1 // SPDX-License-Identifier: GPL-2.0
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
6 * The allocator synchronizes using per slab locks or atomic operatios
7 * and only uses a centralized lock to manage a pool of partial slabs.
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
14 #include <linux/swap.h> /* struct reclaim_state */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
22 #include <linux/proc_fs.h>
23 #include <linux/notifier.h>
24 #include <linux/seq_file.h>
25 #include <linux/kasan.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ctype.h>
30 #include <linux/debugobjects.h>
31 #include <linux/kallsyms.h>
32 #include <linux/memory.h>
33 #include <linux/math64.h>
34 #include <linux/fault-inject.h>
35 #include <linux/stacktrace.h>
36 #include <linux/prefetch.h>
37 #include <linux/memcontrol.h>
38 #include <linux/random.h>
40 #include <trace/events/kmem.h>
46 * 1. slab_mutex (Global Mutex)
48 * 3. slab_lock(page) (Only on some arches and for debugging)
52 * The role of the slab_mutex is to protect the list of all the slabs
53 * and to synchronize major metadata changes to slab cache structures.
55 * The slab_lock is only used for debugging and on arches that do not
56 * have the ability to do a cmpxchg_double. It only protects the second
57 * double word in the page struct. Meaning
58 * A. page->freelist -> List of object free in a page
59 * B. page->counters -> Counters of objects
60 * C. page->frozen -> frozen state
62 * If a slab is frozen then it is exempt from list management. It is not
63 * on any list. The processor that froze the slab is the one who can
64 * perform list operations on the page. Other processors may put objects
65 * onto the freelist but the processor that froze the slab is the only
66 * one that can retrieve the objects from the page's freelist.
68 * The list_lock protects the partial and full list on each node and
69 * the partial slab counter. If taken then no new slabs may be added or
70 * removed from the lists nor make the number of partial slabs be modified.
71 * (Note that the total number of slabs is an atomic value that may be
72 * modified without taking the list lock).
74 * The list_lock is a centralized lock and thus we avoid taking it as
75 * much as possible. As long as SLUB does not have to handle partial
76 * slabs, operations can continue without any centralized lock. F.e.
77 * allocating a long series of objects that fill up slabs does not require
79 * Interrupts are disabled during allocation and deallocation in order to
80 * make the slab allocator safe to use in the context of an irq. In addition
81 * interrupts are disabled to ensure that the processor does not change
82 * while handling per_cpu slabs, due to kernel preemption.
84 * SLUB assigns one slab for allocation to each processor.
85 * Allocations only occur from these slabs called cpu slabs.
87 * Slabs with free elements are kept on a partial list and during regular
88 * operations no list for full slabs is used. If an object in a full slab is
89 * freed then the slab will show up again on the partial lists.
90 * We track full slabs for debugging purposes though because otherwise we
91 * cannot scan all objects.
93 * Slabs are freed when they become empty. Teardown and setup is
94 * minimal so we rely on the page allocators per cpu caches for
95 * fast frees and allocs.
97 * Overloading of page flags that are otherwise used for LRU management.
99 * PageActive The slab is frozen and exempt from list processing.
100 * This means that the slab is dedicated to a purpose
101 * such as satisfying allocations for a specific
102 * processor. Objects may be freed in the slab while
103 * it is frozen but slab_free will then skip the usual
104 * list operations. It is up to the processor holding
105 * the slab to integrate the slab into the slab lists
106 * when the slab is no longer needed.
108 * One use of this flag is to mark slabs that are
109 * used for allocations. Then such a slab becomes a cpu
110 * slab. The cpu slab may be equipped with an additional
111 * freelist that allows lockless access to
112 * free objects in addition to the regular freelist
113 * that requires the slab lock.
115 * PageError Slab requires special handling due to debug
116 * options set. This moves slab handling out of
117 * the fast path and disables lockless freelists.
120 static inline int kmem_cache_debug(struct kmem_cache *s)
122 #ifdef CONFIG_SLUB_DEBUG
123 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
129 void *fixup_red_left(struct kmem_cache *s, void *p)
131 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
132 p += s->red_left_pad;
137 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
139 #ifdef CONFIG_SLUB_CPU_PARTIAL
140 return !kmem_cache_debug(s);
147 * Issues still to be resolved:
149 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
151 * - Variable sizing of the per node arrays
154 /* Enable to test recovery from slab corruption on boot */
155 #undef SLUB_RESILIENCY_TEST
157 /* Enable to log cmpxchg failures */
158 #undef SLUB_DEBUG_CMPXCHG
161 * Mininum number of partial slabs. These will be left on the partial
162 * lists even if they are empty. kmem_cache_shrink may reclaim them.
164 #define MIN_PARTIAL 5
167 * Maximum number of desirable partial slabs.
168 * The existence of more partial slabs makes kmem_cache_shrink
169 * sort the partial list by the number of objects in use.
171 #define MAX_PARTIAL 10
173 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
174 SLAB_POISON | SLAB_STORE_USER)
177 * These debug flags cannot use CMPXCHG because there might be consistency
178 * issues when checking or reading debug information
180 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
185 * Debugging flags that require metadata to be stored in the slab. These get
186 * disabled when slub_debug=O is used and a cache's min order increases with
189 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
192 #define OO_MASK ((1 << OO_SHIFT) - 1)
193 #define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
195 /* Internal SLUB flags */
196 #define __OBJECT_POISON 0x80000000UL /* Poison object */
197 #define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
200 * Tracking user of a slab.
202 #define TRACK_ADDRS_COUNT 16
204 unsigned long addr; /* Called from address */
205 #ifdef CONFIG_STACKTRACE
206 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
208 int cpu; /* Was running on cpu */
209 int pid; /* Pid context */
210 unsigned long when; /* When did the operation occur */
213 enum track_item { TRACK_ALLOC, TRACK_FREE };
216 static int sysfs_slab_add(struct kmem_cache *);
217 static int sysfs_slab_alias(struct kmem_cache *, const char *);
218 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
219 static void sysfs_slab_remove(struct kmem_cache *s);
221 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
222 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
225 static inline void sysfs_slab_remove(struct kmem_cache *s) { }
228 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 #ifdef CONFIG_SLUB_STATS
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
235 raw_cpu_inc(s->cpu_slab->stat[si]);
239 /********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
244 * Returns freelist pointer (ptr). With hardening, this is obfuscated
245 * with an XOR of the address where the pointer is held and a per-cache
248 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
249 unsigned long ptr_addr)
251 #ifdef CONFIG_SLAB_FREELIST_HARDENED
252 return (void *)((unsigned long)ptr ^ s->random ^ swab(ptr_addr));
258 /* Returns the freelist pointer recorded at location ptr_addr. */
259 static inline void *freelist_dereference(const struct kmem_cache *s,
262 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
263 (unsigned long)ptr_addr);
266 static inline void *get_freepointer(struct kmem_cache *s, void *object)
268 return freelist_dereference(s, object + s->offset);
271 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
273 prefetch(object + s->offset);
276 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
278 unsigned long freepointer_addr;
281 if (!debug_pagealloc_enabled())
282 return get_freepointer(s, object);
284 freepointer_addr = (unsigned long)object + s->offset;
285 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
286 return freelist_ptr(s, p, freepointer_addr);
289 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291 unsigned long freeptr_addr = (unsigned long)object + s->offset;
293 #ifdef CONFIG_SLAB_FREELIST_HARDENED
294 BUG_ON(object == fp); /* naive detection of double free or corruption */
297 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
300 /* Loop over all objects in a slab */
301 #define for_each_object(__p, __s, __addr, __objects) \
302 for (__p = fixup_red_left(__s, __addr); \
303 __p < (__addr) + (__objects) * (__s)->size; \
306 #define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
307 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
308 __idx <= __objects; \
309 __p += (__s)->size, __idx++)
311 /* Determine object index from a given position */
312 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
314 return (p - addr) / s->size;
317 static inline int order_objects(int order, unsigned long size, int reserved)
319 return ((PAGE_SIZE << order) - reserved) / size;
322 static inline struct kmem_cache_order_objects oo_make(int order,
323 unsigned long size, int reserved)
325 struct kmem_cache_order_objects x = {
326 (order << OO_SHIFT) + order_objects(order, size, reserved)
332 static inline int oo_order(struct kmem_cache_order_objects x)
334 return x.x >> OO_SHIFT;
337 static inline int oo_objects(struct kmem_cache_order_objects x)
339 return x.x & OO_MASK;
343 * Per slab locking using the pagelock
345 static __always_inline void slab_lock(struct page *page)
347 VM_BUG_ON_PAGE(PageTail(page), page);
348 bit_spin_lock(PG_locked, &page->flags);
351 static __always_inline void slab_unlock(struct page *page)
353 VM_BUG_ON_PAGE(PageTail(page), page);
354 __bit_spin_unlock(PG_locked, &page->flags);
357 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
360 tmp.counters = counters_new;
362 * page->counters can cover frozen/inuse/objects as well
363 * as page->_refcount. If we assign to ->counters directly
364 * we run the risk of losing updates to page->_refcount, so
365 * be careful and only assign to the fields we need.
367 page->frozen = tmp.frozen;
368 page->inuse = tmp.inuse;
369 page->objects = tmp.objects;
372 /* Interrupts must be disabled (for the fallback code to work right) */
373 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
374 void *freelist_old, unsigned long counters_old,
375 void *freelist_new, unsigned long counters_new,
378 VM_BUG_ON(!irqs_disabled());
379 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
380 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
381 if (s->flags & __CMPXCHG_DOUBLE) {
382 if (cmpxchg_double(&page->freelist, &page->counters,
383 freelist_old, counters_old,
384 freelist_new, counters_new))
390 if (page->freelist == freelist_old &&
391 page->counters == counters_old) {
392 page->freelist = freelist_new;
393 set_page_slub_counters(page, counters_new);
401 stat(s, CMPXCHG_DOUBLE_FAIL);
403 #ifdef SLUB_DEBUG_CMPXCHG
404 pr_info("%s %s: cmpxchg double redo ", n, s->name);
410 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
411 void *freelist_old, unsigned long counters_old,
412 void *freelist_new, unsigned long counters_new,
415 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
416 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
417 if (s->flags & __CMPXCHG_DOUBLE) {
418 if (cmpxchg_double(&page->freelist, &page->counters,
419 freelist_old, counters_old,
420 freelist_new, counters_new))
427 local_irq_save(flags);
429 if (page->freelist == freelist_old &&
430 page->counters == counters_old) {
431 page->freelist = freelist_new;
432 set_page_slub_counters(page, counters_new);
434 local_irq_restore(flags);
438 local_irq_restore(flags);
442 stat(s, CMPXCHG_DOUBLE_FAIL);
444 #ifdef SLUB_DEBUG_CMPXCHG
445 pr_info("%s %s: cmpxchg double redo ", n, s->name);
451 #ifdef CONFIG_SLUB_DEBUG
453 * Determine a map of object in use on a page.
455 * Node listlock must be held to guarantee that the page does
456 * not vanish from under us.
458 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
461 void *addr = page_address(page);
463 for (p = page->freelist; p; p = get_freepointer(s, p))
464 set_bit(slab_index(p, s, addr), map);
467 static inline int size_from_object(struct kmem_cache *s)
469 if (s->flags & SLAB_RED_ZONE)
470 return s->size - s->red_left_pad;
475 static inline void *restore_red_left(struct kmem_cache *s, void *p)
477 if (s->flags & SLAB_RED_ZONE)
478 p -= s->red_left_pad;
486 #if defined(CONFIG_SLUB_DEBUG_ON)
487 static int slub_debug = DEBUG_DEFAULT_FLAGS;
489 static int slub_debug;
492 static char *slub_debug_slabs;
493 static int disable_higher_order_debug;
496 * slub is about to manipulate internal object metadata. This memory lies
497 * outside the range of the allocated object, so accessing it would normally
498 * be reported by kasan as a bounds error. metadata_access_enable() is used
499 * to tell kasan that these accesses are OK.
501 static inline void metadata_access_enable(void)
503 kasan_disable_current();
506 static inline void metadata_access_disable(void)
508 kasan_enable_current();
515 /* Verify that a pointer has an address that is valid within a slab page */
516 static inline int check_valid_pointer(struct kmem_cache *s,
517 struct page *page, void *object)
524 base = page_address(page);
525 object = restore_red_left(s, object);
526 if (object < base || object >= base + page->objects * s->size ||
527 (object - base) % s->size) {
534 static void print_section(char *level, char *text, u8 *addr,
537 metadata_access_enable();
538 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
540 metadata_access_disable();
543 static struct track *get_track(struct kmem_cache *s, void *object,
544 enum track_item alloc)
549 p = object + s->offset + sizeof(void *);
551 p = object + s->inuse;
556 static void set_track(struct kmem_cache *s, void *object,
557 enum track_item alloc, unsigned long addr)
559 struct track *p = get_track(s, object, alloc);
562 #ifdef CONFIG_STACKTRACE
563 struct stack_trace trace;
566 trace.nr_entries = 0;
567 trace.max_entries = TRACK_ADDRS_COUNT;
568 trace.entries = p->addrs;
570 metadata_access_enable();
571 save_stack_trace(&trace);
572 metadata_access_disable();
574 /* See rant in lockdep.c */
575 if (trace.nr_entries != 0 &&
576 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
579 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
583 p->cpu = smp_processor_id();
584 p->pid = current->pid;
587 memset(p, 0, sizeof(struct track));
590 static void init_tracking(struct kmem_cache *s, void *object)
592 if (!(s->flags & SLAB_STORE_USER))
595 set_track(s, object, TRACK_FREE, 0UL);
596 set_track(s, object, TRACK_ALLOC, 0UL);
599 static void print_track(const char *s, struct track *t)
604 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
605 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
606 #ifdef CONFIG_STACKTRACE
609 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
611 pr_err("\t%pS\n", (void *)t->addrs[i]);
618 static void print_tracking(struct kmem_cache *s, void *object)
620 if (!(s->flags & SLAB_STORE_USER))
623 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
624 print_track("Freed", get_track(s, object, TRACK_FREE));
627 static void print_page_info(struct page *page)
629 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
630 page, page->objects, page->inuse, page->freelist, page->flags);
634 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
636 struct va_format vaf;
642 pr_err("=============================================================================\n");
643 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
644 pr_err("-----------------------------------------------------------------------------\n\n");
646 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
650 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
652 struct va_format vaf;
658 pr_err("FIX %s: %pV\n", s->name, &vaf);
662 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
663 void **freelist, void *nextfree)
665 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
666 !check_valid_pointer(s, page, nextfree) && freelist) {
667 object_err(s, page, *freelist, "Freechain corrupt");
669 slab_fix(s, "Isolate corrupted freechain");
676 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
678 unsigned int off; /* Offset of last byte */
679 u8 *addr = page_address(page);
681 print_tracking(s, p);
683 print_page_info(page);
685 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
686 p, p - addr, get_freepointer(s, p));
688 if (s->flags & SLAB_RED_ZONE)
689 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
691 else if (p > addr + 16)
692 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
694 print_section(KERN_ERR, "Object ", p,
695 min_t(unsigned long, s->object_size, PAGE_SIZE));
696 if (s->flags & SLAB_RED_ZONE)
697 print_section(KERN_ERR, "Redzone ", p + s->object_size,
698 s->inuse - s->object_size);
701 off = s->offset + sizeof(void *);
705 if (s->flags & SLAB_STORE_USER)
706 off += 2 * sizeof(struct track);
708 off += kasan_metadata_size(s);
710 if (off != size_from_object(s))
711 /* Beginning of the filler is the free pointer */
712 print_section(KERN_ERR, "Padding ", p + off,
713 size_from_object(s) - off);
718 void object_err(struct kmem_cache *s, struct page *page,
719 u8 *object, char *reason)
721 slab_bug(s, "%s", reason);
722 print_trailer(s, page, object);
725 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
726 const char *fmt, ...)
732 vsnprintf(buf, sizeof(buf), fmt, args);
734 slab_bug(s, "%s", buf);
735 print_page_info(page);
739 static void init_object(struct kmem_cache *s, void *object, u8 val)
743 if (s->flags & SLAB_RED_ZONE)
744 memset(p - s->red_left_pad, val, s->red_left_pad);
746 if (s->flags & __OBJECT_POISON) {
747 memset(p, POISON_FREE, s->object_size - 1);
748 p[s->object_size - 1] = POISON_END;
751 if (s->flags & SLAB_RED_ZONE)
752 memset(p + s->object_size, val, s->inuse - s->object_size);
755 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
756 void *from, void *to)
758 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
759 memset(from, data, to - from);
762 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
763 u8 *object, char *what,
764 u8 *start, unsigned int value, unsigned int bytes)
769 metadata_access_enable();
770 fault = memchr_inv(start, value, bytes);
771 metadata_access_disable();
776 while (end > fault && end[-1] == value)
779 slab_bug(s, "%s overwritten", what);
780 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
781 fault, end - 1, fault[0], value);
782 print_trailer(s, page, object);
784 restore_bytes(s, what, value, fault, end);
792 * Bytes of the object to be managed.
793 * If the freepointer may overlay the object then the free
794 * pointer is the first word of the object.
796 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
799 * object + s->object_size
800 * Padding to reach word boundary. This is also used for Redzoning.
801 * Padding is extended by another word if Redzoning is enabled and
802 * object_size == inuse.
804 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
805 * 0xcc (RED_ACTIVE) for objects in use.
808 * Meta data starts here.
810 * A. Free pointer (if we cannot overwrite object on free)
811 * B. Tracking data for SLAB_STORE_USER
812 * C. Padding to reach required alignment boundary or at mininum
813 * one word if debugging is on to be able to detect writes
814 * before the word boundary.
816 * Padding is done using 0x5a (POISON_INUSE)
819 * Nothing is used beyond s->size.
821 * If slabcaches are merged then the object_size and inuse boundaries are mostly
822 * ignored. And therefore no slab options that rely on these boundaries
823 * may be used with merged slabcaches.
826 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
828 unsigned long off = s->inuse; /* The end of info */
831 /* Freepointer is placed after the object. */
832 off += sizeof(void *);
834 if (s->flags & SLAB_STORE_USER)
835 /* We also have user information there */
836 off += 2 * sizeof(struct track);
838 off += kasan_metadata_size(s);
840 if (size_from_object(s) == off)
843 return check_bytes_and_report(s, page, p, "Object padding",
844 p + off, POISON_INUSE, size_from_object(s) - off);
847 /* Check the pad bytes at the end of a slab page */
848 static int slab_pad_check(struct kmem_cache *s, struct page *page)
856 if (!(s->flags & SLAB_POISON))
859 start = page_address(page);
860 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
861 end = start + length;
862 remainder = length % s->size;
866 metadata_access_enable();
867 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
868 metadata_access_disable();
871 while (end > fault && end[-1] == POISON_INUSE)
874 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
875 print_section(KERN_ERR, "Padding ", end - remainder, remainder);
877 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
881 static int check_object(struct kmem_cache *s, struct page *page,
882 void *object, u8 val)
885 u8 *endobject = object + s->object_size;
887 if (s->flags & SLAB_RED_ZONE) {
888 if (!check_bytes_and_report(s, page, object, "Redzone",
889 object - s->red_left_pad, val, s->red_left_pad))
892 if (!check_bytes_and_report(s, page, object, "Redzone",
893 endobject, val, s->inuse - s->object_size))
896 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
897 check_bytes_and_report(s, page, p, "Alignment padding",
898 endobject, POISON_INUSE,
899 s->inuse - s->object_size);
903 if (s->flags & SLAB_POISON) {
904 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
905 (!check_bytes_and_report(s, page, p, "Poison", p,
906 POISON_FREE, s->object_size - 1) ||
907 !check_bytes_and_report(s, page, p, "Poison",
908 p + s->object_size - 1, POISON_END, 1)))
911 * check_pad_bytes cleans up on its own.
913 check_pad_bytes(s, page, p);
916 if (!s->offset && val == SLUB_RED_ACTIVE)
918 * Object and freepointer overlap. Cannot check
919 * freepointer while object is allocated.
923 /* Check free pointer validity */
924 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
925 object_err(s, page, p, "Freepointer corrupt");
927 * No choice but to zap it and thus lose the remainder
928 * of the free objects in this slab. May cause
929 * another error because the object count is now wrong.
931 set_freepointer(s, p, NULL);
937 static int check_slab(struct kmem_cache *s, struct page *page)
941 VM_BUG_ON(!irqs_disabled());
943 if (!PageSlab(page)) {
944 slab_err(s, page, "Not a valid slab page");
948 maxobj = order_objects(compound_order(page), s->size, s->reserved);
949 if (page->objects > maxobj) {
950 slab_err(s, page, "objects %u > max %u",
951 page->objects, maxobj);
954 if (page->inuse > page->objects) {
955 slab_err(s, page, "inuse %u > max %u",
956 page->inuse, page->objects);
959 /* Slab_pad_check fixes things up after itself */
960 slab_pad_check(s, page);
965 * Determine if a certain object on a page is on the freelist. Must hold the
966 * slab lock to guarantee that the chains are in a consistent state.
968 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
976 while (fp && nr <= page->objects) {
979 if (!check_valid_pointer(s, page, fp)) {
981 object_err(s, page, object,
982 "Freechain corrupt");
983 set_freepointer(s, object, NULL);
985 slab_err(s, page, "Freepointer corrupt");
986 page->freelist = NULL;
987 page->inuse = page->objects;
988 slab_fix(s, "Freelist cleared");
994 fp = get_freepointer(s, object);
998 max_objects = order_objects(compound_order(page), s->size, s->reserved);
999 if (max_objects > MAX_OBJS_PER_PAGE)
1000 max_objects = MAX_OBJS_PER_PAGE;
1002 if (page->objects != max_objects) {
1003 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
1004 page->objects, max_objects);
1005 page->objects = max_objects;
1006 slab_fix(s, "Number of objects adjusted.");
1008 if (page->inuse != page->objects - nr) {
1009 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1010 page->inuse, page->objects - nr);
1011 page->inuse = page->objects - nr;
1012 slab_fix(s, "Object count adjusted.");
1014 return search == NULL;
1017 static void trace(struct kmem_cache *s, struct page *page, void *object,
1020 if (s->flags & SLAB_TRACE) {
1021 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1023 alloc ? "alloc" : "free",
1024 object, page->inuse,
1028 print_section(KERN_INFO, "Object ", (void *)object,
1036 * Tracking of fully allocated slabs for debugging purposes.
1038 static void add_full(struct kmem_cache *s,
1039 struct kmem_cache_node *n, struct page *page)
1041 if (!(s->flags & SLAB_STORE_USER))
1044 lockdep_assert_held(&n->list_lock);
1045 list_add(&page->lru, &n->full);
1048 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1050 if (!(s->flags & SLAB_STORE_USER))
1053 lockdep_assert_held(&n->list_lock);
1054 list_del(&page->lru);
1057 /* Tracking of the number of slabs for debugging purposes */
1058 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1060 struct kmem_cache_node *n = get_node(s, node);
1062 return atomic_long_read(&n->nr_slabs);
1065 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1067 return atomic_long_read(&n->nr_slabs);
1070 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1072 struct kmem_cache_node *n = get_node(s, node);
1075 * May be called early in order to allocate a slab for the
1076 * kmem_cache_node structure. Solve the chicken-egg
1077 * dilemma by deferring the increment of the count during
1078 * bootstrap (see early_kmem_cache_node_alloc).
1081 atomic_long_inc(&n->nr_slabs);
1082 atomic_long_add(objects, &n->total_objects);
1085 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1087 struct kmem_cache_node *n = get_node(s, node);
1089 atomic_long_dec(&n->nr_slabs);
1090 atomic_long_sub(objects, &n->total_objects);
1093 /* Object debug checks for alloc/free paths */
1094 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1097 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1100 init_object(s, object, SLUB_RED_INACTIVE);
1101 init_tracking(s, object);
1104 static inline int alloc_consistency_checks(struct kmem_cache *s,
1106 void *object, unsigned long addr)
1108 if (!check_slab(s, page))
1111 if (!check_valid_pointer(s, page, object)) {
1112 object_err(s, page, object, "Freelist Pointer check fails");
1116 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1122 static noinline int alloc_debug_processing(struct kmem_cache *s,
1124 void *object, unsigned long addr)
1126 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1127 if (!alloc_consistency_checks(s, page, object, addr))
1131 /* Success perform special debug activities for allocs */
1132 if (s->flags & SLAB_STORE_USER)
1133 set_track(s, object, TRACK_ALLOC, addr);
1134 trace(s, page, object, 1);
1135 init_object(s, object, SLUB_RED_ACTIVE);
1139 if (PageSlab(page)) {
1141 * If this is a slab page then lets do the best we can
1142 * to avoid issues in the future. Marking all objects
1143 * as used avoids touching the remaining objects.
1145 slab_fix(s, "Marking all objects used");
1146 page->inuse = page->objects;
1147 page->freelist = NULL;
1152 static inline int free_consistency_checks(struct kmem_cache *s,
1153 struct page *page, void *object, unsigned long addr)
1155 if (!check_valid_pointer(s, page, object)) {
1156 slab_err(s, page, "Invalid object pointer 0x%p", object);
1160 if (on_freelist(s, page, object)) {
1161 object_err(s, page, object, "Object already free");
1165 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1168 if (unlikely(s != page->slab_cache)) {
1169 if (!PageSlab(page)) {
1170 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1172 } else if (!page->slab_cache) {
1173 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1177 object_err(s, page, object,
1178 "page slab pointer corrupt.");
1184 /* Supports checking bulk free of a constructed freelist */
1185 static noinline int free_debug_processing(
1186 struct kmem_cache *s, struct page *page,
1187 void *head, void *tail, int bulk_cnt,
1190 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1191 void *object = head;
1193 unsigned long uninitialized_var(flags);
1196 spin_lock_irqsave(&n->list_lock, flags);
1199 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1200 if (!check_slab(s, page))
1207 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1208 if (!free_consistency_checks(s, page, object, addr))
1212 if (s->flags & SLAB_STORE_USER)
1213 set_track(s, object, TRACK_FREE, addr);
1214 trace(s, page, object, 0);
1215 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1216 init_object(s, object, SLUB_RED_INACTIVE);
1218 /* Reached end of constructed freelist yet? */
1219 if (object != tail) {
1220 object = get_freepointer(s, object);
1226 if (cnt != bulk_cnt)
1227 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1231 spin_unlock_irqrestore(&n->list_lock, flags);
1233 slab_fix(s, "Object at 0x%p not freed", object);
1237 static int __init setup_slub_debug(char *str)
1239 slub_debug = DEBUG_DEFAULT_FLAGS;
1240 if (*str++ != '=' || !*str)
1242 * No options specified. Switch on full debugging.
1248 * No options but restriction on slabs. This means full
1249 * debugging for slabs matching a pattern.
1256 * Switch off all debugging measures.
1261 * Determine which debug features should be switched on
1263 for (; *str && *str != ','; str++) {
1264 switch (tolower(*str)) {
1266 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1269 slub_debug |= SLAB_RED_ZONE;
1272 slub_debug |= SLAB_POISON;
1275 slub_debug |= SLAB_STORE_USER;
1278 slub_debug |= SLAB_TRACE;
1281 slub_debug |= SLAB_FAILSLAB;
1285 * Avoid enabling debugging on caches if its minimum
1286 * order would increase as a result.
1288 disable_higher_order_debug = 1;
1291 pr_err("slub_debug option '%c' unknown. skipped\n",
1298 slub_debug_slabs = str + 1;
1303 __setup("slub_debug", setup_slub_debug);
1305 unsigned long kmem_cache_flags(unsigned long object_size,
1306 unsigned long flags, const char *name,
1307 void (*ctor)(void *))
1310 * Enable debugging if selected on the kernel commandline.
1312 if (slub_debug && (!slub_debug_slabs || (name &&
1313 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1314 flags |= slub_debug;
1318 #else /* !CONFIG_SLUB_DEBUG */
1319 static inline void setup_object_debug(struct kmem_cache *s,
1320 struct page *page, void *object) {}
1322 static inline int alloc_debug_processing(struct kmem_cache *s,
1323 struct page *page, void *object, unsigned long addr) { return 0; }
1325 static inline int free_debug_processing(
1326 struct kmem_cache *s, struct page *page,
1327 void *head, void *tail, int bulk_cnt,
1328 unsigned long addr) { return 0; }
1330 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1332 static inline int check_object(struct kmem_cache *s, struct page *page,
1333 void *object, u8 val) { return 1; }
1334 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1335 struct page *page) {}
1336 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1337 struct page *page) {}
1338 unsigned long kmem_cache_flags(unsigned long object_size,
1339 unsigned long flags, const char *name,
1340 void (*ctor)(void *))
1344 #define slub_debug 0
1346 #define disable_higher_order_debug 0
1348 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1350 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1352 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1354 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1357 static bool freelist_corrupted(struct kmem_cache *s, struct page *page,
1358 void **freelist, void *nextfree)
1362 #endif /* CONFIG_SLUB_DEBUG */
1365 * Hooks for other subsystems that check memory allocations. In a typical
1366 * production configuration these hooks all should produce no code at all.
1368 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1370 kmemleak_alloc(ptr, size, 1, flags);
1371 kasan_kmalloc_large(ptr, size, flags);
1374 static inline void kfree_hook(const void *x)
1377 kasan_kfree_large(x);
1380 static inline void *slab_free_hook(struct kmem_cache *s, void *x)
1384 kmemleak_free_recursive(x, s->flags);
1387 * Trouble is that we may no longer disable interrupts in the fast path
1388 * So in order to make the debug calls that expect irqs to be
1389 * disabled we need to disable interrupts temporarily.
1391 #ifdef CONFIG_LOCKDEP
1393 unsigned long flags;
1395 local_irq_save(flags);
1396 debug_check_no_locks_freed(x, s->object_size);
1397 local_irq_restore(flags);
1400 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1401 debug_check_no_obj_freed(x, s->object_size);
1403 freeptr = get_freepointer(s, x);
1405 * kasan_slab_free() may put x into memory quarantine, delaying its
1406 * reuse. In this case the object's freelist pointer is changed.
1408 kasan_slab_free(s, x);
1412 static inline void slab_free_freelist_hook(struct kmem_cache *s,
1413 void *head, void *tail)
1416 * Compiler cannot detect this function can be removed if slab_free_hook()
1417 * evaluates to nothing. Thus, catch all relevant config debug options here.
1419 #if defined(CONFIG_LOCKDEP) || \
1420 defined(CONFIG_DEBUG_KMEMLEAK) || \
1421 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1422 defined(CONFIG_KASAN)
1424 void *object = head;
1425 void *tail_obj = tail ? : head;
1429 freeptr = slab_free_hook(s, object);
1430 } while ((object != tail_obj) && (object = freeptr));
1434 static void setup_object(struct kmem_cache *s, struct page *page,
1437 setup_object_debug(s, page, object);
1438 kasan_init_slab_obj(s, object);
1439 if (unlikely(s->ctor)) {
1440 kasan_unpoison_object_data(s, object);
1442 kasan_poison_object_data(s, object);
1447 * Slab allocation and freeing
1449 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1450 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1453 int order = oo_order(oo);
1455 if (node == NUMA_NO_NODE)
1456 page = alloc_pages(flags, order);
1458 page = __alloc_pages_node(node, flags, order);
1460 if (page && memcg_charge_slab(page, flags, order, s)) {
1461 __free_pages(page, order);
1468 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1469 /* Pre-initialize the random sequence cache */
1470 static int init_cache_random_seq(struct kmem_cache *s)
1473 unsigned long i, count = oo_objects(s->oo);
1475 /* Bailout if already initialised */
1479 err = cache_random_seq_create(s, count, GFP_KERNEL);
1481 pr_err("SLUB: Unable to initialize free list for %s\n",
1486 /* Transform to an offset on the set of pages */
1487 if (s->random_seq) {
1488 for (i = 0; i < count; i++)
1489 s->random_seq[i] *= s->size;
1494 /* Initialize each random sequence freelist per cache */
1495 static void __init init_freelist_randomization(void)
1497 struct kmem_cache *s;
1499 mutex_lock(&slab_mutex);
1501 list_for_each_entry(s, &slab_caches, list)
1502 init_cache_random_seq(s);
1504 mutex_unlock(&slab_mutex);
1507 /* Get the next entry on the pre-computed freelist randomized */
1508 static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1509 unsigned long *pos, void *start,
1510 unsigned long page_limit,
1511 unsigned long freelist_count)
1516 * If the target page allocation failed, the number of objects on the
1517 * page might be smaller than the usual size defined by the cache.
1520 idx = s->random_seq[*pos];
1522 if (*pos >= freelist_count)
1524 } while (unlikely(idx >= page_limit));
1526 return (char *)start + idx;
1529 /* Shuffle the single linked freelist based on a random pre-computed sequence */
1530 static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1535 unsigned long idx, pos, page_limit, freelist_count;
1537 if (page->objects < 2 || !s->random_seq)
1540 freelist_count = oo_objects(s->oo);
1541 pos = get_random_int() % freelist_count;
1543 page_limit = page->objects * s->size;
1544 start = fixup_red_left(s, page_address(page));
1546 /* First entry is used as the base of the freelist */
1547 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1549 page->freelist = cur;
1551 for (idx = 1; idx < page->objects; idx++) {
1552 setup_object(s, page, cur);
1553 next = next_freelist_entry(s, page, &pos, start, page_limit,
1555 set_freepointer(s, cur, next);
1558 setup_object(s, page, cur);
1559 set_freepointer(s, cur, NULL);
1564 static inline int init_cache_random_seq(struct kmem_cache *s)
1568 static inline void init_freelist_randomization(void) { }
1569 static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1573 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1575 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1578 struct kmem_cache_order_objects oo = s->oo;
1584 flags &= gfp_allowed_mask;
1586 if (gfpflags_allow_blocking(flags))
1589 flags |= s->allocflags;
1592 * Let the initial higher-order allocation fail under memory pressure
1593 * so we fall-back to the minimum order allocation.
1595 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1596 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1597 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1599 page = alloc_slab_page(s, alloc_gfp, node, oo);
1600 if (unlikely(!page)) {
1604 * Allocation may have failed due to fragmentation.
1605 * Try a lower order alloc if possible
1607 page = alloc_slab_page(s, alloc_gfp, node, oo);
1608 if (unlikely(!page))
1610 stat(s, ORDER_FALLBACK);
1613 page->objects = oo_objects(oo);
1615 order = compound_order(page);
1616 page->slab_cache = s;
1617 __SetPageSlab(page);
1618 if (page_is_pfmemalloc(page))
1619 SetPageSlabPfmemalloc(page);
1621 start = page_address(page);
1623 if (unlikely(s->flags & SLAB_POISON))
1624 memset(start, POISON_INUSE, PAGE_SIZE << order);
1626 kasan_poison_slab(page);
1628 shuffle = shuffle_freelist(s, page);
1631 for_each_object_idx(p, idx, s, start, page->objects) {
1632 setup_object(s, page, p);
1633 if (likely(idx < page->objects))
1634 set_freepointer(s, p, p + s->size);
1636 set_freepointer(s, p, NULL);
1638 page->freelist = fixup_red_left(s, start);
1641 page->inuse = page->objects;
1645 if (gfpflags_allow_blocking(flags))
1646 local_irq_disable();
1650 mod_lruvec_page_state(page,
1651 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1652 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1655 inc_slabs_node(s, page_to_nid(page), page->objects);
1660 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1662 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1663 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1664 flags &= ~GFP_SLAB_BUG_MASK;
1665 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1666 invalid_mask, &invalid_mask, flags, &flags);
1670 return allocate_slab(s,
1671 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1674 static void __free_slab(struct kmem_cache *s, struct page *page)
1676 int order = compound_order(page);
1677 int pages = 1 << order;
1679 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1682 slab_pad_check(s, page);
1683 for_each_object(p, s, page_address(page),
1685 check_object(s, page, p, SLUB_RED_INACTIVE);
1688 mod_lruvec_page_state(page,
1689 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1690 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1693 __ClearPageSlabPfmemalloc(page);
1694 __ClearPageSlab(page);
1696 page_mapcount_reset(page);
1697 if (current->reclaim_state)
1698 current->reclaim_state->reclaimed_slab += pages;
1699 memcg_uncharge_slab(page, order, s);
1700 __free_pages(page, order);
1703 #define need_reserve_slab_rcu \
1704 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1706 static void rcu_free_slab(struct rcu_head *h)
1710 if (need_reserve_slab_rcu)
1711 page = virt_to_head_page(h);
1713 page = container_of((struct list_head *)h, struct page, lru);
1715 __free_slab(page->slab_cache, page);
1718 static void free_slab(struct kmem_cache *s, struct page *page)
1720 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1721 struct rcu_head *head;
1723 if (need_reserve_slab_rcu) {
1724 int order = compound_order(page);
1725 int offset = (PAGE_SIZE << order) - s->reserved;
1727 VM_BUG_ON(s->reserved != sizeof(*head));
1728 head = page_address(page) + offset;
1730 head = &page->rcu_head;
1733 call_rcu(head, rcu_free_slab);
1735 __free_slab(s, page);
1738 static void discard_slab(struct kmem_cache *s, struct page *page)
1740 dec_slabs_node(s, page_to_nid(page), page->objects);
1745 * Management of partially allocated slabs.
1748 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1751 if (tail == DEACTIVATE_TO_TAIL)
1752 list_add_tail(&page->lru, &n->partial);
1754 list_add(&page->lru, &n->partial);
1757 static inline void add_partial(struct kmem_cache_node *n,
1758 struct page *page, int tail)
1760 lockdep_assert_held(&n->list_lock);
1761 __add_partial(n, page, tail);
1764 static inline void remove_partial(struct kmem_cache_node *n,
1767 lockdep_assert_held(&n->list_lock);
1768 list_del(&page->lru);
1773 * Remove slab from the partial list, freeze it and
1774 * return the pointer to the freelist.
1776 * Returns a list of objects or NULL if it fails.
1778 static inline void *acquire_slab(struct kmem_cache *s,
1779 struct kmem_cache_node *n, struct page *page,
1780 int mode, int *objects)
1783 unsigned long counters;
1786 lockdep_assert_held(&n->list_lock);
1789 * Zap the freelist and set the frozen bit.
1790 * The old freelist is the list of objects for the
1791 * per cpu allocation list.
1793 freelist = page->freelist;
1794 counters = page->counters;
1795 new.counters = counters;
1796 *objects = new.objects - new.inuse;
1798 new.inuse = page->objects;
1799 new.freelist = NULL;
1801 new.freelist = freelist;
1804 VM_BUG_ON(new.frozen);
1807 if (!__cmpxchg_double_slab(s, page,
1809 new.freelist, new.counters,
1813 remove_partial(n, page);
1818 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1819 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1822 * Try to allocate a partial slab from a specific node.
1824 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1825 struct kmem_cache_cpu *c, gfp_t flags)
1827 struct page *page, *page2;
1828 void *object = NULL;
1829 unsigned int available = 0;
1833 * Racy check. If we mistakenly see no partial slabs then we
1834 * just allocate an empty slab. If we mistakenly try to get a
1835 * partial slab and there is none available then get_partials()
1838 if (!n || !n->nr_partial)
1841 spin_lock(&n->list_lock);
1842 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1845 if (!pfmemalloc_match(page, flags))
1848 t = acquire_slab(s, n, page, object == NULL, &objects);
1852 available += objects;
1855 stat(s, ALLOC_FROM_PARTIAL);
1858 put_cpu_partial(s, page, 0);
1859 stat(s, CPU_PARTIAL_NODE);
1861 if (!kmem_cache_has_cpu_partial(s)
1862 || available > slub_cpu_partial(s) / 2)
1866 spin_unlock(&n->list_lock);
1871 * Get a page from somewhere. Search in increasing NUMA distances.
1873 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1874 struct kmem_cache_cpu *c)
1877 struct zonelist *zonelist;
1880 enum zone_type high_zoneidx = gfp_zone(flags);
1882 unsigned int cpuset_mems_cookie;
1885 * The defrag ratio allows a configuration of the tradeoffs between
1886 * inter node defragmentation and node local allocations. A lower
1887 * defrag_ratio increases the tendency to do local allocations
1888 * instead of attempting to obtain partial slabs from other nodes.
1890 * If the defrag_ratio is set to 0 then kmalloc() always
1891 * returns node local objects. If the ratio is higher then kmalloc()
1892 * may return off node objects because partial slabs are obtained
1893 * from other nodes and filled up.
1895 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1896 * (which makes defrag_ratio = 1000) then every (well almost)
1897 * allocation will first attempt to defrag slab caches on other nodes.
1898 * This means scanning over all nodes to look for partial slabs which
1899 * may be expensive if we do it every time we are trying to find a slab
1900 * with available objects.
1902 if (!s->remote_node_defrag_ratio ||
1903 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1907 cpuset_mems_cookie = read_mems_allowed_begin();
1908 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1909 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1910 struct kmem_cache_node *n;
1912 n = get_node(s, zone_to_nid(zone));
1914 if (n && cpuset_zone_allowed(zone, flags) &&
1915 n->nr_partial > s->min_partial) {
1916 object = get_partial_node(s, n, c, flags);
1919 * Don't check read_mems_allowed_retry()
1920 * here - if mems_allowed was updated in
1921 * parallel, that was a harmless race
1922 * between allocation and the cpuset
1929 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1935 * Get a partial page, lock it and return it.
1937 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1938 struct kmem_cache_cpu *c)
1941 int searchnode = node;
1943 if (node == NUMA_NO_NODE)
1944 searchnode = numa_mem_id();
1946 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1947 if (object || node != NUMA_NO_NODE)
1950 return get_any_partial(s, flags, c);
1953 #ifdef CONFIG_PREEMPT
1955 * Calculate the next globally unique transaction for disambiguiation
1956 * during cmpxchg. The transactions start with the cpu number and are then
1957 * incremented by CONFIG_NR_CPUS.
1959 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1962 * No preemption supported therefore also no need to check for
1968 static inline unsigned long next_tid(unsigned long tid)
1970 return tid + TID_STEP;
1973 static inline unsigned int tid_to_cpu(unsigned long tid)
1975 return tid % TID_STEP;
1978 static inline unsigned long tid_to_event(unsigned long tid)
1980 return tid / TID_STEP;
1983 static inline unsigned int init_tid(int cpu)
1988 static inline void note_cmpxchg_failure(const char *n,
1989 const struct kmem_cache *s, unsigned long tid)
1991 #ifdef SLUB_DEBUG_CMPXCHG
1992 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1994 pr_info("%s %s: cmpxchg redo ", n, s->name);
1996 #ifdef CONFIG_PREEMPT
1997 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1998 pr_warn("due to cpu change %d -> %d\n",
1999 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2002 if (tid_to_event(tid) != tid_to_event(actual_tid))
2003 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2004 tid_to_event(tid), tid_to_event(actual_tid));
2006 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2007 actual_tid, tid, next_tid(tid));
2009 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2012 static void init_kmem_cache_cpus(struct kmem_cache *s)
2016 for_each_possible_cpu(cpu)
2017 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2021 * Remove the cpu slab
2023 static void deactivate_slab(struct kmem_cache *s, struct page *page,
2024 void *freelist, struct kmem_cache_cpu *c)
2026 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2027 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2029 enum slab_modes l = M_NONE, m = M_NONE;
2031 int tail = DEACTIVATE_TO_HEAD;
2035 if (page->freelist) {
2036 stat(s, DEACTIVATE_REMOTE_FREES);
2037 tail = DEACTIVATE_TO_TAIL;
2041 * Stage one: Free all available per cpu objects back
2042 * to the page freelist while it is still frozen. Leave the
2045 * There is no need to take the list->lock because the page
2048 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2050 unsigned long counters;
2053 * If 'nextfree' is invalid, it is possible that the object at
2054 * 'freelist' is already corrupted. So isolate all objects
2055 * starting at 'freelist'.
2057 if (freelist_corrupted(s, page, &freelist, nextfree))
2061 prior = page->freelist;
2062 counters = page->counters;
2063 set_freepointer(s, freelist, prior);
2064 new.counters = counters;
2066 VM_BUG_ON(!new.frozen);
2068 } while (!__cmpxchg_double_slab(s, page,
2070 freelist, new.counters,
2071 "drain percpu freelist"));
2073 freelist = nextfree;
2077 * Stage two: Ensure that the page is unfrozen while the
2078 * list presence reflects the actual number of objects
2081 * We setup the list membership and then perform a cmpxchg
2082 * with the count. If there is a mismatch then the page
2083 * is not unfrozen but the page is on the wrong list.
2085 * Then we restart the process which may have to remove
2086 * the page from the list that we just put it on again
2087 * because the number of objects in the slab may have
2092 old.freelist = page->freelist;
2093 old.counters = page->counters;
2094 VM_BUG_ON(!old.frozen);
2096 /* Determine target state of the slab */
2097 new.counters = old.counters;
2100 set_freepointer(s, freelist, old.freelist);
2101 new.freelist = freelist;
2103 new.freelist = old.freelist;
2107 if (!new.inuse && n->nr_partial >= s->min_partial)
2109 else if (new.freelist) {
2114 * Taking the spinlock removes the possiblity
2115 * that acquire_slab() will see a slab page that
2118 spin_lock(&n->list_lock);
2122 if (kmem_cache_debug(s) && !lock) {
2125 * This also ensures that the scanning of full
2126 * slabs from diagnostic functions will not see
2129 spin_lock(&n->list_lock);
2137 remove_partial(n, page);
2139 else if (l == M_FULL)
2141 remove_full(s, n, page);
2143 if (m == M_PARTIAL) {
2145 add_partial(n, page, tail);
2148 } else if (m == M_FULL) {
2150 stat(s, DEACTIVATE_FULL);
2151 add_full(s, n, page);
2157 if (!__cmpxchg_double_slab(s, page,
2158 old.freelist, old.counters,
2159 new.freelist, new.counters,
2164 spin_unlock(&n->list_lock);
2167 stat(s, DEACTIVATE_EMPTY);
2168 discard_slab(s, page);
2177 * Unfreeze all the cpu partial slabs.
2179 * This function must be called with interrupts disabled
2180 * for the cpu using c (or some other guarantee must be there
2181 * to guarantee no concurrent accesses).
2183 static void unfreeze_partials(struct kmem_cache *s,
2184 struct kmem_cache_cpu *c)
2186 #ifdef CONFIG_SLUB_CPU_PARTIAL
2187 struct kmem_cache_node *n = NULL, *n2 = NULL;
2188 struct page *page, *discard_page = NULL;
2190 while ((page = c->partial)) {
2194 c->partial = page->next;
2196 n2 = get_node(s, page_to_nid(page));
2199 spin_unlock(&n->list_lock);
2202 spin_lock(&n->list_lock);
2207 old.freelist = page->freelist;
2208 old.counters = page->counters;
2209 VM_BUG_ON(!old.frozen);
2211 new.counters = old.counters;
2212 new.freelist = old.freelist;
2216 } while (!__cmpxchg_double_slab(s, page,
2217 old.freelist, old.counters,
2218 new.freelist, new.counters,
2219 "unfreezing slab"));
2221 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2222 page->next = discard_page;
2223 discard_page = page;
2225 add_partial(n, page, DEACTIVATE_TO_TAIL);
2226 stat(s, FREE_ADD_PARTIAL);
2231 spin_unlock(&n->list_lock);
2233 while (discard_page) {
2234 page = discard_page;
2235 discard_page = discard_page->next;
2237 stat(s, DEACTIVATE_EMPTY);
2238 discard_slab(s, page);
2245 * Put a page that was just frozen (in __slab_free) into a partial page
2246 * slot if available. This is done without interrupts disabled and without
2247 * preemption disabled. The cmpxchg is racy and may put the partial page
2248 * onto a random cpus partial slot.
2250 * If we did not find a slot then simply move all the partials to the
2251 * per node partial list.
2253 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2255 #ifdef CONFIG_SLUB_CPU_PARTIAL
2256 struct page *oldpage;
2264 oldpage = this_cpu_read(s->cpu_slab->partial);
2267 pobjects = oldpage->pobjects;
2268 pages = oldpage->pages;
2269 if (drain && pobjects > s->cpu_partial) {
2270 unsigned long flags;
2272 * partial array is full. Move the existing
2273 * set to the per node partial list.
2275 local_irq_save(flags);
2276 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2277 local_irq_restore(flags);
2281 stat(s, CPU_PARTIAL_DRAIN);
2286 pobjects += page->objects - page->inuse;
2288 page->pages = pages;
2289 page->pobjects = pobjects;
2290 page->next = oldpage;
2292 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2294 if (unlikely(!s->cpu_partial)) {
2295 unsigned long flags;
2297 local_irq_save(flags);
2298 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2299 local_irq_restore(flags);
2305 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2307 stat(s, CPUSLAB_FLUSH);
2308 deactivate_slab(s, c->page, c->freelist, c);
2310 c->tid = next_tid(c->tid);
2316 * Called from IPI handler with interrupts disabled.
2318 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2320 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2326 unfreeze_partials(s, c);
2330 static void flush_cpu_slab(void *d)
2332 struct kmem_cache *s = d;
2334 __flush_cpu_slab(s, smp_processor_id());
2337 static bool has_cpu_slab(int cpu, void *info)
2339 struct kmem_cache *s = info;
2340 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2342 return c->page || slub_percpu_partial(c);
2345 static void flush_all(struct kmem_cache *s)
2347 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2351 * Use the cpu notifier to insure that the cpu slabs are flushed when
2354 static int slub_cpu_dead(unsigned int cpu)
2356 struct kmem_cache *s;
2357 unsigned long flags;
2359 mutex_lock(&slab_mutex);
2360 list_for_each_entry(s, &slab_caches, list) {
2361 local_irq_save(flags);
2362 __flush_cpu_slab(s, cpu);
2363 local_irq_restore(flags);
2365 mutex_unlock(&slab_mutex);
2370 * Check if the objects in a per cpu structure fit numa
2371 * locality expectations.
2373 static inline int node_match(struct page *page, int node)
2376 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2382 #ifdef CONFIG_SLUB_DEBUG
2383 static int count_free(struct page *page)
2385 return page->objects - page->inuse;
2388 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2390 return atomic_long_read(&n->total_objects);
2392 #endif /* CONFIG_SLUB_DEBUG */
2394 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2395 static unsigned long count_partial(struct kmem_cache_node *n,
2396 int (*get_count)(struct page *))
2398 unsigned long flags;
2399 unsigned long x = 0;
2402 spin_lock_irqsave(&n->list_lock, flags);
2403 list_for_each_entry(page, &n->partial, lru)
2404 x += get_count(page);
2405 spin_unlock_irqrestore(&n->list_lock, flags);
2408 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2410 static noinline void
2411 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2413 #ifdef CONFIG_SLUB_DEBUG
2414 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2415 DEFAULT_RATELIMIT_BURST);
2417 struct kmem_cache_node *n;
2419 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2422 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2423 nid, gfpflags, &gfpflags);
2424 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2425 s->name, s->object_size, s->size, oo_order(s->oo),
2428 if (oo_order(s->min) > get_order(s->object_size))
2429 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2432 for_each_kmem_cache_node(s, node, n) {
2433 unsigned long nr_slabs;
2434 unsigned long nr_objs;
2435 unsigned long nr_free;
2437 nr_free = count_partial(n, count_free);
2438 nr_slabs = node_nr_slabs(n);
2439 nr_objs = node_nr_objs(n);
2441 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2442 node, nr_slabs, nr_objs, nr_free);
2447 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2448 int node, struct kmem_cache_cpu **pc)
2451 struct kmem_cache_cpu *c = *pc;
2454 freelist = get_partial(s, flags, node, c);
2459 page = new_slab(s, flags, node);
2461 c = raw_cpu_ptr(s->cpu_slab);
2466 * No other reference to the page yet so we can
2467 * muck around with it freely without cmpxchg
2469 freelist = page->freelist;
2470 page->freelist = NULL;
2472 stat(s, ALLOC_SLAB);
2481 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2483 if (unlikely(PageSlabPfmemalloc(page)))
2484 return gfp_pfmemalloc_allowed(gfpflags);
2490 * Check the page->freelist of a page and either transfer the freelist to the
2491 * per cpu freelist or deactivate the page.
2493 * The page is still frozen if the return value is not NULL.
2495 * If this function returns NULL then the page has been unfrozen.
2497 * This function must be called with interrupt disabled.
2499 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2502 unsigned long counters;
2506 freelist = page->freelist;
2507 counters = page->counters;
2509 new.counters = counters;
2510 VM_BUG_ON(!new.frozen);
2512 new.inuse = page->objects;
2513 new.frozen = freelist != NULL;
2515 } while (!__cmpxchg_double_slab(s, page,
2524 * Slow path. The lockless freelist is empty or we need to perform
2527 * Processing is still very fast if new objects have been freed to the
2528 * regular freelist. In that case we simply take over the regular freelist
2529 * as the lockless freelist and zap the regular freelist.
2531 * If that is not working then we fall back to the partial lists. We take the
2532 * first element of the freelist as the object to allocate now and move the
2533 * rest of the freelist to the lockless freelist.
2535 * And if we were unable to get a new slab from the partial slab lists then
2536 * we need to allocate a new slab. This is the slowest path since it involves
2537 * a call to the page allocator and the setup of a new slab.
2539 * Version of __slab_alloc to use when we know that interrupts are
2540 * already disabled (which is the case for bulk allocation).
2542 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2543 unsigned long addr, struct kmem_cache_cpu *c)
2551 * if the node is not online or has no normal memory, just
2552 * ignore the node constraint
2554 if (unlikely(node != NUMA_NO_NODE &&
2555 !node_state(node, N_NORMAL_MEMORY)))
2556 node = NUMA_NO_NODE;
2561 if (unlikely(!node_match(page, node))) {
2563 * same as above but node_match() being false already
2564 * implies node != NUMA_NO_NODE
2566 if (!node_state(node, N_NORMAL_MEMORY)) {
2567 node = NUMA_NO_NODE;
2570 stat(s, ALLOC_NODE_MISMATCH);
2571 deactivate_slab(s, page, c->freelist, c);
2577 * By rights, we should be searching for a slab page that was
2578 * PFMEMALLOC but right now, we are losing the pfmemalloc
2579 * information when the page leaves the per-cpu allocator
2581 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2582 deactivate_slab(s, page, c->freelist, c);
2586 /* must check again c->freelist in case of cpu migration or IRQ */
2587 freelist = c->freelist;
2591 freelist = get_freelist(s, page);
2595 stat(s, DEACTIVATE_BYPASS);
2599 stat(s, ALLOC_REFILL);
2603 * freelist is pointing to the list of objects to be used.
2604 * page is pointing to the page from which the objects are obtained.
2605 * That page must be frozen for per cpu allocations to work.
2607 VM_BUG_ON(!c->page->frozen);
2608 c->freelist = get_freepointer(s, freelist);
2609 c->tid = next_tid(c->tid);
2614 if (slub_percpu_partial(c)) {
2615 page = c->page = slub_percpu_partial(c);
2616 slub_set_percpu_partial(c, page);
2617 stat(s, CPU_PARTIAL_ALLOC);
2621 freelist = new_slab_objects(s, gfpflags, node, &c);
2623 if (unlikely(!freelist)) {
2624 slab_out_of_memory(s, gfpflags, node);
2629 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2632 /* Only entered in the debug case */
2633 if (kmem_cache_debug(s) &&
2634 !alloc_debug_processing(s, page, freelist, addr))
2635 goto new_slab; /* Slab failed checks. Next slab needed */
2637 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2642 * Another one that disabled interrupt and compensates for possible
2643 * cpu changes by refetching the per cpu area pointer.
2645 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2646 unsigned long addr, struct kmem_cache_cpu *c)
2649 unsigned long flags;
2651 local_irq_save(flags);
2652 #ifdef CONFIG_PREEMPT
2654 * We may have been preempted and rescheduled on a different
2655 * cpu before disabling interrupts. Need to reload cpu area
2658 c = this_cpu_ptr(s->cpu_slab);
2661 p = ___slab_alloc(s, gfpflags, node, addr, c);
2662 local_irq_restore(flags);
2667 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2668 * have the fastpath folded into their functions. So no function call
2669 * overhead for requests that can be satisfied on the fastpath.
2671 * The fastpath works by first checking if the lockless freelist can be used.
2672 * If not then __slab_alloc is called for slow processing.
2674 * Otherwise we can simply pick the next object from the lockless free list.
2676 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2677 gfp_t gfpflags, int node, unsigned long addr)
2680 struct kmem_cache_cpu *c;
2684 s = slab_pre_alloc_hook(s, gfpflags);
2689 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2690 * enabled. We may switch back and forth between cpus while
2691 * reading from one cpu area. That does not matter as long
2692 * as we end up on the original cpu again when doing the cmpxchg.
2694 * We should guarantee that tid and kmem_cache are retrieved on
2695 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2696 * to check if it is matched or not.
2699 tid = this_cpu_read(s->cpu_slab->tid);
2700 c = raw_cpu_ptr(s->cpu_slab);
2701 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2702 unlikely(tid != READ_ONCE(c->tid)));
2705 * Irqless object alloc/free algorithm used here depends on sequence
2706 * of fetching cpu_slab's data. tid should be fetched before anything
2707 * on c to guarantee that object and page associated with previous tid
2708 * won't be used with current tid. If we fetch tid first, object and
2709 * page could be one associated with next tid and our alloc/free
2710 * request will be failed. In this case, we will retry. So, no problem.
2715 * The transaction ids are globally unique per cpu and per operation on
2716 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2717 * occurs on the right processor and that there was no operation on the
2718 * linked list in between.
2721 object = c->freelist;
2723 if (unlikely(!object || !node_match(page, node))) {
2724 object = __slab_alloc(s, gfpflags, node, addr, c);
2725 stat(s, ALLOC_SLOWPATH);
2727 void *next_object = get_freepointer_safe(s, object);
2730 * The cmpxchg will only match if there was no additional
2731 * operation and if we are on the right processor.
2733 * The cmpxchg does the following atomically (without lock
2735 * 1. Relocate first pointer to the current per cpu area.
2736 * 2. Verify that tid and freelist have not been changed
2737 * 3. If they were not changed replace tid and freelist
2739 * Since this is without lock semantics the protection is only
2740 * against code executing on this cpu *not* from access by
2743 if (unlikely(!this_cpu_cmpxchg_double(
2744 s->cpu_slab->freelist, s->cpu_slab->tid,
2746 next_object, next_tid(tid)))) {
2748 note_cmpxchg_failure("slab_alloc", s, tid);
2751 prefetch_freepointer(s, next_object);
2752 stat(s, ALLOC_FASTPATH);
2755 if (unlikely(gfpflags & __GFP_ZERO) && object)
2756 memset(object, 0, s->object_size);
2758 slab_post_alloc_hook(s, gfpflags, 1, &object);
2763 static __always_inline void *slab_alloc(struct kmem_cache *s,
2764 gfp_t gfpflags, unsigned long addr)
2766 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2769 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2771 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2773 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2778 EXPORT_SYMBOL(kmem_cache_alloc);
2780 #ifdef CONFIG_TRACING
2781 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2783 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2784 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2785 kasan_kmalloc(s, ret, size, gfpflags);
2788 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2792 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2794 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2796 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2797 s->object_size, s->size, gfpflags, node);
2801 EXPORT_SYMBOL(kmem_cache_alloc_node);
2803 #ifdef CONFIG_TRACING
2804 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2806 int node, size_t size)
2808 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2810 trace_kmalloc_node(_RET_IP_, ret,
2811 size, s->size, gfpflags, node);
2813 kasan_kmalloc(s, ret, size, gfpflags);
2816 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2821 * Slow path handling. This may still be called frequently since objects
2822 * have a longer lifetime than the cpu slabs in most processing loads.
2824 * So we still attempt to reduce cache line usage. Just take the slab
2825 * lock and free the item. If there is no additional partial page
2826 * handling required then we can return immediately.
2828 static void __slab_free(struct kmem_cache *s, struct page *page,
2829 void *head, void *tail, int cnt,
2836 unsigned long counters;
2837 struct kmem_cache_node *n = NULL;
2838 unsigned long uninitialized_var(flags);
2840 stat(s, FREE_SLOWPATH);
2842 if (kmem_cache_debug(s) &&
2843 !free_debug_processing(s, page, head, tail, cnt, addr))
2848 spin_unlock_irqrestore(&n->list_lock, flags);
2851 prior = page->freelist;
2852 counters = page->counters;
2853 set_freepointer(s, tail, prior);
2854 new.counters = counters;
2855 was_frozen = new.frozen;
2857 if ((!new.inuse || !prior) && !was_frozen) {
2859 if (kmem_cache_has_cpu_partial(s) && !prior) {
2862 * Slab was on no list before and will be
2864 * We can defer the list move and instead
2869 } else { /* Needs to be taken off a list */
2871 n = get_node(s, page_to_nid(page));
2873 * Speculatively acquire the list_lock.
2874 * If the cmpxchg does not succeed then we may
2875 * drop the list_lock without any processing.
2877 * Otherwise the list_lock will synchronize with
2878 * other processors updating the list of slabs.
2880 spin_lock_irqsave(&n->list_lock, flags);
2885 } while (!cmpxchg_double_slab(s, page,
2893 * If we just froze the page then put it onto the
2894 * per cpu partial list.
2896 if (new.frozen && !was_frozen) {
2897 put_cpu_partial(s, page, 1);
2898 stat(s, CPU_PARTIAL_FREE);
2901 * The list lock was not taken therefore no list
2902 * activity can be necessary.
2905 stat(s, FREE_FROZEN);
2909 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2913 * Objects left in the slab. If it was not on the partial list before
2916 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2917 if (kmem_cache_debug(s))
2918 remove_full(s, n, page);
2919 add_partial(n, page, DEACTIVATE_TO_TAIL);
2920 stat(s, FREE_ADD_PARTIAL);
2922 spin_unlock_irqrestore(&n->list_lock, flags);
2928 * Slab on the partial list.
2930 remove_partial(n, page);
2931 stat(s, FREE_REMOVE_PARTIAL);
2933 /* Slab must be on the full list */
2934 remove_full(s, n, page);
2937 spin_unlock_irqrestore(&n->list_lock, flags);
2939 discard_slab(s, page);
2943 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2944 * can perform fastpath freeing without additional function calls.
2946 * The fastpath is only possible if we are freeing to the current cpu slab
2947 * of this processor. This typically the case if we have just allocated
2950 * If fastpath is not possible then fall back to __slab_free where we deal
2951 * with all sorts of special processing.
2953 * Bulk free of a freelist with several objects (all pointing to the
2954 * same page) possible by specifying head and tail ptr, plus objects
2955 * count (cnt). Bulk free indicated by tail pointer being set.
2957 static __always_inline void do_slab_free(struct kmem_cache *s,
2958 struct page *page, void *head, void *tail,
2959 int cnt, unsigned long addr)
2961 void *tail_obj = tail ? : head;
2962 struct kmem_cache_cpu *c;
2966 * Determine the currently cpus per cpu slab.
2967 * The cpu may change afterward. However that does not matter since
2968 * data is retrieved via this pointer. If we are on the same cpu
2969 * during the cmpxchg then the free will succeed.
2972 tid = this_cpu_read(s->cpu_slab->tid);
2973 c = raw_cpu_ptr(s->cpu_slab);
2974 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2975 unlikely(tid != READ_ONCE(c->tid)));
2977 /* Same with comment on barrier() in slab_alloc_node() */
2980 if (likely(page == c->page)) {
2981 void **freelist = READ_ONCE(c->freelist);
2983 set_freepointer(s, tail_obj, freelist);
2985 if (unlikely(!this_cpu_cmpxchg_double(
2986 s->cpu_slab->freelist, s->cpu_slab->tid,
2988 head, next_tid(tid)))) {
2990 note_cmpxchg_failure("slab_free", s, tid);
2993 stat(s, FREE_FASTPATH);
2995 __slab_free(s, page, head, tail_obj, cnt, addr);
2999 static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
3000 void *head, void *tail, int cnt,
3003 slab_free_freelist_hook(s, head, tail);
3005 * slab_free_freelist_hook() could have put the items into quarantine.
3006 * If so, no need to free them.
3008 if (s->flags & SLAB_KASAN && !(s->flags & SLAB_TYPESAFE_BY_RCU))
3010 do_slab_free(s, page, head, tail, cnt, addr);
3014 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3016 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3020 void kmem_cache_free(struct kmem_cache *s, void *x)
3022 s = cache_from_obj(s, x);
3025 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3026 trace_kmem_cache_free(_RET_IP_, x);
3028 EXPORT_SYMBOL(kmem_cache_free);
3030 struct detached_freelist {
3035 struct kmem_cache *s;
3039 * This function progressively scans the array with free objects (with
3040 * a limited look ahead) and extract objects belonging to the same
3041 * page. It builds a detached freelist directly within the given
3042 * page/objects. This can happen without any need for
3043 * synchronization, because the objects are owned by running process.
3044 * The freelist is build up as a single linked list in the objects.
3045 * The idea is, that this detached freelist can then be bulk
3046 * transferred to the real freelist(s), but only requiring a single
3047 * synchronization primitive. Look ahead in the array is limited due
3048 * to performance reasons.
3051 int build_detached_freelist(struct kmem_cache *s, size_t size,
3052 void **p, struct detached_freelist *df)
3054 size_t first_skipped_index = 0;
3059 /* Always re-init detached_freelist */
3064 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3065 } while (!object && size);
3070 page = virt_to_head_page(object);
3072 /* Handle kalloc'ed objects */
3073 if (unlikely(!PageSlab(page))) {
3074 BUG_ON(!PageCompound(page));
3076 __free_pages(page, compound_order(page));
3077 p[size] = NULL; /* mark object processed */
3080 /* Derive kmem_cache from object */
3081 df->s = page->slab_cache;
3083 df->s = cache_from_obj(s, object); /* Support for memcg */
3086 /* Start new detached freelist */
3088 set_freepointer(df->s, object, NULL);
3090 df->freelist = object;
3091 p[size] = NULL; /* mark object processed */
3097 continue; /* Skip processed objects */
3099 /* df->page is always set at this point */
3100 if (df->page == virt_to_head_page(object)) {
3101 /* Opportunity build freelist */
3102 set_freepointer(df->s, object, df->freelist);
3103 df->freelist = object;
3105 p[size] = NULL; /* mark object processed */
3110 /* Limit look ahead search */
3114 if (!first_skipped_index)
3115 first_skipped_index = size + 1;
3118 return first_skipped_index;
3121 /* Note that interrupts must be enabled when calling this function. */
3122 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3128 struct detached_freelist df;
3130 size = build_detached_freelist(s, size, p, &df);
3134 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3135 } while (likely(size));
3137 EXPORT_SYMBOL(kmem_cache_free_bulk);
3139 /* Note that interrupts must be enabled when calling this function. */
3140 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3143 struct kmem_cache_cpu *c;
3146 /* memcg and kmem_cache debug support */
3147 s = slab_pre_alloc_hook(s, flags);
3151 * Drain objects in the per cpu slab, while disabling local
3152 * IRQs, which protects against PREEMPT and interrupts
3153 * handlers invoking normal fastpath.
3155 local_irq_disable();
3156 c = this_cpu_ptr(s->cpu_slab);
3158 for (i = 0; i < size; i++) {
3159 void *object = c->freelist;
3161 if (unlikely(!object)) {
3163 * We may have removed an object from c->freelist using
3164 * the fastpath in the previous iteration; in that case,
3165 * c->tid has not been bumped yet.
3166 * Since ___slab_alloc() may reenable interrupts while
3167 * allocating memory, we should bump c->tid now.
3169 c->tid = next_tid(c->tid);
3172 * Invoking slow path likely have side-effect
3173 * of re-populating per CPU c->freelist
3175 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3177 if (unlikely(!p[i]))
3180 c = this_cpu_ptr(s->cpu_slab);
3181 continue; /* goto for-loop */
3183 c->freelist = get_freepointer(s, object);
3186 c->tid = next_tid(c->tid);
3189 /* Clear memory outside IRQ disabled fastpath loop */
3190 if (unlikely(flags & __GFP_ZERO)) {
3193 for (j = 0; j < i; j++)
3194 memset(p[j], 0, s->object_size);
3197 /* memcg and kmem_cache debug support */
3198 slab_post_alloc_hook(s, flags, size, p);
3202 slab_post_alloc_hook(s, flags, i, p);
3203 __kmem_cache_free_bulk(s, i, p);
3206 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3210 * Object placement in a slab is made very easy because we always start at
3211 * offset 0. If we tune the size of the object to the alignment then we can
3212 * get the required alignment by putting one properly sized object after
3215 * Notice that the allocation order determines the sizes of the per cpu
3216 * caches. Each processor has always one slab available for allocations.
3217 * Increasing the allocation order reduces the number of times that slabs
3218 * must be moved on and off the partial lists and is therefore a factor in
3223 * Mininum / Maximum order of slab pages. This influences locking overhead
3224 * and slab fragmentation. A higher order reduces the number of partial slabs
3225 * and increases the number of allocations possible without having to
3226 * take the list_lock.
3228 static int slub_min_order;
3229 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3230 static int slub_min_objects;
3233 * Calculate the order of allocation given an slab object size.
3235 * The order of allocation has significant impact on performance and other
3236 * system components. Generally order 0 allocations should be preferred since
3237 * order 0 does not cause fragmentation in the page allocator. Larger objects
3238 * be problematic to put into order 0 slabs because there may be too much
3239 * unused space left. We go to a higher order if more than 1/16th of the slab
3242 * In order to reach satisfactory performance we must ensure that a minimum
3243 * number of objects is in one slab. Otherwise we may generate too much
3244 * activity on the partial lists which requires taking the list_lock. This is
3245 * less a concern for large slabs though which are rarely used.
3247 * slub_max_order specifies the order where we begin to stop considering the
3248 * number of objects in a slab as critical. If we reach slub_max_order then
3249 * we try to keep the page order as low as possible. So we accept more waste
3250 * of space in favor of a small page order.
3252 * Higher order allocations also allow the placement of more objects in a
3253 * slab and thereby reduce object handling overhead. If the user has
3254 * requested a higher mininum order then we start with that one instead of
3255 * the smallest order which will fit the object.
3257 static inline int slab_order(int size, int min_objects,
3258 int max_order, int fract_leftover, int reserved)
3262 int min_order = slub_min_order;
3264 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3265 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3267 for (order = max(min_order, get_order(min_objects * size + reserved));
3268 order <= max_order; order++) {
3270 unsigned long slab_size = PAGE_SIZE << order;
3272 rem = (slab_size - reserved) % size;
3274 if (rem <= slab_size / fract_leftover)
3281 static inline int calculate_order(int size, int reserved)
3289 * Attempt to find best configuration for a slab. This
3290 * works by first attempting to generate a layout with
3291 * the best configuration and backing off gradually.
3293 * First we increase the acceptable waste in a slab. Then
3294 * we reduce the minimum objects required in a slab.
3296 min_objects = slub_min_objects;
3298 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3299 max_objects = order_objects(slub_max_order, size, reserved);
3300 min_objects = min(min_objects, max_objects);
3302 while (min_objects > 1) {
3304 while (fraction >= 4) {
3305 order = slab_order(size, min_objects,
3306 slub_max_order, fraction, reserved);
3307 if (order <= slub_max_order)
3315 * We were unable to place multiple objects in a slab. Now
3316 * lets see if we can place a single object there.
3318 order = slab_order(size, 1, slub_max_order, 1, reserved);
3319 if (order <= slub_max_order)
3323 * Doh this slab cannot be placed using slub_max_order.
3325 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3326 if (order < MAX_ORDER)
3332 init_kmem_cache_node(struct kmem_cache_node *n)
3335 spin_lock_init(&n->list_lock);
3336 INIT_LIST_HEAD(&n->partial);
3337 #ifdef CONFIG_SLUB_DEBUG
3338 atomic_long_set(&n->nr_slabs, 0);
3339 atomic_long_set(&n->total_objects, 0);
3340 INIT_LIST_HEAD(&n->full);
3344 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3346 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3347 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3350 * Must align to double word boundary for the double cmpxchg
3351 * instructions to work; see __pcpu_double_call_return_bool().
3353 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3354 2 * sizeof(void *));
3359 init_kmem_cache_cpus(s);
3364 static struct kmem_cache *kmem_cache_node;
3367 * No kmalloc_node yet so do it by hand. We know that this is the first
3368 * slab on the node for this slabcache. There are no concurrent accesses
3371 * Note that this function only works on the kmem_cache_node
3372 * when allocating for the kmem_cache_node. This is used for bootstrapping
3373 * memory on a fresh node that has no slab structures yet.
3375 static void early_kmem_cache_node_alloc(int node)
3378 struct kmem_cache_node *n;
3380 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3382 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3385 if (page_to_nid(page) != node) {
3386 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3387 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3392 page->freelist = get_freepointer(kmem_cache_node, n);
3395 kmem_cache_node->node[node] = n;
3396 #ifdef CONFIG_SLUB_DEBUG
3397 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3398 init_tracking(kmem_cache_node, n);
3400 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3402 init_kmem_cache_node(n);
3403 inc_slabs_node(kmem_cache_node, node, page->objects);
3406 * No locks need to be taken here as it has just been
3407 * initialized and there is no concurrent access.
3409 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3412 static void free_kmem_cache_nodes(struct kmem_cache *s)
3415 struct kmem_cache_node *n;
3417 for_each_kmem_cache_node(s, node, n) {
3418 s->node[node] = NULL;
3419 kmem_cache_free(kmem_cache_node, n);
3423 void __kmem_cache_release(struct kmem_cache *s)
3425 cache_random_seq_destroy(s);
3426 free_percpu(s->cpu_slab);
3427 free_kmem_cache_nodes(s);
3430 static int init_kmem_cache_nodes(struct kmem_cache *s)
3434 for_each_node_state(node, N_NORMAL_MEMORY) {
3435 struct kmem_cache_node *n;
3437 if (slab_state == DOWN) {
3438 early_kmem_cache_node_alloc(node);
3441 n = kmem_cache_alloc_node(kmem_cache_node,
3445 free_kmem_cache_nodes(s);
3449 init_kmem_cache_node(n);
3455 static void set_min_partial(struct kmem_cache *s, unsigned long min)
3457 if (min < MIN_PARTIAL)
3459 else if (min > MAX_PARTIAL)
3461 s->min_partial = min;
3464 static void set_cpu_partial(struct kmem_cache *s)
3466 #ifdef CONFIG_SLUB_CPU_PARTIAL
3468 * cpu_partial determined the maximum number of objects kept in the
3469 * per cpu partial lists of a processor.
3471 * Per cpu partial lists mainly contain slabs that just have one
3472 * object freed. If they are used for allocation then they can be
3473 * filled up again with minimal effort. The slab will never hit the
3474 * per node partial lists and therefore no locking will be required.
3476 * This setting also determines
3478 * A) The number of objects from per cpu partial slabs dumped to the
3479 * per node list when we reach the limit.
3480 * B) The number of objects in cpu partial slabs to extract from the
3481 * per node list when we run out of per cpu objects. We only fetch
3482 * 50% to keep some capacity around for frees.
3484 if (!kmem_cache_has_cpu_partial(s))
3486 else if (s->size >= PAGE_SIZE)
3488 else if (s->size >= 1024)
3490 else if (s->size >= 256)
3491 s->cpu_partial = 13;
3493 s->cpu_partial = 30;
3498 * calculate_sizes() determines the order and the distribution of data within
3501 static int calculate_sizes(struct kmem_cache *s, int forced_order)
3503 unsigned long flags = s->flags;
3504 size_t size = s->object_size;
3508 * Round up object size to the next word boundary. We can only
3509 * place the free pointer at word boundaries and this determines
3510 * the possible location of the free pointer.
3512 size = ALIGN(size, sizeof(void *));
3514 #ifdef CONFIG_SLUB_DEBUG
3516 * Determine if we can poison the object itself. If the user of
3517 * the slab may touch the object after free or before allocation
3518 * then we should never poison the object itself.
3520 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3522 s->flags |= __OBJECT_POISON;
3524 s->flags &= ~__OBJECT_POISON;
3528 * If we are Redzoning then check if there is some space between the
3529 * end of the object and the free pointer. If not then add an
3530 * additional word to have some bytes to store Redzone information.
3532 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3533 size += sizeof(void *);
3537 * With that we have determined the number of bytes in actual use
3538 * by the object. This is the potential offset to the free pointer.
3542 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3545 * Relocate free pointer after the object if it is not
3546 * permitted to overwrite the first word of the object on
3549 * This is the case if we do RCU, have a constructor or
3550 * destructor or are poisoning the objects.
3553 size += sizeof(void *);
3556 #ifdef CONFIG_SLUB_DEBUG
3557 if (flags & SLAB_STORE_USER)
3559 * Need to store information about allocs and frees after
3562 size += 2 * sizeof(struct track);
3565 kasan_cache_create(s, &size, &s->flags);
3566 #ifdef CONFIG_SLUB_DEBUG
3567 if (flags & SLAB_RED_ZONE) {
3569 * Add some empty padding so that we can catch
3570 * overwrites from earlier objects rather than let
3571 * tracking information or the free pointer be
3572 * corrupted if a user writes before the start
3575 size += sizeof(void *);
3577 s->red_left_pad = sizeof(void *);
3578 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3579 size += s->red_left_pad;
3584 * SLUB stores one object immediately after another beginning from
3585 * offset 0. In order to align the objects we have to simply size
3586 * each object to conform to the alignment.
3588 size = ALIGN(size, s->align);
3590 if (forced_order >= 0)
3591 order = forced_order;
3593 order = calculate_order(size, s->reserved);
3600 s->allocflags |= __GFP_COMP;
3602 if (s->flags & SLAB_CACHE_DMA)
3603 s->allocflags |= GFP_DMA;
3605 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3606 s->allocflags |= __GFP_RECLAIMABLE;
3609 * Determine the number of objects per slab
3611 s->oo = oo_make(order, size, s->reserved);
3612 s->min = oo_make(get_order(size), size, s->reserved);
3613 if (oo_objects(s->oo) > oo_objects(s->max))
3616 return !!oo_objects(s->oo);
3619 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3621 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3623 #ifdef CONFIG_SLAB_FREELIST_HARDENED
3624 s->random = get_random_long();
3627 if (need_reserve_slab_rcu && (s->flags & SLAB_TYPESAFE_BY_RCU))
3628 s->reserved = sizeof(struct rcu_head);
3630 if (!calculate_sizes(s, -1))
3632 if (disable_higher_order_debug) {
3634 * Disable debugging flags that store metadata if the min slab
3637 if (get_order(s->size) > get_order(s->object_size)) {
3638 s->flags &= ~DEBUG_METADATA_FLAGS;
3640 if (!calculate_sizes(s, -1))
3645 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3646 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3647 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3648 /* Enable fast mode */
3649 s->flags |= __CMPXCHG_DOUBLE;
3653 * The larger the object size is, the more pages we want on the partial
3654 * list to avoid pounding the page allocator excessively.
3656 set_min_partial(s, ilog2(s->size) / 2);
3661 s->remote_node_defrag_ratio = 1000;
3664 /* Initialize the pre-computed randomized freelist if slab is up */
3665 if (slab_state >= UP) {
3666 if (init_cache_random_seq(s))
3670 if (!init_kmem_cache_nodes(s))
3673 if (alloc_kmem_cache_cpus(s))
3676 free_kmem_cache_nodes(s);
3678 if (flags & SLAB_PANIC)
3679 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3680 s->name, (unsigned long)s->size, s->size,
3681 oo_order(s->oo), s->offset, flags);
3685 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3688 #ifdef CONFIG_SLUB_DEBUG
3689 void *addr = page_address(page);
3691 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3692 sizeof(long), GFP_ATOMIC);
3695 slab_err(s, page, text, s->name);
3698 get_map(s, page, map);
3699 for_each_object(p, s, addr, page->objects) {
3701 if (!test_bit(slab_index(p, s, addr), map)) {
3702 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3703 print_tracking(s, p);
3712 * Attempt to free all partial slabs on a node.
3713 * This is called from __kmem_cache_shutdown(). We must take list_lock
3714 * because sysfs file might still access partial list after the shutdowning.
3716 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3719 struct page *page, *h;
3721 BUG_ON(irqs_disabled());
3722 spin_lock_irq(&n->list_lock);
3723 list_for_each_entry_safe(page, h, &n->partial, lru) {
3725 remove_partial(n, page);
3726 list_add(&page->lru, &discard);
3728 list_slab_objects(s, page,
3729 "Objects remaining in %s on __kmem_cache_shutdown()");
3732 spin_unlock_irq(&n->list_lock);
3734 list_for_each_entry_safe(page, h, &discard, lru)
3735 discard_slab(s, page);
3739 * Release all resources used by a slab cache.
3741 int __kmem_cache_shutdown(struct kmem_cache *s)
3744 struct kmem_cache_node *n;
3747 /* Attempt to free all objects */
3748 for_each_kmem_cache_node(s, node, n) {
3750 if (n->nr_partial || slabs_node(s, node))
3753 sysfs_slab_remove(s);
3757 /********************************************************************
3759 *******************************************************************/
3761 static int __init setup_slub_min_order(char *str)
3763 get_option(&str, &slub_min_order);
3768 __setup("slub_min_order=", setup_slub_min_order);
3770 static int __init setup_slub_max_order(char *str)
3772 get_option(&str, &slub_max_order);
3773 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3778 __setup("slub_max_order=", setup_slub_max_order);
3780 static int __init setup_slub_min_objects(char *str)
3782 get_option(&str, &slub_min_objects);
3787 __setup("slub_min_objects=", setup_slub_min_objects);
3789 void *__kmalloc(size_t size, gfp_t flags)
3791 struct kmem_cache *s;
3794 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3795 return kmalloc_large(size, flags);
3797 s = kmalloc_slab(size, flags);
3799 if (unlikely(ZERO_OR_NULL_PTR(s)))
3802 ret = slab_alloc(s, flags, _RET_IP_);
3804 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3806 kasan_kmalloc(s, ret, size, flags);
3810 EXPORT_SYMBOL(__kmalloc);
3813 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3818 flags |= __GFP_COMP;
3819 page = alloc_pages_node(node, flags, get_order(size));
3821 ptr = page_address(page);
3823 kmalloc_large_node_hook(ptr, size, flags);
3827 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3829 struct kmem_cache *s;
3832 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3833 ret = kmalloc_large_node(size, flags, node);
3835 trace_kmalloc_node(_RET_IP_, ret,
3836 size, PAGE_SIZE << get_order(size),
3842 s = kmalloc_slab(size, flags);
3844 if (unlikely(ZERO_OR_NULL_PTR(s)))
3847 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3849 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3851 kasan_kmalloc(s, ret, size, flags);
3855 EXPORT_SYMBOL(__kmalloc_node);
3858 #ifdef CONFIG_HARDENED_USERCOPY
3860 * Rejects objects that are incorrectly sized.
3862 * Returns NULL if check passes, otherwise const char * to name of cache
3863 * to indicate an error.
3865 const char *__check_heap_object(const void *ptr, unsigned long n,
3868 struct kmem_cache *s;
3869 unsigned long offset;
3872 /* Find object and usable object size. */
3873 s = page->slab_cache;
3874 object_size = slab_ksize(s);
3876 /* Reject impossible pointers. */
3877 if (ptr < page_address(page))
3880 /* Find offset within object. */
3881 offset = (ptr - page_address(page)) % s->size;
3883 /* Adjust for redzone and reject if within the redzone. */
3884 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3885 if (offset < s->red_left_pad)
3887 offset -= s->red_left_pad;
3890 /* Allow address range falling entirely within object size. */
3891 if (offset <= object_size && n <= object_size - offset)
3896 #endif /* CONFIG_HARDENED_USERCOPY */
3898 static size_t __ksize(const void *object)
3902 if (unlikely(object == ZERO_SIZE_PTR))
3905 page = virt_to_head_page(object);
3907 if (unlikely(!PageSlab(page))) {
3908 WARN_ON(!PageCompound(page));
3909 return PAGE_SIZE << compound_order(page);
3912 return slab_ksize(page->slab_cache);
3915 size_t ksize(const void *object)
3917 size_t size = __ksize(object);
3918 /* We assume that ksize callers could use whole allocated area,
3919 * so we need to unpoison this area.
3921 kasan_unpoison_shadow(object, size);
3924 EXPORT_SYMBOL(ksize);
3926 void kfree(const void *x)
3929 void *object = (void *)x;
3931 trace_kfree(_RET_IP_, x);
3933 if (unlikely(ZERO_OR_NULL_PTR(x)))
3936 page = virt_to_head_page(x);
3937 if (unlikely(!PageSlab(page))) {
3938 BUG_ON(!PageCompound(page));
3940 __free_pages(page, compound_order(page));
3943 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3945 EXPORT_SYMBOL(kfree);
3947 #define SHRINK_PROMOTE_MAX 32
3950 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3951 * up most to the head of the partial lists. New allocations will then
3952 * fill those up and thus they can be removed from the partial lists.
3954 * The slabs with the least items are placed last. This results in them
3955 * being allocated from last increasing the chance that the last objects
3956 * are freed in them.
3958 int __kmem_cache_shrink(struct kmem_cache *s)
3962 struct kmem_cache_node *n;
3965 struct list_head discard;
3966 struct list_head promote[SHRINK_PROMOTE_MAX];
3967 unsigned long flags;
3971 for_each_kmem_cache_node(s, node, n) {
3972 INIT_LIST_HEAD(&discard);
3973 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3974 INIT_LIST_HEAD(promote + i);
3976 spin_lock_irqsave(&n->list_lock, flags);
3979 * Build lists of slabs to discard or promote.
3981 * Note that concurrent frees may occur while we hold the
3982 * list_lock. page->inuse here is the upper limit.
3984 list_for_each_entry_safe(page, t, &n->partial, lru) {
3985 int free = page->objects - page->inuse;
3987 /* Do not reread page->inuse */
3990 /* We do not keep full slabs on the list */
3993 if (free == page->objects) {
3994 list_move(&page->lru, &discard);
3996 } else if (free <= SHRINK_PROMOTE_MAX)
3997 list_move(&page->lru, promote + free - 1);
4001 * Promote the slabs filled up most to the head of the
4004 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4005 list_splice(promote + i, &n->partial);
4007 spin_unlock_irqrestore(&n->list_lock, flags);
4009 /* Release empty slabs */
4010 list_for_each_entry_safe(page, t, &discard, lru)
4011 discard_slab(s, page);
4013 if (slabs_node(s, node))
4021 static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4024 * Called with all the locks held after a sched RCU grace period.
4025 * Even if @s becomes empty after shrinking, we can't know that @s
4026 * doesn't have allocations already in-flight and thus can't
4027 * destroy @s until the associated memcg is released.
4029 * However, let's remove the sysfs files for empty caches here.
4030 * Each cache has a lot of interface files which aren't
4031 * particularly useful for empty draining caches; otherwise, we can
4032 * easily end up with millions of unnecessary sysfs files on
4033 * systems which have a lot of memory and transient cgroups.
4035 if (!__kmem_cache_shrink(s))
4036 sysfs_slab_remove(s);
4039 void __kmemcg_cache_deactivate(struct kmem_cache *s)
4042 * Disable empty slabs caching. Used to avoid pinning offline
4043 * memory cgroups by kmem pages that can be freed.
4045 slub_set_cpu_partial(s, 0);
4049 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4050 * we have to make sure the change is visible before shrinking.
4052 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4056 static int slab_mem_going_offline_callback(void *arg)
4058 struct kmem_cache *s;
4060 mutex_lock(&slab_mutex);
4061 list_for_each_entry(s, &slab_caches, list)
4062 __kmem_cache_shrink(s);
4063 mutex_unlock(&slab_mutex);
4068 static void slab_mem_offline_callback(void *arg)
4070 struct kmem_cache_node *n;
4071 struct kmem_cache *s;
4072 struct memory_notify *marg = arg;
4075 offline_node = marg->status_change_nid_normal;
4078 * If the node still has available memory. we need kmem_cache_node
4081 if (offline_node < 0)
4084 mutex_lock(&slab_mutex);
4085 list_for_each_entry(s, &slab_caches, list) {
4086 n = get_node(s, offline_node);
4089 * if n->nr_slabs > 0, slabs still exist on the node
4090 * that is going down. We were unable to free them,
4091 * and offline_pages() function shouldn't call this
4092 * callback. So, we must fail.
4094 BUG_ON(slabs_node(s, offline_node));
4096 s->node[offline_node] = NULL;
4097 kmem_cache_free(kmem_cache_node, n);
4100 mutex_unlock(&slab_mutex);
4103 static int slab_mem_going_online_callback(void *arg)
4105 struct kmem_cache_node *n;
4106 struct kmem_cache *s;
4107 struct memory_notify *marg = arg;
4108 int nid = marg->status_change_nid_normal;
4112 * If the node's memory is already available, then kmem_cache_node is
4113 * already created. Nothing to do.
4119 * We are bringing a node online. No memory is available yet. We must
4120 * allocate a kmem_cache_node structure in order to bring the node
4123 mutex_lock(&slab_mutex);
4124 list_for_each_entry(s, &slab_caches, list) {
4126 * XXX: kmem_cache_alloc_node will fallback to other nodes
4127 * since memory is not yet available from the node that
4130 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4135 init_kmem_cache_node(n);
4139 mutex_unlock(&slab_mutex);
4143 static int slab_memory_callback(struct notifier_block *self,
4144 unsigned long action, void *arg)
4149 case MEM_GOING_ONLINE:
4150 ret = slab_mem_going_online_callback(arg);
4152 case MEM_GOING_OFFLINE:
4153 ret = slab_mem_going_offline_callback(arg);
4156 case MEM_CANCEL_ONLINE:
4157 slab_mem_offline_callback(arg);
4160 case MEM_CANCEL_OFFLINE:
4164 ret = notifier_from_errno(ret);
4170 static struct notifier_block slab_memory_callback_nb = {
4171 .notifier_call = slab_memory_callback,
4172 .priority = SLAB_CALLBACK_PRI,
4175 /********************************************************************
4176 * Basic setup of slabs
4177 *******************************************************************/
4180 * Used for early kmem_cache structures that were allocated using
4181 * the page allocator. Allocate them properly then fix up the pointers
4182 * that may be pointing to the wrong kmem_cache structure.
4185 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4188 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4189 struct kmem_cache_node *n;
4191 memcpy(s, static_cache, kmem_cache->object_size);
4194 * This runs very early, and only the boot processor is supposed to be
4195 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4198 __flush_cpu_slab(s, smp_processor_id());
4199 for_each_kmem_cache_node(s, node, n) {
4202 list_for_each_entry(p, &n->partial, lru)
4205 #ifdef CONFIG_SLUB_DEBUG
4206 list_for_each_entry(p, &n->full, lru)
4210 slab_init_memcg_params(s);
4211 list_add(&s->list, &slab_caches);
4212 memcg_link_cache(s);
4216 void __init kmem_cache_init(void)
4218 static __initdata struct kmem_cache boot_kmem_cache,
4219 boot_kmem_cache_node;
4221 if (debug_guardpage_minorder())
4224 kmem_cache_node = &boot_kmem_cache_node;
4225 kmem_cache = &boot_kmem_cache;
4227 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4228 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
4230 register_hotmemory_notifier(&slab_memory_callback_nb);
4232 /* Able to allocate the per node structures */
4233 slab_state = PARTIAL;
4235 create_boot_cache(kmem_cache, "kmem_cache",
4236 offsetof(struct kmem_cache, node) +
4237 nr_node_ids * sizeof(struct kmem_cache_node *),
4238 SLAB_HWCACHE_ALIGN);
4240 kmem_cache = bootstrap(&boot_kmem_cache);
4243 * Allocate kmem_cache_node properly from the kmem_cache slab.
4244 * kmem_cache_node is separately allocated so no need to
4245 * update any list pointers.
4247 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4249 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4250 setup_kmalloc_cache_index_table();
4251 create_kmalloc_caches(0);
4253 /* Setup random freelists for each cache */
4254 init_freelist_randomization();
4256 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4259 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%u, Nodes=%d\n",
4261 slub_min_order, slub_max_order, slub_min_objects,
4262 nr_cpu_ids, nr_node_ids);
4265 void __init kmem_cache_init_late(void)
4270 __kmem_cache_alias(const char *name, size_t size, size_t align,
4271 unsigned long flags, void (*ctor)(void *))
4273 struct kmem_cache *s, *c;
4275 s = find_mergeable(size, align, flags, name, ctor);
4280 * Adjust the object sizes so that we clear
4281 * the complete object on kzalloc.
4283 s->object_size = max(s->object_size, (int)size);
4284 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
4286 for_each_memcg_cache(c, s) {
4287 c->object_size = s->object_size;
4288 c->inuse = max_t(int, c->inuse,
4289 ALIGN(size, sizeof(void *)));
4292 if (sysfs_slab_alias(s, name)) {
4301 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
4305 err = kmem_cache_open(s, flags);
4309 /* Mutex is not taken during early boot */
4310 if (slab_state <= UP)
4313 memcg_propagate_slab_attrs(s);
4314 err = sysfs_slab_add(s);
4316 __kmem_cache_release(s);
4321 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4323 struct kmem_cache *s;
4326 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4327 return kmalloc_large(size, gfpflags);
4329 s = kmalloc_slab(size, gfpflags);
4331 if (unlikely(ZERO_OR_NULL_PTR(s)))
4334 ret = slab_alloc(s, gfpflags, caller);
4336 /* Honor the call site pointer we received. */
4337 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4343 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4344 int node, unsigned long caller)
4346 struct kmem_cache *s;
4349 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4350 ret = kmalloc_large_node(size, gfpflags, node);
4352 trace_kmalloc_node(caller, ret,
4353 size, PAGE_SIZE << get_order(size),
4359 s = kmalloc_slab(size, gfpflags);
4361 if (unlikely(ZERO_OR_NULL_PTR(s)))
4364 ret = slab_alloc_node(s, gfpflags, node, caller);
4366 /* Honor the call site pointer we received. */
4367 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4374 static int count_inuse(struct page *page)
4379 static int count_total(struct page *page)
4381 return page->objects;
4385 #ifdef CONFIG_SLUB_DEBUG
4386 static int validate_slab(struct kmem_cache *s, struct page *page,
4390 void *addr = page_address(page);
4392 if (!check_slab(s, page) ||
4393 !on_freelist(s, page, NULL))
4396 /* Now we know that a valid freelist exists */
4397 bitmap_zero(map, page->objects);
4399 get_map(s, page, map);
4400 for_each_object(p, s, addr, page->objects) {
4401 if (test_bit(slab_index(p, s, addr), map))
4402 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4406 for_each_object(p, s, addr, page->objects)
4407 if (!test_bit(slab_index(p, s, addr), map))
4408 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4413 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4417 validate_slab(s, page, map);
4421 static int validate_slab_node(struct kmem_cache *s,
4422 struct kmem_cache_node *n, unsigned long *map)
4424 unsigned long count = 0;
4426 unsigned long flags;
4428 spin_lock_irqsave(&n->list_lock, flags);
4430 list_for_each_entry(page, &n->partial, lru) {
4431 validate_slab_slab(s, page, map);
4434 if (count != n->nr_partial)
4435 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4436 s->name, count, n->nr_partial);
4438 if (!(s->flags & SLAB_STORE_USER))
4441 list_for_each_entry(page, &n->full, lru) {
4442 validate_slab_slab(s, page, map);
4445 if (count != atomic_long_read(&n->nr_slabs))
4446 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4447 s->name, count, atomic_long_read(&n->nr_slabs));
4450 spin_unlock_irqrestore(&n->list_lock, flags);
4454 static long validate_slab_cache(struct kmem_cache *s)
4457 unsigned long count = 0;
4458 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4459 sizeof(unsigned long), GFP_KERNEL);
4460 struct kmem_cache_node *n;
4466 for_each_kmem_cache_node(s, node, n)
4467 count += validate_slab_node(s, n, map);
4472 * Generate lists of code addresses where slabcache objects are allocated
4477 unsigned long count;
4484 DECLARE_BITMAP(cpus, NR_CPUS);
4490 unsigned long count;
4491 struct location *loc;
4494 static void free_loc_track(struct loc_track *t)
4497 free_pages((unsigned long)t->loc,
4498 get_order(sizeof(struct location) * t->max));
4501 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4506 order = get_order(sizeof(struct location) * max);
4508 l = (void *)__get_free_pages(flags, order);
4513 memcpy(l, t->loc, sizeof(struct location) * t->count);
4521 static int add_location(struct loc_track *t, struct kmem_cache *s,
4522 const struct track *track)
4524 long start, end, pos;
4526 unsigned long caddr;
4527 unsigned long age = jiffies - track->when;
4533 pos = start + (end - start + 1) / 2;
4536 * There is nothing at "end". If we end up there
4537 * we need to add something to before end.
4542 caddr = t->loc[pos].addr;
4543 if (track->addr == caddr) {
4549 if (age < l->min_time)
4551 if (age > l->max_time)
4554 if (track->pid < l->min_pid)
4555 l->min_pid = track->pid;
4556 if (track->pid > l->max_pid)
4557 l->max_pid = track->pid;
4559 cpumask_set_cpu(track->cpu,
4560 to_cpumask(l->cpus));
4562 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4566 if (track->addr < caddr)
4573 * Not found. Insert new tracking element.
4575 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4581 (t->count - pos) * sizeof(struct location));
4584 l->addr = track->addr;
4588 l->min_pid = track->pid;
4589 l->max_pid = track->pid;
4590 cpumask_clear(to_cpumask(l->cpus));
4591 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4592 nodes_clear(l->nodes);
4593 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4597 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4598 struct page *page, enum track_item alloc,
4601 void *addr = page_address(page);
4604 bitmap_zero(map, page->objects);
4605 get_map(s, page, map);
4607 for_each_object(p, s, addr, page->objects)
4608 if (!test_bit(slab_index(p, s, addr), map))
4609 add_location(t, s, get_track(s, p, alloc));
4612 static int list_locations(struct kmem_cache *s, char *buf,
4613 enum track_item alloc)
4617 struct loc_track t = { 0, 0, NULL };
4619 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4620 sizeof(unsigned long), GFP_KERNEL);
4621 struct kmem_cache_node *n;
4623 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4626 return sprintf(buf, "Out of memory\n");
4628 /* Push back cpu slabs */
4631 for_each_kmem_cache_node(s, node, n) {
4632 unsigned long flags;
4635 if (!atomic_long_read(&n->nr_slabs))
4638 spin_lock_irqsave(&n->list_lock, flags);
4639 list_for_each_entry(page, &n->partial, lru)
4640 process_slab(&t, s, page, alloc, map);
4641 list_for_each_entry(page, &n->full, lru)
4642 process_slab(&t, s, page, alloc, map);
4643 spin_unlock_irqrestore(&n->list_lock, flags);
4646 for (i = 0; i < t.count; i++) {
4647 struct location *l = &t.loc[i];
4649 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4651 len += sprintf(buf + len, "%7ld ", l->count);
4654 len += sprintf(buf + len, "%pS", (void *)l->addr);
4656 len += sprintf(buf + len, "<not-available>");
4658 if (l->sum_time != l->min_time) {
4659 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4661 (long)div_u64(l->sum_time, l->count),
4664 len += sprintf(buf + len, " age=%ld",
4667 if (l->min_pid != l->max_pid)
4668 len += sprintf(buf + len, " pid=%ld-%ld",
4669 l->min_pid, l->max_pid);
4671 len += sprintf(buf + len, " pid=%ld",
4674 if (num_online_cpus() > 1 &&
4675 !cpumask_empty(to_cpumask(l->cpus)) &&
4676 len < PAGE_SIZE - 60)
4677 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4679 cpumask_pr_args(to_cpumask(l->cpus)));
4681 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4682 len < PAGE_SIZE - 60)
4683 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4685 nodemask_pr_args(&l->nodes));
4687 len += sprintf(buf + len, "\n");
4693 len += sprintf(buf, "No data\n");
4698 #ifdef SLUB_RESILIENCY_TEST
4699 static void __init resiliency_test(void)
4703 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4705 pr_err("SLUB resiliency testing\n");
4706 pr_err("-----------------------\n");
4707 pr_err("A. Corruption after allocation\n");
4709 p = kzalloc(16, GFP_KERNEL);
4711 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4714 validate_slab_cache(kmalloc_caches[4]);
4716 /* Hmmm... The next two are dangerous */
4717 p = kzalloc(32, GFP_KERNEL);
4718 p[32 + sizeof(void *)] = 0x34;
4719 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4721 pr_err("If allocated object is overwritten then not detectable\n\n");
4723 validate_slab_cache(kmalloc_caches[5]);
4724 p = kzalloc(64, GFP_KERNEL);
4725 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4727 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4729 pr_err("If allocated object is overwritten then not detectable\n\n");
4730 validate_slab_cache(kmalloc_caches[6]);
4732 pr_err("\nB. Corruption after free\n");
4733 p = kzalloc(128, GFP_KERNEL);
4736 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4737 validate_slab_cache(kmalloc_caches[7]);
4739 p = kzalloc(256, GFP_KERNEL);
4742 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4743 validate_slab_cache(kmalloc_caches[8]);
4745 p = kzalloc(512, GFP_KERNEL);
4748 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4749 validate_slab_cache(kmalloc_caches[9]);
4753 static void resiliency_test(void) {};
4758 enum slab_stat_type {
4759 SL_ALL, /* All slabs */
4760 SL_PARTIAL, /* Only partially allocated slabs */
4761 SL_CPU, /* Only slabs used for cpu caches */
4762 SL_OBJECTS, /* Determine allocated objects not slabs */
4763 SL_TOTAL /* Determine object capacity not slabs */
4766 #define SO_ALL (1 << SL_ALL)
4767 #define SO_PARTIAL (1 << SL_PARTIAL)
4768 #define SO_CPU (1 << SL_CPU)
4769 #define SO_OBJECTS (1 << SL_OBJECTS)
4770 #define SO_TOTAL (1 << SL_TOTAL)
4773 static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4775 static int __init setup_slub_memcg_sysfs(char *str)
4779 if (get_option(&str, &v) > 0)
4780 memcg_sysfs_enabled = v;
4785 __setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4788 static ssize_t show_slab_objects(struct kmem_cache *s,
4789 char *buf, unsigned long flags)
4791 unsigned long total = 0;
4794 unsigned long *nodes;
4796 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4800 if (flags & SO_CPU) {
4803 for_each_possible_cpu(cpu) {
4804 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4809 page = READ_ONCE(c->page);
4813 node = page_to_nid(page);
4814 if (flags & SO_TOTAL)
4816 else if (flags & SO_OBJECTS)
4824 page = slub_percpu_partial_read_once(c);
4826 node = page_to_nid(page);
4827 if (flags & SO_TOTAL)
4829 else if (flags & SO_OBJECTS)
4840 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4841 * already held which will conflict with an existing lock order:
4843 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4845 * We don't really need mem_hotplug_lock (to hold off
4846 * slab_mem_going_offline_callback) here because slab's memory hot
4847 * unplug code doesn't destroy the kmem_cache->node[] data.
4850 #ifdef CONFIG_SLUB_DEBUG
4851 if (flags & SO_ALL) {
4852 struct kmem_cache_node *n;
4854 for_each_kmem_cache_node(s, node, n) {
4856 if (flags & SO_TOTAL)
4857 x = atomic_long_read(&n->total_objects);
4858 else if (flags & SO_OBJECTS)
4859 x = atomic_long_read(&n->total_objects) -
4860 count_partial(n, count_free);
4862 x = atomic_long_read(&n->nr_slabs);
4869 if (flags & SO_PARTIAL) {
4870 struct kmem_cache_node *n;
4872 for_each_kmem_cache_node(s, node, n) {
4873 if (flags & SO_TOTAL)
4874 x = count_partial(n, count_total);
4875 else if (flags & SO_OBJECTS)
4876 x = count_partial(n, count_inuse);
4883 x = sprintf(buf, "%lu", total);
4885 for (node = 0; node < nr_node_ids; node++)
4887 x += sprintf(buf + x, " N%d=%lu",
4891 return x + sprintf(buf + x, "\n");
4894 #ifdef CONFIG_SLUB_DEBUG
4895 static int any_slab_objects(struct kmem_cache *s)
4898 struct kmem_cache_node *n;
4900 for_each_kmem_cache_node(s, node, n)
4901 if (atomic_long_read(&n->total_objects))
4908 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4909 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4911 struct slab_attribute {
4912 struct attribute attr;
4913 ssize_t (*show)(struct kmem_cache *s, char *buf);
4914 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4917 #define SLAB_ATTR_RO(_name) \
4918 static struct slab_attribute _name##_attr = \
4919 __ATTR(_name, 0400, _name##_show, NULL)
4921 #define SLAB_ATTR(_name) \
4922 static struct slab_attribute _name##_attr = \
4923 __ATTR(_name, 0600, _name##_show, _name##_store)
4925 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4927 return sprintf(buf, "%d\n", s->size);
4929 SLAB_ATTR_RO(slab_size);
4931 static ssize_t align_show(struct kmem_cache *s, char *buf)
4933 return sprintf(buf, "%d\n", s->align);
4935 SLAB_ATTR_RO(align);
4937 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4939 return sprintf(buf, "%d\n", s->object_size);
4941 SLAB_ATTR_RO(object_size);
4943 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4945 return sprintf(buf, "%d\n", oo_objects(s->oo));
4947 SLAB_ATTR_RO(objs_per_slab);
4949 static ssize_t order_store(struct kmem_cache *s,
4950 const char *buf, size_t length)
4952 unsigned long order;
4955 err = kstrtoul(buf, 10, &order);
4959 if (order > slub_max_order || order < slub_min_order)
4962 calculate_sizes(s, order);
4966 static ssize_t order_show(struct kmem_cache *s, char *buf)
4968 return sprintf(buf, "%d\n", oo_order(s->oo));
4972 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4974 return sprintf(buf, "%lu\n", s->min_partial);
4977 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4983 err = kstrtoul(buf, 10, &min);
4987 set_min_partial(s, min);
4990 SLAB_ATTR(min_partial);
4992 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4994 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4997 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5000 unsigned int objects;
5003 err = kstrtouint(buf, 10, &objects);
5006 if (objects && !kmem_cache_has_cpu_partial(s))
5009 slub_set_cpu_partial(s, objects);
5013 SLAB_ATTR(cpu_partial);
5015 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5019 return sprintf(buf, "%pS\n", s->ctor);
5023 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5025 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5027 SLAB_ATTR_RO(aliases);
5029 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5031 return show_slab_objects(s, buf, SO_PARTIAL);
5033 SLAB_ATTR_RO(partial);
5035 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5037 return show_slab_objects(s, buf, SO_CPU);
5039 SLAB_ATTR_RO(cpu_slabs);
5041 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5043 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5045 SLAB_ATTR_RO(objects);
5047 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5049 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5051 SLAB_ATTR_RO(objects_partial);
5053 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5060 for_each_online_cpu(cpu) {
5063 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5066 pages += page->pages;
5067 objects += page->pobjects;
5071 len = sprintf(buf, "%d(%d)", objects, pages);
5074 for_each_online_cpu(cpu) {
5077 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5079 if (page && len < PAGE_SIZE - 20)
5080 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5081 page->pobjects, page->pages);
5084 return len + sprintf(buf + len, "\n");
5086 SLAB_ATTR_RO(slabs_cpu_partial);
5088 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5090 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5093 static ssize_t reclaim_account_store(struct kmem_cache *s,
5094 const char *buf, size_t length)
5096 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5098 s->flags |= SLAB_RECLAIM_ACCOUNT;
5101 SLAB_ATTR(reclaim_account);
5103 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5105 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5107 SLAB_ATTR_RO(hwcache_align);
5109 #ifdef CONFIG_ZONE_DMA
5110 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5112 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5114 SLAB_ATTR_RO(cache_dma);
5117 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5119 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5121 SLAB_ATTR_RO(destroy_by_rcu);
5123 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
5125 return sprintf(buf, "%d\n", s->reserved);
5127 SLAB_ATTR_RO(reserved);
5129 #ifdef CONFIG_SLUB_DEBUG
5130 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5132 return show_slab_objects(s, buf, SO_ALL);
5134 SLAB_ATTR_RO(slabs);
5136 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5138 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5140 SLAB_ATTR_RO(total_objects);
5142 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5144 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5147 static ssize_t sanity_checks_store(struct kmem_cache *s,
5148 const char *buf, size_t length)
5150 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5151 if (buf[0] == '1') {
5152 s->flags &= ~__CMPXCHG_DOUBLE;
5153 s->flags |= SLAB_CONSISTENCY_CHECKS;
5157 SLAB_ATTR(sanity_checks);
5159 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5161 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5164 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5168 * Tracing a merged cache is going to give confusing results
5169 * as well as cause other issues like converting a mergeable
5170 * cache into an umergeable one.
5172 if (s->refcount > 1)
5175 s->flags &= ~SLAB_TRACE;
5176 if (buf[0] == '1') {
5177 s->flags &= ~__CMPXCHG_DOUBLE;
5178 s->flags |= SLAB_TRACE;
5184 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5186 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5189 static ssize_t red_zone_store(struct kmem_cache *s,
5190 const char *buf, size_t length)
5192 if (any_slab_objects(s))
5195 s->flags &= ~SLAB_RED_ZONE;
5196 if (buf[0] == '1') {
5197 s->flags |= SLAB_RED_ZONE;
5199 calculate_sizes(s, -1);
5202 SLAB_ATTR(red_zone);
5204 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5206 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5209 static ssize_t poison_store(struct kmem_cache *s,
5210 const char *buf, size_t length)
5212 if (any_slab_objects(s))
5215 s->flags &= ~SLAB_POISON;
5216 if (buf[0] == '1') {
5217 s->flags |= SLAB_POISON;
5219 calculate_sizes(s, -1);
5224 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5226 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5229 static ssize_t store_user_store(struct kmem_cache *s,
5230 const char *buf, size_t length)
5232 if (any_slab_objects(s))
5235 s->flags &= ~SLAB_STORE_USER;
5236 if (buf[0] == '1') {
5237 s->flags &= ~__CMPXCHG_DOUBLE;
5238 s->flags |= SLAB_STORE_USER;
5240 calculate_sizes(s, -1);
5243 SLAB_ATTR(store_user);
5245 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5250 static ssize_t validate_store(struct kmem_cache *s,
5251 const char *buf, size_t length)
5255 if (buf[0] == '1') {
5256 ret = validate_slab_cache(s);
5262 SLAB_ATTR(validate);
5264 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5266 if (!(s->flags & SLAB_STORE_USER))
5268 return list_locations(s, buf, TRACK_ALLOC);
5270 SLAB_ATTR_RO(alloc_calls);
5272 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5274 if (!(s->flags & SLAB_STORE_USER))
5276 return list_locations(s, buf, TRACK_FREE);
5278 SLAB_ATTR_RO(free_calls);
5279 #endif /* CONFIG_SLUB_DEBUG */
5281 #ifdef CONFIG_FAILSLAB
5282 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5284 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5287 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5290 if (s->refcount > 1)
5293 s->flags &= ~SLAB_FAILSLAB;
5295 s->flags |= SLAB_FAILSLAB;
5298 SLAB_ATTR(failslab);
5301 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5306 static ssize_t shrink_store(struct kmem_cache *s,
5307 const char *buf, size_t length)
5310 kmem_cache_shrink(s);
5318 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5320 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5323 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5324 const char *buf, size_t length)
5326 unsigned long ratio;
5329 err = kstrtoul(buf, 10, &ratio);
5334 s->remote_node_defrag_ratio = ratio * 10;
5338 SLAB_ATTR(remote_node_defrag_ratio);
5341 #ifdef CONFIG_SLUB_STATS
5342 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5344 unsigned long sum = 0;
5347 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5352 for_each_online_cpu(cpu) {
5353 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5359 len = sprintf(buf, "%lu", sum);
5362 for_each_online_cpu(cpu) {
5363 if (data[cpu] && len < PAGE_SIZE - 20)
5364 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5368 return len + sprintf(buf + len, "\n");
5371 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5375 for_each_online_cpu(cpu)
5376 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5379 #define STAT_ATTR(si, text) \
5380 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5382 return show_stat(s, buf, si); \
5384 static ssize_t text##_store(struct kmem_cache *s, \
5385 const char *buf, size_t length) \
5387 if (buf[0] != '0') \
5389 clear_stat(s, si); \
5394 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5395 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5396 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5397 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5398 STAT_ATTR(FREE_FROZEN, free_frozen);
5399 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5400 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5401 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5402 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5403 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5404 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5405 STAT_ATTR(FREE_SLAB, free_slab);
5406 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5407 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5408 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5409 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5410 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5411 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5412 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5413 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5414 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5415 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5416 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5417 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5418 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5419 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5422 static struct attribute *slab_attrs[] = {
5423 &slab_size_attr.attr,
5424 &object_size_attr.attr,
5425 &objs_per_slab_attr.attr,
5427 &min_partial_attr.attr,
5428 &cpu_partial_attr.attr,
5430 &objects_partial_attr.attr,
5432 &cpu_slabs_attr.attr,
5436 &hwcache_align_attr.attr,
5437 &reclaim_account_attr.attr,
5438 &destroy_by_rcu_attr.attr,
5440 &reserved_attr.attr,
5441 &slabs_cpu_partial_attr.attr,
5442 #ifdef CONFIG_SLUB_DEBUG
5443 &total_objects_attr.attr,
5445 &sanity_checks_attr.attr,
5447 &red_zone_attr.attr,
5449 &store_user_attr.attr,
5450 &validate_attr.attr,
5451 &alloc_calls_attr.attr,
5452 &free_calls_attr.attr,
5454 #ifdef CONFIG_ZONE_DMA
5455 &cache_dma_attr.attr,
5458 &remote_node_defrag_ratio_attr.attr,
5460 #ifdef CONFIG_SLUB_STATS
5461 &alloc_fastpath_attr.attr,
5462 &alloc_slowpath_attr.attr,
5463 &free_fastpath_attr.attr,
5464 &free_slowpath_attr.attr,
5465 &free_frozen_attr.attr,
5466 &free_add_partial_attr.attr,
5467 &free_remove_partial_attr.attr,
5468 &alloc_from_partial_attr.attr,
5469 &alloc_slab_attr.attr,
5470 &alloc_refill_attr.attr,
5471 &alloc_node_mismatch_attr.attr,
5472 &free_slab_attr.attr,
5473 &cpuslab_flush_attr.attr,
5474 &deactivate_full_attr.attr,
5475 &deactivate_empty_attr.attr,
5476 &deactivate_to_head_attr.attr,
5477 &deactivate_to_tail_attr.attr,
5478 &deactivate_remote_frees_attr.attr,
5479 &deactivate_bypass_attr.attr,
5480 &order_fallback_attr.attr,
5481 &cmpxchg_double_fail_attr.attr,
5482 &cmpxchg_double_cpu_fail_attr.attr,
5483 &cpu_partial_alloc_attr.attr,
5484 &cpu_partial_free_attr.attr,
5485 &cpu_partial_node_attr.attr,
5486 &cpu_partial_drain_attr.attr,
5488 #ifdef CONFIG_FAILSLAB
5489 &failslab_attr.attr,
5495 static const struct attribute_group slab_attr_group = {
5496 .attrs = slab_attrs,
5499 static ssize_t slab_attr_show(struct kobject *kobj,
5500 struct attribute *attr,
5503 struct slab_attribute *attribute;
5504 struct kmem_cache *s;
5507 attribute = to_slab_attr(attr);
5510 if (!attribute->show)
5513 err = attribute->show(s, buf);
5518 static ssize_t slab_attr_store(struct kobject *kobj,
5519 struct attribute *attr,
5520 const char *buf, size_t len)
5522 struct slab_attribute *attribute;
5523 struct kmem_cache *s;
5526 attribute = to_slab_attr(attr);
5529 if (!attribute->store)
5532 err = attribute->store(s, buf, len);
5534 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5535 struct kmem_cache *c;
5537 mutex_lock(&slab_mutex);
5538 if (s->max_attr_size < len)
5539 s->max_attr_size = len;
5542 * This is a best effort propagation, so this function's return
5543 * value will be determined by the parent cache only. This is
5544 * basically because not all attributes will have a well
5545 * defined semantics for rollbacks - most of the actions will
5546 * have permanent effects.
5548 * Returning the error value of any of the children that fail
5549 * is not 100 % defined, in the sense that users seeing the
5550 * error code won't be able to know anything about the state of
5553 * Only returning the error code for the parent cache at least
5554 * has well defined semantics. The cache being written to
5555 * directly either failed or succeeded, in which case we loop
5556 * through the descendants with best-effort propagation.
5558 for_each_memcg_cache(c, s)
5559 attribute->store(c, buf, len);
5560 mutex_unlock(&slab_mutex);
5566 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5570 char *buffer = NULL;
5571 struct kmem_cache *root_cache;
5573 if (is_root_cache(s))
5576 root_cache = s->memcg_params.root_cache;
5579 * This mean this cache had no attribute written. Therefore, no point
5580 * in copying default values around
5582 if (!root_cache->max_attr_size)
5585 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5588 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5591 if (!attr || !attr->store || !attr->show)
5595 * It is really bad that we have to allocate here, so we will
5596 * do it only as a fallback. If we actually allocate, though,
5597 * we can just use the allocated buffer until the end.
5599 * Most of the slub attributes will tend to be very small in
5600 * size, but sysfs allows buffers up to a page, so they can
5601 * theoretically happen.
5605 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf) &&
5606 !IS_ENABLED(CONFIG_SLUB_STATS))
5609 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5610 if (WARN_ON(!buffer))
5615 len = attr->show(root_cache, buf);
5617 attr->store(s, buf, len);
5621 free_page((unsigned long)buffer);
5625 static void kmem_cache_release(struct kobject *k)
5627 slab_kmem_cache_release(to_slab(k));
5630 static const struct sysfs_ops slab_sysfs_ops = {
5631 .show = slab_attr_show,
5632 .store = slab_attr_store,
5635 static struct kobj_type slab_ktype = {
5636 .sysfs_ops = &slab_sysfs_ops,
5637 .release = kmem_cache_release,
5640 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5642 struct kobj_type *ktype = get_ktype(kobj);
5644 if (ktype == &slab_ktype)
5649 static const struct kset_uevent_ops slab_uevent_ops = {
5650 .filter = uevent_filter,
5653 static struct kset *slab_kset;
5655 static inline struct kset *cache_kset(struct kmem_cache *s)
5658 if (!is_root_cache(s))
5659 return s->memcg_params.root_cache->memcg_kset;
5664 #define ID_STR_LENGTH 64
5666 /* Create a unique string id for a slab cache:
5668 * Format :[flags-]size
5670 static char *create_unique_id(struct kmem_cache *s)
5672 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5679 * First flags affecting slabcache operations. We will only
5680 * get here for aliasable slabs so we do not need to support
5681 * too many flags. The flags here must cover all flags that
5682 * are matched during merging to guarantee that the id is
5685 if (s->flags & SLAB_CACHE_DMA)
5687 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5689 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5691 if (s->flags & SLAB_ACCOUNT)
5695 p += sprintf(p, "%07d", s->size);
5697 BUG_ON(p > name + ID_STR_LENGTH - 1);
5701 static void sysfs_slab_remove_workfn(struct work_struct *work)
5703 struct kmem_cache *s =
5704 container_of(work, struct kmem_cache, kobj_remove_work);
5706 if (!s->kobj.state_in_sysfs)
5708 * For a memcg cache, this may be called during
5709 * deactivation and again on shutdown. Remove only once.
5710 * A cache is never shut down before deactivation is
5711 * complete, so no need to worry about synchronization.
5716 kset_unregister(s->memcg_kset);
5718 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5720 kobject_put(&s->kobj);
5723 static int sysfs_slab_add(struct kmem_cache *s)
5727 struct kset *kset = cache_kset(s);
5728 int unmergeable = slab_unmergeable(s);
5730 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5733 kobject_init(&s->kobj, &slab_ktype);
5737 if (!unmergeable && disable_higher_order_debug &&
5738 (slub_debug & DEBUG_METADATA_FLAGS))
5743 * Slabcache can never be merged so we can use the name proper.
5744 * This is typically the case for debug situations. In that
5745 * case we can catch duplicate names easily.
5747 sysfs_remove_link(&slab_kset->kobj, s->name);
5751 * Create a unique name for the slab as a target
5754 name = create_unique_id(s);
5757 s->kobj.kset = kset;
5758 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5762 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5767 if (is_root_cache(s) && memcg_sysfs_enabled) {
5768 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5769 if (!s->memcg_kset) {
5776 kobject_uevent(&s->kobj, KOBJ_ADD);
5778 /* Setup first alias */
5779 sysfs_slab_alias(s, s->name);
5786 kobject_del(&s->kobj);
5790 static void sysfs_slab_remove(struct kmem_cache *s)
5792 if (slab_state < FULL)
5794 * Sysfs has not been setup yet so no need to remove the
5799 kobject_get(&s->kobj);
5800 schedule_work(&s->kobj_remove_work);
5803 void sysfs_slab_unlink(struct kmem_cache *s)
5805 if (slab_state >= FULL)
5806 kobject_del(&s->kobj);
5809 void sysfs_slab_release(struct kmem_cache *s)
5811 if (slab_state >= FULL)
5812 kobject_put(&s->kobj);
5816 * Need to buffer aliases during bootup until sysfs becomes
5817 * available lest we lose that information.
5819 struct saved_alias {
5820 struct kmem_cache *s;
5822 struct saved_alias *next;
5825 static struct saved_alias *alias_list;
5827 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5829 struct saved_alias *al;
5831 if (slab_state == FULL) {
5833 * If we have a leftover link then remove it.
5835 sysfs_remove_link(&slab_kset->kobj, name);
5836 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5839 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5845 al->next = alias_list;
5850 static int __init slab_sysfs_init(void)
5852 struct kmem_cache *s;
5855 mutex_lock(&slab_mutex);
5857 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5859 mutex_unlock(&slab_mutex);
5860 pr_err("Cannot register slab subsystem.\n");
5866 list_for_each_entry(s, &slab_caches, list) {
5867 err = sysfs_slab_add(s);
5869 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5873 while (alias_list) {
5874 struct saved_alias *al = alias_list;
5876 alias_list = alias_list->next;
5877 err = sysfs_slab_alias(al->s, al->name);
5879 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5884 mutex_unlock(&slab_mutex);
5889 __initcall(slab_sysfs_init);
5890 #endif /* CONFIG_SYSFS */
5893 * The /proc/slabinfo ABI
5895 #ifdef CONFIG_SLABINFO
5896 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5898 unsigned long nr_slabs = 0;
5899 unsigned long nr_objs = 0;
5900 unsigned long nr_free = 0;
5902 struct kmem_cache_node *n;
5904 for_each_kmem_cache_node(s, node, n) {
5905 nr_slabs += node_nr_slabs(n);
5906 nr_objs += node_nr_objs(n);
5907 nr_free += count_partial(n, count_free);
5910 sinfo->active_objs = nr_objs - nr_free;
5911 sinfo->num_objs = nr_objs;
5912 sinfo->active_slabs = nr_slabs;
5913 sinfo->num_slabs = nr_slabs;
5914 sinfo->objects_per_slab = oo_objects(s->oo);
5915 sinfo->cache_order = oo_order(s->oo);
5918 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5922 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5923 size_t count, loff_t *ppos)
5927 #endif /* CONFIG_SLABINFO */