximum number of objects kept in the
3572 * per cpu partial lists of a processor.
3574 * Per cpu partial lists mainly contain slabs that just have one
3575 * object freed. If they are used for allocation then they can be
3576 * filled up again with minimal effort. The slab will never hit the
3577 * per node partial lists and therefore no locking will be required.
3579 * This setting also determines
3581 * A) The number of objects from per cpu partial slabs dumped to the
3582 * per node list when we reach the limit.
3583 * B) The number of objects in cpu partial slabs to extract from the
3584 * per node list when we run out of per cpu objects. We only fetch
3585 * 50% to keep some capacity around for frees.
3587 if (!kmem_cache_has_cpu_partial(s))
3589 else if (s->size >= PAGE_SIZE)
3591 else if (s->size >= 1024)
3593 else if (s->size >= 256)
3594 s->cpu_partial = 13;
3596 s->cpu_partial = 30;
3599 s->remote_node_defrag_ratio = 1000;
3602 /* Initialize the pre-computed randomized freelist if slab is up */
3603 if (slab_state >= UP) {
3604 if (init_cache_random_seq(s))
3608 if (!init_kmem_cache_nodes(s))
3611 if (alloc_kmem_cache_cpus(s))
3614 free_kmem_cache_nodes(s);
3616 if (flags & SLAB_PANIC)
3617 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3618 s->name, (unsigned long)s->size, s->size,
3619 oo_order(s->oo), s->offset, flags);
3623 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3626 #ifdef CONFIG_SLUB_DEBUG
3627 void *addr = page_address(page);
3629 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3630 sizeof(long), GFP_ATOMIC);
3633 slab_err(s, page, text, s->name);
3636 get_map(s, page, map);
3637 for_each_object(p, s, addr, page->objects) {
3639 if (!test_bit(slab_index(p, s, addr), map)) {
3640 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3641 print_tracking(s, p);
3650 * Attempt to free all partial slabs on a node.
3651 * This is called from __kmem_cache_shutdown(). We must take list_lock
3652 * because sysfs file might still access partial list after the shutdowning.
3654 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3657 struct page *page, *h;
3659 BUG_ON(irqs_disabled());
3660 spin_lock_irq(&n->list_lock);
3661 list_for_each_entry_safe(page, h, &n->partial, lru) {
3663 remove_partial(n, page);
3664 list_add(&page->lru, &discard);
3666 list_slab_objects(s, page,
3667 "Objects remaining in %s on __kmem_cache_shutdown()");
3670 spin_unlock_irq(&n->list_lock);
3672 list_for_each_entry_safe(page, h, &discard, lru)
3673 discard_slab(s, page);
3677 * Release all resources used by a slab cache.
3679 int __kmem_cache_shutdown(struct kmem_cache *s)
3682 struct kmem_cache_node *n;
3685 /* Attempt to free all objects */
3686 for_each_kmem_cache_node(s, node, n) {
3688 if (n->nr_partial || slabs_node(s, node))
3691 sysfs_slab_remove(s);
3695 /********************************************************************
3697 *******************************************************************/
3699 static int __init setup_slub_min_order(char *str)
3701 get_option(&str, &slub_min_order);
3706 __setup("slub_min_order=", setup_slub_min_order);
3708 static int __init setup_slub_max_order(char *str)
3710 get_option(&str, &slub_max_order);
3711 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3716 __setup("slub_max_order=", setup_slub_max_order);
3718 static int __init setup_slub_min_objects(char *str)
3720 get_option(&str, &slub_min_objects);
3725 __setup("slub_min_objects=", setup_slub_min_objects);
3727 void *__kmalloc(size_t size, gfp_t flags)
3729 struct kmem_cache *s;
3732 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3733 return kmalloc_large(size, flags);
3735 s = kmalloc_slab(size, flags);
3737 if (unlikely(ZERO_OR_NULL_PTR(s)))
3740 ret = slab_alloc(s, flags, _RET_IP_);
3742 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3744 kasan_kmalloc(s, ret, size, flags);
3748 EXPORT_SYMBOL(__kmalloc);
3751 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3756 flags |= __GFP_COMP | __GFP_NOTRACK;
3757 page = alloc_pages_node(node, flags, get_order(size));
3759 ptr = page_address(page);
3761 kmalloc_large_node_hook(ptr, size, flags);
3765 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3767 struct kmem_cache *s;
3770 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3771 ret = kmalloc_large_node(size, flags, node);
3773 trace_kmalloc_node(_RET_IP_, ret,
3774 size, PAGE_SIZE << get_order(size),
3780 s = kmalloc_slab(size, flags);
3782 if (unlikely(ZERO_OR_NULL_PTR(s)))
3785 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3787 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3789 kasan_kmalloc(s, ret, size, flags);
3793 EXPORT_SYMBOL(__kmalloc_node);
3796 #ifdef CONFIG_HARDENED_USERCOPY
3798 * Rejects objects that are incorrectly sized.
3800 * Returns NULL if check passes, otherwise const char * to name of cache
3801 * to indicate an error.
3803 const char *__check_heap_object(const void *ptr, unsigned long n,
3806 struct kmem_cache *s;
3807 unsigned long offset;
3810 /* Find object and usable object size. */
3811 s = page->slab_cache;
3812 object_size = slab_ksize(s);
3814 /* Reject impossible pointers. */
3815 if (ptr < page_address(page))
3818 /* Find offset within object. */
3819 offset = (ptr - page_address(page)) % s->size;
3821 /* Adjust for redzone and reject if within the redzone. */
3822 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3823 if (offset < s->red_left_pad)
3825 offset -= s->red_left_pad;
3828 /* Allow address range falling entirely within object size. */
3829 if (offset <= object_size && n <= object_size - offset)
3834 #endif /* CONFIG_HARDENED_USERCOPY */
3836 static size_t __ksize(const void *object)
3840 if (unlikely(object == ZERO_SIZE_PTR))
3843 page = virt_to_head_page(object);
3845 if (unlikely(!PageSlab(page))) {
3846 WARN_ON(!PageCompound(page));
3847 return PAGE_SIZE << compound_order(page);
3850 return slab_ksize(page->slab_cache);
3853 size_t ksize(const void *object)
3855 size_t size = __ksize(object);
3856 /* We assume that ksize callers could use whole allocated area,
3857 * so we need to unpoison this area.
3859 kasan_unpoison_shadow(object, size);
3862 EXPORT_SYMBOL(ksize);
3864 void kfree(const void *x)
3867 void *object = (void *)x;
3869 trace_kfree(_RET_IP_, x);
3871 if (unlikely(ZERO_OR_NULL_PTR(x)))
3874 page = virt_to_head_page(x);
3875 if (unlikely(!PageSlab(page))) {
3876 BUG_ON(!PageCompound(page));
3878 __free_pages(page, compound_order(page));
3881 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3883 EXPORT_SYMBOL(kfree);
3885 #define SHRINK_PROMOTE_MAX 32
3888 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3889 * up most to the head of the partial lists. New allocations will then
3890 * fill those up and thus they can be removed from the partial lists.
3892 * The slabs with the least items are placed last. This results in them
3893 * being allocated from last increasing the chance that the last objects
3894 * are freed in them.
3896 int __kmem_cache_shrink(struct kmem_cache *s)
3900 struct kmem_cache_node *n;
3903 struct list_head discard;
3904 struct list_head promote[SHRINK_PROMOTE_MAX];
3905 unsigned long flags;
3909 for_each_kmem_cache_node(s, node, n) {
3910 INIT_LIST_HEAD(&discard);
3911 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3912 INIT_LIST_HEAD(promote + i);
3914 spin_lock_irqsave(&n->list_lock, flags);
3917 * Build lists of slabs to discard or promote.
3919 * Note that concurrent frees may occur while we hold the
3920 * list_lock. page->inuse here is the upper limit.
3922 list_for_each_entry_safe(page, t, &n->partial, lru) {
3923 int free = page->objects - page->inuse;
3925 /* Do not reread page->inuse */
3928 /* We do not keep full slabs on the list */
3931 if (free == page->objects) {
3932 list_move(&page->lru, &discard);
3934 } else if (free <= SHRINK_PROMOTE_MAX)
3935 list_move(&page->lru, promote + free - 1);
3939 * Promote the slabs filled up most to the head of the
3942 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3943 list_splice(promote + i, &n->partial);
3945 spin_unlock_irqrestore(&n->list_lock, flags);
3947 /* Release empty slabs */
3948 list_for_each_entry_safe(page, t, &discard, lru)
3949 discard_slab(s, page);
3951 if (slabs_node(s, node))
3959 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
3962 * Called with all the locks held after a sched RCU grace period.
3963 * Even if @s becomes empty after shrinking, we can't know that @s
3964 * doesn't have allocations already in-flight and thus can't
3965 * destroy @s until the associated memcg is released.
3967 * However, let's remove the sysfs files for empty caches here.
3968 * Each cache has a lot of interface files which aren't
3969 * particularly useful for empty draining caches; otherwise, we can
3970 * easily end up with millions of unnecessary sysfs files on
3971 * systems which have a lot of memory and transient cgroups.
3973 if (!__kmem_cache_shrink(s))
3974 sysfs_slab_remove(s);
3977 void __kmemcg_cache_deactivate(struct kmem_cache *s)
3980 * Disable empty slabs caching. Used to avoid pinning offline
3981 * memory cgroups by kmem pages that can be freed.
3987 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
3988 * we have to make sure the change is visible before shrinking.
3990 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
3994 static int slab_mem_going_offline_callback(void *arg)
3996 struct kmem_cache *s;
3998 mutex_lock(&slab_mutex);
3999 list_for_each_entry(s, &slab_caches, list)
4000 __kmem_cache_shrink(s);
4001 mutex_unlock(&slab_mutex);
4006 static void slab_mem_offline_callback(void *arg)
4008 struct kmem_cache_node *n;
4009 struct kmem_cache *s;
4010 struct memory_notify *marg = arg;
4013 offline_node = marg->status_change_nid_normal;
4016 * If the node still has available memory. we need kmem_cache_node
4019 if (offline_node < 0)
4022 mutex_lock(&slab_mutex);
4023 list_for_each_entry(s, &slab_caches, list) {
4024 n = get_node(s, offline_node);
4027 * if n->nr_slabs > 0, slabs still exist on the node
4028 * that is going down. We were unable to free them,
4029 * and offline_pages() function shouldn't call this
4030 * callback. So, we must fail.
4032 BUG_ON(slabs_node(s, offline_node));
4034 s->node[offline_node] = NULL;
4035 kmem_cache_free(kmem_cache_node, n);
4038 mutex_unlock(&slab_mutex);
4041 static int slab_mem_going_online_callback(void *arg)
4043 struct kmem_cache_node *n;
4044 struct kmem_cache *s;
4045 struct memory_notify *marg = arg;
4046 int nid = marg->status_change_nid_normal;
4050 * If the node's memory is already available, then kmem_cache_node is
4051 * already created. Nothing to do.
4057 * We are bringing a node online. No memory is available yet. We must
4058 * allocate a kmem_cache_node structure in order to bring the node
4061 mutex_lock(&slab_mutex);
4062 list_for_each_entry(s, &slab_caches, list) {
4064 * XXX: kmem_cache_alloc_node will fallback to other nodes
4065 * since memory is not yet available from the node that
4068 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4073 init_kmem_cache_node(n);
4077 mutex_unlock(&slab_mutex);
4081 static int slab_memory_callback(struct notifier_block *self,
4082 unsigned long action, void *arg)
4087 case MEM_GOING_ONLINE:
4088 ret = slab_mem_going_online_callback(arg);
4090 case MEM_GOING_OFFLINE:
4091 ret = slab_mem_going_offline_callback(arg);
4094 case MEM_CANCEL_ONLINE:
4095 slab_mem_offline_callback(arg);
4098 case MEM_CANCEL_OFFLINE:
4102 ret = notifier_from_errno(ret);
4108 static struct notifier_block slab_memory_callback_nb = {
4109 .notifier_call = slab_memory_callback,
4110 .priority = SLAB_CALLBACK_PRI,
4113 /********************************************************************
4114 * Basic setup of slabs
4115 *******************************************************************/
4118 * Used for early kmem_cache structures that were allocated using
4119 * the page allocator. Allocate them properly then fix up the pointers
4120 * that may be pointing to the wrong kmem_cache structure.
4123 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4126 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4127 struct kmem_cache_node *n;
4129 memcpy(s, static_cache, kmem_cache->object_size);
4132 * This runs very early, and only the boot processor is supposed to be
4133 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4136 __flush_cpu_slab(s, smp_processor_id());
4137 for_each_kmem_cache_node(s, node, n) {
4140 list_for_each_entry(p, &n->partial, lru)
4143 #ifdef CONFIG_SLUB_DEBUG
4144 list_for_each_entry(p, &n->full, lru)
4148 slab_init_memcg_params(s);
4149 list_add(&s->list, &slab_caches);
4150 memcg_link_cache(s);
4154 void __init kmem_cache_init(void)
4156 static __initdata struct kmem_cache boot_kmem_cache,
4157 boot_kmem_cache_node;
4159 if (debug_guardpage_minorder())
4162 kmem_cache_node = &boot_kmem_cache_node;
4163 kmem_cache = &boot_kmem_cache;
4165 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4166 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
4168 register_hotmemory_notifier(&slab_memory_callback_nb);
4170 /* Able to allocate the per node structures */
4171 slab_state = PARTIAL;
4173 create_boot_cache(kmem_cache, "kmem_cache",
4174 offsetof(struct kmem_cache, node) +
4175 nr_node_ids * sizeof(struct kmem_cache_node *),
4176 SLAB_HWCACHE_ALIGN);
4178 kmem_cache = bootstrap(&boot_kmem_cache);
4181 * Allocate kmem_cache_node properly from the kmem_cache slab.
4182 * kmem_cache_node is separately allocated so no need to
4183 * update any list pointers.
4185 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4187 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4188 setup_kmalloc_cache_index_table();
4189 create_kmalloc_caches(0);
4191 /* Setup random freelists for each cache */
4192 init_freelist_randomization();
4194 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4197 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4199 slub_min_order, slub_max_order, slub_min_objects,
4200 nr_cpu_ids, nr_node_ids);
4203 void __init kmem_cache_init_late(void)
4208 __kmem_cache_alias(const char *name, size_t size, size_t align,
4209 unsigned long flags, void (*ctor)(void *))
4211 struct kmem_cache *s, *c;
4213 s = find_mergeable(size, align, flags, name, ctor);
4218 * Adjust the object sizes so that we clear
4219 * the complete object on kzalloc.
4221 s->object_size = max(s->object_size, (int)size);
4222 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
4224 for_each_memcg_cache(c, s) {
4225 c->object_size = s->object_size;
4226 c->inuse = max_t(int, c->inuse,
4227 ALIGN(size, sizeof(void *)));
4230 if (sysfs_slab_alias(s, name)) {
4239 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4243 err = kmem_cache_open(s, flags);
4247 /* Mutex is not taken during early boot */
4248 if (slab_state <= UP)
4251 memcg_propagate_slab_attrs(s);
4252 err = sysfs_slab_add(s);
4254 __kmem_cache_release(s);
4259 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4261 struct kmem_cache *s;
4264 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4265 return kmalloc_large(size, gfpflags);
4267 s = kmalloc_slab(size, gfpflags);
4269 if (unlikely(ZERO_OR_NULL_PTR(s)))
4272 ret = slab_alloc(s, gfpflags, caller);
4274 /* Honor the call site pointer we received. */
4275 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4281 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4282 int node, unsigned long caller)
4284 struct kmem_cache *s;
4287 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4288 ret = kmalloc_large_node(size, gfpflags, node);
4290 trace_kmalloc_node(caller, ret,
4291 size, PAGE_SIZE << get_order(size),
4297 s = kmalloc_slab(size, gfpflags);
4299 if (unlikely(ZERO_OR_NULL_PTR(s)))
4302 ret = slab_alloc_node(s, gfpflags, node, caller);
4304 /* Honor the call site pointer we received. */
4305 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4312 static int count_inuse(struct page *page)
4317 static int count_total(struct page *page)
4319 return page->objects;
4323 #ifdef CONFIG_SLUB_DEBUG
4324 static int validate_slab(struct kmem_cache *s, struct page *page,
4328 void *addr = page_address(page);
4330 if (!check_slab(s, page) ||
4331 !on_freelist(s, page, NULL))
4334 /* Now we know that a valid freelist exists */
4335 bitmap_zero(map, page->objects);
4337 get_map(s, page, map);
4338 for_each_object(p, s, addr, page->objects) {
4339 if (test_bit(slab_index(p, s, addr), map))
4340 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4344 for_each_object(p, s, addr, page->objects)
4345 if (!test_bit(slab_index(p, s, addr), map))
4346 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4351 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4355 validate_slab(s, page, map);
4359 static int validate_slab_node(struct kmem_cache *s,
4360 struct kmem_cache_node *n, unsigned long *map)
4362 unsigned long count = 0;
4364 unsigned long flags;
4366 spin_lock_irqsave(&n->list_lock, flags);
4368 list_for_each_entry(page, &n->partial, lru) {
4369 validate_slab_slab(s, page, map);
4372 if (count != n->nr_partial)
4373 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4374 s->name, count, n->nr_partial);
4376 if (!(s->flags & SLAB_STORE_USER))
4379 list_for_each_entry(page, &n->full, lru) {
4380 validate_slab_slab(s, page, map);
4383 if (count != atomic_long_read(&n->nr_slabs))
4384 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4385 s->name, count, atomic_long_read(&n->nr_slabs));
4388 spin_unlock_irqrestore(&n->list_lock, flags);
4392 static long validate_slab_cache(struct kmem_cache *s)
4395 unsigned long count = 0;
4396 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4397 sizeof(unsigned long), GFP_KERNEL);
4398 struct kmem_cache_node *n;
4404 for_each_kmem_cache_node(s, node, n)
4405 count += validate_slab_node(s, n, map);
4410 * Generate lists of code addresses where slabcache objects are allocated
4415 unsigned long count;
4422 DECLARE_BITMAP(cpus, NR_CPUS);
4428 unsigned long count;
4429 struct location *loc;
4432 static void free_loc_track(struct loc_track *t)
4435 free_pages((unsigned long)t->loc,
4436 get_order(sizeof(struct location) * t->max));
4439 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4444 order = get_order(sizeof(struct location) * max);
4446 l = (void *)__get_free_pages(flags, order);
4451 memcpy(l, t->loc, sizeof(struct location) * t->count);
4459 static int add_location(struct loc_track *t, struct kmem_cache *s,
4460 const struct track *track)
4462 long start, end, pos;
4464 unsigned long caddr;
4465 unsigned long age = jiffies - track->when;
4471 pos = start + (end - start + 1) / 2;
4474 * There is nothing at "end". If we end up there
4475 * we need to add something to before end.
4480 caddr = t->loc[pos].addr;
4481 if (track->addr == caddr) {
4487 if (age < l->min_time)
4489 if (age > l->max_time)
4492 if (track->pid < l->min_pid)
4493 l->min_pid = track->pid;
4494 if (track->pid > l->max_pid)
4495 l->max_pid = track->pid;
4497 cpumask_set_cpu(track->cpu,
4498 to_cpumask(l->cpus));
4500 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4504 if (track->addr < caddr)
4511 * Not found. Insert new tracking element.
4513 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4519 (t->count - pos) * sizeof(struct location));
4522 l->addr = track->addr;
4526 l->min_pid = track->pid;
4527 l->max_pid = track->pid;
4528 cpumask_clear(to_cpumask(l->cpus));
4529 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4530 nodes_clear(l->nodes);
4531 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4535 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4536 struct page *page, enum track_item alloc,
4539 void *addr = page_address(page);
4542 bitmap_zero(map, page->objects);
4543 get_map(s, page, map);
4545 for_each_object(p, s, addr, page->objects)
4546 if (!test_bit(slab_index(p, s, addr), map))
4547 add_location(t, s, get_track(s, p, alloc));
4550 static int list_locations(struct kmem_cache *s, char *buf,
4551 enum track_item alloc)
4555 struct loc_track t = { 0, 0, NULL };
4557 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4558 sizeof(unsigned long), GFP_KERNEL);
4559 struct kmem_cache_node *n;
4561 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4564 return sprintf(buf, "Out of memory\n");
4566 /* Push back cpu slabs */
4569 for_each_kmem_cache_node(s, node, n) {
4570 unsigned long flags;
4573 if (!atomic_long_read(&n->nr_slabs))
4576 spin_lock_irqsave(&n->list_lock, flags);
4577 list_for_each_entry(page, &n->partial, lru)
4578 process_slab(&t, s, page, alloc, map);
4579 list_for_each_entry(page, &n->full, lru)
4580 process_slab(&t, s, page, alloc, map);
4581 spin_unlock_irqrestore(&n->list_lock, flags);
4584 for (i = 0; i < t.count; i++) {
4585 struct location *l = &t.loc[i];
4587 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4589 len += sprintf(buf + len, "%7ld ", l->count);
4592 len += sprintf(buf + len, "%pS", (void *)l->addr);
4594 len += sprintf(buf + len, "<not-available>");
4596 if (l->sum_time != l->min_time) {
4597 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4599 (long)div_u64(l->sum_time, l->count),
4602 len += sprintf(buf + len, " age=%ld",
4605 if (l->min_pid != l->max_pid)
4606 len += sprintf(buf + len, " pid=%ld-%ld",
4607 l->min_pid, l->max_pid);
4609 len += sprintf(buf + len, " pid=%ld",
4612 if (num_online_cpus() > 1 &&
4613 !cpumask_empty(to_cpumask(l->cpus)) &&
4614 len < PAGE_SIZE - 60)
4615 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4617 cpumask_pr_args(to_cpumask(l->cpus)));
4619 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4620 len < PAGE_SIZE - 60)
4621 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4623 nodemask_pr_args(&l->nodes));
4625 len += sprintf(buf + len, "\n");
4631 len += sprintf(buf, "No data\n");
4636 #ifdef SLUB_RESILIENCY_TEST
4637 static void __init resiliency_test(void)
4641 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4643 pr_err("SLUB resiliency testing\n");
4644 pr_err("-----------------------\n");
4645 pr_err("A. Corruption after allocation\n");
4647 p = kzalloc(16, GFP_KERNEL);
4649 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4652 validate_slab_cache(kmalloc_caches[4]);
4654 /* Hmmm... The next two are dangerous */
4655 p = kzalloc(32, GFP_KERNEL);
4656 p[32 + sizeof(void *)] = 0x34;
4657 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4659 pr_err("If allocated object is overwritten then not detectable\n\n");
4661 validate_slab_cache(kmalloc_caches[5]);
4662 p = kzalloc(64, GFP_KERNEL);
4663 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4665 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4667 pr_err("If allocated object is overwritten then not detectable\n\n");
4668 validate_slab_cache(kmalloc_caches[6]);
4670 pr_err("\nB. Corruption after free\n");
4671 p = kzalloc(128, GFP_KERNEL);
4674 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4675 validate_slab_cache(kmalloc_caches[7]);
4677 p = kzalloc(256, GFP_KERNEL);
4680 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4681 validate_slab_cache(kmalloc_caches[8]);
4683 p = kzalloc(512, GFP_KERNEL);
4686 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4687 validate_slab_cache(kmalloc_caches[9]);
4691 static void resiliency_test(void) {};
4696 enum slab_stat_type {
4697 SL_ALL, /* All slabs */
4698 SL_PARTIAL, /* Only partially allocated slabs */
4699 SL_CPU, /* Only slabs used for cpu caches */
4700 SL_OBJECTS, /* Determine allocated objects not slabs */
4701 SL_TOTAL /* Determine object capacity not slabs */
4704 #define SO_ALL (1 << SL_ALL)
4705 #define SO_PARTIAL (1 << SL_PARTIAL)
4706 #define SO_CPU (1 << SL_CPU)
4707 #define SO_OBJECTS (1 << SL_OBJECTS)
4708 #define SO_TOTAL (1 << SL_TOTAL)
4711 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4713 static int __init setup_slub_memcg_sysfs(char *str)
4717 if (get_option(&str, &v) > 0)
4718 memcg_sysfs_enabled = v;
4723 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4726 static ssize_t show_slab_objects(struct kmem_cache *s,
4727 char *buf, unsigned long flags)
4729 unsigned long total = 0;
4732 unsigned long *nodes;
4734 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4738 if (flags & SO_CPU) {
4741 for_each_possible_cpu(cpu) {
4742 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4747 page = READ_ONCE(c->page);
4751 node = page_to_nid(page);
4752 if (flags & SO_TOTAL)
4754 else if (flags & SO_OBJECTS)
4762 page = READ_ONCE(c->partial);
4764 node = page_to_nid(page);
4765 if (flags & SO_TOTAL)
4767 else if (flags & SO_OBJECTS)
4778 #ifdef CONFIG_SLUB_DEBUG
4779 if (flags & SO_ALL) {
4780 struct kmem_cache_node *n;
4782 for_each_kmem_cache_node(s, node, n) {
4784 if (flags & SO_TOTAL)
4785 x = atomic_long_read(&n->total_objects);
4786 else if (flags & SO_OBJECTS)
4787 x = atomic_long_read(&n->total_objects) -
4788 count_partial(n, count_free);
4790 x = atomic_long_read(&n->nr_slabs);
4797 if (flags & SO_PARTIAL) {
4798 struct kmem_cache_node *n;
4800 for_each_kmem_cache_node(s, node, n) {
4801 if (flags & SO_TOTAL)
4802 x = count_partial(n, count_total);
4803 else if (flags & SO_OBJECTS)
4804 x = count_partial(n, count_inuse);
4811 x = sprintf(buf, "%lu", total);
4813 for (node = 0; node < nr_node_ids; node++)
4815 x += sprintf(buf + x, " N%d=%lu",
4820 return x + sprintf(buf + x, "\n");
4823 #ifdef CONFIG_SLUB_DEBUG
4824 static int any_slab_objects(struct kmem_cache *s)
4827 struct kmem_cache_node *n;
4829 for_each_kmem_cache_node(s, node, n)
4830 if (atomic_long_read(&n->total_objects))
4837 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4838 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4840 struct slab_attribute {
4841 struct attribute attr;
4842 ssize_t (*show)(struct kmem_cache *s, char *buf);
4843 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4846 #define SLAB_ATTR_RO(_name) \
4847 static struct slab_attribute _name##_attr = \
4848 __ATTR(_name, 0400, _name##_show, NULL)
4850 #define SLAB_ATTR(_name) \
4851 static struct slab_attribute _name##_attr = \
4852 __ATTR(_name, 0600, _name##_show, _name##_store)
4854 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4856 return sprintf(buf, "%d\n", s->size);
4858 SLAB_ATTR_RO(slab_size);
4860 static ssize_t align_show(struct kmem_cache *s, char *buf)
4862 return sprintf(buf, "%d\n", s->align);
4864 SLAB_ATTR_RO(align);
4866 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4868 return sprintf(buf, "%d\n", s->object_size);
4870 SLAB_ATTR_RO(object_size);
4872 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4874 return sprintf(buf, "%d\n", oo_objects(s->oo));
4876 SLAB_ATTR_RO(objs_per_slab);
4878 static ssize_t order_store(struct kmem_cache *s,
4879 const char *buf, size_t length)
4881 unsigned long order;
4884 err = kstrtoul(buf, 10, &order);
4888 if (order > slub_max_order || order < slub_min_order)
4891 calculate_sizes(s, order);
4895 static ssize_t order_show(struct kmem_cache *s, char *buf)
4897 return sprintf(buf, "%d\n", oo_order(s->oo));
4901 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4903 return sprintf(buf, "%lu\n", s->min_partial);
4906 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4912 err = kstrtoul(buf, 10, &min);
4916 set_min_partial(s, min);
4919 SLAB_ATTR(min_partial);
4921 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4923 return sprintf(buf, "%u\n", s->cpu_partial);
4926 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4929 unsigned long objects;
4932 err = kstrtoul(buf, 10, &objects);
4935 if (objects && !kmem_cache_has_cpu_partial(s))
4938 s->cpu_partial = objects;
4942 SLAB_ATTR(cpu_partial);
4944 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4948 return sprintf(buf, "%pS\n", s->ctor);
4952 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4954 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4956 SLAB_ATTR_RO(aliases);
4958 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4960 return show_slab_objects(s, buf, SO_PARTIAL);
4962 SLAB_ATTR_RO(partial);
4964 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4966 return show_slab_objects(s, buf, SO_CPU);
4968 SLAB_ATTR_RO(cpu_slabs);
4970 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4972 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4974 SLAB_ATTR_RO(objects);
4976 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4978 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4980 SLAB_ATTR_RO(objects_partial);
4982 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4989 for_each_online_cpu(cpu) {
4990 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4993 pages += page->pages;
4994 objects += page->pobjects;
4998 len = sprintf(buf, "%d(%d)", objects, pages);
5001 for_each_online_cpu(cpu) {
5002 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
5004 if (page && len < PAGE_SIZE - 20)
5005 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5006 page->pobjects, page->pages);
5009 return len + sprintf(buf + len, "\n");
5011 SLAB_ATTR_RO(slabs_cpu_partial);
5013 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5015 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5018 static ssize_t reclaim_account_store(struct kmem_cache *s,
5019 const char *buf, size_t length)
5021 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5023 s->flags |= SLAB_RECLAIM_ACCOUNT;
5026 SLAB_ATTR(reclaim_account);
5028 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5030 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5032 SLAB_ATTR_RO(hwcache_align);
5034 #ifdef CONFIG_ZONE_DMA
5035 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5037 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5039 SLAB_ATTR_RO(cache_dma);
5042 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5044 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5046 SLAB_ATTR_RO(destroy_by_rcu);
5048 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
5050 return sprintf(buf, "%d\n", s->reserved);
5052 SLAB_ATTR_RO(reserved);
5054 #ifdef CONFIG_SLUB_DEBUG
5055 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5057 return show_slab_objects(s, buf, SO_ALL);
5059 SLAB_ATTR_RO(slabs);
5061 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5063 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5065 SLAB_ATTR_RO(total_objects);
5067 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5069 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5072 static ssize_t sanity_checks_store(struct kmem_cache *s,
5073 const char *buf, size_t length)
5075 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5076 if (buf[0] == '1') {
5077 s->flags &= ~__CMPXCHG_DOUBLE;
5078 s->flags |= SLAB_CONSISTENCY_CHECKS;
5082 SLAB_ATTR(sanity_checks);
5084 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5086 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5089 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5093 * Tracing a merged cache is going to give confusing results
5094 * as well as cause other issues like converting a mergeable
5095 * cache into an umergeable one.
5097 if (s->refcount > 1)
5100 s->flags &= ~SLAB_TRACE;
5101 if (buf[0] == '1') {
5102 s->flags &= ~__CMPXCHG_DOUBLE;
5103 s->flags |= SLAB_TRACE;
5109 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5111 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5114 static ssize_t red_zone_store(struct kmem_cache *s,
5115 const char *buf, size_t length)
5117 if (any_slab_objects(s))
5120 s->flags &= ~SLAB_RED_ZONE;
5121 if (buf[0] == '1') {
5122 s->flags |= SLAB_RED_ZONE;
5124 calculate_sizes(s, -1);
5127 SLAB_ATTR(red_zone);
5129 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5131 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5134 static ssize_t poison_store(struct kmem_cache *s,
5135 const char *buf, size_t length)
5137 if (any_slab_objects(s))
5140 s->flags &= ~SLAB_POISON;
5141 if (buf[0] == '1') {
5142 s->flags |= SLAB_POISON;
5144 calculate_sizes(s, -1);
5149 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5151 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5154 static ssize_t store_user_store(struct kmem_cache *s,
5155 const char *buf, size_t length)
5157 if (any_slab_objects(s))
5160 s->flags &= ~SLAB_STORE_USER;
5161 if (buf[0] == '1') {
5162 s->flags &= ~__CMPXCHG_DOUBLE;
5163 s->flags |= SLAB_STORE_USER;
5165 calculate_sizes(s, -1);
5168 SLAB_ATTR(store_user);
5170 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5175 static ssize_t validate_store(struct kmem_cache *s,
5176 const char *buf, size_t length)
5180 if (buf[0] == '1') {
5181 ret = validate_slab_cache(s);
5187 SLAB_ATTR(validate);
5189 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5191 if (!(s->flags & SLAB_STORE_USER))
5193 return list_locations(s, buf, TRACK_ALLOC);
5195 SLAB_ATTR_RO(alloc_calls);
5197 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5199 if (!(s->flags & SLAB_STORE_USER))
5201 return list_locations(s, buf, TRACK_FREE);
5203 SLAB_ATTR_RO(free_calls);
5204 #endif /* CONFIG_SLUB_DEBUG */
5206 #ifdef CONFIG_FAILSLAB
5207 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5209 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5212 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5215 if (s->refcount > 1)
5218 s->flags &= ~SLAB_FAILSLAB;
5220 s->flags |= SLAB_FAILSLAB;
5223 SLAB_ATTR(failslab);
5226 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5231 static ssize_t shrink_store(struct kmem_cache *s,
5232 const char *buf, size_t length)
5235 kmem_cache_shrink(s);
5243 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5245 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5248 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5249 const char *buf, size_t length)
5251 unsigned long ratio;
5254 err = kstrtoul(buf, 10, &ratio);
5259 s->remote_node_defrag_ratio = ratio * 10;
5263 SLAB_ATTR(remote_node_defrag_ratio);
5266 #ifdef CONFIG_SLUB_STATS
5267 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5269 unsigned long sum = 0;
5272 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5277 for_each_online_cpu(cpu) {
5278 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5284 len = sprintf(buf, "%lu", sum);
5287 for_each_online_cpu(cpu) {
5288 if (data[cpu] && len < PAGE_SIZE - 20)
5289 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5293 return len + sprintf(buf + len, "\n");
5296 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5300 for_each_online_cpu(cpu)
5301 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5304 #define STAT_ATTR(si, text) \
5305 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5307 return show_stat(s, buf, si); \
5309 static ssize_t text##_store(struct kmem_cache *s, \
5310 const char *buf, size_t length) \
5312 if (buf[0] != '0') \
5314 clear_stat(s, si); \
5319 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5320 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5321 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5322 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5323 STAT_ATTR(FREE_FROZEN, free_frozen);
5324 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5325 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5326 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5327 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5328 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5329 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5330 STAT_ATTR(FREE_SLAB, free_slab);
5331 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5332 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5333 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5334 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5335 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5336 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5337 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5338 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5339 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5340 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5341 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5342 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5343 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5344 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5347 static struct attribute *slab_attrs[] = {
5348 &slab_size_attr.attr,
5349 &object_size_attr.attr,
5350 &objs_per_slab_attr.attr,
5352 &min_partial_attr.attr,
5353 &cpu_partial_attr.attr,
5355 &objects_partial_attr.attr,
5357 &cpu_slabs_attr.attr,
5361 &hwcache_align_attr.attr,
5362 &reclaim_account_attr.attr,
5363 &destroy_by_rcu_attr.attr,
5365 &reserved_attr.attr,
5366 &slabs_cpu_partial_attr.attr,
5367 #ifdef CONFIG_SLUB_DEBUG
5368 &total_objects_attr.attr,
5370 &sanity_checks_attr.attr,
5372 &red_zone_attr.attr,
5374 &store_user_attr.attr,
5375 &validate_attr.attr,
5376 &alloc_calls_attr.attr,
5377 &free_calls_attr.attr,
5379 #ifdef CONFIG_ZONE_DMA
5380 &cache_dma_attr.attr,
5383 &remote_node_defrag_ratio_attr.attr,
5385 #ifdef CONFIG_SLUB_STATS
5386 &alloc_fastpath_attr.attr,
5387 &alloc_slowpath_attr.attr,
5388 &free_fastpath_attr.attr,
5389 &free_slowpath_attr.attr,
5390 &free_frozen_attr.attr,
5391 &free_add_partial_attr.attr,
5392 &free_remove_partial_attr.attr,
5393 &alloc_from_partial_attr.attr,
5394 &alloc_slab_attr.attr,
5395 &alloc_refill_attr.attr,
5396 &alloc_node_mismatch_attr.attr,
5397 &free_slab_attr.attr,
5398 &cpuslab_flush_attr.attr,
5399 &deactivate_full_attr.attr,
5400 &deactivate_empty_attr.attr,
5401 &deactivate_to_head_attr.attr,
5402 &deactivate_to_tail_attr.attr,
5403 &deactivate_remote_frees_attr.attr,
5404 &deactivate_bypass_attr.attr,
5405 &order_fallback_attr.attr,
5406 &cmpxchg_double_fail_attr.attr,
5407 &cmpxchg_double_cpu_fail_attr.attr,
5408 &cpu_partial_alloc_attr.attr,
5409 &cpu_partial_free_attr.attr,
5410 &cpu_partial_node_attr.attr,
5411 &cpu_partial_drain_attr.attr,
5413 #ifdef CONFIG_FAILSLAB
5414 &failslab_attr.attr,
5420 static struct attribute_group slab_attr_group = {
5421 .attrs = slab_attrs,
5424 static ssize_t slab_attr_show(struct kobject *kobj,
5425 struct attribute *attr,
5428 struct slab_attribute *attribute;
5429 struct kmem_cache *s;
5432 attribute = to_slab_attr(attr);
5435 if (!attribute->show)
5438 err = attribute->show(s, buf);
5443 static ssize_t slab_attr_store(struct kobject *kobj,
5444 struct attribute *attr,
5445 const char *buf, size_t len)
5447 struct slab_attribute *attribute;
5448 struct kmem_cache *s;
5451 attribute = to_slab_attr(attr);
5454 if (!attribute->store)
5457 err = attribute->store(s, buf, len);
5459 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5460 struct kmem_cache *c;
5462 mutex_lock(&slab_mutex);
5463 if (s->max_attr_size < len)
5464 s->max_attr_size = len;
5467 * This is a best effort propagation, so this function's return
5468 * value will be determined by the parent cache only. This is
5469 * basically because not all attributes will have a well
5470 * defined semantics for rollbacks - most of the actions will
5471 * have permanent effects.
5473 * Returning the error value of any of the children that fail
5474 * is not 100 % defined, in the sense that users seeing the
5475 * error code won't be able to know anything about the state of
5478 * Only returning the error code for the parent cache at least
5479 * has well defined semantics. The cache being written to
5480 * directly either failed or succeeded, in which case we loop
5481 * through the descendants with best-effort propagation.
5483 for_each_memcg_cache(c, s)
5484 attribute->store(c, buf, len);
5485 mutex_unlock(&slab_mutex);
5491 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5495 char *buffer = NULL;
5496 struct kmem_cache *root_cache;
5498 if (is_root_cache(s))
5501 root_cache = s->memcg_params.root_cache;
5504 * This mean this cache had no attribute written. Therefore, no point
5505 * in copying default values around
5507 if (!root_cache->max_attr_size)
5510 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5513 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5516 if (!attr || !attr->store || !attr->show)
5520 * It is really bad that we have to allocate here, so we will
5521 * do it only as a fallback. If we actually allocate, though,
5522 * we can just use the allocated buffer until the end.
5524 * Most of the slub attributes will tend to be very small in
5525 * size, but sysfs allows buffers up to a page, so they can
5526 * theoretically happen.
5530 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5533 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5534 if (WARN_ON(!buffer))
5539 len = attr->show(root_cache, buf);
5541 attr->store(s, buf, len);
5545 free_page((unsigned long)buffer);
5549 static void kmem_cache_release(struct kobject *k)
5551 slab_kmem_cache_release(to_slab(k));
5554 static const struct sysfs_ops slab_sysfs_ops = {
5555 .show = slab_attr_show,
5556 .store = slab_attr_store,
5559 static struct kobj_type slab_ktype = {
5560 .sysfs_ops = &slab_sysfs_ops,
5561 .release = kmem_cache_release,
5564 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5566 struct kobj_type *ktype = get_ktype(kobj);
5568 if (ktype == &slab_ktype)
5573 static const struct kset_uevent_ops slab_uevent_ops = {
5574 .filter = uevent_filter,
5577 static struct kset *slab_kset;
5579 static inline struct kset *cache_kset(struct kmem_cache *s)
5582 if (!is_root_cache(s))
5583 return s->memcg_params.root_cache->memcg_kset;
5588 #define ID_STR_LENGTH 64
5590 /* Create a unique string id for a slab cache:
5592 * Format :[flags-]size
5594 static char *create_unique_id(struct kmem_cache *s)
5596 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5603 * First flags affecting slabcache operations. We will only
5604 * get here for aliasable slabs so we do not need to support
5605 * too many flags. The flags here must cover all flags that
5606 * are matched during merging to guarantee that the id is
5609 if (s->flags & SLAB_CACHE_DMA)
5611 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5613 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5615 if (!(s->flags & SLAB_NOTRACK))
5617 if (s->flags & SLAB_ACCOUNT)
5621 p += sprintf(p, "%07d", s->size);
5623 BUG_ON(p > name + ID_STR_LENGTH - 1);
5627 static void sysfs_slab_remove_workfn(struct work_struct *work)
5629 struct kmem_cache *s =
5630 container_of(work, struct kmem_cache, kobj_remove_work);
5632 if (!s->kobj.state_in_sysfs)
5634 * For a memcg cache, this may be called during
5635 * deactivation and again on shutdown. Remove only once.
5636 * A cache is never shut down before deactivation is
5637 * complete, so no need to worry about synchronization.
5642 kset_unregister(s->memcg_kset);
5644 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5645 kobject_del(&s->kobj);
5646 kobject_put(&s->kobj);
5649 static int sysfs_slab_add(struct kmem_cache *s)
5653 struct kset *kset = cache_kset(s);
5654 int unmergeable = slab_unmergeable(s);
5656 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5659 kobject_init(&s->kobj, &slab_ktype);
5665 * Slabcache can never be merged so we can use the name proper.
5666 * This is typically the case for debug situations. In that
5667 * case we can catch duplicate names easily.
5669 sysfs_remove_link(&slab_kset->kobj, s->name);
5673 * Create a unique name for the slab as a target
5676 name = create_unique_id(s);
5679 s->kobj.kset = kset;
5680 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5684 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5689 if (is_root_cache(s) && memcg_sysfs_enabled) {
5690 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5691 if (!s->memcg_kset) {
5698 kobject_uevent(&s->kobj, KOBJ_ADD);
5700 /* Setup first alias */
5701 sysfs_slab_alias(s, s->name);
5708 kobject_del(&s->kobj);
5712 static void sysfs_slab_remove(struct kmem_cache *s)
5714 if (slab_state < FULL)
5716 * Sysfs has not been setup yet so no need to remove the
5721 kobject_get(&s->kobj);
5722 schedule_work(&s->kobj_remove_work);
5725 void sysfs_slab_release(struct kmem_cache *s)
5727 if (slab_state >= FULL)
5728 kobject_put(&s->kobj);
5732 * Need to buffer aliases during bootup until sysfs becomes
5733 * available lest we lose that information.
5735 struct saved_alias {
5736 struct kmem_cache *s;
5738 struct saved_alias *next;
5741 static struct saved_alias *alias_list;
5743 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5745 struct saved_alias *al;
5747 if (slab_state == FULL) {
5749 * If we have a leftover link then remove it.
5751 sysfs_remove_link(&slab_kset->kobj, name);
5752 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5755 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5761 al->next = alias_list;
5766 static int __init slab_sysfs_init(void)
5768 struct kmem_cache *s;
5771 mutex_lock(&slab_mutex);
5773 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5775 mutex_unlock(&slab_mutex);
5776 pr_err("Cannot register slab subsystem.\n");
5782 list_for_each_entry(s, &slab_caches, list) {
5783 err = sysfs_slab_add(s);
5785 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5789 while (alias_list) {
5790 struct saved_alias *al = alias_list;
5792 alias_list = alias_list->next;
5793 err = sysfs_slab_alias(al->s, al->name);
5795 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5800 mutex_unlock(&slab_mutex);
5805 __initcall(slab_sysfs_init);
5806 #endif /* CONFIG_SYSFS */
5809 * The /proc/slabinfo ABI
5811 #ifdef CONFIG_SLABINFO
5812 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5814 unsigned long nr_slabs = 0;
5815 unsigned long nr_objs = 0;
5816 unsigned long nr_free = 0;
5818 struct kmem_cache_node *n;
5820 for_each_kmem_cache_node(s, node, n) {
5821 nr_slabs += node_nr_slabs(n);
5822 nr_objs += node_nr_objs(n);
5823 nr_free += count_partial(n, count_free);
5826 sinfo->active_objs = nr_objs - nr_free;
5827 sinfo->num_objs = nr_objs;
5828 sinfo->active_slabs = nr_slabs;
5829 sinfo->num_slabs = nr_slabs;
5830 sinfo->objects_per_slab = oo_objects(s->oo);
5831 sinfo->cache_order = oo_order(s->oo);
5834 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5838 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5839 size_t count, loff_t *ppos)
5843 #endif /* CONFIG_SLABINFO */