2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
13 #include <linux/swap.h> /* struct reclaim_state */
14 #include <linux/module.h>
15 #include <linux/bit_spinlock.h>
16 #include <linux/interrupt.h>
17 #include <linux/bitops.h>
18 #include <linux/slab.h>
20 #include <linux/proc_fs.h>
21 #include <linux/notifier.h>
22 #include <linux/seq_file.h>
23 #include <linux/kasan.h>
24 #include <linux/kmemcheck.h>
25 #include <linux/cpu.h>
26 #include <linux/cpuset.h>
27 #include <linux/mempolicy.h>
28 #include <linux/ctype.h>
29 #include <linux/debugobjects.h>
30 #include <linux/kallsyms.h>
31 #include <linux/memory.h>
32 #include <linux/math64.h>
33 #include <linux/fault-inject.h>
34 #include <linux/stacktrace.h>
35 #include <linux/prefetch.h>
36 #include <linux/memcontrol.h>
38 #include <trace/events/kmem.h>
44 * 1. slab_mutex (Global Mutex)
46 * 3. slab_lock(page) (Only on some arches and for debugging)
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
95 * Overloading of page flags that are otherwise used for LRU management.
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
118 static inline int kmem_cache_debug(struct kmem_cache *s)
120 #ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
127 void *fixup_red_left(struct kmem_cache *s, void *p)
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
135 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
137 #ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
145 * Issues still to be resolved:
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 * - Variable sizing of the per node arrays
152 /* Enable to test recovery from slab corruption on boot */
153 #undef SLUB_RESILIENCY_TEST
155 /* Enable to log cmpxchg failures */
156 #undef SLUB_DEBUG_CMPXCHG
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 #define MIN_PARTIAL 5
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
169 #define MAX_PARTIAL 10
171 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
178 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
187 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
190 #define OO_MASK ((1 << OO_SHIFT) - 1)
191 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193 /* Internal SLUB flags */
194 #define __OBJECT_POISON 0x80000000UL /* Poison object */
195 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
198 * Tracking user of a slab.
200 #define TRACK_ADDRS_COUNT 16
202 unsigned long addr; /* Called from address */
203 #ifdef CONFIG_STACKTRACE
204 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
206 int cpu; /* Was running on cpu */
207 int pid; /* Pid context */
208 unsigned long when; /* When did the operation occur */
211 enum track_item { TRACK_ALLOC, TRACK_FREE };
214 static int sysfs_slab_add(struct kmem_cache *);
215 static int sysfs_slab_alias(struct kmem_cache *, const char *);
216 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
218 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
221 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
224 static inline void stat(const struct kmem_cache *s, enum stat_item si)
226 #ifdef CONFIG_SLUB_STATS
228 * The rmw is racy on a preemptible kernel but this is acceptable, so
229 * avoid this_cpu_add()'s irq-disable overhead.
231 raw_cpu_inc(s->cpu_slab->stat[si]);
235 /********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
239 static inline void *get_freepointer(struct kmem_cache *s, void *object)
241 return *(void **)(object + s->offset);
244 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
246 prefetch(object + s->offset);
249 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
253 if (!debug_pagealloc_enabled())
254 return get_freepointer(s, object);
256 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
260 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
262 *(void **)(object + s->offset) = fp;
265 /* Loop over all objects in a slab */
266 #define for_each_object(__p, __s, __addr, __objects) \
267 for (__p = fixup_red_left(__s, __addr); \
268 __p < (__addr) + (__objects) * (__s)->size; \
271 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
272 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
273 __idx <= __objects; \
274 __p += (__s)->size, __idx++)
276 /* Determine object index from a given position */
277 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
279 return (p - addr) / s->size;
282 static inline int order_objects(int order, unsigned long size, int reserved)
284 return ((PAGE_SIZE << order) - reserved) / size;
287 static inline struct kmem_cache_order_objects oo_make(int order,
288 unsigned long size, int reserved)
290 struct kmem_cache_order_objects x = {
291 (order << OO_SHIFT) + order_objects(order, size, reserved)
297 static inline int oo_order(struct kmem_cache_order_objects x)
299 return x.x >> OO_SHIFT;
302 static inline int oo_objects(struct kmem_cache_order_objects x)
304 return x.x & OO_MASK;
308 * Per slab locking using the pagelock
310 static __always_inline void slab_lock(struct page *page)
312 VM_BUG_ON_PAGE(PageTail(page), page);
313 bit_spin_lock(PG_locked, &page->flags);
316 static __always_inline void slab_unlock(struct page *page)
318 VM_BUG_ON_PAGE(PageTail(page), page);
319 __bit_spin_unlock(PG_locked, &page->flags);
322 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
325 tmp.counters = counters_new;
327 * page->counters can cover frozen/inuse/objects as well
328 * as page->_refcount. If we assign to ->counters directly
329 * we run the risk of losing updates to page->_refcount, so
330 * be careful and only assign to the fields we need.
332 page->frozen = tmp.frozen;
333 page->inuse = tmp.inuse;
334 page->objects = tmp.objects;
337 /* Interrupts must be disabled (for the fallback code to work right) */
338 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
339 void *freelist_old, unsigned long counters_old,
340 void *freelist_new, unsigned long counters_new,
343 VM_BUG_ON(!irqs_disabled());
344 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
345 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
346 if (s->flags & __CMPXCHG_DOUBLE) {
347 if (cmpxchg_double(&page->freelist, &page->counters,
348 freelist_old, counters_old,
349 freelist_new, counters_new))
355 if (page->freelist == freelist_old &&
356 page->counters == counters_old) {
357 page->freelist = freelist_new;
358 set_page_slub_counters(page, counters_new);
366 stat(s, CMPXCHG_DOUBLE_FAIL);
368 #ifdef SLUB_DEBUG_CMPXCHG
369 pr_info("%s %s: cmpxchg double redo ", n, s->name);
375 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
380 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
381 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
382 if (s->flags & __CMPXCHG_DOUBLE) {
383 if (cmpxchg_double(&page->freelist, &page->counters,
384 freelist_old, counters_old,
385 freelist_new, counters_new))
392 local_irq_save(flags);
394 if (page->freelist == freelist_old &&
395 page->counters == counters_old) {
396 page->freelist = freelist_new;
397 set_page_slub_counters(page, counters_new);
399 local_irq_restore(flags);
403 local_irq_restore(flags);
407 stat(s, CMPXCHG_DOUBLE_FAIL);
409 #ifdef SLUB_DEBUG_CMPXCHG
410 pr_info("%s %s: cmpxchg double redo ", n, s->name);
416 #ifdef CONFIG_SLUB_DEBUG
418 * Determine a map of object in use on a page.
420 * Node listlock must be held to guarantee that the page does
421 * not vanish from under us.
423 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
426 void *addr = page_address(page);
428 for (p = page->freelist; p; p = get_freepointer(s, p))
429 set_bit(slab_index(p, s, addr), map);
432 static inline int size_from_object(struct kmem_cache *s)
434 if (s->flags & SLAB_RED_ZONE)
435 return s->size - s->red_left_pad;
440 static inline void *restore_red_left(struct kmem_cache *s, void *p)
442 if (s->flags & SLAB_RED_ZONE)
443 p -= s->red_left_pad;
451 #if defined(CONFIG_SLUB_DEBUG_ON)
452 static int slub_debug = DEBUG_DEFAULT_FLAGS;
454 static int slub_debug;
457 static char *slub_debug_slabs;
458 static int disable_higher_order_debug;
461 * slub is about to manipulate internal object metadata. This memory lies
462 * outside the range of the allocated object, so accessing it would normally
463 * be reported by kasan as a bounds error. metadata_access_enable() is used
464 * to tell kasan that these accesses are OK.
466 static inline void metadata_access_enable(void)
468 kasan_disable_current();
471 static inline void metadata_access_disable(void)
473 kasan_enable_current();
480 /* Verify that a pointer has an address that is valid within a slab page */
481 static inline int check_valid_pointer(struct kmem_cache *s,
482 struct page *page, void *object)
489 base = page_address(page);
490 object = restore_red_left(s, object);
491 if (object < base || object >= base + page->objects * s->size ||
492 (object - base) % s->size) {
499 static void print_section(char *level, char *text, u8 *addr,
502 metadata_access_enable();
503 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
505 metadata_access_disable();
508 static struct track *get_track(struct kmem_cache *s, void *object,
509 enum track_item alloc)
514 p = object + s->offset + sizeof(void *);
516 p = object + s->inuse;
521 static void set_track(struct kmem_cache *s, void *object,
522 enum track_item alloc, unsigned long addr)
524 struct track *p = get_track(s, object, alloc);
527 #ifdef CONFIG_STACKTRACE
528 struct stack_trace trace;
531 trace.nr_entries = 0;
532 trace.max_entries = TRACK_ADDRS_COUNT;
533 trace.entries = p->addrs;
535 metadata_access_enable();
536 save_stack_trace(&trace);
537 metadata_access_disable();
539 /* See rant in lockdep.c */
540 if (trace.nr_entries != 0 &&
541 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
544 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
548 p->cpu = smp_processor_id();
549 p->pid = current->pid;
552 memset(p, 0, sizeof(struct track));
555 static void init_tracking(struct kmem_cache *s, void *object)
557 if (!(s->flags & SLAB_STORE_USER))
560 set_track(s, object, TRACK_FREE, 0UL);
561 set_track(s, object, TRACK_ALLOC, 0UL);
564 static void print_track(const char *s, struct track *t)
569 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
570 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
571 #ifdef CONFIG_STACKTRACE
574 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
576 pr_err("\t%pS\n", (void *)t->addrs[i]);
583 static void print_tracking(struct kmem_cache *s, void *object)
585 if (!(s->flags & SLAB_STORE_USER))
588 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
589 print_track("Freed", get_track(s, object, TRACK_FREE));
592 static void print_page_info(struct page *page)
594 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
595 page, page->objects, page->inuse, page->freelist, page->flags);
599 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
601 struct va_format vaf;
607 pr_err("=============================================================================\n");
608 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
609 pr_err("-----------------------------------------------------------------------------\n\n");
611 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
615 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
617 struct va_format vaf;
623 pr_err("FIX %s: %pV\n", s->name, &vaf);
627 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
628 void **freelist, void *nextfree)
630 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
631 !check_valid_pointer(s, page, nextfree) && freelist) {
632 object_err(s, page, *freelist, "Freechain corrupt");
634 slab_fix(s, "Isolate corrupted freechain");
641 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
643 unsigned int off; /* Offset of last byte */
644 u8 *addr = page_address(page);
646 print_tracking(s, p);
648 print_page_info(page);
650 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
651 p, p - addr, get_freepointer(s, p));
653 if (s->flags & SLAB_RED_ZONE)
654 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
656 else if (p > addr + 16)
657 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
659 print_section(KERN_ERR, "Object ", p,
660 min_t(unsigned long, s->object_size, PAGE_SIZE));
661 if (s->flags & SLAB_RED_ZONE)
662 print_section(KERN_ERR, "Redzone ", p + s->object_size,
663 s->inuse - s->object_size);
666 off = s->offset + sizeof(void *);
670 if (s->flags & SLAB_STORE_USER)
671 off += 2 * sizeof(struct track);
673 off += kasan_metadata_size(s);
675 if (off != size_from_object(s))
676 /* Beginning of the filler is the free pointer */
677 print_section(KERN_ERR, "Padding ", p + off,
678 size_from_object(s) - off);
683 void object_err(struct kmem_cache *s, struct page *page,
684 u8 *object, char *reason)
686 slab_bug(s, "%s", reason);
687 print_trailer(s, page, object);
690 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
691 const char *fmt, ...)
697 vsnprintf(buf, sizeof(buf), fmt, args);
699 slab_bug(s, "%s", buf);
700 print_page_info(page);
704 static void init_object(struct kmem_cache *s, void *object, u8 val)
708 if (s->flags & SLAB_RED_ZONE)
709 memset(p - s->red_left_pad, val, s->red_left_pad);
711 if (s->flags & __OBJECT_POISON) {
712 memset(p, POISON_FREE, s->object_size - 1);
713 p[s->object_size - 1] = POISON_END;
716 if (s->flags & SLAB_RED_ZONE)
717 memset(p + s->object_size, val, s->inuse - s->object_size);
720 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
721 void *from, void *to)
723 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
724 memset(from, data, to - from);
727 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
728 u8 *object, char *what,
729 u8 *start, unsigned int value, unsigned int bytes)
734 metadata_access_enable();
735 fault = memchr_inv(start, value, bytes);
736 metadata_access_disable();
741 while (end > fault && end[-1] == value)
744 slab_bug(s, "%s overwritten", what);
745 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
746 fault, end - 1, fault[0], value);
747 print_trailer(s, page, object);
749 restore_bytes(s, what, value, fault, end);
757 * Bytes of the object to be managed.
758 * If the freepointer may overlay the object then the free
759 * pointer is the first word of the object.
761 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
764 * object + s->object_size
765 * Padding to reach word boundary. This is also used for Redzoning.
766 * Padding is extended by another word if Redzoning is enabled and
767 * object_size == inuse.
769 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
770 * 0xcc (RED_ACTIVE) for objects in use.
773 * Meta data starts here.
775 * A. Free pointer (if we cannot overwrite object on free)
776 * B. Tracking data for SLAB_STORE_USER
777 * C. Padding to reach required alignment boundary or at mininum
778 * one word if debugging is on to be able to detect writes
779 * before the word boundary.
781 * Padding is done using 0x5a (POISON_INUSE)
784 * Nothing is used beyond s->size.
786 * If slabcaches are merged then the object_size and inuse boundaries are mostly
787 * ignored. And therefore no slab options that rely on these boundaries
788 * may be used with merged slabcaches.
791 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
793 unsigned long off = s->inuse; /* The end of info */
796 /* Freepointer is placed after the object. */
797 off += sizeof(void *);
799 if (s->flags & SLAB_STORE_USER)
800 /* We also have user information there */
801 off += 2 * sizeof(struct track);
803 off += kasan_metadata_size(s);
805 if (size_from_object(s) == off)
808 return check_bytes_and_report(s, page, p, "Object padding",
809 p + off, POISON_INUSE, size_from_object(s) - off);
812 /* Check the pad bytes at the end of a slab page */
813 static int slab_pad_check(struct kmem_cache *s, struct page *page)
821 if (!(s->flags & SLAB_POISON))
824 start = page_address(page);
825 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
826 end = start + length;
827 remainder = length % s->size;
831 metadata_access_enable();
832 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
833 metadata_access_disable();
836 while (end > fault && end[-1] == POISON_INUSE)
839 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
840 print_section(KERN_ERR, "Padding ", end - remainder, remainder);
842 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
846 static int check_object(struct kmem_cache *s, struct page *page,
847 void *object, u8 val)
850 u8 *endobject = object + s->object_size;
852 if (s->flags & SLAB_RED_ZONE) {
853 if (!check_bytes_and_report(s, page, object, "Redzone",
854 object - s->red_left_pad, val, s->red_left_pad))
857 if (!check_bytes_and_report(s, page, object, "Redzone",
858 endobject, val, s->inuse - s->object_size))
861 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
862 check_bytes_and_report(s, page, p, "Alignment padding",
863 endobject, POISON_INUSE,
864 s->inuse - s->object_size);
868 if (s->flags & SLAB_POISON) {
869 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
870 (!check_bytes_and_report(s, page, p, "Poison", p,
871 POISON_FREE, s->object_size - 1) ||
872 !check_bytes_and_report(s, page, p, "Poison",
873 p + s->object_size - 1, POISON_END, 1)))
876 * check_pad_bytes cleans up on its own.
878 check_pad_bytes(s, page, p);
881 if (!s->offset && val == SLUB_RED_ACTIVE)
883 * Object and freepointer overlap. Cannot check
884 * freepointer while object is allocated.
888 /* Check free pointer validity */
889 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
890 object_err(s, page, p, "Freepointer corrupt");
892 * No choice but to zap it and thus lose the remainder
893 * of the free objects in this slab. May cause
894 * another error because the object count is now wrong.
896 set_freepointer(s, p, NULL);
902 static int check_slab(struct kmem_cache *s, struct page *page)
906 VM_BUG_ON(!irqs_disabled());
908 if (!PageSlab(page)) {
909 slab_err(s, page, "Not a valid slab page");
913 maxobj = order_objects(compound_order(page), s->size, s->reserved);
914 if (page->objects > maxobj) {
915 slab_err(s, page, "objects %u > max %u",
916 page->objects, maxobj);
919 if (page->inuse > page->objects) {
920 slab_err(s, page, "inuse %u > max %u",
921 page->inuse, page->objects);
924 /* Slab_pad_check fixes things up after itself */
925 slab_pad_check(s, page);
930 * Determine if a certain object on a page is on the freelist. Must hold the
931 * slab lock to guarantee that the chains are in a consistent state.
933 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
941 while (fp && nr <= page->objects) {
944 if (!check_valid_pointer(s, page, fp)) {
946 object_err(s, page, object,
947 "Freechain corrupt");
948 set_freepointer(s, object, NULL);
950 slab_err(s, page, "Freepointer corrupt");
951 page->freelist = NULL;
952 page->inuse = page->objects;
953 slab_fix(s, "Freelist cleared");
959 fp = get_freepointer(s, object);
963 max_objects = order_objects(compound_order(page), s->size, s->reserved);
964 if (max_objects > MAX_OBJS_PER_PAGE)
965 max_objects = MAX_OBJS_PER_PAGE;
967 if (page->objects != max_objects) {
968 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
969 page->objects, max_objects);
970 page->objects = max_objects;
971 slab_fix(s, "Number of objects adjusted.");
973 if (page->inuse != page->objects - nr) {
974 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
975 page->inuse, page->objects - nr);
976 page->inuse = page->objects - nr;
977 slab_fix(s, "Object count adjusted.");
979 return search == NULL;
982 static void trace(struct kmem_cache *s, struct page *page, void *object,
985 if (s->flags & SLAB_TRACE) {
986 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
988 alloc ? "alloc" : "free",
993 print_section(KERN_INFO, "Object ", (void *)object,
1001 * Tracking of fully allocated slabs for debugging purposes.
1003 static void add_full(struct kmem_cache *s,
1004 struct kmem_cache_node *n, struct page *page)
1006 if (!(s->flags & SLAB_STORE_USER))
1009 lockdep_assert_held(&n->list_lock);
1010 list_add(&page->lru, &n->full);
1013 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1015 if (!(s->flags & SLAB_STORE_USER))
1018 lockdep_assert_held(&n->list_lock);
1019 list_del(&page->lru);
1022 /* Tracking of the number of slabs for debugging purposes */
1023 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1025 struct kmem_cache_node *n = get_node(s, node);
1027 return atomic_long_read(&n->nr_slabs);
1030 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1032 return atomic_long_read(&n->nr_slabs);
1035 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1037 struct kmem_cache_node *n = get_node(s, node);
1040 * May be called early in order to allocate a slab for the
1041 * kmem_cache_node structure. Solve the chicken-egg
1042 * dilemma by deferring the increment of the count during
1043 * bootstrap (see early_kmem_cache_node_alloc).
1046 atomic_long_inc(&n->nr_slabs);
1047 atomic_long_add(objects, &n->total_objects);
1050 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1052 struct kmem_cache_node *n = get_node(s, node);
1054 atomic_long_dec(&n->nr_slabs);
1055 atomic_long_sub(objects, &n->total_objects);
1058 /* Object debug checks for alloc/free paths */
1059 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1062 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1065 init_object(s, object, SLUB_RED_INACTIVE);
1066 init_tracking(s, object);
1069 static inline int alloc_consistency_checks(struct kmem_cache *s,
1071 void *object, unsigned long addr)
1073 if (!check_slab(s, page))
1076 if (!check_valid_pointer(s, page, object)) {
1077 object_err(s, page, object, "Freelist Pointer check fails");
1081 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1087 static noinline int alloc_debug_processing(struct kmem_cache *s,
1089 void *object, unsigned long addr)
1091 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1092 if (!alloc_consistency_checks(s, page, object, addr))
1096 /* Success perform special debug activities for allocs */
1097 if (s->flags & SLAB_STORE_USER)
1098 set_track(s, object, TRACK_ALLOC, addr);
1099 trace(s, page, object, 1);
1100 init_object(s, object, SLUB_RED_ACTIVE);
1104 if (PageSlab(page)) {
1106 * If this is a slab page then lets do the best we can
1107 * to avoid issues in the future. Marking all objects
1108 * as used avoids touching the remaining objects.
1110 slab_fix(s, "Marking all objects used");
1111 page->inuse = page->objects;
1112 page->freelist = NULL;
1117 static inline int free_consistency_checks(struct kmem_cache *s,
1118 struct page *page, void *object, unsigned long addr)
1120 if (!check_valid_pointer(s, page, object)) {
1121 slab_err(s, page, "Invalid object pointer 0x%p", object);
1125 if (on_freelist(s, page, object)) {
1126 object_err(s, page, object, "Object already free");
1130 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1133 if (unlikely(s != page->slab_cache)) {
1134 if (!PageSlab(page)) {
1135 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1137 } else if (!page->slab_cache) {
1138 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1142 object_err(s, page, object,
1143 "page slab pointer corrupt.");
1149 /* Supports checking bulk free of a constructed freelist */
1150 static noinline int free_debug_processing(
1151 struct kmem_cache *s, struct page *page,
1152 void *head, void *tail, int bulk_cnt,
1155 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1156 void *object = head;
1158 unsigned long uninitialized_var(flags);
1161 spin_lock_irqsave(&n->list_lock, flags);
1164 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1165 if (!check_slab(s, page))
1172 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1173 if (!free_consistency_checks(s, page, object, addr))
1177 if (s->flags & SLAB_STORE_USER)
1178 set_track(s, object, TRACK_FREE, addr);
1179 trace(s, page, object, 0);
1180 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1181 init_object(s, object, SLUB_RED_INACTIVE);
1183 /* Reached end of constructed freelist yet? */
1184 if (object != tail) {
1185 object = get_freepointer(s, object);
1191 if (cnt != bulk_cnt)
1192 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1196 spin_unlock_irqrestore(&n->list_lock, flags);
1198 slab_fix(s, "Object at 0x%p not freed", object);
1202 static int __init setup_slub_debug(char *str)
1204 slub_debug = DEBUG_DEFAULT_FLAGS;
1205 if (*str++ != '=' || !*str)
1207 * No options specified. Switch on full debugging.
1213 * No options but restriction on slabs. This means full
1214 * debugging for slabs matching a pattern.
1221 * Switch off all debugging measures.
1226 * Determine which debug features should be switched on
1228 for (; *str && *str != ','; str++) {
1229 switch (tolower(*str)) {
1231 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1234 slub_debug |= SLAB_RED_ZONE;
1237 slub_debug |= SLAB_POISON;
1240 slub_debug |= SLAB_STORE_USER;
1243 slub_debug |= SLAB_TRACE;
1246 slub_debug |= SLAB_FAILSLAB;
1250 * Avoid enabling debugging on caches if its minimum
1251 * order would increase as a result.
1253 disable_higher_order_debug = 1;
1256 pr_err("slub_debug option '%c' unknown. skipped\n",
1263 slub_debug_slabs = str + 1;
1268 __setup("slub_debug", setup_slub_debug);
1270 unsigned long kmem_cache_flags(unsigned long object_size,
1271 unsigned long flags, const char *name,
1272 void (*ctor)(void *))
1275 * Enable debugging if selected on the kernel commandline.
1277 if (slub_debug && (!slub_debug_slabs || (name &&
1278 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1279 flags |= slub_debug;
1283 #else /* !CONFIG_SLUB_DEBUG */
1284 static inline void setup_object_debug(struct kmem_cache *s,
1285 struct page *page, void *object) {}
1287 static inline int alloc_debug_processing(struct kmem_cache *s,
1288 struct page *page, void *object, unsigned long addr) { return 0; }
1290 static inline int free_debug_processing(
1291 struct kmem_cache *s, struct page *page,
1292 void *head, void *tail, int bulk_cnt,
1293 unsigned long addr) { return 0; }
1295 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1297 static inline int check_object(struct kmem_cache *s, struct page *page,
1298 void *object, u8 val) { return 1; }
1299 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1300 struct page *page) {}
1301 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1302 struct page *page) {}
1303 unsigned long kmem_cache_flags(unsigned long object_size,
1304 unsigned long flags, const char *name,
1305 void (*ctor)(void *))
1309 #define slub_debug 0
1311 #define disable_higher_order_debug 0
1313 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1315 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1317 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1319 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1322 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1323 void **freelist, void *nextfree)
1327 #endif /* CONFIG_SLUB_DEBUG */
1330 * Hooks for other subsystems that check memory allocations. In a typical
1331 * production configuration these hooks all should produce no code at all.
1333 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1335 kmemleak_alloc(ptr, size, 1, flags);
1336 kasan_kmalloc_large(ptr, size, flags);
1339 static inline void kfree_hook(const void *x)
1342 kasan_kfree_large(x);
1345 static inline void *slab_free_hook(struct kmem_cache *s, void *x)
1349 kmemleak_free_recursive(x, s->flags);
1352 * Trouble is that we may no longer disable interrupts in the fast path
1353 * So in order to make the debug calls that expect irqs to be
1354 * disabled we need to disable interrupts temporarily.
1356 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1358 unsigned long flags;
1360 local_irq_save(flags);
1361 kmemcheck_slab_free(s, x, s->object_size);
1362 debug_check_no_locks_freed(x, s->object_size);
1363 local_irq_restore(flags);
1366 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1367 debug_check_no_obj_freed(x, s->object_size);
1369 freeptr = get_freepointer(s, x);
1371 * kasan_slab_free() may put x into memory quarantine, delaying its
1372 * reuse. In this case the object's freelist pointer is changed.
1374 kasan_slab_free(s, x);
1378 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1379 void *head, void *tail)
1382 * Compiler cannot detect this function can be removed if slab_free_hook()
1383 * evaluates to nothing. Thus, catch all relevant config debug options here.
1385 #if defined(CONFIG_KMEMCHECK) || \
1386 defined(CONFIG_LOCKDEP) || \
1387 defined(CONFIG_DEBUG_KMEMLEAK) || \
1388 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1389 defined(CONFIG_KASAN)
1391 void *object = head;
1392 void *tail_obj = tail ? : head;
1396 freeptr = slab_free_hook(s, object);
1397 } while ((object != tail_obj) && (object = freeptr));
1401 static void setup_object(struct kmem_cache *s, struct page *page,
1404 setup_object_debug(s, page, object);
1405 kasan_init_slab_obj(s, object);
1406 if (unlikely(s->ctor)) {
1407 kasan_unpoison_object_data(s, object);
1409 kasan_poison_object_data(s, object);
1414 * Slab allocation and freeing
1416 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1417 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1420 int order = oo_order(oo);
1422 flags |= __GFP_NOTRACK;
1424 if (node == NUMA_NO_NODE)
1425 page = alloc_pages(flags, order);
1427 page = __alloc_pages_node(node, flags, order);
1429 if (page && memcg_charge_slab(page, flags, order, s)) {
1430 __free_pages(page, order);
1437 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1438 /* Pre-initialize the random sequence cache */
1439 static int init_cache_random_seq(struct kmem_cache *s)
1442 unsigned long i, count = oo_objects(s->oo);
1444 /* Bailout if already initialised */
1448 err = cache_random_seq_create(s, count, GFP_KERNEL);
1450 pr_err("SLUB: Unable to initialize free list for %s\n",
1455 /* Transform to an offset on the set of pages */
1456 if (s->random_seq) {
1457 for (i = 0; i < count; i++)
1458 s->random_seq[i] *= s->size;
1463 /* Initialize each random sequence freelist per cache */
1464 static void __init init_freelist_randomization(void)
1466 struct kmem_cache *s;
1468 mutex_lock(&slab_mutex);
1470 list_for_each_entry(s, &slab_caches, list)
1471 init_cache_random_seq(s);
1473 mutex_unlock(&slab_mutex);
1476 /* Get the next entry on the pre-computed freelist randomized */
1477 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1478 unsigned long *pos, void *start,
1479 unsigned long page_limit,
1480 unsigned long freelist_count)
1485 * If the target page allocation failed, the number of objects on the
1486 * page might be smaller than the usual size defined by the cache.
1489 idx = s->random_seq[*pos];
1491 if (*pos >= freelist_count)
1493 } while (unlikely(idx >= page_limit));
1495 return (char *)start + idx;
1498 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1499 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1504 unsigned long idx, pos, page_limit, freelist_count;
1506 if (page->objects < 2 || !s->random_seq)
1509 freelist_count = oo_objects(s->oo);
1510 pos = get_random_int() % freelist_count;
1512 page_limit = page->objects * s->size;
1513 start = fixup_red_left(s, page_address(page));
1515 /* First entry is used as the base of the freelist */
1516 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1518 page->freelist = cur;
1520 for (idx = 1; idx < page->objects; idx++) {
1521 setup_object(s, page, cur);
1522 next = next_freelist_entry(s, page, &pos, start, page_limit,
1524 set_freepointer(s, cur, next);
1527 setup_object(s, page, cur);
1528 set_freepointer(s, cur, NULL);
1533 static inline int init_cache_random_seq(struct kmem_cache *s)
1537 static inline void init_freelist_randomization(void) { }
1538 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1542 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1544 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1547 struct kmem_cache_order_objects oo = s->oo;
1553 flags &= gfp_allowed_mask;
1555 if (gfpflags_allow_blocking(flags))
1558 flags |= s->allocflags;
1561 * Let the initial higher-order allocation fail under memory pressure
1562 * so we fall-back to the minimum order allocation.
1564 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1565 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1566 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1568 page = alloc_slab_page(s, alloc_gfp, node, oo);
1569 if (unlikely(!page)) {
1573 * Allocation may have failed due to fragmentation.
1574 * Try a lower order alloc if possible
1576 page = alloc_slab_page(s, alloc_gfp, node, oo);
1577 if (unlikely(!page))
1579 stat(s, ORDER_FALLBACK);
1582 if (kmemcheck_enabled &&
1583 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1584 int pages = 1 << oo_order(oo);
1586 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1589 * Objects from caches that have a constructor don't get
1590 * cleared when they're allocated, so we need to do it here.
1593 kmemcheck_mark_uninitialized_pages(page, pages);
1595 kmemcheck_mark_unallocated_pages(page, pages);
1598 page->objects = oo_objects(oo);
1600 order = compound_order(page);
1601 page->slab_cache = s;
1602 __SetPageSlab(page);
1603 if (page_is_pfmemalloc(page))
1604 SetPageSlabPfmemalloc(page);
1606 start = page_address(page);
1608 if (unlikely(s->flags & SLAB_POISON))
1609 memset(start, POISON_INUSE, PAGE_SIZE << order);
1611 kasan_poison_slab(page);
1613 shuffle = shuffle_freelist(s, page);
1616 for_each_object_idx(p, idx, s, start, page->objects) {
1617 setup_object(s, page, p);
1618 if (likely(idx < page->objects))
1619 set_freepointer(s, p, p + s->size);
1621 set_freepointer(s, p, NULL);
1623 page->freelist = fixup_red_left(s, start);
1626 page->inuse = page->objects;
1630 if (gfpflags_allow_blocking(flags))
1631 local_irq_disable();
1635 mod_zone_page_state(page_zone(page),
1636 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1637 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1640 inc_slabs_node(s, page_to_nid(page), page->objects);
1645 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1647 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1648 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1649 flags &= ~GFP_SLAB_BUG_MASK;
1650 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1651 invalid_mask, &invalid_mask, flags, &flags);
1654 return allocate_slab(s,
1655 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1658 static void __free_slab(struct kmem_cache *s, struct page *page)
1660 int order = compound_order(page);
1661 int pages = 1 << order;
1663 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1666 slab_pad_check(s, page);
1667 for_each_object(p, s, page_address(page),
1669 check_object(s, page, p, SLUB_RED_INACTIVE);
1672 kmemcheck_free_shadow(page, compound_order(page));
1674 mod_zone_page_state(page_zone(page),
1675 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1676 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1679 __ClearPageSlabPfmemalloc(page);
1680 __ClearPageSlab(page);
1682 page_mapcount_reset(page);
1683 if (current->reclaim_state)
1684 current->reclaim_state->reclaimed_slab += pages;
1685 memcg_uncharge_slab(page, order, s);
1686 __free_pages(page, order);
1689 #define need_reserve_slab_rcu \
1690 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1692 static void rcu_free_slab(struct rcu_head *h)
1696 if (need_reserve_slab_rcu)
1697 page = virt_to_head_page(h);
1699 page = container_of((struct list_head *)h, struct page, lru);
1701 __free_slab(page->slab_cache, page);
1704 static void free_slab(struct kmem_cache *s, struct page *page)
1706 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1707 struct rcu_head *head;
1709 if (need_reserve_slab_rcu) {
1710 int order = compound_order(page);
1711 int offset = (PAGE_SIZE << order) - s->reserved;
1713 VM_BUG_ON(s->reserved != sizeof(*head));
1714 head = page_address(page) + offset;
1716 head = &page->rcu_head;
1719 call_rcu(head, rcu_free_slab);
1721 __free_slab(s, page);
1724 static void discard_slab(struct kmem_cache *s, struct page *page)
1726 dec_slabs_node(s, page_to_nid(page), page->objects);
1731 * Management of partially allocated slabs.
1734 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1737 if (tail == DEACTIVATE_TO_TAIL)
1738 list_add_tail(&page->lru, &n->partial);
1740 list_add(&page->lru, &n->partial);
1743 static inline void add_partial(struct kmem_cache_node *n,
1744 struct page *page, int tail)
1746 lockdep_assert_held(&n->list_lock);
1747 __add_partial(n, page, tail);
1750 static inline void remove_partial(struct kmem_cache_node *n,
1753 lockdep_assert_held(&n->list_lock);
1754 list_del(&page->lru);
1759 * Remove slab from the partial list, freeze it and
1760 * return the pointer to the freelist.
1762 * Returns a list of objects or NULL if it fails.
1764 static inline void *acquire_slab(struct kmem_cache *s,
1765 struct kmem_cache_node *n, struct page *page,
1766 int mode, int *objects)
1769 unsigned long counters;
1772 lockdep_assert_held(&n->list_lock);
1775 * Zap the freelist and set the frozen bit.
1776 * The old freelist is the list of objects for the
1777 * per cpu allocation list.
1779 freelist = page->freelist;
1780 counters = page->counters;
1781 new.counters = counters;
1782 *objects = new.objects - new.inuse;
1784 new.inuse = page->objects;
1785 new.freelist = NULL;
1787 new.freelist = freelist;
1790 VM_BUG_ON(new.frozen);
1793 if (!__cmpxchg_double_slab(s, page,
1795 new.freelist, new.counters,
1799 remove_partial(n, page);
1804 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1805 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1808 * Try to allocate a partial slab from a specific node.
1810 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1811 struct kmem_cache_cpu *c, gfp_t flags)
1813 struct page *page, *page2;
1814 void *object = NULL;
1815 unsigned int available = 0;
1819 * Racy check. If we mistakenly see no partial slabs then we
1820 * just allocate an empty slab. If we mistakenly try to get a
1821 * partial slab and there is none available then get_partials()
1824 if (!n || !n->nr_partial)
1827 spin_lock(&n->list_lock);
1828 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1831 if (!pfmemalloc_match(page, flags))
1834 t = acquire_slab(s, n, page, object == NULL, &objects);
1838 available += objects;
1841 stat(s, ALLOC_FROM_PARTIAL);
1844 put_cpu_partial(s, page, 0);
1845 stat(s, CPU_PARTIAL_NODE);
1847 if (!kmem_cache_has_cpu_partial(s)
1848 || available > s->cpu_partial / 2)
1852 spin_unlock(&n->list_lock);
1857 * Get a page from somewhere. Search in increasing NUMA distances.
1859 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1860 struct kmem_cache_cpu *c)
1863 struct zonelist *zonelist;
1866 enum zone_type high_zoneidx = gfp_zone(flags);
1868 unsigned int cpuset_mems_cookie;
1871 * The defrag ratio allows a configuration of the tradeoffs between
1872 * inter node defragmentation and node local allocations. A lower
1873 * defrag_ratio increases the tendency to do local allocations
1874 * instead of attempting to obtain partial slabs from other nodes.
1876 * If the defrag_ratio is set to 0 then kmalloc() always
1877 * returns node local objects. If the ratio is higher then kmalloc()
1878 * may return off node objects because partial slabs are obtained
1879 * from other nodes and filled up.
1881 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1882 * (which makes defrag_ratio = 1000) then every (well almost)
1883 * allocation will first attempt to defrag slab caches on other nodes.
1884 * This means scanning over all nodes to look for partial slabs which
1885 * may be expensive if we do it every time we are trying to find a slab
1886 * with available objects.
1888 if (!s->remote_node_defrag_ratio ||
1889 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1893 cpuset_mems_cookie = read_mems_allowed_begin();
1894 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1895 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1896 struct kmem_cache_node *n;
1898 n = get_node(s, zone_to_nid(zone));
1900 if (n && cpuset_zone_allowed(zone, flags) &&
1901 n->nr_partial > s->min_partial) {
1902 object = get_partial_node(s, n, c, flags);
1905 * Don't check read_mems_allowed_retry()
1906 * here - if mems_allowed was updated in
1907 * parallel, that was a harmless race
1908 * between allocation and the cpuset
1915 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1921 * Get a partial page, lock it and return it.
1923 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1924 struct kmem_cache_cpu *c)
1927 int searchnode = node;
1929 if (node == NUMA_NO_NODE)
1930 searchnode = numa_mem_id();
1932 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1933 if (object || node != NUMA_NO_NODE)
1936 return get_any_partial(s, flags, c);
1939 #ifdef CONFIG_PREEMPT
1941 * Calculate the next globally unique transaction for disambiguiation
1942 * during cmpxchg. The transactions start with the cpu number and are then
1943 * incremented by CONFIG_NR_CPUS.
1945 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1948 * No preemption supported therefore also no need to check for
1954 static inline unsigned long next_tid(unsigned long tid)
1956 return tid + TID_STEP;
1959 static inline unsigned int tid_to_cpu(unsigned long tid)
1961 return tid % TID_STEP;
1964 static inline unsigned long tid_to_event(unsigned long tid)
1966 return tid / TID_STEP;
1969 static inline unsigned int init_tid(int cpu)
1974 static inline void note_cmpxchg_failure(const char *n,
1975 const struct kmem_cache *s, unsigned long tid)
1977 #ifdef SLUB_DEBUG_CMPXCHG
1978 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1980 pr_info("%s %s: cmpxchg redo ", n, s->name);
1982 #ifdef CONFIG_PREEMPT
1983 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1984 pr_warn("due to cpu change %d -> %d\n",
1985 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1988 if (tid_to_event(tid) != tid_to_event(actual_tid))
1989 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1990 tid_to_event(tid), tid_to_event(actual_tid));
1992 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1993 actual_tid, tid, next_tid(tid));
1995 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1998 static void init_kmem_cache_cpus(struct kmem_cache *s)
2002 for_each_possible_cpu(cpu)
2003 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2007 * Remove the cpu slab
2009 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2012 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2013 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2015 enum slab_modes l = M_NONE, m = M_NONE;
2017 int tail = DEACTIVATE_TO_HEAD;
2021 if (page->freelist) {
2022 stat(s, DEACTIVATE_REMOTE_FREES);
2023 tail = DEACTIVATE_TO_TAIL;
2027 * Stage one: Free all available per cpu objects back
2028 * to the page freelist while it is still frozen. Leave the
2031 * There is no need to take the list->lock because the page
2034 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2036 unsigned long counters;
2039 * If 'nextfree' is invalid, it is possible that the object at
2040 * 'freelist' is already corrupted. So isolate all objects
2041 * starting at 'freelist'.
2043 if (freelist_corrupted(s, page, &freelist, nextfree))
2047 prior = page->freelist;
2048 counters = page->counters;
2049 set_freepointer(s, freelist, prior);
2050 new.counters = counters;
2052 VM_BUG_ON(!new.frozen);
2054 } while (!__cmpxchg_double_slab(s, page,
2056 freelist, new.counters,
2057 "drain percpu freelist"));
2059 freelist = nextfree;
2063 * Stage two: Ensure that the page is unfrozen while the
2064 * list presence reflects the actual number of objects
2067 * We setup the list membership and then perform a cmpxchg
2068 * with the count. If there is a mismatch then the page
2069 * is not unfrozen but the page is on the wrong list.
2071 * Then we restart the process which may have to remove
2072 * the page from the list that we just put it on again
2073 * because the number of objects in the slab may have
2078 old.freelist = page->freelist;
2079 old.counters = page->counters;
2080 VM_BUG_ON(!old.frozen);
2082 /* Determine target state of the slab */
2083 new.counters = old.counters;
2086 set_freepointer(s, freelist, old.freelist);
2087 new.freelist = freelist;
2089 new.freelist = old.freelist;
2093 if (!new.inuse && n->nr_partial >= s->min_partial)
2095 else if (new.freelist) {
2100 * Taking the spinlock removes the possiblity
2101 * that acquire_slab() will see a slab page that
2104 spin_lock(&n->list_lock);
2108 if (kmem_cache_debug(s) && !lock) {
2111 * This also ensures that the scanning of full
2112 * slabs from diagnostic functions will not see
2115 spin_lock(&n->list_lock);
2123 remove_partial(n, page);
2125 else if (l == M_FULL)
2127 remove_full(s, n, page);
2129 if (m == M_PARTIAL) {
2131 add_partial(n, page, tail);
2134 } else if (m == M_FULL) {
2136 stat(s, DEACTIVATE_FULL);
2137 add_full(s, n, page);
2143 if (!__cmpxchg_double_slab(s, page,
2144 old.freelist, old.counters,
2145 new.freelist, new.counters,
2150 spin_unlock(&n->list_lock);
2153 stat(s, DEACTIVATE_EMPTY);
2154 discard_slab(s, page);
2160 * Unfreeze all the cpu partial slabs.
2162 * This function must be called with interrupts disabled
2163 * for the cpu using c (or some other guarantee must be there
2164 * to guarantee no concurrent accesses).
2166 static void unfreeze_partials(struct kmem_cache *s,
2167 struct kmem_cache_cpu *c)
2169 #ifdef CONFIG_SLUB_CPU_PARTIAL
2170 struct kmem_cache_node *n = NULL, *n2 = NULL;
2171 struct page *page, *discard_page = NULL;
2173 while ((page = c->partial)) {
2177 c->partial = page->next;
2179 n2 = get_node(s, page_to_nid(page));
2182 spin_unlock(&n->list_lock);
2185 spin_lock(&n->list_lock);
2190 old.freelist = page->freelist;
2191 old.counters = page->counters;
2192 VM_BUG_ON(!old.frozen);
2194 new.counters = old.counters;
2195 new.freelist = old.freelist;
2199 } while (!__cmpxchg_double_slab(s, page,
2200 old.freelist, old.counters,
2201 new.freelist, new.counters,
2202 "unfreezing slab"));
2204 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2205 page->next = discard_page;
2206 discard_page = page;
2208 add_partial(n, page, DEACTIVATE_TO_TAIL);
2209 stat(s, FREE_ADD_PARTIAL);
2214 spin_unlock(&n->list_lock);
2216 while (discard_page) {
2217 page = discard_page;
2218 discard_page = discard_page->next;
2220 stat(s, DEACTIVATE_EMPTY);
2221 discard_slab(s, page);
2228 * Put a page that was just frozen (in __slab_free) into a partial page
2229 * slot if available. This is done without interrupts disabled and without
2230 * preemption disabled. The cmpxchg is racy and may put the partial page
2231 * onto a random cpus partial slot.
2233 * If we did not find a slot then simply move all the partials to the
2234 * per node partial list.
2236 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2238 #ifdef CONFIG_SLUB_CPU_PARTIAL
2239 struct page *oldpage;
2247 oldpage = this_cpu_read(s->cpu_slab->partial);
2250 pobjects = oldpage->pobjects;
2251 pages = oldpage->pages;
2252 if (drain && pobjects > s->cpu_partial) {
2253 unsigned long flags;
2255 * partial array is full. Move the existing
2256 * set to the per node partial list.
2258 local_irq_save(flags);
2259 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2260 local_irq_restore(flags);
2264 stat(s, CPU_PARTIAL_DRAIN);
2269 pobjects += page->objects - page->inuse;
2271 page->pages = pages;
2272 page->pobjects = pobjects;
2273 page->next = oldpage;
2275 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2277 if (unlikely(!s->cpu_partial)) {
2278 unsigned long flags;
2280 local_irq_save(flags);
2281 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2282 local_irq_restore(flags);
2288 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2290 stat(s, CPUSLAB_FLUSH);
2291 deactivate_slab(s, c->page, c->freelist);
2293 c->tid = next_tid(c->tid);
2301 * Called from IPI handler with interrupts disabled.
2303 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2305 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2311 unfreeze_partials(s, c);
2315 static void flush_cpu_slab(void *d)
2317 struct kmem_cache *s = d;
2319 __flush_cpu_slab(s, smp_processor_id());
2322 static bool has_cpu_slab(int cpu, void *info)
2324 struct kmem_cache *s = info;
2325 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2327 return c->page || c->partial;
2330 static void flush_all(struct kmem_cache *s)
2332 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2336 * Use the cpu notifier to insure that the cpu slabs are flushed when
2339 static int slub_cpu_dead(unsigned int cpu)
2341 struct kmem_cache *s;
2342 unsigned long flags;
2344 mutex_lock(&slab_mutex);
2345 list_for_each_entry(s, &slab_caches, list) {
2346 local_irq_save(flags);
2347 __flush_cpu_slab(s, cpu);
2348 local_irq_restore(flags);
2350 mutex_unlock(&slab_mutex);
2355 * Check if the objects in a per cpu structure fit numa
2356 * locality expectations.
2358 static inline int node_match(struct page *page, int node)
2361 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2367 #ifdef CONFIG_SLUB_DEBUG
2368 static int count_free(struct page *page)
2370 return page->objects - page->inuse;
2373 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2375 return atomic_long_read(&n->total_objects);
2377 #endif /* CONFIG_SLUB_DEBUG */
2379 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2380 static unsigned long count_partial(struct kmem_cache_node *n,
2381 int (*get_count)(struct page *))
2383 unsigned long flags;
2384 unsigned long x = 0;
2387 spin_lock_irqsave(&n->list_lock, flags);
2388 list_for_each_entry(page, &n->partial, lru)
2389 x += get_count(page);
2390 spin_unlock_irqrestore(&n->list_lock, flags);
2393 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2395 static noinline void
2396 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2398 #ifdef CONFIG_SLUB_DEBUG
2399 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2400 DEFAULT_RATELIMIT_BURST);
2402 struct kmem_cache_node *n;
2404 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2407 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2408 nid, gfpflags, &gfpflags);
2409 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2410 s->name, s->object_size, s->size, oo_order(s->oo),
2413 if (oo_order(s->min) > get_order(s->object_size))
2414 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2417 for_each_kmem_cache_node(s, node, n) {
2418 unsigned long nr_slabs;
2419 unsigned long nr_objs;
2420 unsigned long nr_free;
2422 nr_free = count_partial(n, count_free);
2423 nr_slabs = node_nr_slabs(n);
2424 nr_objs = node_nr_objs(n);
2426 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2427 node, nr_slabs, nr_objs, nr_free);
2432 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2433 int node, struct kmem_cache_cpu **pc)
2436 struct kmem_cache_cpu *c = *pc;
2439 freelist = get_partial(s, flags, node, c);
2444 page = new_slab(s, flags, node);
2446 c = raw_cpu_ptr(s->cpu_slab);
2451 * No other reference to the page yet so we can
2452 * muck around with it freely without cmpxchg
2454 freelist = page->freelist;
2455 page->freelist = NULL;
2457 stat(s, ALLOC_SLAB);
2466 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2468 if (unlikely(PageSlabPfmemalloc(page)))
2469 return gfp_pfmemalloc_allowed(gfpflags);
2475 * Check the page->freelist of a page and either transfer the freelist to the
2476 * per cpu freelist or deactivate the page.
2478 * The page is still frozen if the return value is not NULL.
2480 * If this function returns NULL then the page has been unfrozen.
2482 * This function must be called with interrupt disabled.
2484 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2487 unsigned long counters;
2491 freelist = page->freelist;
2492 counters = page->counters;
2494 new.counters = counters;
2495 VM_BUG_ON(!new.frozen);
2497 new.inuse = page->objects;
2498 new.frozen = freelist != NULL;
2500 } while (!__cmpxchg_double_slab(s, page,
2509 * Slow path. The lockless freelist is empty or we need to perform
2512 * Processing is still very fast if new objects have been freed to the
2513 * regular freelist. In that case we simply take over the regular freelist
2514 * as the lockless freelist and zap the regular freelist.
2516 * If that is not working then we fall back to the partial lists. We take the
2517 * first element of the freelist as the object to allocate now and move the
2518 * rest of the freelist to the lockless freelist.
2520 * And if we were unable to get a new slab from the partial slab lists then
2521 * we need to allocate a new slab. This is the slowest path since it involves
2522 * a call to the page allocator and the setup of a new slab.
2524 * Version of __slab_alloc to use when we know that interrupts are
2525 * already disabled (which is the case for bulk allocation).
2527 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2528 unsigned long addr, struct kmem_cache_cpu *c)
2536 * if the node is not online or has no normal memory, just
2537 * ignore the node constraint
2539 if (unlikely(node != NUMA_NO_NODE &&
2540 !node_state(node, N_NORMAL_MEMORY)))
2541 node = NUMA_NO_NODE;
2546 if (unlikely(!node_match(page, node))) {
2548 * same as above but node_match() being false already
2549 * implies node != NUMA_NO_NODE
2551 if (!node_state(node, N_NORMAL_MEMORY)) {
2552 node = NUMA_NO_NODE;
2555 stat(s, ALLOC_NODE_MISMATCH);
2556 deactivate_slab(s, page, c->freelist);
2559 c->tid = next_tid(c->tid);
2565 * By rights, we should be searching for a slab page that was
2566 * PFMEMALLOC but right now, we are losing the pfmemalloc
2567 * information when the page leaves the per-cpu allocator
2569 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2570 deactivate_slab(s, page, c->freelist);
2573 c->tid = next_tid(c->tid);
2577 /* must check again c->freelist in case of cpu migration or IRQ */
2578 freelist = c->freelist;
2582 freelist = get_freelist(s, page);
2586 c->tid = next_tid(c->tid);
2587 stat(s, DEACTIVATE_BYPASS);
2591 stat(s, ALLOC_REFILL);
2595 * freelist is pointing to the list of objects to be used.
2596 * page is pointing to the page from which the objects are obtained.
2597 * That page must be frozen for per cpu allocations to work.
2599 VM_BUG_ON(!c->page->frozen);
2600 c->freelist = get_freepointer(s, freelist);
2601 c->tid = next_tid(c->tid);
2607 page = c->page = c->partial;
2608 c->partial = page->next;
2609 stat(s, CPU_PARTIAL_ALLOC);
2611 c->tid = next_tid(c->tid);
2615 freelist = new_slab_objects(s, gfpflags, node, &c);
2617 if (unlikely(!freelist)) {
2618 slab_out_of_memory(s, gfpflags, node);
2623 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2626 /* Only entered in the debug case */
2627 if (kmem_cache_debug(s) &&
2628 !alloc_debug_processing(s, page, freelist, addr))
2629 goto new_slab; /* Slab failed checks. Next slab needed */
2631 deactivate_slab(s, page, get_freepointer(s, freelist));
2634 c->tid = next_tid(c->tid);
2639 * Another one that disabled interrupt and compensates for possible
2640 * cpu changes by refetching the per cpu area pointer.
2642 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2643 unsigned long addr, struct kmem_cache_cpu *c)
2646 unsigned long flags;
2648 local_irq_save(flags);
2649 #ifdef CONFIG_PREEMPT
2651 * We may have been preempted and rescheduled on a different
2652 * cpu before disabling interrupts. Need to reload cpu area
2655 c = this_cpu_ptr(s->cpu_slab);
2658 p = ___slab_alloc(s, gfpflags, node, addr, c);
2659 local_irq_restore(flags);
2664 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2665 * have the fastpath folded into their functions. So no function call
2666 * overhead for requests that can be satisfied on the fastpath.
2668 * The fastpath works by first checking if the lockless freelist can be used.
2669 * If not then __slab_alloc is called for slow processing.
2671 * Otherwise we can simply pick the next object from the lockless free list.
2673 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2674 gfp_t gfpflags, int node, unsigned long addr)
2677 struct kmem_cache_cpu *c;
2681 s = slab_pre_alloc_hook(s, gfpflags);
2686 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2687 * enabled. We may switch back and forth between cpus while
2688 * reading from one cpu area. That does not matter as long
2689 * as we end up on the original cpu again when doing the cmpxchg.
2691 * We should guarantee that tid and kmem_cache are retrieved on
2692 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2693 * to check if it is matched or not.
2696 tid = this_cpu_read(s->cpu_slab->tid);
2697 c = raw_cpu_ptr(s->cpu_slab);
2698 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2699 unlikely(tid != READ_ONCE(c->tid)));
2702 * Irqless object alloc/free algorithm used here depends on sequence
2703 * of fetching cpu_slab's data. tid should be fetched before anything
2704 * on c to guarantee that object and page associated with previous tid
2705 * won't be used with current tid. If we fetch tid first, object and
2706 * page could be one associated with next tid and our alloc/free
2707 * request will be failed. In this case, we will retry. So, no problem.
2712 * The transaction ids are globally unique per cpu and per operation on
2713 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2714 * occurs on the right processor and that there was no operation on the
2715 * linked list in between.
2718 object = c->freelist;
2720 if (unlikely(!object || !node_match(page, node))) {
2721 object = __slab_alloc(s, gfpflags, node, addr, c);
2722 stat(s, ALLOC_SLOWPATH);
2724 void *next_object = get_freepointer_safe(s, object);
2727 * The cmpxchg will only match if there was no additional
2728 * operation and if we are on the right processor.
2730 * The cmpxchg does the following atomically (without lock
2732 * 1. Relocate first pointer to the current per cpu area.
2733 * 2. Verify that tid and freelist have not been changed
2734 * 3. If they were not changed replace tid and freelist
2736 * Since this is without lock semantics the protection is only
2737 * against code executing on this cpu *not* from access by
2740 if (unlikely(!this_cpu_cmpxchg_double(
2741 s->cpu_slab->freelist, s->cpu_slab->tid,
2743 next_object, next_tid(tid)))) {
2745 note_cmpxchg_failure("slab_alloc", s, tid);
2748 prefetch_freepointer(s, next_object);
2749 stat(s, ALLOC_FASTPATH);
2752 if (unlikely(gfpflags & __GFP_ZERO) && object)
2753 memset(object, 0, s->object_size);
2755 slab_post_alloc_hook(s, gfpflags, 1, &object);
2760 static __always_inline void *slab_alloc(struct kmem_cache *s,
2761 gfp_t gfpflags, unsigned long addr)
2763 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2766 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2768 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2770 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2775 EXPORT_SYMBOL(kmem_cache_alloc);
2777 #ifdef CONFIG_TRACING
2778 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2780 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2781 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2782 kasan_kmalloc(s, ret, size, gfpflags);
2785 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2789 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2791 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2793 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2794 s->object_size, s->size, gfpflags, node);
2798 EXPORT_SYMBOL(kmem_cache_alloc_node);
2800 #ifdef CONFIG_TRACING
2801 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2803 int node, size_t size)
2805 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2807 trace_kmalloc_node(_RET_IP_, ret,
2808 size, s->size, gfpflags, node);
2810 kasan_kmalloc(s, ret, size, gfpflags);
2813 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2818 * Slow path handling. This may still be called frequently since objects
2819 * have a longer lifetime than the cpu slabs in most processing loads.
2821 * So we still attempt to reduce cache line usage. Just take the slab
2822 * lock and free the item. If there is no additional partial page
2823 * handling required then we can return immediately.
2825 static void __slab_free(struct kmem_cache *s, struct page *page,
2826 void *head, void *tail, int cnt,
2833 unsigned long counters;
2834 struct kmem_cache_node *n = NULL;
2835 unsigned long uninitialized_var(flags);
2837 stat(s, FREE_SLOWPATH);
2839 if (kmem_cache_debug(s) &&
2840 !free_debug_processing(s, page, head, tail, cnt, addr))
2845 spin_unlock_irqrestore(&n->list_lock, flags);
2848 prior = page->freelist;
2849 counters = page->counters;
2850 set_freepointer(s, tail, prior);
2851 new.counters = counters;
2852 was_frozen = new.frozen;
2854 if ((!new.inuse || !prior) && !was_frozen) {
2856 if (kmem_cache_has_cpu_partial(s) && !prior) {
2859 * Slab was on no list before and will be
2861 * We can defer the list move and instead
2866 } else { /* Needs to be taken off a list */
2868 n = get_node(s, page_to_nid(page));
2870 * Speculatively acquire the list_lock.
2871 * If the cmpxchg does not succeed then we may
2872 * drop the list_lock without any processing.
2874 * Otherwise the list_lock will synchronize with
2875 * other processors updating the list of slabs.
2877 spin_lock_irqsave(&n->list_lock, flags);
2882 } while (!cmpxchg_double_slab(s, page,
2890 * If we just froze the page then put it onto the
2891 * per cpu partial list.
2893 if (new.frozen && !was_frozen) {
2894 put_cpu_partial(s, page, 1);
2895 stat(s, CPU_PARTIAL_FREE);
2898 * The list lock was not taken therefore no list
2899 * activity can be necessary.
2902 stat(s, FREE_FROZEN);
2906 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2910 * Objects left in the slab. If it was not on the partial list before
2913 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2914 if (kmem_cache_debug(s))
2915 remove_full(s, n, page);
2916 add_partial(n, page, DEACTIVATE_TO_TAIL);
2917 stat(s, FREE_ADD_PARTIAL);
2919 spin_unlock_irqrestore(&n->list_lock, flags);
2925 * Slab on the partial list.
2927 remove_partial(n, page);
2928 stat(s, FREE_REMOVE_PARTIAL);
2930 /* Slab must be on the full list */
2931 remove_full(s, n, page);
2934 spin_unlock_irqrestore(&n->list_lock, flags);
2936 discard_slab(s, page);
2940 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2941 * can perform fastpath freeing without additional function calls.
2943 * The fastpath is only possible if we are freeing to the current cpu slab
2944 * of this processor. This typically the case if we have just allocated
2947 * If fastpath is not possible then fall back to __slab_free where we deal
2948 * with all sorts of special processing.
2950 * Bulk free of a freelist with several objects (all pointing to the
2951 * same page) possible by specifying head and tail ptr, plus objects
2952 * count (cnt). Bulk free indicated by tail pointer being set.
2954 static __always_inline void do_slab_free(struct kmem_cache *s,
2955 struct page *page, void *head, void *tail,
2956 int cnt, unsigned long addr)
2958 void *tail_obj = tail ? : head;
2959 struct kmem_cache_cpu *c;
2963 * Determine the currently cpus per cpu slab.
2964 * The cpu may change afterward. However that does not matter since
2965 * data is retrieved via this pointer. If we are on the same cpu
2966 * during the cmpxchg then the free will succeed.
2969 tid = this_cpu_read(s->cpu_slab->tid);
2970 c = raw_cpu_ptr(s->cpu_slab);
2971 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2972 unlikely(tid != READ_ONCE(c->tid)));
2974 /* Same with comment on barrier() in slab_alloc_node() */
2977 if (likely(page == c->page)) {
2978 void **freelist = READ_ONCE(c->freelist);
2980 set_freepointer(s, tail_obj, freelist);
2982 if (unlikely(!this_cpu_cmpxchg_double(
2983 s->cpu_slab->freelist, s->cpu_slab->tid,
2985 head, next_tid(tid)))) {
2987 note_cmpxchg_failure("slab_free", s, tid);
2990 stat(s, FREE_FASTPATH);
2992 __slab_free(s, page, head, tail_obj, cnt, addr);
2996 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2997 void *head, void *tail, int cnt,
3000 slab_free_freelist_hook(s, head, tail);
3002 * slab_free_freelist_hook() could have put the items into quarantine.
3003 * If so, no need to free them.
3005 if (s->flags & SLAB_KASAN && !(s->flags & SLAB_DESTROY_BY_RCU))
3007 do_slab_free(s, page, head, tail, cnt, addr);
3011 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3013 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3017 void kmem_cache_free(struct kmem_cache *s, void *x)
3019 s = cache_from_obj(s, x);
3022 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3023 trace_kmem_cache_free(_RET_IP_, x);
3025 EXPORT_SYMBOL(kmem_cache_free);
3027 struct detached_freelist {
3032 struct kmem_cache *s;
3036 * This function progressively scans the array with free objects (with
3037 * a limited look ahead) and extract objects belonging to the same
3038 * page. It builds a detached freelist directly within the given
3039 * page/objects. This can happen without any need for
3040 * synchronization, because the objects are owned by running process.
3041 * The freelist is build up as a single linked list in the objects.
3042 * The idea is, that this detached freelist can then be bulk
3043 * transferred to the real freelist(s), but only requiring a single
3044 * synchronization primitive. Look ahead in the array is limited due
3045 * to performance reasons.
3048 int build_detached_freelist(struct kmem_cache *s, size_t size,
3049 void **p, struct detached_freelist *df)
3051 size_t first_skipped_index = 0;
3056 /* Always re-init detached_freelist */
3061 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3062 } while (!object && size);
3067 page = virt_to_head_page(object);
3069 /* Handle kalloc'ed objects */
3070 if (unlikely(!PageSlab(page))) {
3071 BUG_ON(!PageCompound(page));
3073 __free_pages(page, compound_order(page));
3074 p[size] = NULL; /* mark object processed */
3077 /* Derive kmem_cache from object */
3078 df->s = page->slab_cache;
3080 df->s = cache_from_obj(s, object); /* Support for memcg */
3083 /* Start new detached freelist */
3085 set_freepointer(df->s, object, NULL);
3087 df->freelist = object;
3088 p[size] = NULL; /* mark object processed */
3094 continue; /* Skip processed objects */
3096 /* df->page is always set at this point */
3097 if (df->page == virt_to_head_page(object)) {
3098 /* Opportunity build freelist */
3099 set_freepointer(df->s, object, df->freelist);
3100 df->freelist = object;
3102 p[size] = NULL; /* mark object processed */
3107 /* Limit look ahead search */
3111 if (!first_skipped_index)
3112 first_skipped_index = size + 1;
3115 return first_skipped_index;
3118 /* Note that interrupts must be enabled when calling this function. */
3119 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3125 struct detached_freelist df;
3127 size = build_detached_freelist(s, size, p, &df);
3128 if (unlikely(!df.page))
3131 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3132 } while (likely(size));
3134 EXPORT_SYMBOL(kmem_cache_free_bulk);
3136 /* Note that interrupts must be enabled when calling this function. */
3137 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3140 struct kmem_cache_cpu *c;
3143 /* memcg and kmem_cache debug support */
3144 s = slab_pre_alloc_hook(s, flags);
3148 * Drain objects in the per cpu slab, while disabling local
3149 * IRQs, which protects against PREEMPT and interrupts
3150 * handlers invoking normal fastpath.
3152 local_irq_disable();
3153 c = this_cpu_ptr(s->cpu_slab);
3155 for (i = 0; i < size; i++) {
3156 void *object = c->freelist;
3158 if (unlikely(!object)) {
3160 * We may have removed an object from c->freelist using
3161 * the fastpath in the previous iteration; in that case,
3162 * c->tid has not been bumped yet.
3163 * Since ___slab_alloc() may reenable interrupts while
3164 * allocating memory, we should bump c->tid now.
3166 c->tid = next_tid(c->tid);
3169 * Invoking slow path likely have side-effect
3170 * of re-populating per CPU c->freelist
3172 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3174 if (unlikely(!p[i]))
3177 c = this_cpu_ptr(s->cpu_slab);
3178 continue; /* goto for-loop */
3180 c->freelist = get_freepointer(s, object);
3183 c->tid = next_tid(c->tid);
3186 /* Clear memory outside IRQ disabled fastpath loop */
3187 if (unlikely(flags & __GFP_ZERO)) {
3190 for (j = 0; j < i; j++)
3191 memset(p[j], 0, s->object_size);
3194 /* memcg and kmem_cache debug support */
3195 slab_post_alloc_hook(s, flags, size, p);
3199 slab_post_alloc_hook(s, flags, i, p);
3200 __kmem_cache_free_bulk(s, i, p);
3203 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3207 * Object placement in a slab is made very easy because we always start at
3208 * offset 0. If we tune the size of the object to the alignment then we can
3209 * get the required alignment by putting one properly sized object after
3212 * Notice that the allocation order determines the sizes of the per cpu
3213 * caches. Each processor has always one slab available for allocations.
3214 * Increasing the allocation order reduces the number of times that slabs
3215 * must be moved on and off the partial lists and is therefore a factor in
3220 * Mininum / Maximum order of slab pages. This influences locking overhead
3221 * and slab fragmentation. A higher order reduces the number of partial slabs
3222 * and increases the number of allocations possible without having to
3223 * take the list_lock.
3225 static int slub_min_order;
3226 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3227 static int slub_min_objects;
3230 * Calculate the order of allocation given an slab object size.
3232 * The order of allocation has significant impact on performance and other
3233 * system components. Generally order 0 allocations should be preferred since
3234 * order 0 does not cause fragmentation in the page allocator. Larger objects
3235 * be problematic to put into order 0 slabs because there may be too much
3236 * unused space left. We go to a higher order if more than 1/16th of the slab
3239 * In order to reach satisfactory performance we must ensure that a minimum
3240 * number of objects is in one slab. Otherwise we may generate too much
3241 * activity on the partial lists which requires taking the list_lock. This is
3242 * less a concern for large slabs though which are rarely used.
3244 * slub_max_order specifies the order where we begin to stop considering the
3245 * number of objects in a slab as critical. If we reach slub_max_order then
3246 * we try to keep the page order as low as possible. So we accept more waste
3247 * of space in favor of a small page order.
3249 * Higher order allocations also allow the placement of more objects in a
3250 * slab and thereby reduce object handling overhead. If the user has
3251 * requested a higher mininum order then we start with that one instead of
3252 * the smallest order which will fit the object.
3254 static inline int slab_order(int size, int min_objects,
3255 int max_order, int fract_leftover, int reserved)
3259 int min_order = slub_min_order;
3261 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3262 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3264 for (order = max(min_order, get_order(min_objects * size + reserved));
3265 order <= max_order; order++) {
3267 unsigned long slab_size = PAGE_SIZE << order;
3269 rem = (slab_size - reserved) % size;
3271 if (rem <= slab_size / fract_leftover)
3278 static inline int calculate_order(int size, int reserved)
3286 * Attempt to find best configuration for a slab. This
3287 * works by first attempting to generate a layout with
3288 * the best configuration and backing off gradually.
3290 * First we increase the acceptable waste in a slab. Then
3291 * we reduce the minimum objects required in a slab.
3293 min_objects = slub_min_objects;
3295 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3296 max_objects = order_objects(slub_max_order, size, reserved);
3297 min_objects = min(min_objects, max_objects);
3299 while (min_objects > 1) {
3301 while (fraction >= 4) {
3302 order = slab_order(size, min_objects,
3303 slub_max_order, fraction, reserved);
3304 if (order <= slub_max_order)
3312 * We were unable to place multiple objects in a slab. Now
3313 * lets see if we can place a single object there.
3315 order = slab_order(size, 1, slub_max_order, 1, reserved);
3316 if (order <= slub_max_order)
3320 * Doh this slab cannot be placed using slub_max_order.
3322 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3323 if (order < MAX_ORDER)
3329 init_kmem_cache_node(struct kmem_cache_node *n)
3332 spin_lock_init(&n->list_lock);
3333 INIT_LIST_HEAD(&n->partial);
3334 #ifdef CONFIG_SLUB_DEBUG
3335 atomic_long_set(&n->nr_slabs, 0);
3336 atomic_long_set(&n->total_objects, 0);
3337 INIT_LIST_HEAD(&n->full);
3341 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3343 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3344 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3347 * Must align to double word boundary for the double cmpxchg
3348 * instructions to work; see __pcpu_double_call_return_bool().
3350 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3351 2 * sizeof(void *));
3356 init_kmem_cache_cpus(s);
3361 static struct kmem_cache *kmem_cache_node;
3364 * No kmalloc_node yet so do it by hand. We know that this is the first
3365 * slab on the node for this slabcache. There are no concurrent accesses
3368 * Note that this function only works on the kmem_cache_node
3369 * when allocating for the kmem_cache_node. This is used for bootstrapping
3370 * memory on a fresh node that has no slab structures yet.
3372 static void early_kmem_cache_node_alloc(int node)
3375 struct kmem_cache_node *n;
3377 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3379 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3382 if (page_to_nid(page) != node) {
3383 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3384 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3389 page->freelist = get_freepointer(kmem_cache_node, n);
3392 kmem_cache_node->node[node] = n;
3393 #ifdef CONFIG_SLUB_DEBUG
3394 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3395 init_tracking(kmem_cache_node, n);
3397 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3399 init_kmem_cache_node(n);
3400 inc_slabs_node(kmem_cache_node, node, page->objects);
3403 * No locks need to be taken here as it has just been
3404 * initialized and there is no concurrent access.
3406 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3409 static void free_kmem_cache_nodes(struct kmem_cache *s)
3412 struct kmem_cache_node *n;
3414 for_each_kmem_cache_node(s, node, n) {
3415 kmem_cache_free(kmem_cache_node, n);
3416 s->node[node] = NULL;
3420 void __kmem_cache_release(struct kmem_cache *s)
3422 cache_random_seq_destroy(s);
3423 free_percpu(s->cpu_slab);
3424 free_kmem_cache_nodes(s);
3427 static int init_kmem_cache_nodes(struct kmem_cache *s)
3431 for_each_node_state(node, N_NORMAL_MEMORY) {
3432 struct kmem_cache_node *n;
3434 if (slab_state == DOWN) {
3435 early_kmem_cache_node_alloc(node);
3438 n = kmem_cache_alloc_node(kmem_cache_node,
3442 free_kmem_cache_nodes(s);
3447 init_kmem_cache_node(n);
3452 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3454 if (min < MIN_PARTIAL)
3456 else if (min > MAX_PARTIAL)
3458 s->min_partial = min;
3462 * calculate_sizes() determines the order and the distribution of data within
3465 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3467 unsigned long flags = s->flags;
3468 size_t size = s->object_size;
3472 * Round up object size to the next word boundary. We can only
3473 * place the free pointer at word boundaries and this determines
3474 * the possible location of the free pointer.
3476 size = ALIGN(size, sizeof(void *));
3478 #ifdef CONFIG_SLUB_DEBUG
3480 * Determine if we can poison the object itself. If the user of
3481 * the slab may touch the object after free or before allocation
3482 * then we should never poison the object itself.
3484 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3486 s->flags |= __OBJECT_POISON;
3488 s->flags &= ~__OBJECT_POISON;
3492 * If we are Redzoning then check if there is some space between the
3493 * end of the object and the free pointer. If not then add an
3494 * additional word to have some bytes to store Redzone information.
3496 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3497 size += sizeof(void *);
3501 * With that we have determined the number of bytes in actual use
3502 * by the object. This is the potential offset to the free pointer.
3506 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3509 * Relocate free pointer after the object if it is not
3510 * permitted to overwrite the first word of the object on
3513 * This is the case if we do RCU, have a constructor or
3514 * destructor or are poisoning the objects.
3517 size += sizeof(void *);
3520 #ifdef CONFIG_SLUB_DEBUG
3521 if (flags & SLAB_STORE_USER)
3523 * Need to store information about allocs and frees after
3526 size += 2 * sizeof(struct track);
3529 kasan_cache_create(s, &size, &s->flags);
3530 #ifdef CONFIG_SLUB_DEBUG
3531 if (flags & SLAB_RED_ZONE) {
3533 * Add some empty padding so that we can catch
3534 * overwrites from earlier objects rather than let
3535 * tracking information or the free pointer be
3536 * corrupted if a user writes before the start
3539 size += sizeof(void *);
3541 s->red_left_pad = sizeof(void *);
3542 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3543 size += s->red_left_pad;
3548 * SLUB stores one object immediately after another beginning from
3549 * offset 0. In order to align the objects we have to simply size
3550 * each object to conform to the alignment.
3552 size = ALIGN(size, s->align);
3554 if (forced_order >= 0)
3555 order = forced_order;
3557 order = calculate_order(size, s->reserved);
3564 s->allocflags |= __GFP_COMP;
3566 if (s->flags & SLAB_CACHE_DMA)
3567 s->allocflags |= GFP_DMA;
3569 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3570 s->allocflags |= __GFP_RECLAIMABLE;
3573 * Determine the number of objects per slab
3575 s->oo = oo_make(order, size, s->reserved);
3576 s->min = oo_make(get_order(size), size, s->reserved);
3577 if (oo_objects(s->oo) > oo_objects(s->max))
3580 return !!oo_objects(s->oo);
3583 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3585 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3588 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3589 s->reserved = sizeof(struct rcu_head);
3591 if (!calculate_sizes(s, -1))
3593 if (disable_higher_order_debug) {
3595 * Disable debugging flags that store metadata if the min slab
3598 if (get_order(s->size) > get_order(s->object_size)) {
3599 s->flags &= ~DEBUG_METADATA_FLAGS;
3601 if (!calculate_sizes(s, -1))
3606 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3607 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3608 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3609 /* Enable fast mode */
3610 s->flags |= __CMPXCHG_DOUBLE;
3614 * The larger the object size is, the more pages we want on the partial
3615 * list to avoid pounding the page allocator excessively.
3617 set_min_partial(s, ilog2(s->size) / 2);
3620 * cpu_partial determined the maximum number of objects kept in the
3621 * per cpu partial lists of a processor.
3623 * Per cpu partial lists mainly contain slabs that just have one
3624 * object freed. If they are used for allocation then they can be
3625 * filled up again with minimal effort. The slab will never hit the
3626 * per node partial lists and therefore no locking will be required.
3628 * This setting also determines
3630 * A) The number of objects from per cpu partial slabs dumped to the
3631 * per node list when we reach the limit.
3632 * B) The number of objects in cpu partial slabs to extract from the
3633 * per node list when we run out of per cpu objects. We only fetch
3634 * 50% to keep some capacity around for frees.
3636 if (!kmem_cache_has_cpu_partial(s))
3638 else if (s->size >= PAGE_SIZE)
3640 else if (s->size >= 1024)
3642 else if (s->size >= 256)
3643 s->cpu_partial = 13;
3645 s->cpu_partial = 30;
3648 s->remote_node_defrag_ratio = 1000;
3651 /* Initialize the pre-computed randomized freelist if slab is up */
3652 if (slab_state >= UP) {
3653 if (init_cache_random_seq(s))
3657 if (!init_kmem_cache_nodes(s))
3660 if (alloc_kmem_cache_cpus(s))
3663 free_kmem_cache_nodes(s);
3665 if (flags & SLAB_PANIC)
3666 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3667 s->name, (unsigned long)s->size, s->size,
3668 oo_order(s->oo), s->offset, flags);
3672 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3675 #ifdef CONFIG_SLUB_DEBUG
3676 void *addr = page_address(page);
3678 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3679 sizeof(long), GFP_ATOMIC);
3682 slab_err(s, page, text, s->name);
3685 get_map(s, page, map);
3686 for_each_object(p, s, addr, page->objects) {
3688 if (!test_bit(slab_index(p, s, addr), map)) {
3689 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3690 print_tracking(s, p);
3699 * Attempt to free all partial slabs on a node.
3700 * This is called from __kmem_cache_shutdown(). We must take list_lock
3701 * because sysfs file might still access partial list after the shutdowning.
3703 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3706 struct page *page, *h;
3708 BUG_ON(irqs_disabled());
3709 spin_lock_irq(&n->list_lock);
3710 list_for_each_entry_safe(page, h, &n->partial, lru) {
3712 remove_partial(n, page);
3713 list_add(&page->lru, &discard);
3715 list_slab_objects(s, page,
3716 "Objects remaining in %s on __kmem_cache_shutdown()");
3719 spin_unlock_irq(&n->list_lock);
3721 list_for_each_entry_safe(page, h, &discard, lru)
3722 discard_slab(s, page);
3726 * Release all resources used by a slab cache.
3728 int __kmem_cache_shutdown(struct kmem_cache *s)
3731 struct kmem_cache_node *n;
3734 /* Attempt to free all objects */
3735 for_each_kmem_cache_node(s, node, n) {
3737 if (n->nr_partial || slabs_node(s, node))
3743 /********************************************************************
3745 *******************************************************************/
3747 static int __init setup_slub_min_order(char *str)
3749 get_option(&str, &slub_min_order);
3754 __setup("slub_min_order=", setup_slub_min_order);
3756 static int __init setup_slub_max_order(char *str)
3758 get_option(&str, &slub_max_order);
3759 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3764 __setup("slub_max_order=", setup_slub_max_order);
3766 static int __init setup_slub_min_objects(char *str)
3768 get_option(&str, &slub_min_objects);
3773 __setup("slub_min_objects=", setup_slub_min_objects);
3775 void *__kmalloc(size_t size, gfp_t flags)
3777 struct kmem_cache *s;
3780 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3781 return kmalloc_large(size, flags);
3783 s = kmalloc_slab(size, flags);
3785 if (unlikely(ZERO_OR_NULL_PTR(s)))
3788 ret = slab_alloc(s, flags, _RET_IP_);
3790 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3792 kasan_kmalloc(s, ret, size, flags);
3796 EXPORT_SYMBOL(__kmalloc);
3799 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3804 flags |= __GFP_COMP | __GFP_NOTRACK;
3805 page = alloc_pages_node(node, flags, get_order(size));
3807 ptr = page_address(page);
3809 kmalloc_large_node_hook(ptr, size, flags);
3813 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3815 struct kmem_cache *s;
3818 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3819 ret = kmalloc_large_node(size, flags, node);
3821 trace_kmalloc_node(_RET_IP_, ret,
3822 size, PAGE_SIZE << get_order(size),
3828 s = kmalloc_slab(size, flags);
3830 if (unlikely(ZERO_OR_NULL_PTR(s)))
3833 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3835 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3837 kasan_kmalloc(s, ret, size, flags);
3841 EXPORT_SYMBOL(__kmalloc_node);
3844 #ifdef CONFIG_HARDENED_USERCOPY
3846 * Rejects objects that are incorrectly sized.
3848 * Returns NULL if check passes, otherwise const char * to name of cache
3849 * to indicate an error.
3851 const char *__check_heap_object(const void *ptr, unsigned long n,
3854 struct kmem_cache *s;
3855 unsigned long offset;
3858 /* Find object and usable object size. */
3859 s = page->slab_cache;
3860 object_size = slab_ksize(s);
3862 /* Reject impossible pointers. */
3863 if (ptr < page_address(page))
3866 /* Find offset within object. */
3867 offset = (ptr - page_address(page)) % s->size;
3869 /* Adjust for redzone and reject if within the redzone. */
3870 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3871 if (offset < s->red_left_pad)
3873 offset -= s->red_left_pad;
3876 /* Allow address range falling entirely within object size. */
3877 if (offset <= object_size && n <= object_size - offset)
3882 #endif /* CONFIG_HARDENED_USERCOPY */
3884 static size_t __ksize(const void *object)
3888 if (unlikely(object == ZERO_SIZE_PTR))
3891 page = virt_to_head_page(object);
3893 if (unlikely(!PageSlab(page))) {
3894 WARN_ON(!PageCompound(page));
3895 return PAGE_SIZE << compound_order(page);
3898 return slab_ksize(page->slab_cache);
3901 size_t ksize(const void *object)
3903 size_t size = __ksize(object);
3904 /* We assume that ksize callers could use whole allocated area,
3905 * so we need to unpoison this area.
3907 kasan_unpoison_shadow(object, size);
3910 EXPORT_SYMBOL(ksize);
3912 void kfree(const void *x)
3915 void *object = (void *)x;
3917 trace_kfree(_RET_IP_, x);
3919 if (unlikely(ZERO_OR_NULL_PTR(x)))
3922 page = virt_to_head_page(x);
3923 if (unlikely(!PageSlab(page))) {
3924 BUG_ON(!PageCompound(page));
3926 __free_pages(page, compound_order(page));
3929 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3931 EXPORT_SYMBOL(kfree);
3933 #define SHRINK_PROMOTE_MAX 32
3936 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3937 * up most to the head of the partial lists. New allocations will then
3938 * fill those up and thus they can be removed from the partial lists.
3940 * The slabs with the least items are placed last. This results in them
3941 * being allocated from last increasing the chance that the last objects
3942 * are freed in them.
3944 int __kmem_cache_shrink(struct kmem_cache *s)
3948 struct kmem_cache_node *n;
3951 struct list_head discard;
3952 struct list_head promote[SHRINK_PROMOTE_MAX];
3953 unsigned long flags;
3957 for_each_kmem_cache_node(s, node, n) {
3958 INIT_LIST_HEAD(&discard);
3959 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3960 INIT_LIST_HEAD(promote + i);
3962 spin_lock_irqsave(&n->list_lock, flags);
3965 * Build lists of slabs to discard or promote.
3967 * Note that concurrent frees may occur while we hold the
3968 * list_lock. page->inuse here is the upper limit.
3970 list_for_each_entry_safe(page, t, &n->partial, lru) {
3971 int free = page->objects - page->inuse;
3973 /* Do not reread page->inuse */
3976 /* We do not keep full slabs on the list */
3979 if (free == page->objects) {
3980 list_move(&page->lru, &discard);
3982 } else if (free <= SHRINK_PROMOTE_MAX)
3983 list_move(&page->lru, promote + free - 1);
3987 * Promote the slabs filled up most to the head of the
3990 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3991 list_splice(promote + i, &n->partial);
3993 spin_unlock_irqrestore(&n->list_lock, flags);
3995 /* Release empty slabs */
3996 list_for_each_entry_safe(page, t, &discard, lru)
3997 discard_slab(s, page);
3999 if (slabs_node(s, node))
4006 static int slab_mem_going_offline_callback(void *arg)
4008 struct kmem_cache *s;
4010 mutex_lock(&slab_mutex);
4011 list_for_each_entry(s, &slab_caches, list)
4012 __kmem_cache_shrink(s);
4013 mutex_unlock(&slab_mutex);
4018 static void slab_mem_offline_callback(void *arg)
4020 struct kmem_cache_node *n;
4021 struct kmem_cache *s;
4022 struct memory_notify *marg = arg;
4025 offline_node = marg->status_change_nid_normal;
4028 * If the node still has available memory. we need kmem_cache_node
4031 if (offline_node < 0)
4034 mutex_lock(&slab_mutex);
4035 list_for_each_entry(s, &slab_caches, list) {
4036 n = get_node(s, offline_node);
4039 * if n->nr_slabs > 0, slabs still exist on the node
4040 * that is going down. We were unable to free them,
4041 * and offline_pages() function shouldn't call this
4042 * callback. So, we must fail.
4044 BUG_ON(slabs_node(s, offline_node));
4046 s->node[offline_node] = NULL;
4047 kmem_cache_free(kmem_cache_node, n);
4050 mutex_unlock(&slab_mutex);
4053 static int slab_mem_going_online_callback(void *arg)
4055 struct kmem_cache_node *n;
4056 struct kmem_cache *s;
4057 struct memory_notify *marg = arg;
4058 int nid = marg->status_change_nid_normal;
4062 * If the node's memory is already available, then kmem_cache_node is
4063 * already created. Nothing to do.
4069 * We are bringing a node online. No memory is available yet. We must
4070 * allocate a kmem_cache_node structure in order to bring the node
4073 mutex_lock(&slab_mutex);
4074 list_for_each_entry(s, &slab_caches, list) {
4076 * XXX: kmem_cache_alloc_node will fallback to other nodes
4077 * since memory is not yet available from the node that
4080 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4085 init_kmem_cache_node(n);
4089 mutex_unlock(&slab_mutex);
4093 static int slab_memory_callback(struct notifier_block *self,
4094 unsigned long action, void *arg)
4099 case MEM_GOING_ONLINE:
4100 ret = slab_mem_going_online_callback(arg);
4102 case MEM_GOING_OFFLINE:
4103 ret = slab_mem_going_offline_callback(arg);
4106 case MEM_CANCEL_ONLINE:
4107 slab_mem_offline_callback(arg);
4110 case MEM_CANCEL_OFFLINE:
4114 ret = notifier_from_errno(ret);
4120 static struct notifier_block slab_memory_callback_nb = {
4121 .notifier_call = slab_memory_callback,
4122 .priority = SLAB_CALLBACK_PRI,
4125 /********************************************************************
4126 * Basic setup of slabs
4127 *******************************************************************/
4130 * Used for early kmem_cache structures that were allocated using
4131 * the page allocator. Allocate them properly then fix up the pointers
4132 * that may be pointing to the wrong kmem_cache structure.
4135 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4138 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4139 struct kmem_cache_node *n;
4141 memcpy(s, static_cache, kmem_cache->object_size);
4144 * This runs very early, and only the boot processor is supposed to be
4145 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4148 __flush_cpu_slab(s, smp_processor_id());
4149 for_each_kmem_cache_node(s, node, n) {
4152 list_for_each_entry(p, &n->partial, lru)
4155 #ifdef CONFIG_SLUB_DEBUG
4156 list_for_each_entry(p, &n->full, lru)
4160 slab_init_memcg_params(s);
4161 list_add(&s->list, &slab_caches);
4165 void __init kmem_cache_init(void)
4167 static __initdata struct kmem_cache boot_kmem_cache,
4168 boot_kmem_cache_node;
4170 if (debug_guardpage_minorder())
4173 kmem_cache_node = &boot_kmem_cache_node;
4174 kmem_cache = &boot_kmem_cache;
4176 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4177 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
4179 register_hotmemory_notifier(&slab_memory_callback_nb);
4181 /* Able to allocate the per node structures */
4182 slab_state = PARTIAL;
4184 create_boot_cache(kmem_cache, "kmem_cache",
4185 offsetof(struct kmem_cache, node) +
4186 nr_node_ids * sizeof(struct kmem_cache_node *),
4187 SLAB_HWCACHE_ALIGN);
4189 kmem_cache = bootstrap(&boot_kmem_cache);
4192 * Allocate kmem_cache_node properly from the kmem_cache slab.
4193 * kmem_cache_node is separately allocated so no need to
4194 * update any list pointers.
4196 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4198 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4199 setup_kmalloc_cache_index_table();
4200 create_kmalloc_caches(0);
4202 /* Setup random freelists for each cache */
4203 init_freelist_randomization();
4205 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4208 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
4210 slub_min_order, slub_max_order, slub_min_objects,
4211 nr_cpu_ids, nr_node_ids);
4214 void __init kmem_cache_init_late(void)
4219 __kmem_cache_alias(const char *name, size_t size, size_t align,
4220 unsigned long flags, void (*ctor)(void *))
4222 struct kmem_cache *s, *c;
4224 s = find_mergeable(size, align, flags, name, ctor);
4229 * Adjust the object sizes so that we clear
4230 * the complete object on kzalloc.
4232 s->object_size = max(s->object_size, (int)size);
4233 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
4235 for_each_memcg_cache(c, s) {
4236 c->object_size = s->object_size;
4237 c->inuse = max_t(int, c->inuse,
4238 ALIGN(size, sizeof(void *)));
4241 if (sysfs_slab_alias(s, name)) {
4250 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4254 err = kmem_cache_open(s, flags);
4258 /* Mutex is not taken during early boot */
4259 if (slab_state <= UP)
4262 memcg_propagate_slab_attrs(s);
4263 err = sysfs_slab_add(s);
4265 __kmem_cache_release(s);
4270 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4272 struct kmem_cache *s;
4275 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4276 return kmalloc_large(size, gfpflags);
4278 s = kmalloc_slab(size, gfpflags);
4280 if (unlikely(ZERO_OR_NULL_PTR(s)))
4283 ret = slab_alloc(s, gfpflags, caller);
4285 /* Honor the call site pointer we received. */
4286 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4292 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4293 int node, unsigned long caller)
4295 struct kmem_cache *s;
4298 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4299 ret = kmalloc_large_node(size, gfpflags, node);
4301 trace_kmalloc_node(caller, ret,
4302 size, PAGE_SIZE << get_order(size),
4308 s = kmalloc_slab(size, gfpflags);
4310 if (unlikely(ZERO_OR_NULL_PTR(s)))
4313 ret = slab_alloc_node(s, gfpflags, node, caller);
4315 /* Honor the call site pointer we received. */
4316 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4323 static int count_inuse(struct page *page)
4328 static int count_total(struct page *page)
4330 return page->objects;
4334 #ifdef CONFIG_SLUB_DEBUG
4335 static int validate_slab(struct kmem_cache *s, struct page *page,
4339 void *addr = page_address(page);
4341 if (!check_slab(s, page) ||
4342 !on_freelist(s, page, NULL))
4345 /* Now we know that a valid freelist exists */
4346 bitmap_zero(map, page->objects);
4348 get_map(s, page, map);
4349 for_each_object(p, s, addr, page->objects) {
4350 if (test_bit(slab_index(p, s, addr), map))
4351 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4355 for_each_object(p, s, addr, page->objects)
4356 if (!test_bit(slab_index(p, s, addr), map))
4357 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4362 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4366 validate_slab(s, page, map);
4370 static int validate_slab_node(struct kmem_cache *s,
4371 struct kmem_cache_node *n, unsigned long *map)
4373 unsigned long count = 0;
4375 unsigned long flags;
4377 spin_lock_irqsave(&n->list_lock, flags);
4379 list_for_each_entry(page, &n->partial, lru) {
4380 validate_slab_slab(s, page, map);
4383 if (count != n->nr_partial)
4384 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4385 s->name, count, n->nr_partial);
4387 if (!(s->flags & SLAB_STORE_USER))
4390 list_for_each_entry(page, &n->full, lru) {
4391 validate_slab_slab(s, page, map);
4394 if (count != atomic_long_read(&n->nr_slabs))
4395 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4396 s->name, count, atomic_long_read(&n->nr_slabs));
4399 spin_unlock_irqrestore(&n->list_lock, flags);
4403 static long validate_slab_cache(struct kmem_cache *s)
4406 unsigned long count = 0;
4407 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4408 sizeof(unsigned long), GFP_KERNEL);
4409 struct kmem_cache_node *n;
4415 for_each_kmem_cache_node(s, node, n)
4416 count += validate_slab_node(s, n, map);
4421 * Generate lists of code addresses where slabcache objects are allocated
4426 unsigned long count;
4433 DECLARE_BITMAP(cpus, NR_CPUS);
4439 unsigned long count;
4440 struct location *loc;
4443 static void free_loc_track(struct loc_track *t)
4446 free_pages((unsigned long)t->loc,
4447 get_order(sizeof(struct location) * t->max));
4450 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4455 order = get_order(sizeof(struct location) * max);
4457 l = (void *)__get_free_pages(flags, order);
4462 memcpy(l, t->loc, sizeof(struct location) * t->count);
4470 static int add_location(struct loc_track *t, struct kmem_cache *s,
4471 const struct track *track)
4473 long start, end, pos;
4475 unsigned long caddr;
4476 unsigned long age = jiffies - track->when;
4482 pos = start + (end - start + 1) / 2;
4485 * There is nothing at "end". If we end up there
4486 * we need to add something to before end.
4491 caddr = t->loc[pos].addr;
4492 if (track->addr == caddr) {
4498 if (age < l->min_time)
4500 if (age > l->max_time)
4503 if (track->pid < l->min_pid)
4504 l->min_pid = track->pid;
4505 if (track->pid > l->max_pid)
4506 l->max_pid = track->pid;
4508 cpumask_set_cpu(track->cpu,
4509 to_cpumask(l->cpus));
4511 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4515 if (track->addr < caddr)
4522 * Not found. Insert new tracking element.
4524 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4530 (t->count - pos) * sizeof(struct location));
4533 l->addr = track->addr;
4537 l->min_pid = track->pid;
4538 l->max_pid = track->pid;
4539 cpumask_clear(to_cpumask(l->cpus));
4540 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4541 nodes_clear(l->nodes);
4542 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4546 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4547 struct page *page, enum track_item alloc,
4550 void *addr = page_address(page);
4553 bitmap_zero(map, page->objects);
4554 get_map(s, page, map);
4556 for_each_object(p, s, addr, page->objects)
4557 if (!test_bit(slab_index(p, s, addr), map))
4558 add_location(t, s, get_track(s, p, alloc));
4561 static int list_locations(struct kmem_cache *s, char *buf,
4562 enum track_item alloc)
4566 struct loc_track t = { 0, 0, NULL };
4568 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4569 sizeof(unsigned long), GFP_KERNEL);
4570 struct kmem_cache_node *n;
4572 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4575 return sprintf(buf, "Out of memory\n");
4577 /* Push back cpu slabs */
4580 for_each_kmem_cache_node(s, node, n) {
4581 unsigned long flags;
4584 if (!atomic_long_read(&n->nr_slabs))
4587 spin_lock_irqsave(&n->list_lock, flags);
4588 list_for_each_entry(page, &n->partial, lru)
4589 process_slab(&t, s, page, alloc, map);
4590 list_for_each_entry(page, &n->full, lru)
4591 process_slab(&t, s, page, alloc, map);
4592 spin_unlock_irqrestore(&n->list_lock, flags);
4595 for (i = 0; i < t.count; i++) {
4596 struct location *l = &t.loc[i];
4598 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4600 len += sprintf(buf + len, "%7ld ", l->count);
4603 len += sprintf(buf + len, "%pS", (void *)l->addr);
4605 len += sprintf(buf + len, "<not-available>");
4607 if (l->sum_time != l->min_time) {
4608 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4610 (long)div_u64(l->sum_time, l->count),
4613 len += sprintf(buf + len, " age=%ld",
4616 if (l->min_pid != l->max_pid)
4617 len += sprintf(buf + len, " pid=%ld-%ld",
4618 l->min_pid, l->max_pid);
4620 len += sprintf(buf + len, " pid=%ld",
4623 if (num_online_cpus() > 1 &&
4624 !cpumask_empty(to_cpumask(l->cpus)) &&
4625 len < PAGE_SIZE - 60)
4626 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4628 cpumask_pr_args(to_cpumask(l->cpus)));
4630 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4631 len < PAGE_SIZE - 60)
4632 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4634 nodemask_pr_args(&l->nodes));
4636 len += sprintf(buf + len, "\n");
4642 len += sprintf(buf, "No data\n");
4647 #ifdef SLUB_RESILIENCY_TEST
4648 static void __init resiliency_test(void)
4652 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4654 pr_err("SLUB resiliency testing\n");
4655 pr_err("-----------------------\n");
4656 pr_err("A. Corruption after allocation\n");
4658 p = kzalloc(16, GFP_KERNEL);
4660 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4663 validate_slab_cache(kmalloc_caches[4]);
4665 /* Hmmm... The next two are dangerous */
4666 p = kzalloc(32, GFP_KERNEL);
4667 p[32 + sizeof(void *)] = 0x34;
4668 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4670 pr_err("If allocated object is overwritten then not detectable\n\n");
4672 validate_slab_cache(kmalloc_caches[5]);
4673 p = kzalloc(64, GFP_KERNEL);
4674 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4676 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4678 pr_err("If allocated object is overwritten then not detectable\n\n");
4679 validate_slab_cache(kmalloc_caches[6]);
4681 pr_err("\nB. Corruption after free\n");
4682 p = kzalloc(128, GFP_KERNEL);
4685 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4686 validate_slab_cache(kmalloc_caches[7]);
4688 p = kzalloc(256, GFP_KERNEL);
4691 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4692 validate_slab_cache(kmalloc_caches[8]);
4694 p = kzalloc(512, GFP_KERNEL);
4697 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4698 validate_slab_cache(kmalloc_caches[9]);
4702 static void resiliency_test(void) {};
4707 enum slab_stat_type {
4708 SL_ALL, /* All slabs */
4709 SL_PARTIAL, /* Only partially allocated slabs */
4710 SL_CPU, /* Only slabs used for cpu caches */
4711 SL_OBJECTS, /* Determine allocated objects not slabs */
4712 SL_TOTAL /* Determine object capacity not slabs */
4715 #define SO_ALL (1 << SL_ALL)
4716 #define SO_PARTIAL (1 << SL_PARTIAL)
4717 #define SO_CPU (1 << SL_CPU)
4718 #define SO_OBJECTS (1 << SL_OBJECTS)
4719 #define SO_TOTAL (1 << SL_TOTAL)
4721 static ssize_t show_slab_objects(struct kmem_cache *s,
4722 char *buf, unsigned long flags)
4724 unsigned long total = 0;
4727 unsigned long *nodes;
4729 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4733 if (flags & SO_CPU) {
4736 for_each_possible_cpu(cpu) {
4737 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4742 page = READ_ONCE(c->page);
4746 node = page_to_nid(page);
4747 if (flags & SO_TOTAL)
4749 else if (flags & SO_OBJECTS)
4757 page = READ_ONCE(c->partial);
4759 node = page_to_nid(page);
4760 if (flags & SO_TOTAL)
4762 else if (flags & SO_OBJECTS)
4773 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4774 * already held which will conflict with an existing lock order:
4776 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4778 * We don't really need mem_hotplug_lock (to hold off
4779 * slab_mem_going_offline_callback) here because slab's memory hot
4780 * unplug code doesn't destroy the kmem_cache->node[] data.
4783 #ifdef CONFIG_SLUB_DEBUG
4784 if (flags & SO_ALL) {
4785 struct kmem_cache_node *n;
4787 for_each_kmem_cache_node(s, node, n) {
4789 if (flags & SO_TOTAL)
4790 x = atomic_long_read(&n->total_objects);
4791 else if (flags & SO_OBJECTS)
4792 x = atomic_long_read(&n->total_objects) -
4793 count_partial(n, count_free);
4795 x = atomic_long_read(&n->nr_slabs);
4802 if (flags & SO_PARTIAL) {
4803 struct kmem_cache_node *n;
4805 for_each_kmem_cache_node(s, node, n) {
4806 if (flags & SO_TOTAL)
4807 x = count_partial(n, count_total);
4808 else if (flags & SO_OBJECTS)
4809 x = count_partial(n, count_inuse);
4816 x = sprintf(buf, "%lu", total);
4818 for (node = 0; node < nr_node_ids; node++)
4820 x += sprintf(buf + x, " N%d=%lu",
4824 return x + sprintf(buf + x, "\n");
4827 #ifdef CONFIG_SLUB_DEBUG
4828 static int any_slab_objects(struct kmem_cache *s)
4831 struct kmem_cache_node *n;
4833 for_each_kmem_cache_node(s, node, n)
4834 if (atomic_long_read(&n->total_objects))
4841 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4842 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4844 struct slab_attribute {
4845 struct attribute attr;
4846 ssize_t (*show)(struct kmem_cache *s, char *buf);
4847 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4850 #define SLAB_ATTR_RO(_name) \
4851 static struct slab_attribute _name##_attr = \
4852 __ATTR(_name, 0400, _name##_show, NULL)
4854 #define SLAB_ATTR(_name) \
4855 static struct slab_attribute _name##_attr = \
4856 __ATTR(_name, 0600, _name##_show, _name##_store)
4858 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4860 return sprintf(buf, "%d\n", s->size);
4862 SLAB_ATTR_RO(slab_size);
4864 static ssize_t align_show(struct kmem_cache *s, char *buf)
4866 return sprintf(buf, "%d\n", s->align);
4868 SLAB_ATTR_RO(align);
4870 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4872 return sprintf(buf, "%d\n", s->object_size);
4874 SLAB_ATTR_RO(object_size);
4876 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4878 return sprintf(buf, "%d\n", oo_objects(s->oo));
4880 SLAB_ATTR_RO(objs_per_slab);
4882 static ssize_t order_store(struct kmem_cache *s,
4883 const char *buf, size_t length)
4885 unsigned long order;
4888 err = kstrtoul(buf, 10, &order);
4892 if (order > slub_max_order || order < slub_min_order)
4895 calculate_sizes(s, order);
4899 static ssize_t order_show(struct kmem_cache *s, char *buf)
4901 return sprintf(buf, "%d\n", oo_order(s->oo));
4905 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4907 return sprintf(buf, "%lu\n", s->min_partial);
4910 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4916 err = kstrtoul(buf, 10, &min);
4920 set_min_partial(s, min);
4923 SLAB_ATTR(min_partial);
4925 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4927 return sprintf(buf, "%u\n", s->cpu_partial);
4930 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4933 unsigned int objects;
4936 err = kstrtouint(buf, 10, &objects);
4939 if (objects && !kmem_cache_has_cpu_partial(s))
4942 s->cpu_partial = objects;
4946 SLAB_ATTR(cpu_partial);
4948 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4952 return sprintf(buf, "%pS\n", s->ctor);
4956 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4958 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4960 SLAB_ATTR_RO(aliases);
4962 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4964 return show_slab_objects(s, buf, SO_PARTIAL);
4966 SLAB_ATTR_RO(partial);
4968 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4970 return show_slab_objects(s, buf, SO_CPU);
4972 SLAB_ATTR_RO(cpu_slabs);
4974 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4976 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4978 SLAB_ATTR_RO(objects);
4980 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4982 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4984 SLAB_ATTR_RO(objects_partial);
4986 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4993 for_each_online_cpu(cpu) {
4994 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4997 pages += page->pages;
4998 objects += page->pobjects;
5002 len = sprintf(buf, "%d(%d)", objects, pages);
5005 for_each_online_cpu(cpu) {
5006 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
5008 if (page && len < PAGE_SIZE - 20)
5009 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5010 page->pobjects, page->pages);
5013 return len + sprintf(buf + len, "\n");
5015 SLAB_ATTR_RO(slabs_cpu_partial);
5017 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5019 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5022 static ssize_t reclaim_account_store(struct kmem_cache *s,
5023 const char *buf, size_t length)
5025 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5027 s->flags |= SLAB_RECLAIM_ACCOUNT;
5030 SLAB_ATTR(reclaim_account);
5032 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5034 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5036 SLAB_ATTR_RO(hwcache_align);
5038 #ifdef CONFIG_ZONE_DMA
5039 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5041 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5043 SLAB_ATTR_RO(cache_dma);
5046 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5048 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
5050 SLAB_ATTR_RO(destroy_by_rcu);
5052 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
5054 return sprintf(buf, "%d\n", s->reserved);
5056 SLAB_ATTR_RO(reserved);
5058 #ifdef CONFIG_SLUB_DEBUG
5059 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5061 return show_slab_objects(s, buf, SO_ALL);
5063 SLAB_ATTR_RO(slabs);
5065 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5067 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5069 SLAB_ATTR_RO(total_objects);
5071 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5073 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5076 static ssize_t sanity_checks_store(struct kmem_cache *s,
5077 const char *buf, size_t length)
5079 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5080 if (buf[0] == '1') {
5081 s->flags &= ~__CMPXCHG_DOUBLE;
5082 s->flags |= SLAB_CONSISTENCY_CHECKS;
5086 SLAB_ATTR(sanity_checks);
5088 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5090 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5093 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5097 * Tracing a merged cache is going to give confusing results
5098 * as well as cause other issues like converting a mergeable
5099 * cache into an umergeable one.
5101 if (s->refcount > 1)
5104 s->flags &= ~SLAB_TRACE;
5105 if (buf[0] == '1') {
5106 s->flags &= ~__CMPXCHG_DOUBLE;
5107 s->flags |= SLAB_TRACE;
5113 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5115 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5118 static ssize_t red_zone_store(struct kmem_cache *s,
5119 const char *buf, size_t length)
5121 if (any_slab_objects(s))
5124 s->flags &= ~SLAB_RED_ZONE;
5125 if (buf[0] == '1') {
5126 s->flags |= SLAB_RED_ZONE;
5128 calculate_sizes(s, -1);
5131 SLAB_ATTR(red_zone);
5133 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5135 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5138 static ssize_t poison_store(struct kmem_cache *s,
5139 const char *buf, size_t length)
5141 if (any_slab_objects(s))
5144 s->flags &= ~SLAB_POISON;
5145 if (buf[0] == '1') {
5146 s->flags |= SLAB_POISON;
5148 calculate_sizes(s, -1);
5153 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5155 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5158 static ssize_t store_user_store(struct kmem_cache *s,
5159 const char *buf, size_t length)
5161 if (any_slab_objects(s))
5164 s->flags &= ~SLAB_STORE_USER;
5165 if (buf[0] == '1') {
5166 s->flags &= ~__CMPXCHG_DOUBLE;
5167 s->flags |= SLAB_STORE_USER;
5169 calculate_sizes(s, -1);
5172 SLAB_ATTR(store_user);
5174 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5179 static ssize_t validate_store(struct kmem_cache *s,
5180 const char *buf, size_t length)
5184 if (buf[0] == '1') {
5185 ret = validate_slab_cache(s);
5191 SLAB_ATTR(validate);
5193 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5195 if (!(s->flags & SLAB_STORE_USER))
5197 return list_locations(s, buf, TRACK_ALLOC);
5199 SLAB_ATTR_RO(alloc_calls);
5201 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5203 if (!(s->flags & SLAB_STORE_USER))
5205 return list_locations(s, buf, TRACK_FREE);
5207 SLAB_ATTR_RO(free_calls);
5208 #endif /* CONFIG_SLUB_DEBUG */
5210 #ifdef CONFIG_FAILSLAB
5211 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5213 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5216 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5219 if (s->refcount > 1)
5222 s->flags &= ~SLAB_FAILSLAB;
5224 s->flags |= SLAB_FAILSLAB;
5227 SLAB_ATTR(failslab);
5230 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5235 static ssize_t shrink_store(struct kmem_cache *s,
5236 const char *buf, size_t length)
5239 kmem_cache_shrink(s);
5247 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5249 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5252 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5253 const char *buf, size_t length)
5255 unsigned long ratio;
5258 err = kstrtoul(buf, 10, &ratio);
5263 s->remote_node_defrag_ratio = ratio * 10;
5267 SLAB_ATTR(remote_node_defrag_ratio);
5270 #ifdef CONFIG_SLUB_STATS
5271 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5273 unsigned long sum = 0;
5276 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5281 for_each_online_cpu(cpu) {
5282 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5288 len = sprintf(buf, "%lu", sum);
5291 for_each_online_cpu(cpu) {
5292 if (data[cpu] && len < PAGE_SIZE - 20)
5293 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5297 return len + sprintf(buf + len, "\n");
5300 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5304 for_each_online_cpu(cpu)
5305 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5308 #define STAT_ATTR(si, text) \
5309 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5311 return show_stat(s, buf, si); \
5313 static ssize_t text##_store(struct kmem_cache *s, \
5314 const char *buf, size_t length) \
5316 if (buf[0] != '0') \
5318 clear_stat(s, si); \
5323 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5324 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5325 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5326 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5327 STAT_ATTR(FREE_FROZEN, free_frozen);
5328 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5329 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5330 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5331 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5332 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5333 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5334 STAT_ATTR(FREE_SLAB, free_slab);
5335 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5336 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5337 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5338 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5339 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5340 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5341 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5342 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5343 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5344 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5345 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5346 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5347 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5348 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5351 static struct attribute *slab_attrs[] = {
5352 &slab_size_attr.attr,
5353 &object_size_attr.attr,
5354 &objs_per_slab_attr.attr,
5356 &min_partial_attr.attr,
5357 &cpu_partial_attr.attr,
5359 &objects_partial_attr.attr,
5361 &cpu_slabs_attr.attr,
5365 &hwcache_align_attr.attr,
5366 &reclaim_account_attr.attr,
5367 &destroy_by_rcu_attr.attr,
5369 &reserved_attr.attr,
5370 &slabs_cpu_partial_attr.attr,
5371 #ifdef CONFIG_SLUB_DEBUG
5372 &total_objects_attr.attr,
5374 &sanity_checks_attr.attr,
5376 &red_zone_attr.attr,
5378 &store_user_attr.attr,
5379 &validate_attr.attr,
5380 &alloc_calls_attr.attr,
5381 &free_calls_attr.attr,
5383 #ifdef CONFIG_ZONE_DMA
5384 &cache_dma_attr.attr,
5387 &remote_node_defrag_ratio_attr.attr,
5389 #ifdef CONFIG_SLUB_STATS
5390 &alloc_fastpath_attr.attr,
5391 &alloc_slowpath_attr.attr,
5392 &free_fastpath_attr.attr,
5393 &free_slowpath_attr.attr,
5394 &free_frozen_attr.attr,
5395 &free_add_partial_attr.attr,
5396 &free_remove_partial_attr.attr,
5397 &alloc_from_partial_attr.attr,
5398 &alloc_slab_attr.attr,
5399 &alloc_refill_attr.attr,
5400 &alloc_node_mismatch_attr.attr,
5401 &free_slab_attr.attr,
5402 &cpuslab_flush_attr.attr,
5403 &deactivate_full_attr.attr,
5404 &deactivate_empty_attr.attr,
5405 &deactivate_to_head_attr.attr,
5406 &deactivate_to_tail_attr.attr,
5407 &deactivate_remote_frees_attr.attr,
5408 &deactivate_bypass_attr.attr,
5409 &order_fallback_attr.attr,
5410 &cmpxchg_double_fail_attr.attr,
5411 &cmpxchg_double_cpu_fail_attr.attr,
5412 &cpu_partial_alloc_attr.attr,
5413 &cpu_partial_free_attr.attr,
5414 &cpu_partial_node_attr.attr,
5415 &cpu_partial_drain_attr.attr,
5417 #ifdef CONFIG_FAILSLAB
5418 &failslab_attr.attr,
5424 static struct attribute_group slab_attr_group = {
5425 .attrs = slab_attrs,
5428 static ssize_t slab_attr_show(struct kobject *kobj,
5429 struct attribute *attr,
5432 struct slab_attribute *attribute;
5433 struct kmem_cache *s;
5436 attribute = to_slab_attr(attr);
5439 if (!attribute->show)
5442 err = attribute->show(s, buf);
5447 static ssize_t slab_attr_store(struct kobject *kobj,
5448 struct attribute *attr,
5449 const char *buf, size_t len)
5451 struct slab_attribute *attribute;
5452 struct kmem_cache *s;
5455 attribute = to_slab_attr(attr);
5458 if (!attribute->store)
5461 err = attribute->store(s, buf, len);
5463 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5464 struct kmem_cache *c;
5466 mutex_lock(&slab_mutex);
5467 if (s->max_attr_size < len)
5468 s->max_attr_size = len;
5471 * This is a best effort propagation, so this function's return
5472 * value will be determined by the parent cache only. This is
5473 * basically because not all attributes will have a well
5474 * defined semantics for rollbacks - most of the actions will
5475 * have permanent effects.
5477 * Returning the error value of any of the children that fail
5478 * is not 100 % defined, in the sense that users seeing the
5479 * error code won't be able to know anything about the state of
5482 * Only returning the error code for the parent cache at least
5483 * has well defined semantics. The cache being written to
5484 * directly either failed or succeeded, in which case we loop
5485 * through the descendants with best-effort propagation.
5487 for_each_memcg_cache(c, s)
5488 attribute->store(c, buf, len);
5489 mutex_unlock(&slab_mutex);
5495 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5499 char *buffer = NULL;
5500 struct kmem_cache *root_cache;
5502 if (is_root_cache(s))
5505 root_cache = s->memcg_params.root_cache;
5508 * This mean this cache had no attribute written. Therefore, no point
5509 * in copying default values around
5511 if (!root_cache->max_attr_size)
5514 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5517 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5520 if (!attr || !attr->store || !attr->show)
5524 * It is really bad that we have to allocate here, so we will
5525 * do it only as a fallback. If we actually allocate, though,
5526 * we can just use the allocated buffer until the end.
5528 * Most of the slub attributes will tend to be very small in
5529 * size, but sysfs allows buffers up to a page, so they can
5530 * theoretically happen.
5534 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5535 !IS_ENABLED(CONFIG_SLUB_STATS))
5538 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5539 if (WARN_ON(!buffer))
5544 len = attr->show(root_cache, buf);
5546 attr->store(s, buf, len);
5550 free_page((unsigned long)buffer);
5554 static void kmem_cache_release(struct kobject *k)
5556 slab_kmem_cache_release(to_slab(k));
5559 static const struct sysfs_ops slab_sysfs_ops = {
5560 .show = slab_attr_show,
5561 .store = slab_attr_store,
5564 static struct kobj_type slab_ktype = {
5565 .sysfs_ops = &slab_sysfs_ops,
5566 .release = kmem_cache_release,
5569 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5571 struct kobj_type *ktype = get_ktype(kobj);
5573 if (ktype == &slab_ktype)
5578 static const struct kset_uevent_ops slab_uevent_ops = {
5579 .filter = uevent_filter,
5582 static struct kset *slab_kset;
5584 static inline struct kset *cache_kset(struct kmem_cache *s)
5587 if (!is_root_cache(s))
5588 return s->memcg_params.root_cache->memcg_kset;
5593 #define ID_STR_LENGTH 64
5595 /* Create a unique string id for a slab cache:
5597 * Format :[flags-]size
5599 static char *create_unique_id(struct kmem_cache *s)
5601 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5605 return ERR_PTR(-ENOMEM);
5609 * First flags affecting slabcache operations. We will only
5610 * get here for aliasable slabs so we do not need to support
5611 * too many flags. The flags here must cover all flags that
5612 * are matched during merging to guarantee that the id is
5615 if (s->flags & SLAB_CACHE_DMA)
5617 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5619 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5621 if (!(s->flags & SLAB_NOTRACK))
5623 if (s->flags & SLAB_ACCOUNT)
5627 p += sprintf(p, "%07d", s->size);
5629 BUG_ON(p > name + ID_STR_LENGTH - 1);
5633 static int sysfs_slab_add(struct kmem_cache *s)
5637 int unmergeable = slab_unmergeable(s);
5641 * Slabcache can never be merged so we can use the name proper.
5642 * This is typically the case for debug situations. In that
5643 * case we can catch duplicate names easily.
5645 sysfs_remove_link(&slab_kset->kobj, s->name);
5649 * Create a unique name for the slab as a target
5652 name = create_unique_id(s);
5654 return PTR_ERR(name);
5657 s->kobj.kset = cache_kset(s);
5658 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5662 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5667 if (is_root_cache(s)) {
5668 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5669 if (!s->memcg_kset) {
5676 kobject_uevent(&s->kobj, KOBJ_ADD);
5678 /* Setup first alias */
5679 sysfs_slab_alias(s, s->name);
5686 kobject_del(&s->kobj);
5690 void sysfs_slab_remove(struct kmem_cache *s)
5692 if (slab_state < FULL)
5694 * Sysfs has not been setup yet so no need to remove the
5700 kset_unregister(s->memcg_kset);
5702 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5703 kobject_del(&s->kobj);
5704 kobject_put(&s->kobj);
5708 * Need to buffer aliases during bootup until sysfs becomes
5709 * available lest we lose that information.
5711 struct saved_alias {
5712 struct kmem_cache *s;
5714 struct saved_alias *next;
5717 static struct saved_alias *alias_list;
5719 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5721 struct saved_alias *al;
5723 if (slab_state == FULL) {
5725 * If we have a leftover link then remove it.
5727 sysfs_remove_link(&slab_kset->kobj, name);
5728 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5731 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5737 al->next = alias_list;
5742 static int __init slab_sysfs_init(void)
5744 struct kmem_cache *s;
5747 mutex_lock(&slab_mutex);
5749 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5751 mutex_unlock(&slab_mutex);
5752 pr_err("Cannot register slab subsystem.\n");
5758 list_for_each_entry(s, &slab_caches, list) {
5759 err = sysfs_slab_add(s);
5761 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5765 while (alias_list) {
5766 struct saved_alias *al = alias_list;
5768 alias_list = alias_list->next;
5769 err = sysfs_slab_alias(al->s, al->name);
5771 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5776 mutex_unlock(&slab_mutex);
5781 __initcall(slab_sysfs_init);
5782 #endif /* CONFIG_SYSFS */
5785 * The /proc/slabinfo ABI
5787 #ifdef CONFIG_SLABINFO
5788 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5790 unsigned long nr_slabs = 0;
5791 unsigned long nr_objs = 0;
5792 unsigned long nr_free = 0;
5794 struct kmem_cache_node *n;
5796 for_each_kmem_cache_node(s, node, n) {
5797 nr_slabs += node_nr_slabs(n);
5798 nr_objs += node_nr_objs(n);
5799 nr_free += count_partial(n, count_free);
5802 sinfo->active_objs = nr_objs - nr_free;
5803 sinfo->num_objs = nr_objs;
5804 sinfo->active_slabs = nr_slabs;
5805 sinfo->num_slabs = nr_slabs;
5806 sinfo->objects_per_slab = oo_objects(s->oo);
5807 sinfo->cache_order = oo_order(s->oo);
5810 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5814 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5815 size_t count, loff_t *ppos)
5819 #endif /* CONFIG_SLABINFO */